A novel algorithm for multi-node load forecasting based on big data of distribution network

Transcription

1 International Conerence on Advanced Electronic Science and Technology (AEST 206) A novel algorithm or multi-node load orecasting based on big data o distribution network Guangsong Hou, Ke Xu, Shoubin Yin, Yang Wang, Yan Han, Zhiguo Wang, Yiei Mao 2 Zhengxin Lei 2,a and State Grid Shandong electric power company Heze power supply company, Heze, , China Key Laboratory o Control o Power Transmission and Conversion, Ministry o Education, Shanghai Jiao Tong University, Shanghai, , China 2 Abstract. Eective load orecasting or dierent scales o loads is essential or the planning and operation o distribution network. The extending scale o the network, the diversity o load types and the rapid increase o data volume are some o the acing problems. In this paper, we propose a novel method or multi-node load orecasting, AR-ANN, which takes those problems into consideration. This new algorithm makes a combination o AR method and BP neural network method while eliminating their disadvantages. Comparing to traditional bottom-up method, a top-down method is more applicable when considering the limitation o measurement equipment in distribution network. Both top-down method and bottom-up method are tested in this paper by using AR-ANN algorithm. The data processing speed and the orecasting accuracy o AR-ANN is validated by several tests: an ordinary single-node load orecast, two multi-node load orecasts by traditional bottom-up method and by the new top-down method. Keywords: Multi-Node Load Forecast; BP-ANN; AR; big data; distribution network. Introduction Because o the increasing load o China and the modiication o the distribution network structure, dierent kinds o system analysis, which mainly consider load low analysis, N- contingency analysis and emergency power supply, should have more detailed data to support. Thus multi-node load orecasting or changeable distribution networks based on big data is needed. Traditional multi-node load orecasting can reduce the prediction error due to the more characteristically obvious nodes compared to the whole. Nowadays, the load o active users becomes more random with the insert o dierent new kinds o power generating and consuming devices, thus load orecasting o some terminal nodes and users is particularly emphasized. Meanwhile, the developing distribution network involves wide range and big volume o data, which makes ast data processing important in both traditional and big data-based load orecasting; while additional installation o measurement equipment is economically unavorable, it is necessary to orecast load by using the existing data rom substations, electric power lines and smart meters. Researchers have been doing some optimization and modiication to the classical multi-node load orecasting method based on distributed computing theory. An improved methodology based on general a Corresponding author : iomeiya@sjtu.edu.cn 206. The authors - Published by Atlantis Press 655

2 AEST206 regression neural networks or short-term multi-node load orecasting is introduced in [] and it also proposed another method which aims to limit calculation time by reducing input variables. Reerence [] compared the two methods with traditional general regression neural networks by tests on a New Zealand distribution subsystem. Additionally, the author divided the methods into local load orecasting, global load orecasting and participation actor orecasting. Reerence [2] constructed a ramework o sel-adapting dynamic load models o ultra-short-term orecasting or multi-node active and reactive load based on the ideas o "hierarchy" and sub-area. Reerence [3] used uzzy-rough sets to determine the initial weights o artiicial neural networks or short term load orecasting in order to select the most signiicant input variables. Among the improved conventional orecasting techniques, the stochastic time series approach is most popular[4] because o its accuracy. AR (Auto-Regressive Model), one o the time series analysis approach characterized by added linear ilter, has high short-term orecasting accuracy however is unable to process data o multiple sources. On the other hand, artiicial intelligence techniques have been much studied these years to apply or load orecasting. The models that