Estimation of Spatially Detailed Electricity Demands Using Spatial Statistical Downscaling Techniques

Available online at www.sciencedirect.com

ScienceDirect Energy Procedia 75 (2015) 2751 – 2756

The 7th International Conference on Applied Energy – ICAE2015

Estimation of spatially detailed electricity demands using spatial statistical downscaling techniques Daisuke Murakami*a, Yoshiki Yamagataa, Hajime Seyab a

National Institue for Environmental Studies, 16-2 Onogawa, Tsukuba 305-8506, Japan a Hiroshima University, 1-3-2 Kagamiyama, Higashihiroshima 739-8511, Japan

Abstract

Although city scale aggregate electricity demands are usually estimated by multiplying intensity data by floor space, in Japan there are few available sources for municipality level building stock (floor space) data. Hence in this study, we attempt to create municipality level building stock data using the techniques of spatial statistical downscaling. Firstly, this study compares predictive accuracy of several downscaling methods including both deterministic and statistical ones. The results support the use of statistical downscaling methods, which consider spatial autocorrelation or spatial heterogeneity. Secondly, it actually creates building stock data of Japan at municipality level (1803) in 2005 by downscaling prefectural level (49) data employing one of the spatial statistical downscaling methods. Thirdly, using the estimated building stock data, it empirically estimates electricity demand at municipality level. © 2015 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license

© 2015 The Authors. Published by Elsevier Ltd. (http://creativecommons.org/licenses/by-nc-nd/4.0/). Selection and/or peer-review under responsibility of ICAE

Peer-review under responsibility of Applied Energy Innovation Institute Keywords: Electricity; Downscaling; Spatial statistics; Building stocks; Urban resilience

1. Introduction Especially after the great east Japan earthquake in 2011, it has been increasingly recognized the importance of estimating electricity demands accurately at local scale such as municipality level in Japan. Unfortunately, however, hourly electricity demands data by sector (residential, commercial, transport, etc.) at municipality level are not publicly available in Japan, and some kind of estimation is required. Conventionally, regional electricity demands have been estimated by multiplying electricity intensity data (demand per floor) with floor space (or building stock amount) in each region. The applicability of this estimation approach has dramatically been raised after Japanese Ministry of Land, Infrastructure,

* Corresponding author. Tel.: +81-29-850-2419; fax: +81-29-850-2960. E-mail address: [email protected].

1876-6102 © 2015 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). Peer-review under responsibility of Applied Energy Innovation Institute doi:10.1016/j.egypro.2015.07.515

2752

Daisuke Murakami et al. / Energy Procedia 75 (2015) 2751 – 2756

Transport and Tourism (MLIT) begins to publish the building stock survey in 2010, which including data on building stock amount of each sector. However, the statistics is only available at prefecture level, while more spatially finer data are required for regional electricity planning. Hence, we estimate municipality level building stock (floor space) data by downscaling (DS) the prefectural level building stock data. Firstly, this study compares estimation accuracy of several downscaling methods including both deterministic and statistical ones. Secondly, it actually creates building stock data of Japan at municipality level (1803) in 2005 by downscaling prefectural level (49) data employing one of the spatial statistical downscaling methods. Thirdly, using the estimated building stock data, it empirically estimates electricity demand at municipality level. 2. DS approaches 2.1. Deterministic DS methods The dasymetric mapping (DM) method [1] is representative downscaling methods developed in geography. Suppose that y P is the building stock in prefecture P and ym is the unknown stock in municipality m, then, DM method estimates ym with Eq.(1):

ym

wm

¦

m P

wm

yP ,

(1)

where wm is a weight assigned to the municipality m. This is given using ancillary data; for instance, when residential building stocks are downscaled, wm can be given by the population of municipality m. 2.2. Stochastic DS methods Statistical DS methods typically assume that municipality level building stock for m divided by wm, say ym/wm has distribution as Eq.(2):

y m / wm

P ( x m ,q , ȕ )  H m

E[H m ]

0

Var[H m ] V 2V (ș) ,

(2)

where ȝ(xm,q, ȕ) is a mean function depending on Q explanatory variables xm,q where q  {1,...Q} and ȕ is a vector of parameters. ı2 is a variance parameter and V(ș) is a (co-)variance function depending on parameters in ș . Examples of the mean and the variance functions are shown in Table 1. Different from the DM methods, models in Table 1 can consider the introduction of multiple auxiliary variables as explanatory variables. In order to consider possible spatial autocorrelation (i.e., building stocks in nearby municipalities tend to be similar), the geostatistical (GS) method [2] [3], which models whose covariance with a distance decay function; for example, exp(–dm,m'/r) shown in Table 1, is employed. Also, to capture possible spatial heterogeneity, we propose applying the geographically weighted regression (GWR) model [4], which allows local trend parameters ȕm,q that varies smoothly across geographic region, as shown in Table 1. We compare the results of these models. Downscaling must be performed while considering the volume preserving property. Specifically, in our case, aggregations of municipal level stock estimates must equal to the prefecture level actual stocks. This property is satisfied by imposing the following constraint: yP

¦

m P

ym .

