Power System Short-term Load Forecasting

With the development of the state power syStem the electric netWork managementmodemizes day by day and load forecasting arouses increasing more and more interests fromresearchers. Load forecasting, a major foundation for the research of power system planning andPower system economic oPeration and automatic dispach, is an important taSk of the modernpower system opefation research.This paper is concerned with a series of problems in the aPPlication of forward-back neuralnetWork to the short-term load forecasting of power system. The main achievements are asfollows:l. The load of power system is an unsteady stochastic process, whose observed value mayexist some "unhealthy data" due to the effect of various factors. These unhealthy data,participating the training of neural network intermingled with normal data. badIy affect theaccuracy of load fOrecasting. This paper finds out the mean value and variance of load Sequence ina period of tim9 based on statistics and then works out the bias ratio of every point in loadspquence using corresponding calculation formula and at the same time compares it with thresholdvalue so that "unhealthy data" can be removed and accurate and effective load forecasting can beensured.2. Through the analysis of the regularity of historical load data, the conclusion is drawn thatthe variance of performance capacity is due to the rule that "large cycIe period" overlaps "smallcycle period". As to the selection of neuraI network input node, not only is related historical loadwas introduced as the drilling sample, but also influence of temperature and weather sensitivefactors to the load variance is considered.3. ExPerimental proof method based on the Kolmogorov theorem was proposed in theselection of neural network implicit layer node. The implicit lnyer structure of neural netWorkmodel is determined by experiments.4. Two algorithms of the application of the forward-back neural network to the power systemshort-term load forecasting are proposed. That is, time sequence method and synthetic algorithm,which are both based on the neural network. The main idea of time sequence method is thataccording to the related characters of the historical load and forecasting load of the power system,the time sequence model based on neural network can be established through the combination ofthe principles of neural network and time sequence method. In the input of neural netWork, besidesthe related historical load data, the conception of daily parameters must be introduced to enhancethe model's adaptation ability to the daily cycle; the main idea of the synthetic algorithm is that onthe basis of neural network time sequence model, the highest and Iowest temperature on theforecasting day as well as weather sensitive factors must be added to the neural network inputnode to enhance the load forecasting accuracy.5. In this Paper, simulation experiments on the two models using standard BP algorithm andimproved BP aIgorithm with changeable step qand variable factor a are carried out. The resultsshow that the improved BP aIgorithm can raise the conversence velocity of the neural network;under the condition of the same algorithm, time sequence method based on neural network has ashorter period of convergence due to its simple structure, but its forecasting accuracy is relativelypoor; synthetic algorithm based on neural network has a more advanced forecaning accuracy dueto its consideration of all kinds of load effects. which is, however, at the cost of longer trainingtl me.6. The neural net''ork tbrecasting method presented in tl1e paper has been appIied in "AnalogMarket QperatiQn Management System of Xian Yang Power SuPPly Bureau ---Subsyst