Forecasting ground-level irradiance over short horizons: Time series, meteorological, and time-varying parameter models

Forecasting ground-level irradiance over short horizons: Time series, meteorological, and time-varying parameter models

Accepted Manuscript Forecasting ground-level irradiance over short horizons: Time series, meteorological, and time-varying models Gordon Reikard, Sue ...

3MB Sizes 3 Downloads 81 Views

Accepted Manuscript Forecasting ground-level irradiance over short horizons: Time series, meteorological, and time-varying models Gordon Reikard, Sue Ellen Haupt, Tara Jensen PII:

S0960-1481(17)30404-4

DOI:

10.1016/j.renene.2017.05.019

Reference:

RENE 8782

To appear in:

Renewable Energy

Received Date: 30 December 2016 Revised Date:

28 March 2017

Accepted Date: 5 May 2017

Please cite this article as: Reikard G, Haupt SE, Jensen T, Forecasting ground-level irradiance over short horizons: Time series, meteorological, and time-varying models, Renewable Energy (2017), doi: 10.1016/j.renene.2017.05.019. This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

ACCEPTED MANUSCRIPT 1 1 2 3

RI PT

4 5 6 7

SC

8 9

M AN U

10 11 12

Forecasting ground-level irradiance over short horizons:

13

Time series, meteorological, and time-varying models

14

25 26

TE D

Gordon Reikard U.S. Cellular

EP

Sue Ellen Haupt National Center for Atmospheric Research Tara Jensen National Center for Atmospheric Research

AC C

15 16 17 18 19 20 21 22 23 24

Revised, March 27, 2017

ACCEPTED MANUSCRIPT 2 27 ABSTRACT

28 29

One of the key enabling technologies for integrating solar energy into the grid is short-range

31

forecasting. Two issues have emerged in the literature. The first has to do with the relative merits of

32

physics-based versus time series models. The second is how to parameterize short-term variability. One

33

promising approach is time-varying parameter models. Time series models can be updated using moving

34

windows. Meteorological models can be adjusted to match the data more closely. This study evaluates

35

several types of models over forecast horizons ranging from 15 minutes to 4 hours, using data from two

36

locations in the United States. The Weather Research Forecast (WRF) model is a state-of-the art

37

numerical weather prediction system. The Dynamic Integrated Forecast (DICast) system combines

38

meteorological models with statistical adjustments. The primary time series model is the ARIMA.

39

Several other techniques are also tested, cloud advection, smart persistence forecasts and regression trees.

40

Each type of model is found to have particular strengths and weaknesses. Among time series models,

41

ARIMAs with time-varying coefficients are superior to fixed coefficient methods. In a direct comparison

42

of meteorological and time series models, the ARIMA is more accurate at short horizons, while the

43

numerical weather prediction models are more accurate as the horizon extends. The convergence point, at

44

which the two methods achieve similar degrees of accuracy, is in the range of 1-3 hours. Adjusting

45

meteorological model output using statistical corrections at regular intervals, as in the DICast,

46

consistently outperforms the alternatives at horizons of 2-4 hours, and is highly competitive at 1 hour.

48 49 50

SC

M AN U

TE D

EP

AC C

47

RI PT

30

Keywords: solar irradiance, meteorological models, time series models, forecasting

ACCEPTED MANUSCRIPT 3 51

1. Introduction

52 53

One of the key enabling technologies for integrating solar energy into the grid is short-term forecasting. In most utilities in North America, balancing reserves, used to buffer imbalances between supply

55

and demand, are calculated at the 1 hour horizon. The Federal Energy Regulatory Commission has also

56

mandated 15 minute transmission scheduling to assist in integrating variable sources [1]. Some independent

57

system operators are scheduling at horizons of as little as 5 minutes. Forecasts at somewhat longer horizons

58

are used in operational planning, peak load matching, switching sources and planning.

SC

59

RI PT

54

An extensive literature on forecasting has emerged over the last two decades. Several classes of forecasting models have demonstrated value over particular horizons [2]. At the shortest time scales, less

61

than 15 minutes, models based on sky image data can often outperform other methods [3-8]. Time series

62

models have also been found to perform well over short horizons, ranging from a few minutes to several

63

hours. At longer horizons, on the order of 4 hours and beyond, meteorological or Numerical Weather

64

Prediction (NWP) models have been found to yield the most accurate predictions [9-13]. Two issues have emerged in the literature. The first has to do with the relative merits of physics-

TE D

65

M AN U

60

based versus time series methods. In principle, meteorological models are attractive because they can capture

67

the factors influencing ground-level irradiance: cloud cover, precipitation, humidity and aerosols. The

68

corresponding disadvantage is that these variables may be difficult to predict accurately. In time series

69

models, the causal factors are captured only implicitly, through lag coefficients. Despite this, they often

70

predict more accurately at short horizons.

AC C

71

EP

66

The second issue is how to model short-term variability. The atmosphere is known to have

72

multifractal properties: It is characterized by high degrees of intermittency and irregular outliers [14-15].

73

One approach is time varying models. Evidence from other fields, primarily econometrics, has demonstrated

74

that stochastic parameter models can often predict more accurately than their fixed-coefficient counterparts

75

[16-17]. Meteorological models can be made time varying by adjusting the forecasts to more closely match

76

the recent data.

ACCEPTED MANUSCRIPT 4 This study evaluates meteorological and time series models over horizons ranging from 15 minutes to

78

4 hours. The models are reviewed in Section 2. The databases and forecasting methodology are reviewed in

79

Section 3. The empirical findings are reported in Section 4. Further analysis is conducted in in Sections 5-6.

80

Section 7 concludes.

RI PT

77

81 82

2. The forecasting models

83

Table 1 provides a glossary of the forecasting models. The meteorological forecasts were generated

SC

84

using systems developed at the National Center for Atmospheric Research (NCAR). These include two

86

versions of the Weather Research Forecast model and the Dynamic Integrated Forecast system. Two cloud

87

advection models are also tested. The primary time series model tested here is the well-known

88

autoregressive, integrated, moving average (ARIMA) class. The other statistical models include persistence

89

and regression forecasts based on the clearness index, i.e., the ratio of the irradiance reaching the surface of

90

the earth to irradiance impinging on the top of the atmosphere. The DICast and ARIMA incorporate time-

91

varying parameters. The other techniques are fixed coefficient models. [TABLE 1 ABOUT HERE]

TE D

92

M AN U

85

2.1 Meteorological and combined models

94

The Weather Research and Forecasting (WRF) model uses the Navier Stokes equations of fluid

EP

93

flow [18-19]. It simulates advection in the atmosphere using the initial conditions established by

96

observations and boundary conditions from a global model, with a series of parameterizations for the

97

unresolved and external processes [20].

98 99

AC C

95

A recent version known as WRF-Solar incorporates several new features that enhance its ability

to predict irradiance [21]. The solar tracking algorithm accounts for changes in the eccentricity of the

100

Earth’s orbit and the obliquity of the Earth’s axis. The model output includes diffuse and direct normal as

101

well as global horizontal irradiance. Interpolation algorithms are used to account for irradiance between

102

model runs, and a fast radiative transfer algorithm calculates the surface irradiance. A new

103

parameterization is used to to capture absorption and scattering of radiation by aerosols. Three

ACCEPTED MANUSCRIPT 5 104

dimensional aerosols are allowed to interact with the cloud microphysics. The model also accounts for

105

the feedbacks in shortwave irradiance from smaller clouds. The forecasts are from two different versions of WRF-Solar. The first is the now-casting version

107

of WRF-Solar (hereafter WRF-Solar-Now), which was run hourly at 9 km resolution over the contiguous

108

United States. The initial and boundary conditions are calculated using the National Centers for

109

Environmental Prediction (NCEP) Rapid Refresh model. The second version, WRF-Solar Day Ahead

110

(hereafter WRF-Solar-DA), uses a higher resolution of 3 km. It is run once per day, targeting day-ahead

111

decision points. This version is initialized with NCEP’s High Resolution Rapid Refresh system for the

112

first 15 hours.

SC

Prior studies have found that models combining atmospheric physics and statistical adjustment can

M AN U

113

RI PT

106

predict more accurately than meteorological models alone, over horizons of 1-5 hours [22-26]. The forecasts

115

are from the Dynamic Integrated forecast (DICast®) system, which adjusts NWP model output based on

116

recent data [27]. Forecasts have been created for several measures of irradiance at selected locations where

117

observational data was provided by private utilities [28, 13]. DICast uses a two-step process. The NWP

118

model forecasts are post-processed using a Model Output Statistics (MOS) approach [29]: the model output

119

is adjusted upward or downward to match the most recent actual values. Then, several of these adjusted

120

forecasts are combined using optimized weights. In effect, DICast is an artificial intelligence system that

121

continually updates the forecasts so as to more closely approximate the observed data.

EP

TE D

114

122

124

2.2 Cloud advection models

AC C

123

Two cloud advection models are also tested. Analyses of satellite-derived advection schemes

125

have found that these fill a gap between short-term sky-imaging methods and methods based on

126

meteorological models alone, as well as capturing ramps in irradiance [30-31]. Empirical tests of

127

satellite-derived advection forecasts have concluded that these are effective primarily within very short

128

time frames [32-42].

129 130

The CIRACast algorithm identifies cloud fields from satellite observations [43-45]. The Colorado State University’s Cooperative Institute for Research in the Atmosphere (CIRA) team has

ACCEPTED MANUSCRIPT 6 developed an operational satellite-derived forecast for predicting irradiance over intervals of 15 minutes

132

to 3 hours. The CIRA method utilizes real-time cloud imagery obtained from the Pathfinder Atmospheres

133

Extended (PTMOS-x) retrieval suite, which is based on geostationary satellite observations [46]. The

134

cloud motion is then predicted using wind forecasts derived from the Global Forecast System (GFS).

135

Ground level irradiance is computed using a radiative transfer model.

136

RI PT

131

The second forecast uses the Multisensor Advection Diffusion (MADCast) algorithm of Auligné [47-48]. The method is in three stages. The first is to retrieve observations and express them as cloud

138

fractions. Irradiance is calculated using a radiative transfer model, under the assumption of clear skies.

139

Departures between the irradiance measured by the satellites and irradiance estimated by the model are

140

computed. The cloud fraction is based on the difference between the two. The cloud fraction profiles are

141

then interpolated to the model grid points. The interpolations are enhanced using the information from

142

multiple satellite platforms, to obtain better estimates of their horizontal and vertical resolutions. The

143

second stage is to forecast the cloud fraction, using the dynamic core of the WRF model. The third stage

144

is to convert these forecasts into ground-level irradiance.

M AN U

SC

137

TE D

145

2.3 Time series models

147

Time series models can be implemented using commercially available software, and the

148

programming is straightforward. The ARIMA class of models is well-established [49]. Using standard

149

notation, let φ(L) be the autoregressive operator, represented as a polynomial in the backshift operator:

150

φ(L) = 1 – φ1L - … - φpLp. Let Φ(L) be the cyclical autoregressive operator, defined the same way. Let

151

θ(L) be the moving average operator: θ(L) = 1 + θ1L + … + θqLq, and Θ(L) be the cyclical moving

152

average operator. Let the superscript ξ denote the order of differencing, and the superscript ζ denote the

153

order of cyclical differencing. Let the superscript f denote the cyclical frequency, for the hourly data, 24

154

hours. Let Yt denote ground-level irradiance and the t-subscript denote time variation. The model is then

155

of the form:

156

(1-L)ξ(1-Lf)ζYt = [θt(L) Θt(L) / φt(L) Φt(L)] εt

AC C

EP

146

(1)

ACCEPTED MANUSCRIPT 7 157

where ε is the residual and the coefficients are stochastic. The success of the ARIMA and related models in forecasting irradiance traces back to their

159

ability to reproduce the diurnal cycle. An early review of time series techniques found that ARIMA

160

models were able to outperform most of the alternatives, including neural networks and unobserved

161

components models, which use trigonometric terms to capture cyclical behavior [50]. This has been

162

confirmed in more recent studies [51].

