Collaboration protocols for sustainable wind energy distribution networks

Author’s Accepted Manuscript Collaboration protocols for sustainable wind energy distribution networks Ehsan Jahanpour, Hoo Sang Ko, Shimon Y. Nof

www.elsevier.com/locate/ijpe

PII: DOI: Reference:

S0925-5273(16)30240-7 http://dx.doi.org/10.1016/j.ijpe.2016.09.010 PROECO6529

To appear in: Intern. Journal of Production Economics Received date: 7 July 2015 Revised date: 5 July 2016 Accepted date: 15 September 2016 Cite this article as: Ehsan Jahanpour, Hoo Sang Ko and Shimon Y. Nof, Collaboration protocols for sustainable wind energy distribution networks, Intern. Journal of Production Economics, http://dx.doi.org/10.1016/j.ijpe.2016.09.010 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 galley proof before it is published in its final citable 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.

Collaboration protocols for sustainable wind energy distribution networks Ehsan Jahanpoura, Hoo Sang Kob,*, Shimon Y. Nofc a

Monsanto Company, Chesterfield, MO 63017, USA

b

Department of Mechanical and Industrial Engineering, Southern Illinois University Edwardsville, Edwardsville, IL 62026, USA

c

PRISM Center and School of Industrial Engineering, Purdue University, West Lafayette, IN 47907, USA *

Corresponding author. Tel.: +1-618-650-5962. E-mail addresses: [email protected] (H.S. Ko), [email protected] (E. Jahanpour), [email protected] (S.Y. Nof).

Abstract: Wind energy has attracted more attentions in recent decades as a green and sustainable electricity generation resource with little or no pollution. Fluctuations in output of wind turbines and their dependency on environmental conditions, however, limit their penetration into power grids. This research proposes a collaboration platform to overcome these challenges by collaboration among communities in the energy distribution network. First, a second order trigonometric regression model is applied to forecast energy demands, and a multiple linear regression model is used to predict energy output of a wind farm. Based on this information, communities in the network can initiate the collaboration process through the proposed platform, which is controlled by two collaboration protocols: Demand and Capacity Sharing Protocol (DCSP) and Best Matching Protocol (BMP). The protocols are applied to optimize communities' profit through matching the communities with excessive capacity with the ones with energy shortage so that they can create a sustainable distribution network. A simulation of three communities with three wind farms is conducted to measure the impact of the platform. The results show the sustainability of the energy network can be achieved by reduced demand loss and holding cost and improved collaborative capacity.

1

Keywords: Best matching protocol; collaborative control theory; demand and capacity sharing protocol; regression analysis; sustainable network; wind energy.

2

1. Introduction Wind energy has been gaining more attention recently as an environmentally-friendly alternative to traditional fossil fuel or nuclear energy. Wind power generates electricity in a clean, sustainable, and affordable manner without leaving any environmental pollution or wastes. Therefore it will play an important role in protecting the environment while keeping up with the increasing electricity demands. The performance of wind turbines in terms of electricity generation, however, depends on numerous environmental, mechanical, and electrical parameters such as operational status of wind turbines, wind direction, wind speed, and weather conditions. As a result, wind power output is highly variable and only partially controllable [1]. In spite of the high uncertainties in energy generation and distribution, wind resources have been integrated gradually into the electric grid as one of the major players. Various programs and studies have been established to increase the share of renewable resources into the electric grid [2-4]. The U.S. Department of Energy’s report “20% Wind Energy by 2030” envisioned that wind power could supply 20% of all electricity nationwide [5]. The UK Government expects offshore wind energy to be a major contributor to its target to generate 15% of UK electricity from renewable sources by 2020 [6]. The high uncertainty in wind power generation, however, is the major obstacle in achieving these goals. There are a variety of sources causing the uncertainty, which can be classified in two ways: 1) controllable (e.g. failure and maintenance) and uncontrollable sources (e.g. environmental factors such as wind speed/direction, temperature, and humidity), or 2) sources at the supply side (energy capacity affected by the controllable or uncontrollable sources) and at the demand side (energy consumption affected by weather, season, social and cultural factors, income, education, etc.). These various sources of uncertainty in wind energy discourage its penetration and integration into the electric grid, and thus the majority of the electricity demand is still fulfilled by traditional energy sources with well-established infrastructure and stable supplies. Part of the fluctuations in daily energy output of a wind turbine depends on environmental changes such as wind speed, direction, and temperature. In addition to wind power’s dependency on weather conditions, turbine failures influence the electricity generation. Since the energy output from a wind power system highly

3

depends on the performance and failure rates of its wind turbines, numerous techniques have been investigated in literatures in order to detect and predict wind turbine failures [7-12]. The fluctuating characteristic of wind power is often referred to as intermittency. It is important to note that intermittency is not the same as unpredictability. To the extent that intermittency in wind power is driven directly by weather, it is somewhat predictable, and could be managed in part through the use of near-term wind power forecasts [13]. There have been a number of studies concerning variability reduction in wind farm outputs through reducing the uncertainties in energy generation forecasts, yet the uncertainty in energy output from wind farms is still challenging. Since the uncertainty is also condition-based, it cannot be entirely eliminated in spite of being predictable; however, statistical and machine learning methods have been developed to handle the uncertainty to certain extent. Moreover, advanced control schemes have been developed to increase the efficiency and reliability of wind turbines. For instance, [14] proposed a nonlinear and adaptive algorithm to control wind turbine speed. Modern collaborative control theory algorithms are applied in this research to accelerate wind energy penetration in power grids. Another major source of uncertainty is fluctuations in energy demand that not only limits sustainable energy expansion, but also impacts the reliability of current power grids that hold wind farms, solar farms, nuclear, and fossil power plants. Numerous models have been developed to forecast energy demand. These models can be divided into two major groups based on their forecasting techniques: a) models based on time series analysis such as multiple regression models, autoregressive moving average, exponential smoothing, and decomposed time series [15-18], and b) knowledge based techniques based on artificial intelligence, artificial neural network, fuzzy logic, etc. [19-21]. Moreover, hybrid techniques have been proposed in the literature to improve energy demand forecasts [22]. According to these studies, multiple parameters have impact on electricity demands including seasonal or yearly trend, population, general income, temperature, holidays, culture, etc. Although understanding all these parameters are helpful, we should avoid over-fitting issues that could prevent us from modelling the general trend of the problem.

