Research On Intelligent Forecasting Method Of Short-Term Electricity Load

When the power systems are controlled on line, short-term load forecasting should be used for realizing reasonable distribution of generating and supplying electricity. Short-term load forecasting is important basis of safely assigning and economically running. The forecasting precision will directly affect the reliability, economy running and supplying power quality of power system. The primary work done in short-term load forecasting in the paper is as follows.Firstly, via vertical and horizontal pretreatment, the historical load data can furthermore show the load changing trend, which establishes the basis for short-term load forecasting model using these historical data. By using the autocorrelation function on input variables selection for short-term load forecasting model, the principle which effect is big at near place and is small at far place and autocorrelation function theory are combined to select input variables. Then a group of input variables selected can effectively explain load varying relation. This approach has clear theory foundation and very strong maneuverability, and it is a systemic and scientific method for input variables selection. By use of this method, we can get small input variables sets, the correlation of adopted input variables and forecasting hour load is most strong, and the forecasting effect is more exact.Secondly, a combined load forecasting model is presented by integrating neural networks and fuzzy logic. In the model, neural networks only settle historical load information. This not only shortens the learning time of neural networks, but also avoids improper study produced by slowness of neural networks to other information. Moreover, fuzzy logic deals with the factors which have great effect to load varying, such as air temperature and holidays, etc. According to the own characteristic of load varying, the memberships and fuzzy rules base are constructed, and the modifying of basic load heft is realized by fuzzy logic. Comparing with conventional neural networks forecasting models, the combined model adequately makes use of the neural network learning ability and absorbing of fuzzy logic to subjective experience. It fully considers the effect of factors such as air temperature and holidays to power load, which can enhance the load forecasting results veracity to a certain extent, especially weekends and holidays.Thirdly, in order to describe the non-linear relation of power load varying, a fuzzy neural networks short-term load forecasting model based on chaos mechanism is presented. For the sake of extending the scope of weight acting, the chaotic learning algorithm of weight values is introduced. The non-liner feedback item of weight value is employed in dynamic system equation of fuzzy neural networks weights space learning algorithm, which make the neural networks learning dynamics turn into chaos dynamics.So the system can find the global minimal point or its approximation quickly. The fuzzy inference and defuzzification of the model are both realized by neural networks. The selected membership function made neural network weight values have definite knowledge meaning, and it can be analyzed and understood. In the model, two different fuzzy inference algorithms are put forward to finish fuzzy inference, and it is confirmed that the fuzzy multiplication inference algorithm can get better forecasting effect. The conventional centroid method is replaced by sample adding authority calculation in fuzzy inference. Under not reducing forecasting precision, the learning time is shortened, and the running efficiency is increased. The model preferably settles the slow convergence speed and low forecasting precision limitation of general BP algorithm, and has preferable forecasting capability for non-linear load varying.Fourthly, in order to overcome the shortcomings that forecasting error will increase evidently while weather varies rapidly on the forecasting day, a new on line real time load forecasting model with feedback and recurrent structure is proposed. In the model, the fo...