RI PT

158

The ARIMA can also include causal inputs, in which case it is referred to as a transfer function.

163

Inputs that have been considered include a range of meteorological variables such as temperature,

165

precipitation, cloud cover, and processed satellite images [52-56]. Let Mt denote the input, λ and δ denote

166

the moving average and autoregressive polynomials for the input. The transfer function is of the form:

167

(1-L)ξ(1-Lf)ζYt = [θt (L)t Θ t (L) / φt (L) Φt (L)] εt + [λt (L) / δt (L) ] Mt

M AN U

SC

164

(2)

Other popular techniques include neural networks [57-65]. Training neural networks directly on

168

irradiance has been found to forecast effectively only at very short horizons. Instead, the approach used by

170

many of these studies has been to train the net on the clearness index. The advantage of this method is that it

171

identifies the effect of aerosols and cloud cover, and implicitly takes into account the solar angle, which is

172

computed explicitly. The corresponding disadvantage is that cloud cover can be highly intermittent, making

173

it difficult to predict.

TE D

169

Two forecasts using the clearness index are also tested. These were run at NCAR as part of the

EP

174

Sun4Cast solar forecasting system [13], and are reproduced here. The first is simply a “smart”

176

persistence forecast, which is often used as the baseline for the other forecasts to beat: The clearness

177

index is assumed to be equal to its previous value, while the solar angle changes as computed for the

178

specific time of day and day of year. The second is a regression tree forecast based on the open source

179

Cubist software [53, 66]. These forecasts were run hourly to predict 15 minute interval forecasts out to 3

180

hours.

181 182 183

AC C

175

ACCEPTED MANUSCRIPT 8 184

3. The databases and the forecasting methodology

185 3.1 The data

187

Seven data sets at two locations in the United States were used in the forecasting tests. Table 2

RI PT

186

shows the datasets, the interval spanned, and the number of usable observations, i.e., observations for

189

which there are meteorological model forecasts, less missing and nighttime values. The data are available

190

at two resolutions, 15-minute and hourly. All the databases include global horizontal irradiance (GHI)

191

and the clearness index. Irradiance is denominated in watts per meter squared (W/m2). The clearness

192

index ranges between values of 0 and 1, where 1 indicates completely clear skies.

Figures 1-2 show the locations on maps. Four sites are from the Sacramento Municipal Utility

M AN U

193

SC

188

District (SMUD) in California (latitude 38.33 N, longitude 121.28 W). The district spans about 1400

195

square kilometers. Site 67 lies east of Sacramento. Sites 68-69 are northeast of Sacramento, near the

196

towns of Folsom and Roseville respectively. Site 70 is further south of the city, and west of the Interstate

197

5 highway. At the 15-minute resolution, the data run from January 2, 2015 through April 28, 2016, and

198

consist of 17,701 to 19,636 usable values. At the hourly resolution, the data run from January 17, 2015

199

through June 6, 2016. The number of usable observations ranges from 3,056 to 5,232.

200

TE D

194

Three sites are at Brookhaven National Laboratory, in Upton, New York (latitude 40.52 N, longitude 72.53 W). The Long Island Solar Farm (LISF) is a 32-megawatt solar photovoltaic plant built

202

as a joint venture between the Department of Energy and the Long Island Power Authority. Since the

203

power plant spans only about 200 acres, these sites are closer together. The data run from February 6,

204

2015 through April 12, 2016. The number of usable values ranges from 3,659 to 3,815 for the hourly

205

data, and 15,170 to 15,808 for the 15-minute data. [FIGURES 1-2 ABOUT HERE]

206

ABOUT HERE]

207

AC C

EP

201

[TABLE 2

Figures 3-4 show ground level irradiance at the 24-hour resolution, at Sacramento site 69, and

208

Brookhaven site 13. Irradiance is dominated by the diurnal cycle, but also shows high degrees of

209

nonlinear variability. The data at Brookhaven is much more volatile, due to higher precipitation and

210

cloud cover. [FIGURES 3-4 ABOUT HERE]

ACCEPTED MANUSCRIPT 9 In all the data sets, there were large numbers of missing observations. Missing values are not a

211

problem for the meteorological models, since all of the forecasts correspond with actual observations.

213

However, they are an issue for the ARIMAs, which need to be estimated over continuous data streams.

214

Several interpolation methods were tried. The optimal method for Sacramento was to estimate an

215

ARIMA over the actual observations, and use the fitted values to fill in the missing data points. At

216

Brookhaven, this was impossible, since with more missing observations, the likelihood function could not

217

be maximized. Instead, the procedure was to use data from Brookhaven’s solar bay station, where there

218

were no missing values. The bay station is located less than 1 kilometer from the power plant. The

219

observations for irradiance were similar to those at the three sites, making these values suitable to use as

220

interpolations.

M AN U

SC

RI PT

212

In the tests of forecast accuracy, however, all the interpolated values were omitted. To insure

221 222

comparability, the time series forecasts are evaluated only for the same intervals as for the meteorological

223

models.

224 3.2 Design of the tests

226

The forecasting experiments for Sacramento were set up as follows. For the 15-minute data, the

TE D

225

following models were used: the smart persistence and regression tree, the WRF-Solar-Now, the ARIMA,

228

the transfer function, the CIRACast and MADCast, and a weighted average of the WRF-Solar-Now and

229

cloud advection models. For the hourly data, the models also include the WRF-Solar-DA, and the

230

DICast. Two measures of forecast accuracy are used, the mean absolute error (MAE), in W/m2, and the

231

root mean squared error (RMSE). The RMSE assigns a stronger penalty to large errors.

AC C

232

EP

227

The ARIMA models were specified as ARIMA (1,0,0)(1,1,0), i.e., the model is differenced at

233

the cyclical horizon; it includes one proximate lag and one lag corresponding to the diurnal cycle. For the

234

hourly data, the interval of differencing is of course 24 hours. For the 15-minute data, the interval of

235

differencing is 96 periods. Various specifications were essayed, including longer lags. However, the

236

simpler specification produced the lowest forecast errors.

ACCEPTED MANUSCRIPT 10 In the transfer function, the clearness index is used as an input. The issues involved in estimating

238

the transfer function were complex. First, missing values in the clearness index had to be interpolated. A

239

battery of interpolation algorithms was run; the best results were obtained using a regression with

240

stochastic coefficients. Spurious interpolations such as negative values or values in excess of unity were

241

constrained to lie between 0 and 1. Second, the clearness index itself needs to be forecasted. Several

242

methods were tried, but none were able to predict effectively beyond very short horizons. The model

243

used here was a regression on lags. A neural network produced very similar results.

In the time series models for the 15-minute data, the first 2,000 observations were used as a

SC

244

RI PT

237

training sample. The irradiance and clearness index series were then forecasted iteratively, over horizons

246

of 15, 30 and 45 minutes. In the tests for the hourly data, the first 500 observations were used as the

247

training sample, and the forecasts were run over horizons of 1-4 hours. In each instance, the models were

248

estimated over prior values, forecasted, then re-estimated over the most recent value and forecasted again,

249

until the end of the time series. All the predictions are true out-of-sample forecasts, in that they only use

250

data prior to the start of the horizon. The forecasts for horizons beyond one observation are for this

251

interval only, and skip the intervening values.

TE D

252

M AN U

245

Time-varying parameter regressions can be estimated either using a Kalman filter [67] or a moving window. With an unrestricted Kalman filter, the coefficients behave as a random walk, reducing

254

predictive accuracy, so the moving window was used instead. Narrower widths allow high degrees of

255

coefficient variation, while wider widths make the coefficients more inertial [68]. Preliminary tests were

256

run over a range of moving windows. For the 15-minute data, the lowest errors were found over window

257

width of 1400-1700 observations. The width used in the tests was 1580 observations (365 hours or

258

roughly 15 days). For the hourly data, the smallest errors were found for widths in the range of 400-600

259

hours. In the tests, a window width of 480 hours (20 days) was used.

260 261 262 263

AC C

EP

253

ACCEPTED MANUSCRIPT 11 264

4. The forecasting tests

265 266

Because of the number of models tested, it is useful to report the salient findings up front. At horizons of 15-45 minutes, the ARIMA is generally superior. At horizons of 2-4 hours, the DICast

268

achieves the most accurate forecasts. At 1 hour, the contest between the ARIMA and the DICast is close

269

at the California sites, but the DICast is superior at Brookhaven. The ARIMA is generally more accurate

270

than WRF-Solar models at short horizons, but as the horizon extends, the meteorological models achieve

271

greater accuracy.

SC

RI PT

267

272 4.1 Sacramento

274

Table 3 shows the results for the 15-minute data at Sacramento. Parts 1 and 2 report the MAE

275

and RMSE respectively. At this resolution, the ARIMA easily achieves the most accurate forecasts in

276

terms of the MAE. The average error for the ARIMA is 41.1 W/m2 at the 15-minute horizon, increasing

277

to 53.6 W/m2 at 30 minutes and 60.2 W/m2 at 45 minutes. The error for the transfer function is slightly

278

higher. At first sight, this might appear counterintuitive: including the clearness index should make the

279

model more sensitive to cloud cover. However, including more terms on the right hand side can cause the

280

model to become “over-parameterized”. The clearness index adds a second term for behavior that is

281

already captured by the lags. As a result, the model terms are not independent and may interfere with

282

each other, reducing predictive accuracy.

TE D

EP

At the 15-minute horizon, the results from tseveral of the other models are similar. The MAEs

AC C

283

M AN U

273

284

for the regression tree, WRF-Solar and CIRACast all lie in a range of 74.4 to 77.3 W/m2. However, at

285

30 and 45 minutes, the WRF-Solar-Now model is clearly superior. The MADCast model generates lower

286

MAEs at 15 minutes, although at 45 minutes the two cloud advection models achieve similar degrees of

287

accuracy. At all horizons, these models are consistently able to beat the smart persistence forecast. The

288

regression tree is somewhat better, but the errors remain prohibitively high.

289 290

The findings for the RMSE confirm that the ARIMA and transfer function are more accurate at short horizons. When this measure is used, the transfer function is more competitive. Taking the average

ACCEPTED MANUSCRIPT 12 of all four sites, the transfer function error is only negligibly higher than the ARIMA at 15-30 minutes,

292

although at 45 minutes the ARIMA is clearly better. At Site 67, the transfer function is actually slightly

293

better at 30-45 minutes, but this is an anomalous result. The more typical finding is that the ARIMA and

294

transfer function results are generally close, with the ARIMA slightly more accurate.

RI PT

291

At 15 minutes, the weighted average of three models runs in third place, while the MADCast and

296

persistence forecasts run fourth and fifth. The MADCast model continues to do reasonably well at 30-45

297

minutes, when it is substantially better than the CIRACast. The accuracy of the persistence forecast falls

298

away very quickly. [TABLE 3 ABOUT HERE]

299

SC

295

Table 4 shows the results for the hourly data. Using the MAE, at the 1 hour horizon, the contest between the DICast and the ARIMA is fairly close, with DICast winning in two cases, while the ARIMA

301

is more accurate in the other two. Averaging the four sites, the mean absolute error for the DICast is 42.8

302

W/m2, while the mean absolute error for the ARIMA is 46.5 W/m2. The other models all show much

303

higher errors. The WRF-Solar-Now shows an MAE of 70.5 W/m2. The WRF-Solar-DA is more

304

accurate: the MAE is 67.5 W/m2. The cloud advection models achieve MAEs in the 83-86 W/m2 range;

305

the MADCast is slightly better than the CIRACast. The transfer function MAE averages 51.6 W/m2,

306

higher than the ARIMA. Again, the smart persistence and regression tree do poorly, consistently

307

showing the highest errors.