have received the largest share o attention are the artiicial neural networks (NNs)[5]. A great number o papers have done practical tests with these nonlinear algorithms and compared the results. Typical NN methods are BP-ANN (Back Propagation Artiicial Neural Network), ELMAN Neural Network, FNN (Fuzzy Neural Network) and WNN (Wavelet Neural Network). A recent study o these our NN method compared their orecasting accuracy through dierent indicators[6]. The results o this study is shown in the table below. Table. Perormance comparison o our NN methods[6] Unprocessed data Processed data Perormance index Mean squared error (MSE) (%) Root-mean-sq uare error (RMSE) (%) Maximum error (%) Mean squared error (MSE) (%) Root-mean-sq uare error (RMSE) (%) Maximum error (%) BP-ANN ELMAN NN FNN WNN We can see that BP-ANN has relatively better perormance among theses NN methods both or unprocessed data and processed data. BP-ANN method can handle multi-source data but the prediction accuracy is aected by the setting o similar days. Considering the disadvantages o AR time series model and BP-ANN method, this paper proposes a load orecasting method, AR-ANN, which complements the advantages o AR and BP-ANN by adding AR into the training part o BP-ANN. For the load requirements, this paper presents a top-down multi-node big data based load orecasting method to ace to the contradiction between the limitation o data collection and the need o node load orecasting o distribution network. The main part o this paper is divided into three parts. The irst part introduces load orecasting models and raises the new orecasting method. The second part mainly presents the AR method and 656

3 AEST206 BP-ANN method and the last part gives some tests o global load orecasting and two types o multi-node load orecasting by using AR-ANN method. The global load orecasting test validates the advantage o AR-ANN method in prediction error when comparing to BP-ANN method, and the traditional multi-node load orecasting test conirms that AR-ANN method is less time consuming than global load orecasting. In the third test where AR-ANN is applied to the new multi-node load orecasting, it is shown that the AR-ANN is more accurate than BP-ANN. 2 Models o multi-node load orecasting According to the related papers[-3]with author's own research experience, the deinition o multi-node load orecasting can be summarized as ollows: multi-node load orecasting is a distributed load orecasting in units o node which takes dierent sources o actors into consideration like the types o load, the electricity price and traic condition. A node is a part o the network with common eatures (single user, whole o part o eeder line, substation etc.) The "nodes" in the multi-node load orecasting are mutually exclusive and the set o the nodes orms the "whole". As shown in Figure, the structure o the nodes is branched rom top to bottom and can be divided into several layers D D2 D3 Figure. Structure o multi-node load orecast In the actual load orecasting o distribution network, the structure is normally constructed by the layers o regional power grid, provincial power grid, main transormer line and the lower level voltage lines. The number o the arrows that lead to the node represents the layer o the node. Those lightest blue circles (node no.7 to no. 24) in the bottom represent the lines o 35kV and 0kV. How to deinite "whole" and "node" depends on the concrete orecasting problem. Take Figure or example, we can choose those regional power grid as "whole" and the provincial power grid as "nodes". We can otherwise choose D - D3 parts as orecasting part so that those topmost nodes o these three parts are "wholes" and the others are "nodes". Traditional methods mainly based on the bottom-up idea: start rom the measured data o a single node, distributed orecast the changing load o the other loads, then get the whole load orecasting by integrating the results o all the nodes within. The mathematical expression can be summarized as: n F( X, X,, X ) = η ( x, x,..., x ), i =,..., n () 2 m i, i,2 im, i= where F represents the orecasting algorithm o the whole and X,, X m are the inputs o F; represents the orecasting algorithm o the node and x i,,...,x i,m are its inputs; i means the ith node in the network; n is the number o nodes, η is the ratio between the sum o the and F due to the line loss or other reasons. However, this bottom-up method might not be eicient in all cases. The data rom the bottom or the terminal nodes o the distribution network can be incomplete or diicult to achieve in real-time due to lots o reasons like the independent data bases o dierent layers and permissions, the lack o measurement equipment or the instability o devices. Facing all this disadvantages, we proposes another method o multi-node load orecasting, "rom the whole to the nodes", where the orecasting o the 657