Besides, to consider the non-negativity of building stocks, Eq.(4) is also imposed:

(3)

2753

Daisuke Murakami et al. / Energy Procedia 75 (2015) 2751 – 2756

ym t 0

.

(4)

The unknown ym, satisfying both the volume preserving property and the non-negativity, is estimated by minimizing Hˆm2 under Eqs.(3), (4).

¦

m

Table 1. Municipal level statistical models Model Mean: ȝ(xm,q, ȕ) Variance: V(ș) Regression model Ȉmxm,qȕq 1 if m = m', and 0 otherwise GS model Ȉmxm,qȕq exp(–dm,m'/r) GWR model Ȉmxm,qȕm,q 1 if m = m', and 0 otherwise ȕq is a trend parameter; ȕm,q is a local trend parameter that varies across geographical space smoothly; dm,m' is the Euclidean distance between the geometric centers of municipalities m and m', and r is a range parameter.

3. Application 3.1. Outline This section applies above discussed DS methods to downscale 46 prefectural wooden/non-wooden residential stocks in 2005 (Fig.1; Source: Building Stock Statistics) into 1,803 municipal units, and these accuracies are compared. Compared methods include two DS methods (DM_P and DM_B), whose weights are populations and building land areas, respectively, the GS method, and the GWR method. Features of these methods are summarized in Table 2. Following most geostatistical DS studies, wm in the GS method is given by area, and population and building land area, which are used as weights in other methods, are considered as explanatory variables. 2

2

Value (km /100km )㻌 䣘䣣䣮䣷䣧䢢䢪䢳䢲䣭䣯䢴䢱䣯䢴

䢺䢲䢲䢢䢯 8.00-㻌 䢶䢲䢲䢢䢯䢢䢺䢲䢲 4.00-8.00㻌 䢴䢷䢲䢢䢯䢢䢶䢲䢲 2.50-4.00㻌 䢳䢷䢲䢢䢯䢢䢴䢷䢲 1.50-2.50㻌 䢳䢲䢲䢢䢯䢢䢳䢷䢲 1.00-1.50㻌 䢶䢲䢢䢯䢢䢳䢲䢲 0.40-1.00㻌 䢴䢲䢢䢯䢢䢶䢲 0.20-0.40㻌 䢳䢲䢢䢯䢢䢴䢲 䢷䢢䢯䢢䢳䢲 0.10-0.20㻌 䢲䢢䢯䢢䢷 0.05-0.10㻌

0.00-0.05㻌

Wooden

Non-wooden㻌

Fig. 1. Prefectural residential electricity demands. Table 2. Auxiliary variables used in each DS methods. Model DM_P DM_B GS GWR

Weights: wm

Explanatory variables: xm,p

Population N.A. Building land area N.A. Area Population density; ratio of building lands; Density of railway stations; Road density; Distance to the nearest city designated by ordinance Population Density of railway stations; Road density; Building land area Distance to the nearest city designated by ordinance

Spatial dependence

Spatial heterogeneity

‫ݱ‬ ‫ݱ‬

2754

Daisuke Murakami et al. / Energy Procedia 75 (2015) 2751 – 2756

3.2. Accuracy comparison Municipal residential stock data is required to evaluate downscaling accuracy. Fortunately, [5] collected Property Tax Ledger data in 241 municipalities in Tokyo metropolitan area in 2005 including residential stock data, and we can access this data. Hence, while stocks in 1,803 municipalities are estimated by downscaling, the estimates in 241 municipalities are compared with the actual stock amounts, and accuracies of each method are evaluated. Accuracy comparison result can change significantly depending on the error statistic used; hence, the following four error statistics are used for accuracy evaluation:

1 241

RMSE

MAE

1 241

RMSPE

MAPE

¦ ( yˆ

 ym |

m

m

1 241

¦

¦ m

 ym ) 2

m

¦| yˆ 1 241

m

m

,

(5)

,

§ yˆ m  y m ¨¨ © ym

(6) 2

· ¸¸ ¹ ,

(7)

yˆ m  y m . ym

(8)