310

TE D

EP

309

At 2 hours, the DICast comes in first, with an error of 55 W/m2. The ARIMA runs second, with an MAE of 60.3 W/m2. The WRF-Solar-DA runs third, and the WRF-Solar-Now runs fourth. At 3-4 hours, DICast wins unambiguously, with MAEs in the range of 55-57 W/m2, while the

AC C

308

M AN U

300

311

ARIMA MAE increases to 65.8 W/m2 and 68.7 W/m2. At the 3 hour horizon, the WRF-Solar-DA and the

312

ARIMA are tied almost exactly. At 4 hours however, the WRF-Solar-DA is slightly better.

313

There is of course some variation at the individual sites. For instance, at Site 68, the results for

314

the WRF-Solar-DA, ARIMA and DICast are very similar at 2-3 hours. By comparison, at Site 69, the

315

WRF-Solar-DA is better at all horizons beyond one hour, while at Site 70, the ARIMA is better for the

316

first two hours. [TABLE 4 ABOUT HERE]

ACCEPTED MANUSCRIPT 13 The finding for the RMSE, reported in Part II of Table 4, differ in certain respects, but on the

318

whole produce similar findings. The contest is again primarily between the DICast and the ARIMA.

319

Using an average of the four Sacramento sites, the ARIMA achieves the lowest error at the 1 hour

320

horizon. The convergence point, at which the two methods achieve comparable degrees of accuracy, is 2

321

hours. The DICast is better at 3-4 hours. The WRF-Solar-DA and WRF-Solar-NOW achieve comparable

322

degrees of accuracy at 1 hour, but at all other horizons, the WRF-Solar-DA is better. The MADCast is

323

slightly better than CIRACast at 1 hour, but at these horizons the cloud advection models are not

324

particularly competitive.

SC

RI PT

317

Figures 5-6 show the errors from the DICast and ARIMA for Sacramento site 69 in a scatterplot,

325

with ground level irradiance on the horizontal axis. For the DICast, the majority of the errors fall in an

327

intermediate range, generally under 30 W/m2. However, a substantial number lie in a higher range, and

328

there are occasional extreme outliers. For the ARIMA, the distribution is more diffuse. There is no

329

obvious relationship between forecast accuracy and the level of irradiance. [FIGURES 5-6 ABOUT

330

HERE]

331

TE D

M AN U

326

332

4.2 Brookhaven

333

Tables 5-6 show the results for Brookhaven. Not all the models were available for this site. Instead, the methods used here are the smart persistence, the NWP-Solar-Now, CIRAcast, MADcast, the

335

weighted average and the ARIMA. The transfer function also could not be tested, due to extended gaps in

336

the clearness index.

AC C

337

EP

334

At the 15 minute resolution, the ARIMA achieves the highest degree of accuracy, but the errors

338

here are considerably higher than at the Sacramento sites. The weighted average achieves the second

339

smallest error, followed by WRF-Solar-Now. As the horizon increases, the accuracy of all the models

340

falls away very quickly. The ARIMA does poorly, running fourth at 30 minutes. At 45 minutes, the

341

WRF-Solar-Now does better than the alternatives. The accuracy of CIRACast falls away very quickly

342

between 30 and 45 minutes. The MADCast achieves similar degrees of accuracy at all three horizons.

343

[TABLE 5 ABOUT HERE]

ACCEPTED MANUSCRIPT 14 344

At the hourly resolution, the models are the smart persistence, MADCast, WRF-Solar-Now, WRF-Solar-DA, DICast and ARIMA. The DICast achieves the highest degree of accuracy at all

346

horizons. The forecast error from the DICast increases sharply between 1 and 2 hours, but then levels off.

347

The ARIMA places second at 1 hour, but again the accuracy falls away very rapidly. The WRF-Solar-

348

DA runs in third place at 1 hour and second at 2-4 hours. The WRF-Solar-Now runs in fourth place at 1

349

hour and third thereafter. [TABLE 6 ABOUT HERE]

351

Figures 7-8 show the DICast and ARIMA errors for Brookhaven site 13. The errors for the DICast are visibly lower than for the ARIMA. [FIGURES 7-8 ABOUT HERE]

SC

350

5. Findings from the forecasting tests

354 355

M AN U

352 353

RI PT

345

Among the statistical models, the ARIMA is superior to the two models based on the clearness index, and perhaps surprisingly to the transfer function as well. This outcome is at variance with other studies, so it

357

requires some explanation. Several other studies comparing neural networks with ARIMAs have used long

358

training periods, extended holdout periods for the forecasts, and fixed coefficients [70]. This procedure,

359

however, massively understates the power of the ARIMA. Allowing regression coefficients to vary over time

360

can capture a great deal of nonlinear variability. The main caveat is that the ARIMA works better at

361

Sacramento. At Brookhaven, the ARIMA is much less effective, except at the shortest horizons. The higher

362

errors from the regression tree are probably attributable to the use of fixed coefficients.

EP

The tests also provide further evidence on how well statistical methods compare with

AC C

363

TE D

356

364

meteorological models. At Sacramento, the convergence point between the WRF-Solar models and the

365

ARIMA is in the range of 3 hours, using the MAE. At horizons beyond 3 hours, the WRF-Solar models

366

are more accurate. At horizons of 15 minutes to 1 hour, ARIMA models are clearly superior. At 2 hours,

367

the contest is much closer, with the ARIMA slightly better for the California data sets. Conversely, at

368

Brookhaven, the convergence point appears to be about an hour. At all horizons beyond 1 hour, the

369

WRF-solar models are superior.

ACCEPTED MANUSCRIPT 15 370

The contest between the ARIMA and the DIcast is fairly close at short horizons. At the

371

Sacramento sites, the ARIMA is competitive at 1 hour using the MAE at Sacramento, and over somewhat

372

longer horizons using the RMSE. The DICast, however, is more accurate at all horizons at Brookhaven.

374

RI PT

373 6. Regressions for the physics models

375 376

As a further gauge of the relative strengths of the meteorological and cloud advection models, the actual values were regressed on the model forecasts at 1 hour, in natural logs, using a grid-search

378

correction for serial correlation. Tables 7-8 present these results for Sacramento and Brookhaven

379

respectively. The coefficients are elasticities: they express the percent change in the data in response to

380

the percent change in the model forecast. Ideally, the elasticity should equal to 1. Rho (ρ) is the

381

coefficient of fractional differencing, estimated by the serial correlation correction. Lower values of rho

382

are consistent with less structure in the residuals. [TABLES 7-8 ABOUT HERE]

M AN U

SC

377

The cloud advection models do not fit the data particularly well. the CIRAcast elasticities range

384

from 0.70 to 0.77. The adjusted R-square ranges from 0.65 to 0.71, indicating that a substantial share of

385

the variance remains unexplained. The MADcast is only somewhat better: the elasticities range from 0.71

386

to 0.83 at Sacramento, and 0.73 to 0.76 at Brookhaven. The weighted average also shows poor results.

387

The elasticities are in the range of 0.68 to 0.75.

EP

TE D

383

The WRF models yield rather good results at Sacramento. The elasticities are close to unity at

389

three of the sites, although site 67 shows a lower value. At Brookhaven, however, the WRF elasticities

390

are considerably lower. Similarly, the R-squares are high at three of the Sacramento sites, but lower at

391

Brookhaven.

392

AC C

388

The DICast shows the best results. At Sacramento, the elasticity for the DICast ranges from 0.95

393

to 1.06, averaging out to roughly unity. At Brookhaven, the elasticities are just short of 1. The R-squares

394

are in the range of 0.88 to 0.91 at Sacramento, and 0.96 at Brookhaven. The constant terms and

395

coefficients of fractional differencing are also lower, pointing to less serial correlation in the residuals.

396

ACCEPTED MANUSCRIPT 16 397 398

7. Conclusions

399 Each model has particular strengths and weaknesses. Time series models are able to predict more

RI PT

400

accurately at the shortest horizons primarily because the data is dominated by dependence between

402

proximate time points. As this dependence falls away, over longer horizons, the time series models

403

become less effective. Expressed another way, the atmosphere has moved beyond the Lagrangian integral

404

time scale at this point, i.e., that time span over which the atmosphere “remembers” its prior state.

405

Among time series models, ARIMAs with stochastic parameters are superior to fixed coefficient models.

406

While the persistence and regression tree models showed much larger errors, this does not necessarily

407

invalidate the idea of converting irradiance to an index. In this respect, there is an emerging literature on

408

clear sky models, which quantify the impact of factors ranging from the air mass, to cloud cover and

409

atmospheric turbidity [71]. A detailed comparison of clear sky models with ARIMAs, both using

410

stochastic parameters, will be the topic of a future study.

M AN U

SC

401

Cloud advection models are not found to fit the data closely, and their predictive accuracy is

412

generally lower than large-scale meteorological models. The reasons most likely have to do with the

413

larger number of causal factors taken into account in numerical weather prediction systems. Large-scale meteorological models do better for time scales beyond about 2 hours. The reason is

EP

414

TE D

411

their ability to predict the synoptic and mesoscale changes in cloud cover and weather patterns. Their

416

main limitation is that they do not capture all the short-term variability in the data. Statistical adjustment

417

enables the models to track the data more effectively. The DICast combines the physics embodied in

418

meteorological models with the ability to adjust to changing atmospheric conditions.

419

AC C

415

In a direct comparison of time series and physics-based models, the ARIMA is more accurate at

420

short horizons, while the meteorological models are more accurate as the horizon extends. The

421

convergence points lie in a range of only 1-3 hours. Comparisons for wind and wave energy have found

422

longer convergence points, in the range of 5-6 hours [72-74]. By implication, the expected integral time

423

scale for wave energy is longer than for solar. The shorter convergence points may also reflect the fact

ACCEPTED MANUSCRIPT 17 that the WRF-Solar-Now model was run in nowcasting mode, updating every hour, which is seldom

425

available for other NWP models. When NWP models are statistically adjusted, they are able to

426

outperform all the alternatives at horizons of 2-4 hours, while they are competitive at the 1-hour horizon.

427

The more general implication is that over horizons where the data exhibits high degrees of variability,

428

enabling models to adapt to changing conditions raises forecast accuracy.

429 430

433

Acknowledgements

M AN U

432

SC

431

RI PT

424

The authors thank Sacramento Municipal Utility District and Brookhaven National Laboratory for use of their solar irradiance measurements. They also thank the modelers who produced the forecasts

435

used as a comparison for this work and Jared Lee for the maps used in Figures 1-2. Finally, the authors

436

thank the reviewers, whose comments have made this a better paper.

EP AC C

437

TE D

434

ACCEPTED MANUSCRIPT 18 438

REFERENCES

439 440 441 442 443 444

[1] Federal Energy Regulatory Commission (FERC), 2012. Order No. 764. Washington, DC. Federal Energy Regulatory Center.

445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480

[3] Chow, C. W., Urquhart, N., Lave, M., Dominquez, A., Kleissl, J., Shields, J., Washom, B., 2011: Intra-hour forecasting with a total sky imager at the UC San Diego solar energy testbed. Solar Energy, 85, 2881-2893.

481 482

[14] Schertzer, D., Lovejoy, S., Schmitt, F., Chigirinskaya, Y., Marsan, D., 1997. Multifractal cascade dynamics and turbulent intermittency. Fractals. 5, 427–471.