4 AEST206 "whole" load can be an auxiliary manner to orecast the node load by use o the historical data o node loads when the real-time data o the nodes are unavailable. Its mathematical expression is as ollows: ( x, x,..., x ) = β F( X, X, X ), i =,..., n (2) i, i,2 im, i 2 m where β represents the ratio o node load to global load diagonal matrix. i Since equations () and (2) speciy neither the algorithm o load orecasting nor the time span o load orecasting, we now start modeling o the pre-day load orecasting by the AR-ANN method. 2. Multi-node load orecasting by AR-ANN Load orecasts can be divided into three categories: ultra-short-term [7], short-term [8] and long-term orecasts [9]. Short-term orecasts are real-time orecasts which orecast every iteen minutes. Short-term orecasts are usually rom one hour to one day, and long-term orecasts are by month or by year. Pre-day load orecasting, which has a orecasting time span o 24 hours (rom 0h to 24h), belongs to short-term orecasts though it orecasts every hal hour or per hour which is more requent. Thus short-term load orecasting methods can be applied to pre-day load orecasting and because o its orecasting requency, lots o researches have been done in the selection o similar days [0-2]. Considering the need o rapid processing o big data, this paper simpliies the algorithm o selecting similar days and chooses experienced similar days which base on day type [3]. Load orecasting algorithms, according to how historical load derives to orecasting load, can be classiied under extrapolation or correlation. ARMA [4], as one o the most common method o extrapolation, establishes a linear unction o orecasting load, historical load and historical error. ANN (artiicial neural networks) [5], as an example o correlation method, builds non-linear circuits between orecasting load, historical load, weather data and other multi-source actors. Load orecasting methods can otherwise be divided into statistical or artiicial intelligence techniques. Statistical techniques include regression methods [6], time series, exponential smoothing model [7], Kalman ilter based method [8], state estimate method [9]. Artiicial intelligence techniques are such as ANN, uzzy logic [6] and gray system theory[20]. The choice o similar days in this paper has already taken seasons and other attributes into consideration, so that the established time series orecasting model does not relect seasonal variations. Under such condition, ARMA model takes only AR (auto-regressive) part and needs no MA (moving average) part [2]. Although AR method has high accuracy o ultra-short-term load orecasting, it cannot process data o multiple sources. Meanwhile the prediction error augments over time which becomes a disadvantage or pre-day orecasting. On the other hand, BP-ANN can handle multi-source data, however its prediction error is largely aected by the selection o similar days and network parameters. Comparing this two methods, we adjust the training inputs o BP-ANN by the results o AR model, which orms the novel algorithm AR-ANN proposed in this paper. Figure 2 shows the basic lowchart o multi-node load orecasting applying AR-ANN algorithm. The overall process contains mainly our steps: data reading, mode judgment and parameter setting, AR-ANN iteration and inal result calculation with output display. In the step o data reading, the whole, numbered 0, is the higher level node as described in Figure. In AR-ANN, both AR orecasting results and multi-source data o same period o time are used or training BP-ANN model, so that pre-day data o orecasting should be provided or data reading. Mode judgment and parameter setting decide the number o iterations and the ormula with its parameters. To improve the operational eiciency, we set η in equation () to be and let β i in equation (2) be the ratio diagonal matrix o similar days by chronological order. AR-ANN iteration could be either bottom-up which orecasts the load o node 0 to node i by distributed computing, or top-down which calculates directly the global load only. Since AR-ANN is a modiied algorithm o BP-ANN, the inal result calculation ollows the same way o BP-ANN which will be described urther later in Section