4

Wind farms are naturally subject to fluctuations in their power generation rates that vary with seasonal, environmental, and operational factors. Besides, breakdowns of wind turbines are inevitable and could affect the performance of the wind farm in terms of power generation. Thus, a wind farm may encounter two situations regarding its promised amount of supply: 1) capacity shortage, 2) excess capacity (here capacity refers to the amount of power generated over time). The ongoing technological innovations and research works being conducted on smart grid areas, such as supplier and consumer behaviors in an automated fashion, are discussed by [23]. Recently, Collaborative Control Theory (CCT) have been widely applied to deal with the uncertainties in the market demand and capacity [24, 25]. [26] combined two CCT-based protocols— Demand and Capacity Sharing Protocol (DCSP) and Best Matching Protocol (BMP) to maximize profit and resource utilization of supply enterprises, to enable timely delivery to customers in spite of uncertain market demands and unexpected capacity shortages, and to maximize the overall stability of the supply network. They apply fuzzy mixed integer programming to model and optimize the decisions and queueing theory for model validation. [27] proposed an adaptive and collaborative DCSP to deal with volatile product demand and rapidly changing manufacturing processes for sustainable returns. This paper facilitates the use of CCT to handle these uncertainties in supply and demand of the wind energy. Assuming the participants in the network are willing to share their demand and capacity, this article proposes a CCT-based collaboration platform, which supports communities to fulfil their electricity demand within the network by the collaboration in an effort to create a self-sustainable energy network. In order to facilitate collaboration between participants for dynamic demand and capacity sharing and matching, the collaboration platform is developed based on two CCT-based protocols: 1) DCSP [28, 29] and 2) BMP [30, 31]. To summarize, this research aims at: 

Reducing uncertainty at supply by accurate output prediction;



Designing collaboration protocols (DCSP and BMP) to handle excessive energy and shortage caused by uncertainty at both supply and demand; and

5



Creating a sustainable wind energy distribution network through the proposed collaboration platform.

This paper is organized as follows. Section 2 applies statistical techniques to improve energy demand and capacity predictions. A second-order trigonometric model is applied in Section 2.1 for demand forecasting and a multiple linear regression model is developed in Section 2.2 for predicting energy output of a sustainable energy provider. In Section 3, a collaborative wind energy distribution network is developed on top of DCSP and BMP protocols. A simulation study with three communities is illustrated in Section 4. Finally, Section 5 concludes this article.

2. Statistical analysis of community-based energy demand and capacity Power grid networks are not built in conformance with a plan like the interstate highway system. Each grid is built by individual communities as isolated transmission islands to meet local needs (Fig. 1). These small networks are unsystematically linked when communities find it beneficial to jointly own energy sources or to connect to neighboring communities to facilitate power sales [32]. Collaborative energy sharing between communities may increase the uncertainty and complexity of the model by increasing the number of parameters and estimations in the model. However, communities will improve their profit through selling their excessive electricity; meanwhile they will increase their service reliability by being able to use other community’s energy resource during electricity shortage. Two major obstacles against expanding community-based sustainable energy provision are the uncertainties in estimating energy demand and energy capacity. It is necessary to get the information on energy demand and supply in each community in order to develop an efficient sharing method. This section presents statistical analyses to deal with the uncertainties in both areas. Once the variables are estimated, collaborative control theory can be employed, as described in Section 3, to select the best energy sharing scenarios that simultaneously improve community profits while increasing energy supply reliability.

6

Fig. 1. Community-based power grid network including three sustainable energy providers.

2.1 Wind energy demand forecasting using a trigonometric regression model As discussed in Section 1, there have been numerous time series analyses and knowledge based techniques studied to reduce the uncertainty and improve accuracy of electricity demand forecasting for a community. There are pros and cons for each forecasting method. For instance, in spite of machine learning techniques’ competency in forecasting electricity demands, their “black box” nature is limiting its applicability. We use time series analysis to forecast the energy demand of a community. In order to develop a statistical model for demand forecasting, hourly energy load data in different zones are collected from the "Historical Metered Load Data" report by PJM Interconnection, a regional transmission organization that coordinates the movement of wholesale electricity in a number of states in US [33]. Fig. 2(a) illustrates the daily energy amount consumed in a zone of interest from 2011 to 2013. As indicated in Fig. 2(a), the energy consumption in a zone shows seasonality; the consumers tend to use more energy in summer and winter times when the temperature is high or low. Also, the data also shows a periodic behavior which is repeated over the years.

7

Therefore, a trigonometric regression model is used in this study in order to model the fluctuations (seasonality) of energy consumption in the area. This model not only uses the benefits of Fourier analysis to model energy consumption, but also it is computationally simpler. Since energy usage during a year peaks in two seasons, a second-order trigonometric regression model is effective to estimate the daily energy consumption. Equation (1) shows the trigonometric regression model used to estimate daily demands: (

)

(

)

(

)

(

)

(1)

where Dt is the total demand of day t, β0, …, β4 are the regression model coefficients, T = 365 denotes the number of days in a yearly period. Fig. 2(b) plots the residual histogram calculated by subtracting estimated values from real observations. Shapiro-Wilk hypothesis test is used to check the normality of the residuals. The Wilk statistics and p-value of the test are W = 0.9862 and p-value = 0.1327. Thus, the test shows that residuals follow a normal distribution with mean of zero and standard deviation of 4283, thus, ϵ ~ N(0, 4283) is the error (uncertainty) associated with energy estimations. The regression model, shown as the blue line in Fig. 2(a), is as follows: (

)

(

)

(

)

(

)

(2)

A part of variation in daily energy demand might be as a result of randomness and no systematic reason could be assigned to it. However, such random variability could be reduced by combining daily demands into weekly ones. Fig. 3 illustrates the weekly demand along with trigonometric regression model (D = 52, ϵ ~ N(0,2399)) of the same zone during 2011-2013. The regression line (blue) is obtained by minimizing the sum of square errors between estimated values and observations and is the best estimate for the mean demands at different time. Equation (3) represents the secondorder trigonometric model with estimated parameter values: (

)

(

)

(

)

(

)

(3)

Since the parameters of time-dependent demands are known, a set of random series of demands (green up-triangles) are shown by adding a random normal error to the value of regression line at time t. We will apply this model in Section 4 to simulate random demands over time. Figs 4 and 5 show the

8

observed and simulated energy demands of a zone from January 2014 until June 2014. The secondorder trigonometric regression model is developed using 5-year historical data. The Root Mean Square Error (RMSE) parameter is used as a scale-dependent statistics to compare the accuracy of both models [34, 35]. RMSE measures the difference between estimates and observed values and is a good indicator of the accuracy of the estimations. Since RMSE has the same unit as the variables, it should be compared based on the variables’ magnitude. For comparing accuracy of multiple predictive models on a set of observations, smaller values of RMSE indicate better estimations. The RMSE values for daily and weekly demands are 6582.137 and 4444.878, respectively. Comparing the RMSE values to the mean and standard deviation of the demands (mean = 63820 and stdev = 12715) shows that both estimated regression model is a good fit for the model. In addition, the uncertainty is smaller when data are aggregated into longer periods because the residual standard error is smaller. However, management of energy generation and distribution gets more complicated as the estimation horizon increases. Thus, the energy demands are analyzed on a daily basis in this study.