483 484

[15] Lovejoy, S., Schertzer, D., 2013. The weather and climate: emergent laws and multifractal cascades. Cambridge, Cambridge University Press.

RI PT

[2] Tuohy, A., Zack, J., Haupt, S.E., Sharp, J., Ahlstrom, M., Dise, S., Grimit, E., Möhlren, C., Lange, M., Casado, M.G., Black, J., Marquis, M., Collier, Craig., 2015. Solar Forecasting: Methods, Challenges and Performance. IEEE Power and Energy Magazine. 13, 50-59.

SC

[4] Marquez, R., Coimbra, C.F.M., 2013: Intra-hour DNI forecasting based on cloud tracking image analysis. Solar Energy, 91, 327-336.

M AN U

[5] Marquez, R., Pedro, H.T.C., Coimbra, C.F.M., 2013: Hybrid solar forecasting method uses satellite imaging and ground telemetry as inputs to ANNs. Solar Energy, 92, 176-188. [6] Quesada-Ruiz, S., Chu, Y., Tovar-Pescador, J., Pedro, H. T. C., Coimbra, C.F.M., 2014: Cloudtracking methodology for intra-hour DNI forecasting. Solar Energy, 102, 267-275. [7] Chu, Y., Pedro, H.T.C., Coimbra, C.F.M., 2013: Hybrid intra-hour DNI forecasts with sky image processing enhanced by stochastic learning. Solar Energy, 98, 592-603.

TE D

[8] Chu, Y., Pedro, H.T.C., Li, M., Coimbra, C.F.M., 2015: Real-time forecasting of solar irradiance ramps with smart image processing. Solar Energy, 114, 91-104. [9] Lorenz, E., Kuhnert, J., Heinemann, D., 2012: Overview on irradiance and photovoltaic power prediction. Weather Matters for Energy, 429-454. [10] Kleissl, J. 2013: Solar Energy Forecast and Resource Assessment. Academic Press.

EP

[11] Mathiesen, P., Kleissl, J., 2011: Evolution of numerical weather prediction for intra-day solar forecasting in the continental United States. Solar Energy 85, 967-997.

AC C

[12] Zhang, J., Hodge, B.M., Lu, S., Hamann, H.F., Lehman, B., Simmons, J., Campos, E., Banunarayanan, V. Black, J., Tedesco, J. , 2015: Baseline and target values for regional and point PV power forecasts: Toward improved solar forecasting. Solar Energy, 122, 804-819. [13] Haupt, S.E., B. Kosovic, T. Jensen, J. Lee, P. Jimenez, J. Lazo, J. Cowie, T. McCandless, J. Pearson, G. Weiner, S. Alessandrini, L. Delle Monache, D. Yu, Z. Peng, D. Huang, J. Heiser, S. Yoo, P. Kalb, S. Miller, M. Rogers, and L. Hinkleman, 2016: The SunCast Solar Power Forecasting System: The Results of the Public-Private-Academic Partnership to Advance Solar Power Forecasting. NCAR TechnicalReport TN-526+STR, 307 pp, doi:10.5065/D6N58JR2.

ACCEPTED MANUSCRIPT 19

[16]Granger, C.W.J., 2008. Non-Linear Models: Where Do We Go Next – Time Varying Parameter Models? Studies in Nonlinear Dynamics and Econometrics 12: Article 1. http://www.bepress.cxom/snde/vol12/iss3/art1

489

[17] Bunn, D.W., 2004. Modelling Prices in Competitive Electricity Markets. New York, Wiley.

490 491 492 493 494 495 496 497 498

[18] Dutton, J.A., 1976: The Ceaseless Wind: An Introduction to the Theory of Atmospheric Motion, New York, McGraw-Hill.

499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516

[21] Jimenez, P., Hacker, J. Dudhia, J., Haupt, S.E., Ruiz-Arias, J., Gueymard, C., Thompson, G., Eidhammer, T., and A. Deng, A., 2015: WRF-Solar: An augmented NWP model for solar power prediction. Model description and clear sky assessment. Bulletin of the American Meteorological Society. 97, 7, 1249-1264. doi:10.1175/BAMS-D-14-00279.1.

517 518 519 520 521 522

[26] Delle Monache, L., Eckel, F.A., Rife, D.L., Nagarajan, B., and Searight, K., 2013: Probabilistic weather predictions with an analog ensemble. Monthly Weather Review. 141, 3489–3516, doi:10.1175/MWR-D-12-00281.1.

523 524

[28] Haupt, S.E. and B. Kosovic, 2016: Variable Generation Power Forecasting as a Big Data Problem, IEEE Transactions on Sustainable Energy. Forthcoming.

525 526 527 528

[29] Glahn, H.R., Lowry, D.A., 1972. The use of model output statistics (MOS) in objective weather Forecasting. Journal of Applied Meteorology, 11, 1203-1211.

RI PT

485 486 487 488

[19] Vallis, G.K., 2006: Atmospheric and Oceanic Fluid Dynamics: Fundamentals and Large-Scale Circulation, Cambridge University Press, Cambridge, UK.

M AN U

SC

[20] Skamarock, W.C., J.B. Klemp, J. Dudhia, D.O. Gill, D.M. Barker, M.G. Duda, X.-Y. Huang, W. Wang, and J.G. Powers, 2008: A description of the Advanced Research WRF Version 3. NCAR Tech. Note NCAR/TN-475+STR. 113 pp. doi:10.5065/D68S4MVH.

[22] Bouzerdoum, M., Mellit, A., and Massi Pavan, A., 2013: A hybrid model (SARIMA-SVM) for short-term power forecasting of a small-scale grid-connected photovoltaic plant. Solar Energy, 98, 226235.

TE D

[23] Voyant, C., Muselli, M., Paoli, C., Nivet, M.L., 2014: Numerical Weather Prediction (NWP) and hybrid ARMA/ANN to predict global radiation. Energy, 39:1, 341-355. [24] Voyant, C., Muselli, M., Paoli, C., Nivet, M.L., 2013: Hybrid methodology for hourly global radiation forecasting in Mediterranean area. Renewable Energy, 53, 1-11.

AC C

EP

[25] Myers, W., G. Wiener, S. Linden, and S. E. Haupt, 2011: A consensus forecasting approach for improved turbine hub height wind speed predictions, in Proceedings, Wind Power, 2011. Anaheim, CA, May 24, 2011.

[27] Haupt, S.E. and W.P. Mahoney, 2015: Wind Power Forecasting, IEEE Spectrum, Nov 2015, pp. 4652.

ACCEPTED MANUSCRIPT 20 529 530 531 532 533 534

[30] Peng, Z, Yoo, S., Yu, D., Huang, D., 2013. Solar irradiance forecast system based on geostationary satellites. 2013 IEEE International Conference on Smart Grid Communications, 708–713.

535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574

[32] Perez, R., Kivalov, S., Schlemmer, J. Hemker, K., Renne, D., Hoff, T.E., 2010: Validation of short and medium-term operational solar radiation forecasts for the U.S. Solar Energy, 84, 2161-2172.

575 576 577 578

[44] Miller, S.D., Forsythe, J.M., Partain, P.T., Haynes, J.M., Bankert, R.L., Sengupta, M., Mitrescu, C., Hawkins, J.D., Vonder Haar, T.H., 2014: Estimating Three-Dimensional Cloud Structure via Statistically Blended Satellite Observations. Journal of Applied Meteorology and Climatology, 53, 437–455.

RI PT

[31] Peng, Z., Yoo, S., Yu, D., Huang, D., Kalb, P, Heiser, J. 2014. 3D cloud detection and tracking for solar forecast using multiple sky imagers. In Proceedings of the 29th Annual ACM Symposium on Applied Computing, 512–517. Association for Computing Machinery 2014.

[33] Perez, R., Lorenz, E., Pelland, S., Beautharnois, M., Van Knowe, G., Hemker, K., Heinemann, D., Remund, J., Muller, S.C. Traumiller W., et al., 2013: Comparison of numerical weather prediction solar irradiance forecasts in the U.S., Canada and Europe. Solar Energy, 94, 305-326.

SC

[34] Beyer, H. G., Costanzo, C., Heinemann, D., 1996: Modifications of the Heliosat procedure for irradiance estimates from satellite data. Solar Energy, 56, 121–207.

M AN U

[35] Bilionis, I., Constantinescu, E. M., and Anitescu, M., 2014: Data-driven model for solar irradiation based on satellite observations. Solar Energy, 110, 22-38. [36] Cros, S., Liandrat, O., Sébastien, N., Schmutz, N., 2014: Extracting cloud motion vectors from satellite images for solar power forecasting. Geoscience and Remote Sensing Symposium (IGARSS), 2014 IEEE International, 4123-4126.

[37] Stauffer, D.R., Seaman, N.L., 1990: Use of four-dimensional data assimilation in a limited area mesoscale model. Part I: Experiments with synoptic-scale data. Monthly Weather Review, 181, 1250-1277.

TE D

[38] Stauffer, D.R., Seaman, N.L., 1994: Multiscale four-dimensional data assimilation, Journal of Applied Meteorology. 33, 416-434 [39] Tapakis, R., Charalambides, A.G., 2013: Equipment and methodologies for cloud detection and classification: A review. Solar Energy, 95, 392-430.

EP

[40] Zagouras, A., Kazantzidis, A., Nikitidou, E., Argiriou, A.A., 2013: Determination of measuring sites for solar irradiance, based on cluster analysis of satellite-derived cloud estimations. Solar Energy, 97, 111.

AC C

[41] Hammer, A., Heinemann, D., Lorenz, E., Lückehe, B., 1999: Short-term forecasting of solar radiation: a statistical approach using satellite data. Solar Energy, 67, 139-150. [42] Lorenz, E., Hammer, A., Heinemann, D., 2004: Short term forecasting of solar radiation based on satellite data. EUROSUN2004, ISES Europe Solar Congress, 841-848. [43] Miller, S. D., Rogers, M.A., Heidinger, A.K., Laszlo, I., Sengupta, M., 2012. Cloud advection schemes for short-term satellite-based insolation forecasts. Proceedings of the World Renewable Energy Forum, 17 May 2012, Denver, CO, 1963-1967.

ACCEPTED MANUSCRIPT 21 [45] Rogers, M.A., Miller, S.D., Haynes, J.M., Heidinger, A. K., Haupt, S.E., Sengupta, M., 2015. Improvements in satellite-derived short-term insolation forecasting: Statistical comparisons, challenges for advection-based forecasts, and new techniques. Sixth conference on Weather, Climate, and the New Energy Economy at the 95th American Meteorological Society Annual Meeting, Phoenix, AZ, 6 Jan 2015.

584 585 586

[46] Heidinger, A. K., Foster, M.J., Walther, A., Zhao, X., 2013: The Pathfinder Atmospheres Extended AVHRR Climate Data Set. Bulletin of the American Meteorological Society. doi:10.1175/BAMS-D-12-00246.1.

SC

[47] Auligné, T., 2014a: Multivariate minimum residual method for cloud retrieval. Part I: Theoretical aspects and simulated observations experiments. Monthly Weather Review, 142, 4383-4398. doi: 10.1175/ mWR-D-13-00172.1. [48] Auligné, T., 2014b: Multivariate minimum residual method for cloud retrieval. Part II: real observations experiments. Monthly Weather Review, 142, 4399-4415. doi: 10.1175/ mWR-D-1300173.1.

M AN U

587 588 589 590 591 592 593 594

RI PT

579 580 581 582 583

595 596

[49] Box, G.E.P., Jenkins, G.M., 1976. Time Series Analysis: Forecasting and Control. San Francisco, Holden-Day.

597 598 599 600 601 602

[50] Reikard, G., 2009. Predicting Solar Radiation at High Resolutions: A Comparison of Time Series Forecasts. Solar Energy. 83, 342-349.