5 AEST206 Start Data Reading Read: data o three days - similar day, orecasting day, preday o orecasting day Design the algorithm o ormula (k): F/=AR-ANN Number nodes : -n; Deine the whole as node 0 Apply orecasting algorithm or node 0 Apply orecasting algorithm or node i Distributed Iteration No Bottom-up method? Mode Judgment and Parameter Setting Initialize the parameters: i=0;k=2 Yes Initialize the parameters: i=;k= Calculate the inal orecasting value with ormula (k) Result Calculation and Output Display Chart the data Deine/calculate the parameters o ormula (k) End Figure 2. Flow chart o AR-ANN multi-node load orecast The detailed AR-ANN based load orecasting process is as ollows: Step : Use time series AR algorithm to achieve the irst predictive value o the orecasting day by the load curve o the day beore. (The AR algorithm will be explained in Section 3..) For the non-pre-day load orecasting, the predictive value o AR algorithm will not still be the irst point o the orecasting day since the real-time value is available during the prediction. We treat this situation as a short-term orecasting with increasing real-time data, which will be calculated in a test in Section 4.. Step 2: Set the predictive value o AR and the other sources o data (such as electricity price and weather data) as a new group o training sample points which replaces the time-corresponding group o training points in BP-ANN. Finish load orecasting by BP-ANN algorithm (details in Section 3.2.). Step 3: Record results and calculate prediction error. 3 Forecasting algorithms This chapter is mainly committed to two algorithms: AR and BP-ANN, or that AR-ANN is based on the previous two. 3. AR orecasting algorithms AR method is a time series method which bases on the assumption that there is an internal structure o autocorrelation in data. AR model can be used or stationary processes and it orecasts the next time point value by ixing the order and the parameters o the model. Since load can be greatly inluenced by weather conditions and some social actors, time series method maintains its stability in short period o time. AR method is usually applied or short-term or ultra-short-term load orecasting only. The three steps o AR algorithm: Step : Data preprocessing. Dierence processes are needed or those short-term load time series with inadequate stability. The second dierence is usually enough. Ater that, calculate the mean value and variance o the time series and standardize it. Step 2: Fix the order and estimate the parameters o the model. The AR model is illustrated in equation (3). Y = c+ φy + φ Y + + φ Y + ε (3) t t 2 t 2... p t p t 659

6 AEST206 εt is white noise; c represents a constant related to the model and its deault φ,..., φp are the parameters o each order. The establishment o AR model is to ixε t, the Where the process value is 0. order p and the parametersφ p. εt usually takes its mean value 0; order p can be estimated by plotting the partial autocorrelation unctions and then considering those unctions or the residuals o the model. AIC (Akaike Inormation Criterion) [22] is applied or inding p. Finally the values o the parameters are calculated to minimize the error term. Step 3: Calculate the orecasting value Yt rom the AR model with inverse dierence to replace part o the training samples o BP-ANN. 3.2 BP-ANN Algorithm or multi-node load orecasting BP-ANN is an approach in supervised learning inspired by biological neural networks. The actual numerical weights assigned to element inputs are determined by matching historical data to desired outputs or several times in a pre-operational training session. BP-ANN is a non-linear method, contrast to AR, and it can process massive parallel computing with multiple sources o data[23], which makes it more popular than AR and can be applied to longer term orecasting. However the results o BP-ANN can be aected by the selection o similar days in pre-day load orecasting. Figure 3 shows the structure