Short-term Power Load Forecasting Based On Phase Space Reconstruction And Support Vector Machine

In recent years, with the demand for electricity increasingly strong in almost all industries.Therefore, highly accurate predictions of the load on the grid can support more precise workof power plant fuel supply and can make the power plant system secure and stable.This thesis focuses on the power load of a city in Guilin. First, analyze the load timeseries using the chaos theory, and construct the training data set and the prediction data setemploying phase space reconstruction method. Embedding dimension and delay time areobtained by improved C-C method. This method can make the solving process of delay timeand embedding dimension simple as well as improve the precision and accuracy of the loadforecasting. The maximum Lyapunov index obtained by the small amount of data algorithmproves that the load time series has the characteristics of chaos.Then conduct model training and power load forecasting are using the support vectorregression algorithm. And searching for the optimal value influencing on the key parametersof support vector regression employing the particle swarm optimization algorithm. Comparethe optimal parameters obtained by the particle swarm optimization algorithm with thatobtained by the cross validation-step grid search method. The comparative results show thatthe two algorithms can both obtain the optimal parameters meanwhile the PSO algorithm cansave a lot of computing time.By comparing the result of power load forecasting using the method of chaotic phasespace reconstruction and the support vector regression, the result of power load forecastingused the method of BP neural network and the result of power load forecasting used themethods of and time series analysis, the following conclusions are obtained: Threeforecasting methods can all achieve better forecasting results on weekdays. The method ofchaotic phase space reconstruction and the support vector regression for power loadforecasting has a distinguish advantage over the other two methods on Sundays and majorholidays when load change is large. Finally, considering the defect of the numerous complicated operations of analysis andforecasting process for the data of the power load in Matlab platform, a kind of Matlab GUIforecasting system is designed in this thesis. This power load forecasting system can quicklyimport, display and manipulate data, and greatly improves accuracy of the power loadforecasting.