Fig. 2. (a) Second-order trigonometric regression model along with daily electricity usage in a zone during 2011-2013; (b) Histogram of the regression model residuals.

9

Fig. 3. Weekly electricity usage in a zone during 2011-2013.

Fig. 4. Observed daily energy demand of a zone during January 2014-June 2014 (black circles) along with the trigonometric regression model (blue line) and estimated energy (green up-triangles).

10

Fig. 5. Observed weekly energy demand of a zone during January 2014-June 2014 (black circles) along with the trigonometric regression model (blue line) and estimated energy (green up-triangles).

2.2 Wind farm capacity forecasting using a multiple linear regression model Forecasting the energy output of a wind farm is the second source of uncertainty that limits the growth of wind energy usage. Numerous approaches such as ensemble forecasting, physical methods, statistical analyses and machine learning approaches have been proposed to improve wind energy output forecasts [36-39]. [40] have reviewed different tools and methods being used in wind speed and wind energy generation forecasting. They classify the wind energy forecasting methodologies into physical, statistical, and hybrid approaches. Another thorough study [41] divided the wind energy forecasting models into two main groups: 1) forecasting based on time series analysis of historical wind data and 2) forecasting using numerical weather prediction (NWP) models. According to their study, the first group of models provides better results for forecasts under higher temporal scales (monthly, quarterly, or annual mean) while the second group models are more effective for short-term forecasting horizons (hourly, daily, or weekly mean). Since the capacity information should be available for short-term horizons in our collaboration platform, we employ NWP models for capacity forecasting that reduce uncertainty over short-term horizons.

11

Wind energy prediction models make estimates based on weather parameters forecasts. Thus, the accuracy of weather parameter forecasts largely influences the accuracy of wind energy predictions. According to Mohrlen [42], wind speed (including density and vertical wind shear) plays the most important role in predicting energy output. In addition, wind turbines’ sudden and unpredicted failures could inflate the uncertainty in predictions. The magnitude of prediction uncertainty depends on the severity of the failure and time required to fix it. Numerous studies have been proposed to improve wind turbine failure prediction and detection. [10] applied pattern recognition model to classify the existence of a failure based on multiple variables including rotation speed of the blade, actual torque, DC bus power, and wind speed. The maximum capacity of a wind farm is also an important parameter in modeling energy capacity. Wind farms can increase their maximum capacity by adding more wind turbines to the farm. In addition to the structural and operational costs of new turbines, [43] stated that interaction between wind turbines and “wind shadow” impacts could change the upper limit. Since the major contribution of this article is developing a collaboration-based sustainable wind energy network in existence of uncertainty, rather than improving the prediction algorithms, we introduce a NWP-based model being used in the industry. The EirGrid model being used in Irish energy business uses a multiple linear regression model to forecast the wind energy outputs [44]. This model uses weather forecasts, seasonal variation, holidays and special events, and lifestyle changes to forecast energy outputs using historical data. Fig. 6 illustrates energy output of the Irish wind farms during January 2014-June 2014 along with the EirGrid model forecasted values. In addition to visual analysis, RMSE is used to measure the accuracy of the model. The RMSE of the EirGrid model is 2,833.64 MWh. Since the average daily energy output is 16,019.38 MWh and standard deviation is 11,049.64 MWh, the computed RMSE shows that the model is appropriate to forecast the wind turbine energy output.

12

Fig. 6. Daily energy output observations and forecasts used by Irish energy business during January 1, 2014-June 6, 2014.

The EirGrid model is shown to be appropriate for wind generation forecasting, however the model is not publically available. Similar to the widely used EirGrid model, a multiple linear regression is applied in this research to identify the important parameters and their relationship for forecasting wind energy outputs. Eight parameters including temperature (˚C), effective temperature (Feels) (˚C), rain (cm), cloud (%), wind speed (m/s), gust (m/s), humidity (%), pressure (cmHg) are studied in this research. A regression analysis is used to evaluate the statistical relationships between these variables and the response variable (energy output). The p-value for each variable indicates if the variable has a significant impact on energy output. The p-values for temperature, effective temperature, rain, cloud, wind speed, gust, humidity, and pressure are 0.5690, 0.0323, 0.0232, 0.2496, 0.9828, 6.62e-11, 0.8205, 0.1716, and 0.7873, respectively. Therefore, the regression analysis shows that temperature, effective temperature, and wind speed are more important than the others in forecasting wind energy output, assuming type I error () of 0.05. However, because of the intercorrelation between temperature and effective temperature, we selected temperature and wind speed for forecasting wind turbine energy output. The study does not show a significant interaction between temperature and wind speed. Equation (4) shows the multiple linear regression model used to forecast the outputs:

13

(4) Where Et denotes the energy output at day (time period) t, δt is the average temperature of day t, and γt is the average wind speed on day t. The coefficient of determination (R2) for the multiple linear regression model is 85.74% that shows that wind speed and temperature could define 85.74% of the variations in a wind turbine energy. The p-value of the model is smaller than 2.2×10-16 and the RMSE is 4,173.29 MWh (the average daily energy output is 16,019.38 MWh and the standard deviation is 11,049.64 MWh).

So the analysis shows that the multiple linear regression model denoted as

Equation (4) is appropriate for forecasting the wind turbine energy output although the model could be improved by including seasonality and special daily effects into model. Fig. 7 shows the actual energy output observations along with the multiple linear regression model outputs used to forecast those observations.

Fig. 7. Daily energy output observations and multiple linear regression model forecasts during January 1, 2014- June 6, 2014.

3. Collaborative wind energy distribution network The proposed collaboration platform and protocols are motivated by the stochastic nature of the electricity demands, the dynamic changes in power generation over time, and the ability to overcome such uncertainty through collaboration between energy providers to build a sustainable energy

14

distribution network. In this network, each energy provider is deemed to be a self-operative organization who is willing to collaborate with one another to achieve higher benefits. For example, when a wind farm cannot fulfil a customer order, the demand will be shared with other collaborating providers who have excess capacity. As a result, the possibly unfulfilled demand can be delivered by other farms and the remaining capacity of collaborating farms can be utilized, such that mutual benefits can be achieved. This collaboration can not only encourage the energy providers to improve their benefits by selling their excessive capacity obtained from the renewable sources but also motivate the customers to join the sustainable network to relieve the shortage for more economical prices. A framework using two protocols is developed to control the collaboration process and ultimately to create a sustainable energy distribution network. The two protocols are: 1) Demand and Capacity Sharing Protocol (DCSP) and 2) Best Matching Protocol (BMP). This sustainable energy network is characterized by a heterarchical framework, such that a farm does not have total control over the other farms. However, a collaborative platform is established between farms to improve their energy production profit and reliability (Section 3.1). The collaborating farms achieve their goals only through collaborative decision making processes, which involve demand and capacity forecasts, information exchange, negotiation, and coordination, governed by DCSP (Section 3.2). Once the demand and capacity information including available capacity, requested amount, price, transportation and delivery costs are shared through DCSP, the subsequent collaboration process will be managed by the BMP (Section 3.3). The collaboration planning horizon is variable depending upon the forecast time window.