o Artiicial Neural Network arranged in three layers, one input layer, one hidden layer and one output layer. Each segment in Figure 3 represents a non-linear unction to its let side element, which is composed by a linear part o layer weights and a non-linear part o activation unction. The linear unction between the input layer I and the hidden layer J is shown in equation (4). Where y is the output o node i; i j y= ( wx) (4) i ij j j x is the output o the upper node j; wij is the layer weight orm node i to node j; is the activation unction. We have a similar linear unction between the hidden layer J and the output layer K where the corresponding weight is w in Figure 3. X X2 X3 jk... w ij wjk... Y Yk Xi Figure 3. Structure o ANN Inputs Input Layer I Hidden layer J Output Layer K Outputs There are a number o common activation unctions. This paper chooses a standard log-sigmoid unction (also known as a logistic unction), given by (5). 660

7 AEST206 yi = + exp(- wx ij j ) j (5) With the architecture o ANN in Figure 3, the task o BP is to ind the values o weights. In short, the process o BP-ANN can be summarized in three parts: data preprocessing, data training and data orecasting. For the time eiciency o big data processing, the ANN applied in this paper is o three layers and the number o hidden nodes is determined by Kolmogorov s Theorem [24]. Data training part iterates the local gradients o the layers and the weight vectors until the MSE (minimum square error) reaches the acceptable range or the maximum number o iterations is reached, so that the weights are corrected. Ater that the parameters o ANN structure are determined, data orecasting part calculates the normalized output rom the normalized input and then gets inal orecasting results by a reverse o normalization process. The low chart o BN-ANN is presented in Figure 4. In particular, or multi-node load orecasting, the training data is deined to be the all types o data o the same time on the similar day. Thus the orecasting data is o the corresponding type. The complete BP-ANN multi-node load orecasting process is similar to Figure 2, except that the algorithm should be BP-ANN and it needs no pre-day load. Start Maximum iteration time attained? Yes Set parameters: hidden layer number, node number or each layer, maximum iteration time, learning speed, allowable errors... Reverse iterate the weight vectors Calculate the local gradient o each layer No Training vectors get the outputs o each hidden layer and output layer by the activation unctions Data training Data Preprocessing Read data: time, temperature, price, week number, load. Decide training vectors No Within the maximum allowable error? Yes Normalize training data (min-max normalization) Read real load data and the orecasting vector (consists o time, temperature, week and price) Decide initial values o weight vectors Normalize orecasting vector Iterate once the orecasting vector with the trained weight vector; get the orecast value ater reverse normalization Data orecasting Chart the data; Analyze the results End Figure 4. Flow chart o BP-ANN 4 Simulation tests and results According to ormula () and (2) in Section 2, multi-node load orecasting is the outcome o linear superposition o single-node load orecasting results. Since the errors in linear calculations are cumulative, we test the accuracy o AR-ANN algorithm by comparing the single-node load orecasting results o AR-ANN and BP-ANN, which is test. Test 2 tries to compare the traditional bottom-up method with global load orecasting (regard the whole load as a single node) by both using AR-ANN algorithm. This test can also veriy the application o AR-ANN algorithm to this type o multi-node load orecasting. Finally in test 3, AR-ANN is applied to the top-down load orecasting case and compared with 66

