The Research On Short-term Load Forecasting Of Smart Grid Based On Multiple Neural Network

With the development of socio-economic and technology, protecting the power systemand providing the high-quality supply are the key instruments for the economy growth and thedaily life. Many operating planning and decisions in the smart grid, for example, dispatchscheduling of generating capacity, reliability analysis and maintenance plan for generators, arebased on the load forecasting. Therefore, load forecasting is to guarantee the quality ofelectricity supply. Load Forecasting can be divided into four categories according to differenttime horizon, very short term, short term, medium term, long term. In particular, day-to-dayoperation of the smart grid must be based on short-term load forecasting.In general, the load in future is predicted according to historical load data. Statisticalmethod assumes that the load follows a particular distribution. The advantages are simple andeasy to understand. However, the statistical model cannot fit most situations properly,especially smart grid owing to its high complexity. As a result, many studies apply artificialintelligence techniques to load forecasting to increase the accuracy and flexibility. Artificialintelligence based method achieve satisfying results on nonlinear mapping problems.The load of microgrid is low but with large fluctuation. Therefore, its modeling andprediction are more complex and difficult. Thus, a novel multiple classifier system has beenproposed in this thesis. Radial Basis Function Neural Network is used as basic classifiers.Localized Generalization Error Model is applied to select the architecture and parameter forRBFNNs. To adapt the fluctuation of load and solve the problem of massive computing, weapply parallel computation on matrixes and propose a new K-means strategy and dynamicweighting method for on-line learning, which adjusts the weights of base classifiers accordingto the last prediction results and returns feedback to adjust the multiple classifier systemstructure and fusion parameters. The real historical load data has been used in theexperimental study. Experimental results show that more accurate forecasting can be achievedby the proposed method in comparison to some existing artificial intelligence methods.