3.1. Sustainable collaborative platform using DCSP+BMP As discussed earlier, the collaboration control theory is used in this research to reduce the uncertainty of sustainable energy provision. The DCSP+BMP works on the collaborative platform as follows: 1) Start at period T. Periods can be daily or weekly, depending on the possibility in terms of available resources (for forecasting, for example) and the anticipated precision in decisions.

15

2) At the beginning of each period, define sets A (who need to fulfil their demands by demand sharing) and B (who can share their excessive capacity) based on energy output forecasts and demand rates associated with each community. 3) The information on Sets A and B are shared with all wind farms in the network. 4) Wind farms in Set A receive or forecast community orders, evaluate, and based on their available capacity either 1) accept the order, or 2) generate and send a demand sharing proposal to the Set B wind farms and wait for their demand sharing proposals. 5) Wind farms in Set B receive different proposals (if any) from Set A, evaluate the community orders and the capacity that they can promise to each proposal, prepare their capacity sharing proposals including provision price, penalty cost, and demand loss costs and send them back to the wind farms in Set A. 6) BMP is employed to match the energy providers and in-need communities and define energy transmission amount based on the pre-approved promised and real-time prices and penalty costs to optimize the profitability and reliability of the sustainable energy network. 7) Wind farms in Set A receive the BMP proposals and accordingly either 1) reject the proposal, or 2) accept and send the allocation results to corresponding wind farms in Set B. 8) Wind farms in Set B receive the allocation results and accept the shared demand. 9) At the end of period T, compute the real-time demand and capacity and update the forecasting models based on the demand and capacity variation. 10) Set T = T + 1 and go to step 1.

Fig. 8 illustrates the overall collaboration process controlled by DCSP and BMP, which are described in more details in subsequent sections.

16

Forecast daily demand

Negative

Forecast daily supply

Compute capacity-demand

Send demand sharing proposal

Evaluate capacity and demand sharing proposals

Non-Negative

Send capacity sharing proposal

Find the best matching solution to optimize community profits and reliability

Update the distribution network

Send capacity-demand matching results

Wait for the next forecast

Demand Sharing Protocol

Update the distribution network

Wait for the next forecast

Best Matching Protocol

Capacity Sharing Protocol

Fig. 8. DCSP + BMP for the collaboration platform.

3.2. Demand and capacity sharing protocol Consider a set of collaborative wind farms F = {f1, f2, ..., fn}, where each farm serves its own community C = {c1, c2, …, cn}. The collaborative model includes two stochastic inputs; energy demand and wind energy output. The trigonometric regression model and the multiple linear regression models have been developed in Sections 2.1 and 2.2 to forecast the energy demand and capacity. Suppose that the electricity demand of k-th community on period t is forecasted to be dkt. The corresponding wind farm needs to evaluate whether it can serve the whole or a part of the community demands on the given period based on the farm’s capacity constraints and predicted environmental/operational conditions (Section 2.2). If the capacity constraint is violated, a portion of dkt cannot be accepted by the farm. By sharing the demands and capacities among collaborative wind

17

farms dynamically, it would be possible for that portion of the demand to be fulfilled by a set of collaborating energy providers, such that the mutual benefits can be achieved; the demand sharing farm fulfils its own community demand, and the capacity sharing farm receives the additional demand. This collaboration will be controlled by the DCSP, which will increase the reliability of demand fulfilment in the network and motivate communities to participate in the network. Let CAkt be the capacity of the wind farm that belongs to community k on period t, and dkt be the forecasted demand of the same community. The shared capacity (SCkt) and shared demand (SDkt) of community k on period t can be calculated as: t t CAk  d k SC   0 t k

t t d  CAk SDkt   k 0

if CAkt  d kt

(5)

otherwise if CAkt  d kt

(6)

otherwise

At the beginning of each period, collaborating communities could be divided into (a) demand sharing communities who need electricity, (b) capacity sharing communities who own excessive energy, and (c) communities who neither needs nor owns extra energy source during that period. Set A and set B contains demand sharing and capacity sharing communities, respectively. The outputs of DCSP at the beginning of period are electricity bidding price provided from community Bjt to community Ai (̂

), damage cost for not being able to fulfil shared capacity (

), and penalty cost for the

unfulfilled demand that have to be paid by buyer Ai for not purchasing the electricity (

).

3.3. Best matching protocol Due to technical limitations, a portion of the generated electricity will be wasted during the transmission and will increase energy distribution costs. This waste of energy is an increasing function of the length of transmission line, from the source to the sink nodes. This phenomenon could affect the efficiency of DCSP. In other words, fulfilling partial demands of a specific community by a wind farm in another community may be inefficient in practice due to the pairwise distances between communities. In addition, the reliability of the energy providers and accuracy of demand and capacity

18

forecasts affects the DCSP efficiency. The goal of BMP is to identify optimal matching between communities (or buyers and sellers) dynamically under these challenges. In a collaborative market, it may more beneficial that a given farm, despite having enough capacity to fulfil its own orders, would be better off leaving a part of these orders to be satisfied by a second wind farm, thus having enough capacity to fulfil an order from a third wind farm. In this setting, all wind farms may belong to both sets A (buyers) and B (sellers), and the BMP will match the best (Ai  A, Bj  B) based on the distance criterion. The BMP, which is indeed an optimization problem, matches the shared demand with the shared capacity so that the demand is fulfilled and the provider’s profit is maximized. The matching problem in this study is a simple two-sided one-to-one assignment which could be solved by using polynomial time algorithms such as Hungarian method. The decision variable is the demand set {̂

̂

̂

} where ̂

is the demand variable that has been assigned to deliver to the

buyer Ai from seller Bj on one period ahead of period . The bidding price offered to buyer Ai from each provider is denoted as ̂

̂

̂

. Different parameters might influence the

bidding price offered by a seller such as market demand, distribution cost, on-hand supply of energy, maintenance schedule, etc.; however, determination of bidding, pricing, and any related penalty costs is beyond the scope of this article. For simplicity, we assume that bidding price is determined by transmission distance between electricity buyer and seller. Other notations used in the model include: real-time electricity price bid offered from seller Bj to buyer Ai ̂

electricity price bid offered from seller Bj to buyer Ai

Ci

i-th community; Ci  A or Ci  B, or neither real-time electricity demand of buyer Ai

̂

estimated electricity demand of buyer Ai real-rime electricity demand supplied by seller Bj to buyer Ai