8 AEST206 BP-ANN algorithm, to validate the advantages o AR-ANN and top-down method in multi-node load orecasting. These three tests which are implemented on MATLAB include the load orecasting o holidays like st January and ordinary working days. The accuracy o the results is relected by maximum relative error, average and root mean square error. 4. Test : AR-ANN and BP-ANN or single-node load orecasting Today the big data with complex structures and giant size[25] updates its database and its volume quickly[26]. The big data load orecasting based on AR-ANN can utilize the latest measured data as AR in AR-ANN method can take advantage o the latest updated weather data to orecast the load. Based on the pre-day orecasting, we orecast the 24 points o the orecasting day ( point/hour). While adding the measured load data into the algorithm and orecasting the rest points o the day, there are 24 situations which correspond to dierent numbers o points to orecast. The data comes rom PJM Markets and Operations Database[27], and the load o the region run by AP Company is chosen to be the data or the whole load orecasting. The special orecasting day is st January 205 and its similar day is st January 204. The ordinary orecasting day sets to be 3 th January 205 (Tuesday) with its similar day 6 th January 205, Tuesday. PJM provides the hourly load (24 values/day) with the electricity price. We also obtain the weather data (the highest and lowest temperatures) o the load center or the last two years. Considering the parameter setting, we take pre-day load orecasting or example: the node number or the input layer o ANN is 96 (); node number or the hidden layer is 93 (); the output layer has 24 nodes; maximum iteration number is 000 and maximum allowable error is BP-ANN and AR-ANN are respectively applied to orecasts these 24 situations mentioned beore. The orecasting results o the special day (Figure 5 Figure 7) and the ordinary day (Figure 8 Figure 0) show that in most cases AR-ANN is more accurate than BP-ANN with smaller maximum relative error, smaller average and smaller root mean square error. This means that or the most widely used single-node load orecasting, AR-ANN improves the traditional BP-ANN method with smaller errors or pre-day load orecasting. Even under the conditions o data update, AR-ANN adjusts well and maintains its advantages. Figure 5. Comparison o MRE (maximum ) o load orecast on 0/0/205 (special day) under multiple scenarios Figure 6. Comparison o ARE (average relative error) o load orecast on 0/0/205 (special day) under multiple scenarios 662

9 AEST206 Figure 7. Comparison o RMSE o load orecast on 0/0/205 (special day) under multiple scenarios Figure 8. Comparison o MRE o load orecast on 3/0/205 (ordinary day) under multiple scenarios Figure 9. Comparison o ARE o load orecast on 3/0/205 (ordinary day) under multiple scenarios Figure 0. Comparison o RMSE o load orecast on 3/0/205 (ordinary day) under multiple scenarios 4.2 Test 2: AR-ANN or bottom-up multi-node load orecasting The data needed or load orecasting are possible to be in dierent databases [28]. Big data based load orecasting can be carried out by distributed methods with divide and conquer techniques [25]. Bottom-up multi-node load orecasting is exactly one o the distributed applications in the context o big data. Test 2, based on the irst test, orecasts the load rom bottom to top. The whole is the PJM-WEST part which includes AP region. PJM-WEST has 8 nodes in total and AP region is one o those. The other parameters o test 2 are the same as the irst one. Figure and Figure 2 give the orecasting results; table 2 and table 3 show the prediction errors o the special day and the ordinary day. We can see rom the prediction errors that AR-ANN algorithm is eective or bottom-up multi-node load orecasting and this kind o load orecasting is slightly better than global load orecasting. Even though the results might be sometimes close to each other, multi-node orecasting consumes relatively less time by adapting the technique o distributed computing. In actual applications, AR-ANN multi-node load orecasting algorithm can not only detective the load variation o dierent nodes in distribution networks with accuracy, but also shorten the orecasting time by predicting local loads. 663

