1、英文原文Short Term Load Forecasting of IranNational Power SystemUsing Artificial Neural NetworkGeneration TwoR. Barzamini,M. B. Menhaj, Sh. Kamalvand, A. TajbakhshAbstractThis paper presents a neuro-based short termload forecasting (STLF) method for Iran national power system (INPS) and its regions. Thi
2、s is an improved version of the one given in 1. The architecture of the proposed network is a three-layer feed forward neural network whose parameters are tuned by Levenberg-Marquardt BP (LMBP) augmented by an Early Stopping (ES) method tried out for increasing the speed of convergence. Instead of s
3、easonal training, an input as a month indicator is added to the input vectors. The short term load forecasting simulator developed so far presents satisfactory and better results for one hour up to a week prediction of INPS loads and region of INPS,Bakhtar Region Electric Co (BREC).I. INTRODUCTIONLo
4、ad forecasting has always been the essential part of an efficient power system planning and operation.Generally there are two groups of forecasting models, traditional models (model-based techniques) and modern technique (known as model-free techniques). Traditional load forecasting models are time
5、series and regression analysis. In recent years, computational intelligence methods are more commonly used for load forecasting 2-10.Multilayer feed forward neural networks as universal approximates are very suitable for load forecasting because they have remarkable ability to approximate nonlinear
6、functions with any desired accuracy. Selection of the input-output training data and input vector of the neural network plays a crucial role. Essentially in our case (load forecasting problem) the MLP-based networks are greatly affected by selection of inputs. Day type, Month type, historical load d
7、ata and weather information. How to choose the hourly load inputs for each weekly group plays an important role in improving networks performance (section II). The second Niroo Research Institute (NRI) STLF(NSTLFII) program is based on a three-layer feed forward neural network building block. For th
8、e training of this MLP, instead of conventional back propagation (BP) methods, the Levenberg-Marquardt BP (LMBP) and Early Stopping (ES) methods was employed in order to reach the optimum networks parameters faster, and also instead of seasonal training the month input was added to the input vectors
9、 (section III).Some examples of the NSTLFII performance are presented using INPS actual load and temperature data of the year 2000 and region of INPS, Bakhtar region electrical Co (BREC) actual load and temperature data of the year 2002 (section IV). Future works can address the fuzzy system applica
10、tion for special conditions and reshaping the load shapes by charging the peak load (section V).II. INPUT SELECTIONSelection of proper and optimal number of inputs would result a higher accuracy and convergence speed in a multiplayer feed forward neural network. Most of load forecasting methods use
11、weather information, load power of the days before forecast day, month day and day for input variables. Day type is justified by correlation analysis on the historical load information. By correlation analysis the load shapes of INPS, and its regions a week is divided into 4 groups: Saturdays (first
12、 day of the week), Sundays to Wednesdays (workdays), Thursdays and Fridays (weekends). For each group, most effective lags (load of the previous hours) on hourly load were selected by correlation analysis 11. Table 1 shows the selected lags for each weekly group. For example first row in table 1 say
13、s that for Group “Saturdays” loads of one hour, two hours, and finally 169 hours earlier are used to predict the future load. Temperature is the most effective weather information on hourly load. In this work, temperature of three cities, representing hot, moderate and cold cities of the region of i
14、nterest constitute a part of the MLPs input vector. For example, the selected cities of Bakhtar region are Hamedan (cold), Arak (moderate) and Khoramabad (hot) and Instead of using temperatures of all cities in Iran, temperatures of Ahvaz, Tehran and Tabriz were selected, by correlation analysis, as
15、 representatives of hot, moderate and cold region temperatures, respectively 11. The number of month is the last input used. TABLE 1SELECTED LOAD LAGS FOR WEEKLY GROUPS III. NSTLFIIIn the first generation of NSTLF (NSTLFI), MLPs of each weekly group were trained for each season separately, while in
16、the NSTLFII by using month number input, only one MLP per a weekly group is needed. The selected training method is LMBP using ES. LMBP method increases the speed of convergence from ten to one hundred times faster16. ES method is used for improving generalization. In this technique data is divided into three subsets: training, validation and test sets. Training set is used for c