603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625

[52] Lopez, G., Batlles, F.J., Tovar-Pescador, J., 2005: Selection of input parameters to model direct solar irradiance by using artificial neural networks. Energy, 30, 1675-1684.

TE D

[51] Nobre, A.M., Severiano, C.A., Karthik, S. Marek Kubis, M., Zhao , L., Martins, F.R., Pereira, E.B., Rüther R., Reindl T., 2016. PV power conversion and short-term forecasting in a tropical, densely-built environment in Singapore. Renewable Energy, 94, 496-509.

EP

[53] McCandless, T. C., Haupt, S. E., Young, G.S., 2014: Short Term Solar Radiation Forecasts Using Weather Regime-Dependent Artificial Intelligence Techniques, 12th Conference on Artificial Intelligence: Applications of Artificial Intelligence Methods for Energy, Atlanta, GA, American Meteorological Society, J3.5.

AC C

[54] McCandless, T. C., Haupt, S. E., Young, G.S., 2016: A Regime-Dependent Artificial Neural Network Technique for Short-Range Solar Irradiance Forecasting. Renewable Energy, 89, 351-359. [55] McCandless, T.C., Young, G.S., Haupt, S.E., Hinkelman, L.M., 2016. Regime-Dependent Short-Range Solar Irradiance Forecasting. Journal of Applied Meteorology and Climatology, 55, 15991613. [56] Fu, C.L., and Cheng, H.Y., 2013: Predicting solar irradiance with all-sky image features via regression. Solar Energy, 97, 537-550. [57]Mellit, A., 2008: Artificial Intelligence Technique for Modeling and Forecasting of Solar Radiation Data: A Review. International Journal Artificial Intelligence and Soft Computing, 1, 52-76. [58] Mellit, A., Massi Pavan, A., Lughi, V., 2014: Short-Term Forecasting of Power Production in a Large-Scale Photovoltaic Plant, Solar Energy, 105, 401-413.

ACCEPTED MANUSCRIPT 22 626 627 628 629

[59] Notton, G., Paoli, C., Vasileva, S., Nivet, M.L., Canaletti, J-L., Cristofari, C., 2012: Estimation of hourly global solar irradiation on tilted planes from horizontal one using artificial neural networks. Energy, 39, 166-179. [60] Bhardwaj, S., Sharma, V., Srivastava, S., Sastry, O. S., Bandyopadhyay, B., Chandel, S. S., Gupta, J.R.P., 2013: Estimation of solar radiation using a combination of Hidden Markov Model and generalized Fuzzy model. Solar Energy, 93, 43-54.

654 655

[67] Kalman, R.E., 1960. A new approach to linear filtering and prediction problems. Transactions of the American Society of Mechanical Engineers, Journal of Basic Engineering. 83D, 35-45.

656 657

[68] Rossi, B., Inoue, A., 2012. Out-of-sample forecast tests robust to the choice of window size. Journal of Business and Economic Statistics. 30, 432-453.

658 659 660 661 662 663 664 665 666 667 668 669 670 671 672

[70] Pedro, H.T.C., Coimbra, C.F.M., 2012: Assessment of forecasting techniques for solar power prediction with no exogenous inputs. Solar Energy, 86, 2017-2028.

RI PT

630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653

[61] Diagne, M., David, M., Lauret, P., Boland, J., Schmutz, N., 2013: Review of solar irradiance forecasting methods and a proposition for small-scale insular grids. Renewable and Sustainable Energy Reviews, 27, 65-76.

SC

[62] Inman, R. H., Pedro, H.T.C., Coimbra, C.F.M., 2013: Solar forecasting methods for renewable energy integration. Progress in Energy and Combustion Science, 39, 535-576.

M AN U

[63] Fernandez, E., F. Almonacid, N. Sarmah, P. Rodrigo, T.K. Mallick, Perez-Higueras, P., 2014: A model based on artificial neuronal network for the prediction of the maximum power of a low concentration photovoltaic module for building integration. Solar Energy, 100, 148-158. [64] Almonacid, F., Pérez-Higueras, P. J., Fernández, E. F., Hontoria, L., 2014: A methodology based on dynamic artificial neural network for short-term forecasting of the power output of a PV generator. Energy Conversion and Management, 85, 389-398. [65] Morf, H., 2014: Sunshine and cloud cover prediction based on Markov processes. Solar Energy, 110, 615-626.

AC C

EP

TE D

[66] Quinlan, J.R., 1996: Improved use of continuous attributes in C4.5. Journal of Artificial Intelligence Research, 4, 77-90.

[71] Reno, M.J., Hansen, C.W., Stein, J.S., 2012. Global Horizontal Clear Sky Models: Implementation and Analysis. Sandia National Laboratory report SAND 2012-2389. [72] Reikard, G., Pinson, P., Bidlor, J.R., 2011. Forecasting Ocean Wave Energy: The ECMWF wave model and time series methods. Ocean Engineering. 38, 1089-1099. [73] Foley, A.M, Leahy, P.G., Marvuglia, A., McKeough, E.J., 2012. Current methods and advances in forecasting of wind power generation. Renewable Energy, 37, 1-8. [74] Reikard, G., Robertson, B., Bidlot, J.R., 2015. Combining wave energy with wind and solar: Shortterm forecasting. Renewable Energy. 81, 2015, 442-456.

ACCEPTED MANUSCRIPT

Model

RI PT

Table 1: Glossary of the Forecasting Models

Description

SC

Numerical Weather Prediction models

Temporal Resolution

Weather Research and Forecasting model. Now-casting version, 9 km resolution. Initial and boundary conditions calculated using the National Centers for Environmental Prediction (NCEP) Rapid Refresh model.

15 minutes, 1 hour

WRF-Solar-DA

Weather Research and Forecasting model. Day ahead, 3 km resolution. Initial and boundary conditions calculated using NCEP’s High Resolution Rapid Refresh system for the first 15 hours, transitioning to the Rapid Refresh at longer horizons.

1 hour

DICast

Dynamic Integrated Forecast system, using statistical adjustment and blending 1 hour of multiple NWP model outputs. The model forecasts are post-processed to correct the bias using the recent forecast error. Then, several of the adjusted NWP forecasts are combined using optimized weights.

Cloud Advection Models

AC C

EP

TE D

M AN U

WRF-Solar-NOW

CIRACast

Combines cloud fields identified by satellite observations with wind forecasts from the NOAA Global Forecast System.

15 minutes, 1 hour

MADCast

Multisensor Advection Diffusion algorithm.

15 minutes, 1 hour

ACCEPTED MANUSCRIPT

Combined Models Weighted average of the WRF-Solar-Now, CIRACast and MADCast.

RI PT

Weighted Average Time Series Models

15 minutes, 1 hour

Autoregressive, integrated, moving average model, with time-varying coefficients. Specification: ARIMA (1,0,0)(1,1,0). Coefficient variation estimated using a moving window.

15 minutes, 1 hour

Transfer Function

ARIMA with the clearness index as an input. The clearness index is forecasted using a regression on proximate lags.

15 minutes, 1 hour

Smart Persistence

The clearness index is set equal to its previous value, while the solar angle changes In relation to the specific time of day and day of year.

15 minutes, 1 hour

Regression tree

The clearness index, forecasted using a regression tree.

15 minutes, 1 hour

AC C

EP

TE D

M AN U

SC

ARIMA

ACCEPTED MANUSCRIPT

Table 2: The Data Sacramento Site

Start date

End date

Usable Observations

January 2, 2015, 02:00 hours January 2, 2015, 02:00 hours January 2, 2015, 02:00 hours January 2, 2015, 02:00 hours

April 28, 2016, 23:00 hours April 28, 2016, 23:00 hours April 28, 2016, 23:00 hours April 28, 2016, 23:00 hours

19441 19523 19536 17701

January 17, 2015, 20:00 hours January 17, 2015, 20:00 hours January 17, 2015, 20:00 hours January 17, 2015, 20:00 hours

June 6, 2016, 0:100 hours June 6, 2016, 0:100 hours June 6, 2016, 0:100 hours April 14, 2016, 10:00 hours

5232 5305 5215 3056

67 68 69 70

Site Site Site Site

67 68 69 70

M AN U

Hourly resolution

Brookhaven 15-minute resolution

February 7, 2015, 01:00 hours February 7, 2015, 01:00 hours February 7, 2015, 01:00 hours

Hourly resolution

EP

February 6, 2015, 20:00 hours February 6, 2015, 20:00 hours February 6, 2015, 20:00 hours

AC C

Site 13 Site 18 Site 24

TE D

Site 13 Site 18 Site 24

SC

Site Site Site Site

RI PT

15-minute resolution

April 12, 2016, 14:00 hours April 12, 2016, 14:00 hours April 12, 2016, 14:00 hours

15714 15170 15808

April 12, 2016, 14:00 hours April 12, 2016, 14:00 hours April 12, 2016, 14:00 hours

3797 3659 3815

ACCEPTED MANUSCRIPT

Table 3: Model Accuracy, 15 minute resolution, Sacramento

Site

30 minutes

Sacramento site 67 Smart Persistence Regression tree NWP-Solar-NOW CIRACast MADCast Weifhted average ARIMA Transfer Function

86.5 78.1 67.6 70.7 75.6 77.7 42.4 42.9

99.8 85.7 67.8 71.4 81.9 86.6 54.8 57.2

Sacramento site 68 Smart Persistence Regression tree NWP-Solar-NOW CIRACast MADCast Weifhted average ARIMA Transfer Function

85.9 73.6 83.4 84.8 70.8 71.1 41.1 42.7

45 minutes

AC C

EP

TE D

M AN U

15 minutes

SC

Forecast Horizon

RI PT

Part 1: The mean absolute error (watts per meter squared)

98.5 80.7 80.6 85.6 74.5 84.7 54.3 57.6

106.2 94.2 66.5 70.9 84.5 88.6 61.5 65.2

111.8 88.3 80.4 84.6 80.8 91.2 60.7 65.6

ACCEPTED MANUSCRIPT

Table 3, continued Forecast Horizon

Sacramento site 69 Smart Persistence Regression tree NWP-Solar-NOW CIRACast MADCast Weifhted average ARIMA Transfer Function

81.6 68.5 68.1 71.9 80.9 72.5 41.4 42.7

94.8 75.2 68.6 72.8 83.1 80.1 54.2 56.9

Sacramento site 70 Smart Persistence Regression tree NWP-Solar-NOW CIRACast MADCast Weifhted average ARIMA Transfer Function

79.1 67.6 65.1 74.3 60.6 70.8 42.2 43.4

TE D

EP AC C

Average, Four sites Smart Persistence Regression tree NWP-Solar-NOW CIRACast MADCast Weifhted average ARIMA

83.3 72.0 71.1 75.4 72.0 73.0 41.8

45 minutes

RI PT

30 minutes

M AN U

15 minutes

SC

Site

105.7 83.7 68.5 73.2 75.5 85.7 60.9 64.8

92.2 72.3 64.3 75.5 65.6 79.4 53.1 56.9

104.9 81.6 64.5 77.3 69.5 86.9 60.7 65.1

96.3 78.5 70.3 76.3 76.3 82.7 54.1

107.2 87.0 70.0 76.5 77.6 88.1 61.0

ACCEPTED MANUSCRIPT

57.2

65.2

EP

TE D

M AN U

SC

RI PT

42.9

AC C

Transfer Function

ACCEPTED MANUSCRIPT

Table 3, continued

Site

30 minutes

Sacramento site 67 Smart Persistence Regression tree NWP-Solar-NOW CIRACast MADCast Weifhted average ARIMA Transfer Function