̂

electricity demand promised to be transported from seller Bj to buyer Ai

19

excessive electricity of the wind turbine in community Ci from previous period t electricity generation cost of the wind turbine in community Ci at period t holding cost of the excessive electricity stored by buyer Ai penalty cost for the unfulfilled demand paid by buyer Ai electricity price sold to community Ci by its corresponding provider at period t electricity amount supplied by the wind turbine in community Ci at period t electricity shortage of the wind turbine in community Ci from previous period t maximum capacity of wind turbine in community Ci at period t damage cost for not being able to fulfill shared capacity paid by seller Bj satisfaction profit per unit electricity earned by avoiding energy shortage in buyer Ai total profit of demand sharing communities (A) at period t total profit of capacity sharing communities (B) at period t

The BMP tries to motivate all communities to collaborate more in the network by maximizing the profits of demand sharing and capacity sharing communities considering the maximum capacity of each wind turbine. Different parameters impact the maximum capacity of a wind farm including geographical and environmental conditions, number of wind turbines, wind shadow effect, and reliability of wind turbines. The revenue of a community is obtained by selling electricity to the customers in the network and/or receiving damage cost paid by other communities for not fulfilling their promised shared capacity. The cost for each community in this BMP model includes: 

Holding cost for excessive electricity;



Cost for demand loss (due to losing customer satisfaction and trust); and



Cost for purchasing electricity from other communities.

Equations (7) and (8) show the formula to compute electricity excessive and shortage, respectively: {

∑

}

(7)

20

∑

{

}

(8)

Thus, total profits of demand sharing and capacity sharing communities (

and

) are

computed using the following model: ∑ (

}) })

∑

(̂

∑

∑

)(

{

∑

{

})

)̂

∑ )̂

[(

∑

∑

)(

(

{

}) (

})

{̂

( ∑ ̂

{

{ ̂

{

∑

)(

)(

[(

(

(

})

(9)

})]

∑

(̂

{

∑ (10)

})

(

∑

{

})

∑

(

{̂

})]

(̂

s.t.

) ∑ (11)

∑

̂

̂

All parameters Note that the parameters of the model are defined in two steps: 1) ̂

,̂

are defined before period T based on forecasted demands and capacities and 2)

, and model constraints ,

, and real-

time values are submitted once the real-time demands and capacities are known. Bidding prices (̂

and

) can be decided dynamically by techniques such as bidding mechanisms and

21

multi-agent systems on sequential or simultaneous auctions ([45- 47]); however, this was out of scope, so we assume all the required information is shared among communities and they are willing to collaborate for overall benefits of the entire network rather than competing for its own benefit. Thus, we simply determined bidding prices based on transmission distance. In other words, communities, either demand sharing or capacity sharing, tend to collaborate with closer communities. So they offer the lowest price to the closest community and the price increases as the distance increases. The impact of the proposed platform on profitability of the wind farms through their bargaining power and bidding strategies can be a potential future work. Equation (9) computes the total profit of demand sharing communities A and Equation (10) denotes the total profit of capacity sharing communities B. The role of each community is dynamic, and a community provider, Ci, could be of type A on a day but of type B on another day. Equation (11) shows the constraints against maximizing BMP functions. The first constraint shows the relationship between real-time and estimated demands. The standard error of the distribution

highly depends

on the accuracy of predictions and is smaller for more accurate predictions. The second constraint computes the excessive energy at the end of period t or limits the community not to sell more than what it owns and produces during a period. The third constraint limits the electricity amount supplied by each community to be less than its maximum capacity. The fourth constraint assures that each wind farm does not promise to transport energy more than its maximum capacity after providing energy to its community. The last constraint assures the non-negativity of the variables. More constraints might be added to the program based on participant agreements, e.g., the community fulfilling its own demand before collaborating with others. The output of the BMP model is a dynamic network, where communities who should share their demand and capacity on a specific day are paired. 4. Simulation A simulation study including three communities is developed to evaluate the effect of the proposed DCSP+BMP. Both energy outputs and community demands are stochastic. Actual demands of three PJM subzones including Jersey Central (Zone A), MetED (Zone B), and PennElec (Zone C) are used

22

to develop the trigonometric regression model. Daily and weekly demands along with the regression model are available as illustrated in Fig. 9; however, this simulation study is conducted based on daily figures. Equation (12) shows the trigonometric model used for each subzone: ̂ ̂ ̂

(

)

(

)

(

)

( (

) )

(

( (

)

) )

(

( (

)

) )

(

(12) )

The model developed in Section 2.2 is used to forecast the energy capacity of the wind farms. Lets assume that operational performance of the wind farms located in zones A, B, and C are similar to the ones provided by EirGrid. Thus, the regression model developed with the EirGrid model is used to forecast energy capacity of the simulated wind farms. Historical weather information is used to generate random wind speed and temperature values. One year time-window is considered to simulate the DCSP+BMP protocol on wind energy provision of community-based wind farms. Using one-year simulation period provides enough timeline to evaluate the proposed platform’s performance including seasonal impacts and correlation between energy supply and demand of different zones. According to the DCSP+BMP protocol, the communities have two types of collaborations: 1) before the operation time when they trade electricity based on their forecasted demands and capacities and 2) during the operation when they observe their real-time demands and change their orders after paying the agreed penalties. Each community could reduce their penalty costs by reducing forecasting uncertainties. Since one of the purposes of this study is to increase the reliability of the distribution network to make it sustainable, the communities are considered to sell their excessive energy after serving their own community.

23

Fig. 9. Daily and weekly energy demand of zones A, B, and C during 2011-2013 and estimates by trigonometric regression curves.

Communities try to increase their profit while improving their service reliability. Energy transmission cost is an important criterion that affect the profit. This cost is reflected in both pre-offered and realtime bidding price. Thus, closer communities have higher priority for electricity trade than farther ones. We assume that Distance (A, B) < Distance (B, C) < Distance (A, C). Fig. 10 provides a visual comparison between the collaboration case and the non-collaboration case. The blue line in Fig. 10 illustrates the demand of the three simulated zones in a year. The red line shows how each community could satisfy their own demand using their own daily capacity while green line shows the communities’ ability after using DCSP+BMP. The fluctuations in the capacity of energy providers are

24

as a result of 1) environmental conditions and 2) wind turbines’ operational performances. Although the fluctuations in wind turbine energy outputs are reducible through applying more accurate environmental and fault prediction models, they are unavoidable. The goal of each community is to set its capacity close to the actual demand to avoid holding cost for excessive demand and to minimize the capacity shortage. As Fig. 10 illustrates, the range of fluctuations are smaller and closer to the actual demand when collaboration platform is employed. So, using the collaborative platform could overcome the adverse effects caused by the uncertainties. Note that collaborative platform takes advantage of the systematic fluctuations, both environmental and mechanical ones, to reduce the overall long-term fluctuations that a community deals with; i.e., the platform enables communities to sell their excessive capacity at the time of abundance and/or buy their required energy during the time of shortage to reduce the capacity fluctuations and improve the reliability of sustainable energy network. Community A

Community B

100

200

period

300

200000 150000

energy

100000 0

0 0

Demand Noncollaborative Capacity Collaborative Capacity

50000

100000

energy

150000

200000

Demand Noncollaborative Capacity Collaborative Capacity

50000

100000 0

50000

energy

150000

200000

Demand Noncollaborative Capacity Collaborative Capacity

Community C

0

100

200

300

period

0

100

200

300

period

Fig. 10. Daily demand of three simulated zones (A, B, C) along with daily capacity of the wind farms before and after employing DCSP+BMP.