10 AEST206 Figure. Comparison o dierent orecasting methods on 0/0/205 (special day) in PJM-WEST Table 2. Forecast statistics o Test 2 on a special day ( st January 205) Type o error Maximum (MRE) Average (ARE) AR-ANN global orecast AR-ANN multi-node orecast 3.50% 3.33%.95%.84% RMSE 2.29% 2.7% Figure 2. Comparison o dierent orecasting methods on 3/0/205 (ordinary day) in PJM-WEST Table 3. Forecast statistics o Test 2 on an ordinary day (3 th January 205) Type o error Maximum (MRE) Average (ARE) AR-ANN global orecast AR-ANN multi-node orecast 6.4% 6.20% 2.35% 2.39% RMSE 2.89% 2.89% 4.3 Test 3: Top-Down multi-node load orecasting In the environment o big data, the quick access between each database is not ideal [29] due to the actors like dierent permissions, dissimilar data structures and interace instability. With the developing size o distribution networks and higher density o measurements, missing data and abnormal data caused by terminal hardware ailure increase. These lead to a growing demand or the eiciency o load orecasting. Thus this paper proposes a top-down method or multi-node load orecasting. The data o test 3 comes rom the SCADA system o Heze city, Shandong Province. The load o City Line I (35 kv) is considered as the whole, and its Distribution Transormer Line I (DTI) carries the load as one o the nodes. Since not all the lines o the DTI have measuring devices, we orecast the load o these non-measure lines with the help o whole load. In this test the special day chosen is the 5 th August 205(Saturday) with its similar day the 8 th o the same month and same year; the ordinary orecasting day or test 3 is the th August (Tuesday) in the same year and its similar day is the 4 th. The measured load data o City Line I contains all the points including those o the pre-day, while DTI provides only the data o the similar day. Combing these data with the TOU (Time O Use) price policy o Heze city and its extreme temperatures o each day, we construct a similar ANN structure to test with the same parameters to orecast all the 24 load points ( point/hour) o DTI on the 5 th August by using AR-ANN algorithm and BP-ANN algorithm respectively. The orecasting load curve o two methods with the actual load curve are illustrated in Figure 3 and Figure 4. The average o AR-ANN is 7.96%, compared to 8.36% o BP-ANN (in Table 4). Thus top-down AR-ANN multi-node load orecasting produce better results than direct BP-ANN orecasting. Though with the reduce o user number, the increase o random actors and the occurrence o some special circumstances, the results might turn to be deviated like the ones o 20h and 2h in the 664

11 AEST206 test; the remains within 0% or the most cases. Figure 3. Comparison between orecast load and actual load o DTI o City Line I on 5/08/205 (special day) Table 4. Forecast statistics o Test 3 on a special day (5 th August 205, Saturday) Type o error Maximum (MRE) Average relative err(are) BP-ANN load orecast AR-ANN multi-node orecast 35.99% 35.78% 8.36% 7.96% RMSE.63%.59% Figure 4. Comparison between orecast load and actual load o DTI o City Line I on /08/205 (ordinary day) Table 5. Forecast statistics o Test 3 on an ordinary day (thaugust 205, Tuesday) Type o error Maximum (MRE) Average (ARE) BP-ANN load orecast AR-ANN multi-node orecast 34.75% 36.04% 4.23% 2.83% RMSE 5.88% 5.73% The results above show that AR part o AR-ANN method has quick access to the nearest database to orecast the "whole" load, and also orecasts the "node" load by their proportion. Even in the case o lack o data, the "node" load can be orecasted by AR-ANN method with the help o the "whole" database, and thereore acilitates dierent kinds o system analysis. 5 Conclusion This paper proposes a novel algorithm or multi-node load orecasting, AR-ANN, which combines the advantages o AR time series method and BP-ANN non-linear algorithm. The single-node load orecasting in test shows that in most cases AR-ANN obtains smaller errors than BP-ANN method; when applying to bottom-up multi-node load orecasting, this algorithm can shorten the processing time and reduce errors in some cases compared to global load orecasting. A backward top-down method or multi-node load orecasting is also presented in this paper and tested by AR-ANN algorithm. The results show the possibility o load orecasting by AR-ANN when there is no access to some o the databases and/or a lack o data. Meanwhile, the inaccuracy o some orecasting points appeared in the test gives the space o urther improvement. This top-down short-term load orecasting algorithm, in conclusion, shows some advantages in orecasting precision and processing time eiciency. Furthermore, it can be utilized in the distribution network load orecasting to provide data or next step N- contingency analysis and the study emergency power supply. 665