126.5 146.7 159.9 130.1 124.6 117.4 72.4 72.9

152.9 158.4 155.5 155.2 123.8 136.8 92.7 91.5

Sacramento site 68 Smart Persistence Regression tree NWP-Solar-NOW CIRACast MADCast Weifhted average ARIMA Transfer Function

128.1 133.4 169.4 135.2 115.7 119.3 71.8 72.4

45 minutes

AC C

EP

TE D

M AN U

15 minutes

SC

Forecast Horizon

RI PT

Part 2: The root mean squared error

142.8 139.8 172.4 140.1 116.2 125.1 90.5 92.1

157.5 170.2 148.7 145.8 124.8 130.1 103.2 100.9

156.8 150.7 169.1 166.1 115.4 126.7 99.6 102.7

ACCEPTED MANUSCRIPT

Table 3, continued. Forecast Horizon

Sacramento site 69 Smart Persistence Regression tree NWP-Solar-NOW CIRACast MADCast Weifhted average ARIMA Transfer Function

121.1 137.1 140.1 146.6 119.2 111.2 70.2 70.8

143.7 150.4 149.8 156.7 123.3 124.8 88.3 89.9

Sacramento site 70 Smart Persistence Regression tree NWP-Solar-NOW CIRACast MADCast Weifhted average ARIMA Transfer Function

123.8 127.7 170.1 129.2 129.5 114.9 72.0 72.5

TE D

EP AC C

Average, Four sites Smart Persistence Regression tree NWP-Solar-NOW CIRACast MADCast Weifhted average ARIMA Transfer Function

124.9 136.2 159.9 135.3 122.3 115.7 71.6 72.1

45 minutes

RI PT

30 minutes

M AN U

15 minutes

SC

Site

158.3 265.4 161.1 168.9 124 129.7 98.1 100.7

145.5 134.6 167.2 140.4 123.2 126.9 89.3 90.8

167.4 145.2 165.7 163.1 130.4 136.2 99.1 110.7

146.2 145.8 161.2 148.1 121.6 128.4 90.2 91.1

160.0 182.9 161.2 161.0 123.7 130.7 100.0 103.8

ACCEPTED MANUSCRIPT

Table 4: Model Accuracy, hourly resolution, Sacramento

Forecast Horizon 1 hour

2 hours

Site

125.1 100.1 85.0 87.9 95.1 66.1 66.0 49.1 47.4 53.4

SC

153.9 130.5 98.5 73.8 91.7 65.6 … 43.1 61.1 70.6

M AN U

TE D

EP

Sacramento 68 Smart Persistence Regression tree CIRACast MADCast Weighted average NWP-Solar-NOW NWP-Solar-DA DICast ARIMA Transfer Function

120.4 103.9 88.5 72.1 89.9 67.3 … 38.9 46.6 52.4

AC C

Sacramento 67 Smart Persistence Regression tree CIRACast MADCast Weighted average NWP-Solar-NOW NWP-Solar-DA DICast ARIMA Transfer Function

RI PT

Part 1: Mean absolute error (watts per meter squared)

158.8 128.2 91.2 89.2 104.8 81.7 65.6 66.2 62.1 70.6

3 hours

4 hours

182.7 … … 75.6 99.1 64.8 … 44.5 65.5 75.5

195.4 … … 78.1 90.3 64.3 … 44.6 68.3 80.2

186.4 … … 90.1 112.7 80.8 66.6 66.1 68.1 79.4

198.6 … … 93.3 106.7 81.7 67.1 66.8 71.8 84.4

ACCEPTED MANUSCRIPT

Table 4, continued.

2 hours

3 hours

Site

117.8 89.4 74.9 75.2 90.7 67.3 63.5 44.5 43.3 49.9

SC

Sacramento70 Smart Persistence Regression tree CIRACast MADCast Weighted average NWP-Solar-NOW NWP-Solar-DA DICast ARIMA Transfer Function

EP AC C

157.4 122.3 95.8 73.2 92.4 67.1 49.1 43.2 62.8 70.1

M AN U

130.5 93.3 90.1 71.7 87.4 67.9 49.8 36.2 48.5 54.1

TE D

Sacramento 69 Smart Persistence Regression tree CIRACast MADCast Weighted average NWP-Solar-NOW NWP-Solar-DA DICast ARIMA Transfer Function

RI PT

Forecast Horizon 1 hour

156.3 113.3 77.6 76.1 95.4 63.3 61.5 58.9 55.9 66.9

4 hours

187.2 … … 75.4 100.4 68.1 49.9 44.1 69.1 78.3

201.9 … … 77.9 93.2 66.1 50.1 49.3 72.1 83.4

186.8 … … 77.1 103.6 63.1 67.5 58.6 60.6 67.4

202.1 … … 77.6 92.7 63.2 67.1 58.9 62.7 68.8

ACCEPTED MANUSCRIPT

Table 4, continued

TE D

Part 2: Root mean squared error

163.2 145.2 128.1 114.1 136.2 114.3 … 73.8 76.8 90.7

AC C

Sacramento 67 Smart Persistence Regression tree CIRACast MADCast Weighted average NWP-Solar-NOW NWP-Solar-DA DICast ARIMA Transfer Function

EP

Forecast Horizon 1 hour

Site

2 hours

199.1 174.4 144.9 116.6 139.9 111.6 … 74.9 96.7 113.1

185.8 … … 79.6 104.0 69.2 61.3 53.3 65.8 75.2

199.5 … … 81.7 95.7 68.8 61.4 54.9 68.7 79.2

3 hours

4 hours

230.6 … … 119.2 151.8 116.4 … 75.4 106.1 123.8

247.4 … … 119.9 152.9 119.7 … 76.1 112.5 132.5

RI PT

156.6 123.6 90.8 78.1 96.1 69.4 58.7 52.9 60.5 69.6

SC

123.5 96.7 84.6 76.7 90.8 67.2 59.8 42.2 46.4 52.5

M AN U

Average of four sites Smart Persistence Regression tree CIRACast MADCast Weighted average NWP-Solar-NOW NWP-Solar-DA DICast ARIMA Transfer Function

ACCEPTED MANUSCRIPT

Table 4, continued

2 hours

3 hours

Site

130.2 136.2 129.5 112.8 121.9 91.2 80.7 62.9 77.3 90.8

SC

Sacramento 69 Smart Persistence Regression tree CIRACast MADCast Weighted average NWP-Solar-NOW NWP-Solar-DA DICast ARIMA Transfer Function

EP AC C

211.4 177.1 140.6 146.4 162.3 140.5 132.3 129.1 101.4 114.8

M AN U

177.8 147.7 129.8 145.5 147.7 131.2 131.1 90.2 77.1 87.5

TE D

Sacramento 68 Smart Persistence Regression tree CIRACast MADCast Weighted average NWP-Solar-NOW NWP-Solar-DA DICast ARIMA Transfer Function

RI PT

Forecast Horizon 1 hour

203.6 169.1 142.7 114.6 139.3 110.3 83.9 63.8 100.7 117.4

4 hours

241.2 … … 148.2 174.2 148.8 135.6 129.7 113.8 129.4

256.3 … … 149.6 181.2 148.7 138.3 137.8 120.2 137.6

236.1 … … 116.9 153.2 108.1 88.5 73.9 113.8 129.4

255.2 … … 117.8 155.1 105.3 91.5 79.3 118.0 139.3

ACCEPTED MANUSCRIPT Forecast Horizon 1 hour

2 hours

180.9 159.3 131.9 161.2 152.1 143.3 156.1 120.8 74.6 82.2

221.3 187.2 140.1 160.7 174.9 153.7 143.6 121.6 96.5 105.2

Average of four sites Smart Persistence Regression tree CIRACast MADCast Weighted average NWP-Solar-NOW NWP-Solar-DA DICast ARIMA Transfer Function

163.0 147.1 129.8 133.4 139.5 120.0 122.6 86.9 76.5 87.8

208.9 177.0 142.1 134.6 154.1 129.0 119.9 97.4 98.8 112.6

AC C

EP

TE D

M AN U

SC

Sacramento70 Smart Persistence Regression tree CIRACast MADCast Weighted average NWP-Solar-NOW NWP-Solar-DA DICast ARIMA Transfer Function

3 hours

4 hours

251.4 … … 160.4 185.9 150.8 143.7 123.6 104.5 118.2

267.3 … … 161.5 185.9 146.3 143.7 124.5 113.2 126.3

239.8 … … 136.2 166.3 131.0 122.6 100.6 109.6 125.2

256.6 … … 137.2 168.8 130.0 124.5 104.4 116.0 133.9

RI PT

Table 4, continued.

ACCEPTED MANUSCRIPT

Table 5: Model Accuracy, 15-minute resolution, Brookhaven

Forecast Horizon

Brookhaven 24 Smart Persistence NWP-Solar-NOW CIRACast MADCast Weighted average ARIMA

81.6 81.5 …

94.2 79.1 65.1

87.2 81.3



90.8 78.7 64.9

45 minutes

SC

95.1 81.8 66.4

100.1 85.2 82.6 97.7 89.6 89.1

M AN U



TE D

Brookhaven 18 Smart Persistence NWP-Solar-NOW CIRACast MADCast Weighted average ARIMA

30 minutes

87.9 84.2

EP

Brookhaven 13 Smart Persistence NWP-Solar-NOW CIRACast MADCast Weighted average ARIMA

15 minutes

AC C

Site

RI PT

Part 1: Mean absolute error (watts per meter squared)

109.1 86.8 93.3 96.8 93.2 101.8

93.9 83.4 83.5 96.1 86.4 88.4

101.1 84.3 93.8 95.2 89.4 100.9

93.5 82.1 84.4 91.7 85.1 88.2

101.1 82.9 95.1 90.9 87.7 100.5

ACCEPTED MANUSCRIPT

Table 5, continued

Forecast Horizon

Site

15 minutes

Average of all sites Smart Persistence NWP-Solar-NOW CIRACast MADCast Weighted average ARIMA

45 minutes

95.8 83.6 83.5 95.2 87.0 88.6

103.8 84.7 94.1 94.3 90.1 101.1

30 minutes

45 minutes

147.7 135.1 124.8 155.8 134.1 136.1

156.5 134.7 149.3 155.7 135.7 139.6

142.6 142.2 125.3 154.9 132.2 134.1

149.5 143.1 149.8 155.1 132.9 138.5

85.6 82.3 93.4 79.9 65.5

Site

15 minutes

Brookhaven 13 Smart Persistence NWP-Solar-NOW CIRACast MADCast Weighted average ARIMA

SC

Forecast Horizon

RI PT



Part 2: Root mean squared error

M AN U

132.9 126.5



TE D

155.4 126.5 103.7

125.2 138.1



AC C

EP

Brookhaven 18 Smart Persistence NWP-Solar-NOW CIRACast MADCast Weighted average ARIMA

30 minutes

154.6 125.5 102.6

ACCEPTED MANUSCRIPT

Table 5, continued

Forecast Horizon

Site

15 minutes

Brookhaven 24 Smart Persistence NWP-Solar-NOW CIRACast MADCast Weighted average ARIMA

126.2 137.6 150.7 124.6 102.3

Average of all sites Smart Persistence NWP-Solar-NOW CIRACast MADCast Weighted average ARIMA