A Monte-Carlo simulation is conducted for 365 days to show the impact of the proposed platform compared to non-collaborative wind energy distribution networks. We assume that the wind farms are

25

always ready to operate during the simulation period and no maintenance activity or sudden failures occur during the simulation. The simulation is developed in R statistical software and runs 1000 times. In order to assess the platform, total electricity each community sold and purchased during the simulation period, and total demand kept unfulfilled by each community before and after DCSP+BMP are computed. We also measure the total extra electricity each wind farm has to store at the end of the day after serving all its daily demands. In addition to unfulfilled demand and excessive capacity, Root Mean Squared Deviation (RMSD) is introduced to evaluate the platform. Since both excessive capacity and unfulfilled demands are undesirable and costly for the wind farms, RMSD measures the average difference between wind farm’s capacity and demand. Smaller values of RMSD indicate that wind farms are successful in balancing their demand and capacity such that they can reduce the total cost by decreasing unfulfilled demand or holding less excessive capacity. Equation (13) shows the formula to calculate RMSD: √∑

(

)

(13)

where Ct is the capacity of the farm on day t, Dt is the demand on day t, and n is the number of days for the simulation. Table 1 shows the simulation results. Mean and standard deviation of unfulfilled demand, excessive capacity, and RMSD are compared for non-collaborative wind energy distribution networks versus collaborative ones at each zone. According to the results, the average unfulfilled demand has been reduced by 1,728,681, 1,033,526, and 544,986 units and average excessive capacity has been reduced by 927,076, 1,384,907, and 995,211 units for Zones A, B, and C, respectively. Average RMSD, which is an indication of the wind farms’ performance under a certain platform, is also reduced by 5,065, 4,555, and 2,752 units for zones A, B, and C, respectively, which indicates significant cost savings. Fig. 11 illustrates the difference between unfulfilled demand, excessive capacity, and RMSD of the zones before and after employing DCSP+BMP.

26

Fig. 11. Average Unfulfilled Demand, Excessive Capacity, and RMSD of simulated communities (A, B, C) under collaborative (green) and non-collaborative (red) platforms

Smaller unfulfilled demand, excessive capacity, and RMSD under collaborative platform indicate that the wind farms are balancing their demand and capacity more effectively through collaboration. In addition, the standard deviation in Table 1 shows low variability in unfulfilled demand, excessive capacity, and RMSD among 1000 iterations. The mean values for all variables and regions are larger than standard deviation; the ratio of standard deviation to mean varies between 0.03 and 0.48, which represents low variability between scenario results and high precision of the simulation. The results indicate that the communities would increase their benefits through reducing holding costs while fulfilling more customer demands. In addition to the large differences in average unfulfilled demand, excessive capacity, and RMSD of large samples (1000 data points), hypothesis tests are used to show the impact of the collaborative platform on unfulfilled demand, excessive capacity, and RMSD. Equation (14) shows the hypothesis tests where

is non-collaborative network’s mean and

is the collaborative network one. Since

the sample size is large (1000 iterations), t-test is a plausible test. Table 2 shows the p-values and 95% confidence intervals of hypothesis tests. Since the null hypothesis asserts that collaborative platform

27

has no impact on wind farms’ unfulfilled demand, excessive capacity, and RMSD, smaller p-values mean that the proposed platform would significantly improve the wind farms’ performance. As shown in Table 2, p-values of unfulfilled demand, excessive capacity, and RMSD are near zero for all three zones. Thus, the test shows significant improvements in the three measures achieved by collaboration. Also, the largely positive confidence intervals indicate that the proposed collaboration platform significantly reduces the mean unfulfilled demand, excessive capacity, and RMSD for any zone in the energy distribution networks.

(14)

Table 1 Performance measures with and without DCSP+BMP. Zone A Total Electricity transmitted to (365 days) 1,989,876.0 Total Electricity transmitted from (365 days) 1,188,271.0 mean 8,247,421.0 NonCollaborative stdev 180,107.0 Unfulfilled Demand mean 6,518,740.0 Collaborative stdev 167,506.2 mean 6,339,519.0 NonCollaborative stdev 1,799,005.0 Excessive Capacity mean 5,412,443.0 Collaborative stdev 1,844,718.0 mean 52,151.6 NonCollaborative stdev 9,590.1 RMSD mean 47,086.3 Collaborative stdev 10,362.1

Zone B 1,210,468.0 1,561,849.0 5,498,245.0 136,314.0 4,464,719.0 123,032.6 5,532,662.0 1,728,970.0 4,147,755.0 1,784,394.0 40,408.2 8,930.4 35,853.3 9,730.7

Zone C 691,150.0 1,141,375.0 4,575,136.0 133,496.2 4,030,150.0 98,631.0 6,154,340.0 2,429,363.0 5,159,129.0 2,467,106.0 41,103.4 13,007.0 38,351.8 13,478.5

Table 2 Hypothesis test results and confidence intervals for the DCSP+BMP impacts on communities’ unfulfilled demands, excessive capacity and RMSD. Zone A Zone B Zone C Unfulfilled Demand Excessive Capacity

RMSD

P-value Confidence Interval P-value Confidence Interval P-value Confidence Interval

< 2.2e-16 (8,247,421.0, 6,518,740.0) 0.0002026 (6,339,519.0, 5,412,443.0) 0.0002106 (52151.6, 47,086.3)

< 2.2e-16 (5,498,245.0, 4,464,719.0) 4.044e-08 (5,532,662.0, 4,147,755.0) 0.0003446 (40408.2, 35,853.3)

< 2.2e-16 (4,575,136.0, 4,030,150.0) 0.002246 (6,154,340.0, 5,159,129.0) 0.03171 (41,103.4, 38,351.8)