RMSE (Root mean squared error) is more sensitive to a large error than MAE (Mean absolute error). RMSPE (Root mean squared percentage error) and MAPE (Mean absolute percentage error), which are standardized version of RMSE and MAE, respectively, are large if the predictive error is large relative to the stock amount. Predictive error comparison result is summarized in Table 3. In both wooden and non-wooden stocks, GS method outperforms the others in terms of RMSE, and GWR method outperforms the others in terms of MAE, RMSE, and MAE. Thus, it is verified that, while the accurateness of DM method has been demonstrated frequently in geography, spatial statistical approaches, which utilize multiple auxiliary variables and spatial information, are further accurate. The next section applies the GWR method, which was the most accurate in 3 in 4 error statistics, to the electricity demand estimation. Table 3. Predictive error comparison result. In each comparison, the smallest values are highlighted by gray marker. Wooden residential stocks

Non-wooden residential stocks

DM_P

DM_B

GS

GWR

DM_P

DM_B

GS

GWR

RMSE

1.55×106

8.39×105

8.98×105

9.16×105

8.26×105

1.91×106

4.64×105

4.69×105

MAE

6.89×105

5.56×105

4.88×105

4.22×105

4.34×105

9.75×105

2.93×105

2.69×105

RMSPE

3.91×10-1

5.11×10-1

3.89×10-1

3.17×10-1

1.80

6.63

1.68

9.97×10-1

-1

-1

-1

-1

MAE

2.99×10

3.41×10

2.62×10

3.3. Municipal electricity demand estimation

1.98×10

9.56×10

-1

2.68

7.67×10

-1

5.23×10-1

Daisuke Murakami et al. / Energy Procedia 75 (2015) 2751 – 2756

Wooden/non-wooden and residential/non-residential municipal building stocks are estimated by the GWR method-based downscaling, and hourly electricity demands in each month in each municipality are estimated by multiplying energy consumption rates provided by [6] with the estimated stocks. Estimated electricity demands in 5AM 0PM, and 8PM in January and July are plotted in Fig.2 (residential sector) and Fig.3 (non-residential sector). These figures show the following features: Residential electricity demands in January are nearly constant across hours; Residential demands in July have a peak at night; Non-residential electricity demands have a peak at 0PM; Demands are intense around principal cities including Tokyo and Osaka. These results are intuitively reasonable.

5AM (January) Osaka

5AM (July)

0PM (January)

8PM (January)

0PM (July)

8PM (July)

Tokyo

Fig. 2. Estimated residential electricity demands.

5AM (January)

5AM (July)

0PM (January)

0PM (July)

Fig. 3. Estimated non-residential electricity demands.

8PM (January)

8PM (July)

2755

2756

Daisuke Murakami et al. / Energy Procedia 75 (2015) 2751 – 2756

4. Conclusions This study first compared geographical and spatial statistical downscaling approaches in municipal building stock estimation, and demonstrated accurateness of the latter. Then, the GWR method, whose accuracy was the best, was utilized to estimate municipal electricity demands. Downscale study would especially be important in developing countries, where spatially detailed building stock data or other data explaining electricity demands are not publicly available. Thus, applying spatial statistical downscaling for developing countries would be an important future study. Besides, in the future where increase of some natural hazards under the climate change is projected (e.g., [7]), application of downscaling to discuss urban resilience and adaptation would also be increasingly important. Acknowledgements This work is an achievement of the Global Climate Risk Management Strategies (S10) project. References [1] Fisher PF, Langford M. Modeling the errors in areal interpolation between zonal systems by Monte Carlo simulation. Env Plan A 1995; 27: 211–224. [2] Cressie N. Statistics for Spatial Data, Revised Edition. John Wiley & Sons; 1993. [3] Gotway CA, Young LJ. Geostatistical approach to linking geographically aggregated data from different sources. J Comput Graph Stat 2007; 16: 115–135. [4] Brunsdon C, Fotheringham AS, Charlton ME. Geographically weighted regression: a method for exploring spatial nonstationarity. Geogr Anal 1996; 28: 281–298. [5] Tsutsumi M, Miyagi T, Yamasaki K. Potential of computable urban economic model formalizing building market. J JSCE Div D 2012; 68, 333–343 [in Japanese]. [6] The Japan Institute of Energy. Cogeneration plan and design manual 2008. Tokyo: Japan Industrial Publishing Co. Ltd.; 2008 [in Japanese]. [7] Hirabayashi Y, Mahendran R, Koirela S, Konoshima L, Yamazaki D, Watanabe S, Kim H, Kanae S. Global flood risk under climate change. Nature Climate Change 2013; 3: 816–821.

Biography Daisuke Murakami is a research associate of the National Institute for Environmental Studies, Japan. He received his Ph.D. in engineering from University of Tsukuba in 2014. His research topics include downscaling of social economic data, spatial statistics, quantitative geography etc.