12 AEST206 Reerences. K. Nose-Filho, A. D. P. Lotuo, and C. R. Minussi, Short-term multinodal load orecasting using a modiied general regression neural network, Power Delivery, IEEE Transactions on, 26, (20) 2. X. Han, L. Han, H. Gooi, and Z. Pan, Ultra-short-term multi-node load orecasting-a composite approach, Generation, Transmission & Distribution, IET, 6, (202) 3. Z. Wang, C. Guo, and Y. Cao, A new method or short-term load orecasting integrating uzzy-rough sets with artiicial neural network, in Power Engineering Conerence, IPEC The 7th International, -73 (2005) 4. S. Khatoon and A. Singh, Analysis and comparison o various methods available or load orecasting: An overview, CIPECH, 204 Innovative Applications o, (204) 5. H. S. Hippert, C. E. Pedreira, and R. C. Souza, Neural networks or short-term load orecasting: A review and evaluation, Power Systems, IEEE Transactions on, 6, (200) 6. J. Wang and Q. Zhu, Short-term electricity load orecast perormance comparison based on our neural network models, CCDC, (205) 7. X. Wang and L. Meng, Ultra-short-term load orecasting based on EEMD-LSSVM, (205) 8. M. Zhe and S. Qin, Short term load orecasting based on ESPRIT integrated algorithm, (205) 9. B. Lu, S. Zhao, Y. Tian, Y. Yang, B. Li, X. Chen, et al., Mid-long term electricity consumption orecasting based on improved NGM (,,k) gray model, Power System Protection and Control, (205) 0. W. Mo, B. Zhang, H. Sun, and Z. Hu, Method to select similar days or short-term load orecasting, Journal Tsinghua University, 44, (2004). H. Lin, J. Liu, Z.-. Hao, F.-. Zhu, and G.-c. Wu, Short-term load orecasting or holidays based on the similar days load modiication, Relay, 7, 47-5 (200) 2. C. Li, X. Li, R. Zhao, J. Li, and X. Liu, "A Novel Algorithm o Selecting Similar Days or Short-term Power Load Forecasting [J]," Automation o Electric Power Systems, 9, 08 (2008) 3. Z.-l. Yang, Y. Tian, G.-t. Zhang, and K.-y. Lin, "Nonlinear Theoretical Foundation and Improvement o Similar Days Method or Short Term Load Forecasting [J]," Power System Technology, 6, 04 (2006) 4. G. Ye, Y. Luo, Y. Liu, and H. Jin, "Research on Method o Power System Load Forecasting Based on ARMA Model [J]," Electronic Technology (2002) 5. Z. Zhou, J. Li, and X. Zhang, "The application o ANN in middle term load orecasting o power system", Proceedings o the CSU-EPSA, (2003) 6. J. He, G. Wei, L. Xiong, "Fuzzy improvement o linear regression analysis or load orecasting [J]," East China Electric Power,, 2-23 (2003) 7. J. Lian, H. Liu, H. Xie, X. Gong, and X. Xu, "Classiied load balancing algorithm based on prediction mechanism," Computer Engineering and Applications, 67-7,98 (205) 8. M. Zhang, H. Bao, L. Yan, J.-p. Cao, and J.-g. Du, "Research on processing o short-term historical data o daily load based on Kalman ilter [J]", Power System Technology, 0, 009, G. Chen, F. Yan, X. Gong, and Y. Wang, "State estimate based on parameter-optimized least square support vector machines," Power System Protection and Control, 39, (20) 20. J.-. Zhang, Y.-a. Wu, and J.-j. Wu, "Application o gray system theory in load orecasting [J]," Electric Power Automation Equipment, 5, 005 (2004) 2. J. W. Taylor and P. E. McSharry, "Short-term load orecasting methods: An evaluation based on European data," Power Systems, IEEE Transactions on, 22, (2007) 22. Y. Guo, "The methods o spectrum density estimation and prediction o multidimensional ARMA(p,q) models" (Southwest Jiaotong University, 2008) 23. C. Lv, "Short-term load orecasting based on BP artiicial neural network" (Huazhong University o Science and Technology, 2007) 24. Y. Shao, "Research on electric power system short-term load orecasting using a neural network",(harbin University o Science and Technology, 2005) 666

13 AEST S. Zhang, B. Zhao, F. Wang, and D. Zhang "Short-term power load orecasting based on big data," Proceedings o the CSEE, (205) 26. D. Wang and Z. Sun, "Big Data Analysis and Parallel Load Forecasting o Electric Power User Side," Proceedings o the CSEE, 3, 004 (205) 27. PJM - Markets & Operations. Available: (206) 28. K. Liu, W. Sheng, D. Zhang, D. Jia, L. Hu, and K. He, "Big data application requirements and scenario analysis in smart distribution network," Proceedings o the CSEE, 35, (205) 29. Y. Song, G. Zhou, Y. Zhu, L. Li, L. Wang, and D. Wang, "Storage Optimization and Parallel Processing o Condition Monitoring Big Data o Transmission and Transorming Equipment Based on Cloud Platorm," Proceedings o the CSEE, 2, 00 (205) 667