28

5. Conclusions This article addresses a major challenge in creating a sustainable and reliable wind energy distribution network without substantial dependence on traditional resources. The uncertainties in forecasting energy demand and power generation have been a huge obstacle in introducing the cost-effective renewable energy. The major sources of variability in output forecasting include uncertainty in the environmental conditions and forecasting techniques. In this article, a multiple linear regression model is developed for capacity forecasting by using the weather information and energy output of the EirGrid industry. The black-box model used by the company is more accurate and complex than the used model, but since the goal of this research is not to improve forecasting method, we ran the simulation using the simple model. A trigonometric model is also used to deal with demand forecasting as the other major source of uncertainty. Each community tries to reduce the difference between its predicted energy capacity and demand. A part of such differences is caused by the systematic issues that limit the capacity of a wind turbine such as environmental condition and wind turbine operational performance. The other part is due to the uncertainty incorporated in prediction methods. The uncertainty incorporated in both capacity and demand forecasting is inevitable because we are not able to model every single aspect of the nature and human behavior, yet it is manageable through more accurate prediction and detection methods. Moreover, we can take advantage of the collaboration between different communities to reduce the range of fluctuations in a wind farm’s energy capacity and improve the reliability of the sustainable energy network as described in Section 4. Two CCT-based protocols are developed in this research to reduce the negative impact of the uncertainty by collaboration and to create a sustainable distribution network. In the DCSP+BMP platform introduced in Section 3, forecasted demand and capacity are used, as model inputs, to define communities who need or provide energy for the following period. Once communities are divided into demand sharing or capacity sharing groups, the best trade strategy that maximizes the community profit is developed using the BMP. BMP records both predicted and real-time capacity/demands

29

during the collaboration periods and provides a reliability trajectory system that enables users trace how system reliability changes over time. The simulation results show that collaboration protocols can improve the sustainability of the network by reducing the amount of unfulfilled demands and excessive capacity and also by decreasing the variability of the difference between forecasted demands and capacities. The accuracy of our prediction model is comparable to the industry-level model; however, the collaboration protocols can be more effective in dealing with deviations from predictions that could arise by using a less accurate model. Reduced uncertainty by accurate prediction and enhanced collaboration can make the wind energy more available, reliable and thus responsive to demand. As discussed by [48, 49], reliability, availability, maintainability, and demand response play vital role in sustainability and profitability of wind energy systems. However, measuring the effect of the collaborative network in terms of the total profit is challenging as it may involve additional uncertainty. [50] conducted a thorough study on wind farm’s total costs including operational and non-operational costs. Their study demonstrates that the price of operating farm depends greatly on many assumptions regarding the operational flexibility of the generation fleet. In addition, a large fraction of the regulation price in this analysis was derived from the assumed generator bid prices. Information related to the regulation prices and generators’ ability to provide service is not publicly available and prone to various uncertainties which is out of the scope of this research. Energy holding costs and unfulfilled demands are large components of the wind farm’s operational costs that introduce large uncertainty. Therefore, reproducing the cost of reserves in a production cost model involves significant uncertainty. The comparisons between collaborative wind energy communities vs. non-collaborative ones, discussed in Section 4, shows that average unfulfilled demand and excessive capacity have been reduced by applying the proposed collaborative platform. Thus, unfulfilled demand costs and holding costs of the wind farm will be decreased and the availability of the wind energy will be increased. The saving amount of the total costs for each wind farm is different and depends on various factors, such as energy market price, labor costs, energy storage technology, and market competition. As a future study, we will analytically compare the impacts of the proposed protocol, on demand fulfillment costs, with other

30

optimization methods such as the ones discussed in Section 1 ([13] and [23]). Also, the analyses will consider modeling the operational and non-operational costs of wind farms and analyzing sensitivity of the total costs to demand and holding costs. A well-defined community-based network of hybrid renewable energy can enhance smart grid development and improve the sustainability of power generation processes while reducing dependence on fossil fuel and nuclear energy. In the future, this study will extend the scope of the model by including precise failure prediction models in forecasting energy generation as well as including wind shadow impacts on wind farm’s maximum capacity. The proposed trigonometric model has significantly improved the accuracy of predictions; however, a further research work is being conducted to improve the model by including more variables such as year index and net income of the people in a community. In addition, other statistical and machine learning models will be evaluated to improve the accuracy of energy output predictions with more factors involved such as seasonality. The DCSP+BMP platform can be extended for other sustainable energy resources, such as water turbines and solar panels, to be integrated into the sustainable smart grid network. In addition, other matching methods and algorithms will be investigated to improve the efficiency of BMP and the collaboration platform. A dynamic pricing mechanism will be developed in the platform to determine the bidding price, real time price, penalty cost, and other damage costs. For this purposes, cooperative and non-cooperative game theory strategies can be adopted.

References [1] K. Coughlin, J.H. Eto, 2010, Analysis of wind power and load data at multiple time scales, Lawrence Berkeley National Laboratory Technical Report. [2] J. Enslin, 2009, Grid impacts and solutions of renewables at high penetration levels, Quanta Technology Technical Report. [3] H. Lund, 2005, Large-scale integration of wind power into different energy systems, Energy, 30(13), 2402-2412. [4] H. Lund, W. Kempton, 2008, Integration of renewable energy into the transport and electricity sectors through V2G, Energy Policy, 36(9), 3578-3587. [5] U.S. Department of Energy, 2008, 20% wind energy by 2030 - increasing wind energy’s contribution to U.S. electricity supply, EERE Technical Report.

31

[6] E. Miguelanez, D. Lane, 2010, Predictive diagnosis for offshore wind turbines using holistic condition monitoring, Proceedings of OCEANS, 1-7. [7] S. Arabgol, H.S. Ko, S. Esmaeili, 2014, Artificial neural network and EWMA-based fault prediction in wind turbines, Proceedings of the Industrial and Systems Engineering Research Conference. [8] R.G. Fernandes, D.C. Silva, L.H. de Oliveira, A.D.D. Neto, 2007, Faults detection and isolation based on neural networks applied to a levels control system, Proceedings of the International Joint Conference on Neural Networks, 1859-1864. [9] D.R. Hush, C.T. Abdallah, G.L. Heileman, D. Docampo, 1997, Neural networks in fault detection: a case study, Proceedings of the American Control Conference, 918-921. [10] E. Jahanpour, H.S. Ko, S.Y. Nof, 2014, Collaboration platform for sustainable wind energy distribution network, Proceedings of the 19th IFAC World Congress, 4266-4271. [11] O. Uluyol, G. Parthasarathy, 2012, Multi-turbine associative model for wind turbine performance monitoring, Proceedings of the Annual Conference of the Prognostics and Health Management Society, 1-10. [12] X. Xu, J.W. Hines, R.E. Uhrig, 1999, Sensor validation and fault detection using neural networks, Proceedings of the Maintenance and Reliability Conference. [13] J. Matevosyan, L. Soder, 2006, Minimization of imbalance cost trading wind power on the shortterm power market, IEEE Transactions on Power Systems, 21(3), 1396-1404. [14] Y.D. Song, B. Dhinakaran, X. Bao, 2000, Control of wind turbines using nonlinear adaptive field excitation algorithms, Proceedings of the American Control Conference, 1551-1555. [15] L. Magnano, J.W. Boland, 2007, Generation of synthetic sequences of electricity demand: application in South Australia, Energy, 32, 2230–2243. [16] R. Ramanathan, R. Engle, C.W.J. Granger, F. Vahid-Araghi, C. Brace, 1997, Short-run forecasts of electricity loads and peaks, International Journal of Forecast, 13, 161–174. [17] J. Taylor, 2003, Short-term electricity demand forecasting using double seasonal exponential smoothing, Journal of Operations Research Society, 54, 799–805. [18] C. Wang, G. Grozev, S. Seo, 2012, Decomposition and statistical analysis for regional electricity demand forecasting, Energy, 41(1), 313-325. [19] A. Azadeh, S.F. Ghaderi, S. Sohrabkhani, 2008, A simulated-based neural network algorithm for forecasting electrical energy consumption in Iran, Energy Policy, 36(7), 2637–2644. [20] S. Kucukali, K. Baris, 2010, Turkey's short-term gross annual electricity demand forecast by fuzzy logic approach, Energy Policy, 38(5), 2438–2445. [21] K. Metaxiotis, A. Kagiannas, D. Askounis, J. Psarras, 2003, Artificial intelligence in short term electrical load forecasting: a state-of-the-art survey for the researcher, Energy Conversion and Management, 44(9), 1525–1534. [22] H. Cho, Y. Goude, X. Brossat, Q. Yao, 2013, Modeling and forecasting daily electricity load curves: a hybrid approach, Journal of the American Statistical Association, 108(501), 7-21. [23] N. Gensollen, V. Gauthier, M. Marot, M. Becker, 2013, Modeling and optimizing a distributed power network: A complex system approach of the prosumer management in the smart grid, Survey arXiv:1305.4096v1. [24] S.Y. Nof, 2007, Collaborative control theory for e-work, e-production, and e-service, Annual Reviews in Control, 31(2), 281-292. [25] H.S. Ko, S.Y. Nof, 2012, Design and application of task administration protocols for collaborative production and service systems, International Journal of Production Economics, 135(1), 177-189.