140.9 139.6 127.3 150.2 129.6 133.1

148.4 139.9 152.9 150.8 130.1 138.3

143.7 139.0 125.8 153.6 132.0 134.4

151.5 139.2 150.7 153.9 132.9 138.8

SC

128.1 134.1

45 minutes

RI PT



30 minutes

AC C

EP

TE D

M AN U

153.6 125.5 102.9

ACCEPTED MANUSCRIPT

Table 6: Model Accuracy, hourly resolution, Brookhaven

Forecast Horizon 2 hours

115.1 98.6 88.4 76.7 51.1 71.4

Brookhaven 18 Smart Persistence MADCast NWP-Solar-NOW NWP-Solar-DA DICast ARIMA

109.6 96.5 86.4 76.1 49.6 70.1

EP AC C

Brookhaven 24 Smart Persistence MADCast NWP-Solar-NOW NWP-Solar-DA DICast ARIMA

TE D

Brookhaven 13 Smart Persistence MADCast NWP-Solar-NOW NWP-Solar-DA DICast ARIMA

110.6 92.1 84.6 73.5 48.2 70.2

3 hours

4 hours

171.6 102.1 89.1 77.7 63.8 145.1

187.4 104.7 88.5 79.1 64.2 151.4

142.6 98.1 85.7 76.3 60.3 112.2

165.1 100.5 85.5 76.9 60.7 124.4

180.1 102.8 86.4 78.2 61.4 149.1

142.7 94.4 83.9 73.7 58.8 113.6

165.6 96.9 85.8 74.3 59.3 126.2

171.1 99.1 85.9 75.3 59.8 150.3

153.9 99.9 87.1 77.1 63.2 114.8

SC

1 hour

M AN U

Site

RI PT

Part 1: Mean absolute error (watts per meter squared)

ACCEPTED MANUSCRIPT

Table 6, continued Forecast Horizon Average of all sites Smart Persistence MADCast NWP-Solar-NOW NWP-Solar-DA DICast ARIMA

2 hours

Part 2: Root mean squared error Forecast Horizon

Brookhaven 18 Smart Persistence MADCast NWP-Solar-NOW NWP-Solar-DA DICast ARIMA

TE D

163.4 155.1 143.4 124.7 82.7 122.7

159.6 155.7 146.6 124.9 81.3 119.1

4 hours

167.4 99.8 86.8 76.3 61.3 131.9

179.5 102.2 86.9 77.5 61.8 150.3

3 hours

4 hours

200.6 155.5 140.8 125.1 101.8 179.3

226.8 158.1 144.1 125.6 102.9 198.8

248.7 159.4 144.3 127.6 103.8 219.8

196.2 156.6 145.3 125.2 99.2 172.4

222.1 159.1 146.7 125.7 99.4 186.5

243.1 160.6 146.6 127.4 100.6 215.7

2 hours

EP

Brookhaven 13 Smart Persistence MADCast NWP-Solar-NOW NWP-Solar-DA DICast ARIMA

1 hour

AC C

Site

146.4 97.5 85.6 75.7 60.8 113.5

SC

111.8 95.7 86.5 75.4 49.6 70.6

3 hours

RI PT

1 hour

M AN U

Site

ACCEPTED MANUSCRIPT

Table 6, continued

Forecast Horizon 1 hour

2 hours

160.4 151.8 142.2 122.5 79.6 119.6

195.1 152.9 139.8 122.8 96.8 169.2

Average of all sites Smart Persistence MADCast NWP-Solar-NOW NWP-Solar-DA DICast ARIMA

161.1 154.2 144.1 124.0 81.2 120.5

197.3 155.0 142.0 124.4 99.3 173.6

AC C

EP

TE D

M AN U

SC

Brookhaven 24 Smart Persistence MADCast NWP-Solar-NOW NWP-Solar-DA DICast ARIMA

3 hours

4 hours

220.2 156.8 144.3 123.1 97.3 185.7

240.5 157.3 144.9 124.8 98.1 217.6

223.0 158.0 145.0 124.8 99.9 190.3

244.1 159.1 145.3 126.6 100.8 217.7

RI PT

Site

ACCEPTED MANUSCRIPT

Table 7: Regression coefficients, Sacramento sites

CIRAcast

MADcast

Weighted Average NWP-Solar-NOW

NWP-Solar-DA

DICast

SC

Site

RI PT

Statistics are regression coefficients, unless otherwise indicated. Rho is the coefficient of fractional differencing.

Constant

0.31

0.84

0.82

Elasticity

0.89

0.82

0.74

R-bar-square

0.65

0.86

0.57

Rho

0.81

0.75

TE D

M AN U

Sacramento 67

-0.39

0.75 …

1.05

0.81 …

0.90

0.71 …

0.11

EP

0.86

1.21 …

Constant

1.45

Elasticity

0.71

R-bar-square Rho

AC C

Sacramento 68 1.04

0.89

-0.41

-0.41

0.16

0.78

0.75

1.03

1.02

0.95

0.69

0.84

0.59

0.87

0.87

0.89

0.91

0.91

0.89

0.89

0.81

0.77

ACCEPTED MANUSCRIPT

Table 7, continued

CIRAcast

MADcast

Weighted Average NWP-Solar-NOW

Sacramento 69 Constant

1.05

0.71

0.86

Elasticity

0.77

0.83

0.72

R-bar-square

0.70

0.85

0.58

Rho

0.91

0.84

0.88

Constant

1.66

1.58

Elasticity

0.70

0.71

R-bar-square

0.66

Rho

0.92

NWP-Solar-DA

RI PT

Site

-0.43

1.05

1.06

0.89

0.90

0.91

0.78

0.31

0.31

0.71

1.96

0.11

0.68

0.95

0.63

0.95

0.81

0.58

0.87

0.74

0.88

0.95

0.91

0.91

0.93

0.61

SC

-0.43

M AN U

0.96

TE D EP

1.43

AC C

Sacramento 70

-0.14

DICast

ACCEPTED MANUSCRIPT Table 8: Regression coefficients, Brookhaven sites Statistics are regression coefficients, unless otherwise indicated. Rho is the coefficient of fractional differencing.

MADcast

Weighted Average NWP-Solar-NOW

Brookhaven 13 Constant

1.04

1.27

Elasticity

0.74

0.71

R-bar-square

0.84

0.79

Rho

0.91

0.84

DICast

RI PT

Site

-0.05

0.71

0.96

0.79

0.95

0.84

0.44

1.27

1.30

-0.12

0.71

0.71

0.96

0.84

0.79

0.80

0.96

0.91

0.83

0.83

0.31

1.12

1.33

1.33

-0.06

0.73

0.71

0.71

0.97

R-bar-square

0.84

0.79

0.81

0.96

Rho

0.91

0.82

0.83

0.29

0.92

Elasticity

0.76

Rho

Constant Elasticity

AC C

Brookhaven 24

EP

R-bar-square

TE D

Constant

M AN U

Brookhaven 18

SC

1.27

AC C

EP

TE D

M AN U

SC

RI PT

ACCEPTED MANUSCRIPT

AC C

EP

TE D

M AN U

SC

RI PT

ACCEPTED MANUSCRIPT

RI PT

ACCEPTED MANUSCRIPT

1200

SC

1000

M AN U

800 600

EP

0

AC C

200

TE D

400

RI PT

ACCEPTED MANUSCRIPT

1200

M AN U

SC

1000 800

0

EP

200

AC C

400

TE D

600

ACCEPTED MANUSCRIPT

600

RI PT

500

SC

400

M AN U

300

200

200

400

EP

0

AC C

0

TE D

100

600

800

1000

1200

ACCEPTED MANUSCRIPT

600

RI PT

500

SC

400

M AN U

300

200

200

400

EP

0

AC C

0

TE D

100

600

800

1000

1200

ACCEPTED MANUSCRIPT

800

RI PT

700 600

SC

500 400

M AN U

300 200

200

400

EP

0

AC C

0

TE D

100

600

800

1000

1200

ACCEPTED MANUSCRIPT

900

RI PT

800 700 600

SC

500

M AN U

400 300 200

200

400

EP

0

AC C

0

TE D

100

600

800

1000

1200

ACCEPTED MANUSCRIPT

Figure captions and titles.

RI PT

Title: Figure 1: Sites 67-70, Sacramento Municipal Utility District (SMUD).

SC

Title: Figure 2: The Long Island Solar Farm, at Brookhaven National Laboratory, New York.

Title: Title: Figure 3: Global Horizontal Irradiance at Sacramento site 69.

M AN U

Caption: Left scale: watts per meter squared. Resolution: Hourly. Time span: May 1 to May 31, 2015. Source: Sacramento Municipal District and National Center for Atmospheric Research.

Title: Figure 4: Global Horizontal Irradiance at Brookhaven site 13.

TE D

Caption: Left scale: watts per meter squared. Resolution: Hourly. Time span: May 1 to May 31, 2015. Source: Brookhaven National Laboratory.

Title: Figure 5: The DICast forecast error versus irradiance, Sacramento site 69

AC C

EP

Caption: Forecast horizon: 1 hour. Left scale: absolute error, in watts per meter squared. Horizontal scale: ground level irradiance, in watts per meter squared.

Title: Figure 6: The ARIMA forecast error versus irradiance, Sacramento site 69 Caption: Forecast horizon: 1 hour. Left scale: absolute error, in watts per meter squared. Horizontal scale: ground level irradiance, in watts per meter squared.

ACCEPTED MANUSCRIPT

Title: Figure 7: The DICast forecast error versus irradiance, Brookhaven site 13

RI PT

Caption: Forecast horizon: 1 hour. Left scale: absolute error, in watts per meter squared. Horizontal scale: ground level irradiance, in watts per meter squared.

Title: Figure 8: The ARIMA forecast error versus irradiance, Brookhaven site 13

AC C

EP

TE D

M AN U

SC

Caption: Forecast horizon: 1 hour. Left scale: absolute error, in watts per meter squared. Horizontal scale: ground level irradiance, in watts per meter squared.

ACCEPTED MANUSCRIPT

Meteorological and time series are used to forecast. Models with fixed and time-varying coefficients are tested.

RI PT

Time series models are more accurate at short horizons, while meteorological models are more effective at longer horizons.

AC C

EP

TE D

M AN U

SC

The optimal model is the Dynamic Integrated Forecast (DICast) system, which combines meteorological models with statistical adjustments.

ACCEPTED MANUSCRIPT 1

RENE-D-16-3668: Responses to the Reviewers

RI PT

In responding to the reviewers, we face the following dilemma. Three of the reviews are highly favorable, one is moderately favorable, and one is critical. For this reason, we are responding to the more favorable reviews first. We have prepared detailed comments in response to the critical review; these come at the end.

SC

Several reviewers commented on the titles and labels for the figures. We chose to put these into a separate document, since the titles and figure captions are typeset separately.

Reviewer #1:

M AN U

Reviewer 1: The abstract should include a brief introduction of research motivation and purpose, some description of the research process and major methods used and some highlight on the conclusions. The research motivation and process were not summarized clearly. Response: We agree, and have revised the abstract.

TE D

Reviewer 1: The introduction should be reorganized to emphasize the significance and innovation of the study comparing to the other researchers. Response: We agree, and have revised the introduction.

EP

Reviewer 1: In the models review sections (i.e. sections 2-6), it is suggested that we add a little more explanations on comparing the weakness of each model and the expected improvements on the combine usage of the models.

AC C

Response: We agree, and have inserted some additional text. The entire paper has been reorganized, as requested by one of the reviewers. The new Section 2, on the models, incorporates some material that was previously in the introduction. The evaluation of the strengths and weaknesses of the models comes in the conclusions, rather than in the main text. Further, in the new Section 6, we run sensitivity tests to evaluate the relationship between the models and the data. Reviewer 1: Please check if there's any typo on the End date in Table 1, and add label to figure 1 and figure 2. Response: The end dates did not always match in the 15-minute and 1-hour data sets, which accounts for the discrepancies in the dates and times. We could not add the labels directly to

ACCEPTED MANUSCRIPT 2

RI PT

the figures with the existing software, so we put these into a separate document. This issue is noted by several of the other reviewers as well. In the production process, Elsevier often typesets the figure and table titles and captions separately, and then inserts them. We chose to prepare a separate document with the titles and captions to facilitate this. This also makes it much easier to revise titles or captions.