32

[26] M. Moghaddam, S.Y. Nof, 2014, Combined demand and capacity sharing with best matching decisions in enterprise collaboration, International Journal of Production Economics, 148, 93– 109. [27] H. Seok, S.Y. Nof, 2013, Dynamic coalition reformation for adaptive demand and capacity sharing, International Journal of Production Economics, 147, 136–146. [28] S.W. Yoon, S.Y. Nof, 2010, Demand and capacity sharing decisions and protocols in a collaborative network of enterprises, Decision Support Systems, 49(4), 442-450. [29] S.W. Yoon, S.Y. Nof, 2011, Affiliation/dissociation decisions in collaborative enterprise networks, International Journal of Production Economics, 130(2), 135-143. [30] J.D. Velasquez, S.Y. Nof, 2008, A best-matching protocol for collaborative e-work and emanufacturing, International Journal of Computer Integrated Manufacturing, 21(8), 943-956. [31] J.D. Velasquez, S.Y. Nof, 2009, Best-matching protocols for assembly in e-work networks, International Journal of Production Economics, 122(1), 508-516. [32] S. M. Kaplan, 2009, Electrical power transmission: background and policy issues, Congressional Research Service. [33] PJM, http://www.pjm.com/about-pjm.aspx, accessed on 7/29/2014. [34] R. J. Hyndman, A.B. Koehler, 2006, Another look at measures of forecast accuracy, International Journal of Forecasting, 22(4), 679-688. [35] E. Jahanpour, X. Chen, 2014, Improving accuracy of complex network modeling using maximum likelihood estimation and expectation-maximization, Discontinuity, Nonlinearity, and Complexity, 3(2), 169-221. [36] I.G. Damousis, M.C. Alexiadis, J.B. Theocharis, P.S. Dokopoulos, 2004, A fuzzy model for wind speed prediction and power generation in wind parks using spatial correlation, IEEE Transactions on Energy Conversion, 19(2), 352–361. [37] U. Focken, M. Lange, K. Mönnich, H.P. Waldl, H.G. Beyer, 2002, A. Luig, Short-term prediction of the aggregated power output of wind farms – a statistical analysis of the reduction of the prediction error by spatial smoothing effects, Journal of Wind Engineering and Industrial Aerodynamics, 9, 231–246. [38] A. Kusiak, H. Zheng, Z. Song, 2009, Short-term prediction of wind farm power: a data mining approach, IEEE Transaction on Energy Conversion, 24(1), 125-136. [39] J.L. Torres, A. García, M. de Blas, A. de Francisco, 2005, Forecast of hourly averages wind speed with ARMA models in Navarre, Solar Energy, 79(1), 65–77. [40] X. Wang, P. Guo, X. Huang, 2011, A review of wind power forecasting models, Energy Procedia, 12, 770-778. [41] A. M. Foley, P.G. Leahy, A. Marvuglia, E. McKeogh, 2012, Current methods and advances in forecasting of wind power generation, Renewable Energy, 37(1), 1-8. [42] C. Mohrlen, 2004, Uncertainty in wind energy forecasting, PhD Thesis, University College Cork. [43] A. Adams, D. Keith, 2013, Are global wind power resource estimates overstated?, Environmental Research Letter, 8(1), 1-9. [44] J. McGarry, D. Beasant, 2008, Load Forecast, Scheduling and Dispatch Workshop. [45] R. K. Dash, P. Vytelingum, A. Rogers, E. David, N.R. Jennings, 2007, Market-based task allocation mechanisms for limited capacity suppliers. IEEE Transactions on Systems, Man and Cybernetics: Part A, 37(3), 391–405. [46] E. H. Gerding, R. K. Dash, A. Byde, N. R. Jennings, 2008, Optimal strategies for bidding agents participating in simultaneous Vickrey auctions with perfect substitutes, Journal of Artificial Intelligence Research, 32, 939-982.

33

[47] G. Xiong, S. Okuma, H. Fujita, 2004, Multi-agent based experiments on uniform price and payas-bid electricity auction markets, Proceedings of the 2004 IEEE International Conference on Electric Utility Deregulation, Restructuring and Power Technologies, 1, 72-76. [48] B. Dusi, D. Emery, Demand response strength sustainability, grid reliability, and profitability, 2011, Cement Americas Technical Report. [49] F. Besnard, K. Fischer, L. Bertling, 2010, Reliability-Centred Asset Maintenance — A step towards enhanced reliability, availability, and profitability of wind power plants, IEEE PES Innovative Smart Grid Technologies Conference Europe, 1-8. [50] M. Hummon, P. Denholm, J. Jorgenson, D. Palchak, B. Kirby, O. Ma, Fundamental drivers of the cost and price of operating reserves, National Laboratory of the U.S. Department of Energy Technical Report.

Highlights 

We propose a collaboration platform & protocols for sustainable wind energy networks.



Two protocols are designed: demand and capacity sharing and best matching protocols.



Regression analysis is conducted to estimate energy demand and capacity.



Simulation is conducted to assess impact of protocols on collaborative performance.



The collaboration protocols improve network sustainability by handling variability.

34