Reviewer #4:

SC

Reviewer 4: The assessment of the results was realized by using only two accuracy indicators: MAE and RMSE. The authors must also use some graphical methods to characterize the models performance such as scatterplots, boxplots, Taylor diagrams, etc.

M AN U

Response: We agree, and have generated several additional graphs that illustrate the properties of the models. The titles and captions are in the separate document. Reviewer 4: Line 118: which data have been used to correct the forecast errors? Response: These calculations were done as part of the DICast system run at NCAR. We have revised the text accordingly.

TE D

Reviewer 4: Table 1: provide geographical coordinates of the stations mentioned in the table. Response: We put the latitude and longitude into the text, rather than the table. Since the sites are fairly close together, reporting this for each site would have been redundant. Further, one of the other reviewers requested a map, which we provided.

EP

Reviewer 4: Add axis titles to figure 1 and figure 2.

AC C

Response: The titles and captions are in the separate document.

Reviewer #5:

Reviewer 5: Provide a map with the geography of the two locations. Response: Maps are included.

Reviewer 5: Clarify why cloud advection models are treated separately and not under the section 2, Meteorological models.

ACCEPTED MANUSCRIPT 3

Response: Reviewer 3 makes the same point, and argues for a reorganization of the text with different subtitles. We have moved the section on cloud advection models forward, so that it follows directly after the other meteorological models.

RI PT

Reviewer 5: Clarify if time series models are considered statistical methods.

SC

Response: This issue appears to be one of semantics, and we have encountered it before. Time series models lie in a middle ground between what are usually referred to as statistical models and other fields such as econometrics. There is a great deal of overlap between the two. We have used the term statistical to differentiate time series methods from meteorological models. However, in the revised version, we have replaced the term statistical with time series on at several junctures.

M AN U

Reviewer 5: Table 3 should be provided as supplementary material. Graphs could be much more relevant.

EP

Reviewer #3:

TE D

Response: We prefer to leave the tables as part of the paper. In other studies by some of the same authors, also published in Renewable Energy, extensive tabular material was provided to support the narrative. This did not in any way interfere with narrative coherence. However, since the other reviewers also mentioned this, we included summary of the salient findings at the start of the new section on empirical findings. We considered using a graph or a table here, but decided instead to use a short text description.

AC C

Reviewer 3: The results reported in sections 8 and 9 are detailed, but there are no explanations regarding how and why models perform differently. The authors should provide a sensitivity analysis, or a similar procedure, to deepen the differences among the proposed models, and to find which parameters are the most important. Response: We agree, and this issue came up as we were drafting the paper. We have included some new text on this issue. In Section 6, we include two new tables, along with regressions of the data on the model forecasts. There is also an extended discussion of the results. These clearly indicate that the DICast output shows a closer relationship to the data than the meteorological and cloud advection model results. Reviewer 3: I suggest to provide a diagram rather than Table 2 (or both of them), to highlight the differences between meteorological, statistical (time series), and combined models.

ACCEPTED MANUSCRIPT 4

RI PT

Response: Here, we ran into some trouble. There was no obvious way to provide a visual guide to the models. I admit to a strong preference for English prose rather than visual images. Instead, after reading through these sections several times, we concluded that the text provides an adequate explanation for the material in Table 2. We did, however, expand considerably on the relative strengths and weaknesses of the models in the new Section 5.

SC

Reviewer 3: I suggest to convert sections 2, 3, 4, and 5 into subsections, and to gather them under a section 2 named "Forecasting models" or a similar title. Then, I suggest to convert sections 6 and 7 into subsections, and to gather them under a section 3 called "Simulation initialization" or similar. Finally, I suggest to convert sections 8 and 9 into subsections, and to gather them under a section 4 called "Results and discussion" or a similar title.

M AN U

Response: Reviewer 5 also argues for reorganizing some of the text. We did recognize in the initial drafting process that some of the titles read more like subtitles. We have reorganized the paper as into long sections, with subtitles for the sub-sections. Reviewer 3: In my opinion, the conclusions are too long. There should be a discussion section, or some findings should be directly reported in sections 8 and 9. Response: This has been done as part of the overall reorganization. The discussion is in Section 5, and the conclusions are shorter.

TE D

Reviewer 3: I suggest reporting all of figures and tables in the main text, in proximity of their citation. This greatly helps both readers and reviewers. Figures 1 and 2 have no caption, and no labels with units of measurement are reported.

AC C

EP

Response: We prefer to keep all the tables and figures in separate documents. These are typeset separately as part of the production process. Journals differ here with respect to this policy. Some journals encourage combining all the text, tables and figures into a single PDF, while others prefer to keep text, tables and figures in separate documents. Because of the number of additional figures and tables, we prefer to use the latter approach. Also, I do not have access to Latex software. Everything is written in Microsoft Word. Similarly, we prefer to place the figure titles and captions in a separate document. We hope that this does not unduly inconvenience the reviewers.

Reviewer 3: There is no Nomenclature section. Response: We did not consider this to be necessary. Most of the terminology used in the paper is fairly standard. Anyone in the solar energy community is familiar with acronyms like GHI, and in fact we always wrote out the names in prose before using the acronym. Similarly,

ACCEPTED MANUSCRIPT 5

RI PT

we wrote out the names of models like DICast in the text before switching to the acronym. Finally, in Table, 2, we give the names of the models in acronym form, followed by a description of what the models do.

Reviewer #2:

SC

Reviewer 2: The study has very weak originality and significance, and thus provided minimum useful information to the interested readers. Therefore, the manuscript is not recommended for publication.

M AN U

Response: We note upfront that the other reviewers do not share this assessment. Reviewers 4 and 5 characterized the study as scientifically rigorous and original. We did note some similarity between this comment and the request by Reviewer 1, who requested a stronger justification for the paper upfront. For this reason, we have substantially revised the introduction, to place greater emphasis on the new material.

TE D

We acknowledge that some of the findings here build on previous studies, and provide additional support for known findings. However, this does not disqualify the paper. In this respect, a great deal of scientific progress takes place in the form of incremental studies, rather than quantum leaps. Although this study is more incremental in nature, we view it as original and useful for the following reasons:

EP

1] The findings for meteorological, combined, cloud advection and time series models have rarely been compared so systematically, at the same sites and over the same data sets. By making this comparison more rigorous, we add new empirical findings to the literature.

AC C

2] There is an ongoing discussion in the literature on renewable energy as to the relative merits of physics-based and statistical or time series methods. This paper bears directly on these issues. It provides both systematic empirical evidence and insights as to the relative strengths and weaknesses of the two approaches. 3] There has been only limited discussion in the renewable energy literature as to the power of time-varying coefficient models. A great deal of this literature has used models with fixed parameters, or in the case of neural networks, fixed input and bias weights. However, it has been established in other fields, notably econometrics, that when the data is highly stochastic, time-varying parameter models are more effective. This study looks at this issue in greater depth than most of the existing works. The finding that a system which combines physicsbased model output with statistical adjustments and artificial intelligence can produce more

ACCEPTED MANUSCRIPT 6

accurate forecasts than meteorological models with fixed coefficients and in some instances ARIMAs with time-varying coefficients is a new one. This finding is worth making explicit.

SC

RI PT

Reviewer 2: The detailed reasons are itemized as follows. 1. Four conclusions were made from the provided data: A. The combined model is the best; B. ARIMA is better than other included time series methods; C. In very short horizons, statistical models are better than meteorological models; in long horizons, the opposite is true; D. Using time-dependent parameters instead fixed ones can improve the model accuracies. The conclusions A and C were not new, which simply reconfirm the findings presented in existing publications.

M AN U

Response: This paper actually does go further than earlier publications with respect to findings A and C. Specifically, we quantify the transition points, i.e., the forecast horizons at which meteorological and combined models are able to achieve greater accuracy than time series models. It is interesting that in the case of solar energy, the transition point is relatively short, in the range of 2-3 hours. By comparison, in other areas like wave energy, the transition point is in the range of 5-6 hours. We make this point explicitly, and view it as a new contribution.

EP

TE D

Reviewer 2: So, the only contributions of the study were the second and last conclusions. However, the last conclusion was undermined by the approach that the study took. The timevarying coefficients were used only in two models: DICast and ARIMA. Since DICast is the only model of its kind (the combined model), it is very difficult to tell whether its good accuracy is only due to the use of time-varying coefficients since the authors also claimed that combining of Meteorology models and time series models itself can improve the accuracy.

AC C

Response: We explicitly conclude in the revised paper that the DICast is superior primarily because of the statistical adjustment. We note that the weighted average reported in the tables does not do particularly well. In a more general sense, DICast has two key features. First, the model parameters are adjusted, as described in the text. This makes it superior to meteorological models with fixed parameters, cloud advection models and other techniques. Second, it incorporates information on weather and climate, which the ARIMA does not. The finding is that despite the fact that DICast and ARIMA both incorporate time-varying coefficients, DICast is better at all but the shortest horizons because it includes the meteorological information. Finally, a direct comparison of the ARIMA with the meteorological models with fixed parameters demonstrates that allowing the coefficients to vary over time makes the ARIMA very competitive. This

ACCEPTED MANUSCRIPT 7

provides strong evidence that time-varying coefficients enhance predictive accuracy in this kind of environment.

RI PT

Reviewer 2: The manuscript was loaded with data, which is encouraged. However, the authors just simply narrated observations from the data (especially when the conclusions did not apply to some data entries), instead of providing enough insights or explanations from the data. Therefore, there is no transformative information for the readers to either build a better model or use an existing model in a smarter way.

M AN U

SC

Response: Some of the other reviewers made similar points, although their phraseology was more supportive. We have expanded the discussion of the results, to take all these considerations into account, and believe that the recommendations do in fact contribute to building more effective models. The principle underlying the DICast – adjusting meteorological forecasts to more closely match the recent data – yields the highest degree of accuracy. Reviewer 2: The authors noticed that, in the RMSE data sheet of 15 min resolution at Site 67 in Sacramento, the transfer function was better than ARIMA, which conflicted with the second conclusion. A valid explanation will be very useful, probably even more useful than the data itself. However, the authors failed to provide one.

TE D

Response: This was an anomalous result, which occurred in one instance, at one site. It is clearly not typical. In other words, this is the exception that justifies the rule. However, we have provided some text as to why this should not be viewed as representative. Reviewer 2: The combined model is a key study object. However, the data from DICast was missing in all of the 15 min resolution data set.

AC C

EP

Response: Unfortunately, there was nothing that could really be done about this. DICast simulations were not available at higher frequencies. We would have dearly loved to evaluate the model at the higher resolutions. Due to the limitations of the database, this will have to be left to a future study. However, since the ARIMA is often better than DICast at the 1-hour horizon, it is reasonable to assume that it will be superior at horizons of 15-45 minutes.

Reviewer 2: Relative errors in percentage may serve better than the mean absolute errors and root mean squared errors. Response: There have been several studies of the relative merits of different measures of forecast accuracy. For wind and wave energy, we would prefer the mean absolute percent error. However, for solar, one crucial problem is that the percent error can become very high during periods of low irradiance, i.e., early morning and late evening. The reason of course is

ACCEPTED MANUSCRIPT 8

AC C

EP

TE D

M AN U

SC

RI PT

that while the error in W / m2 may be small, the low denominator can raise the percent error. Most of the literature has therefore preferred to use the mean absolute error and the RMSE. We followed this convention. Since there is already a great deal of tabular material, we did not want to expand the tables by one-third simply to provide a third measure of accuracy.