Research On Short-term Load Frorecasting Of Electric Vehicle Charging Station

With the environment and energy problems have become increasingly prominent, governments have paid more attention on electric vehicle (EV). The use of large-scale electric cars can reduce greenhouse gas emissions and promote the adjustment of energy structure. A large number of electric vehicles charging during the peak load will affect the safety, dispatching and economic operation of grid. So the more precise electric load forecasting is needed to orderly control the charging of electric vehicle to participate in the peak shaving and even the frequency regulation.This paper discusses the daily load forecasting method by the support vector machine based on similar days. Through researching on charging load of pure electric bus charging station in Beijing, the characteristics of pure electric bus charging load is analyzed. The factors that affect the electric bus charging load are determined by the correlation analysis. Gray relational analysis method is used to establish small sample and then multi-input and single-output support vector machine model is established. The two-stage method determining the parameters of the model is proposed when it comes to terms with the SVM model. The first step is to directly determine the parameter e, and the next step is to find the kernel parameter and the regularization parameter by genetic algorithm in order to improve the prediction accuracy when the selected range of e is large. Finally, the result shows the prediction accuracy and stability of the method in this paper are better.Because of the volatility of load of electric bus station, the stability of single forecasting model is poor. A novel combined forecasting model is proposed in this paper. Firstly, the load characteristic of the electric bus BSS is analyzed. The traditional load forecasting model is to be improved in two aspects:1) the time characteristics of data will be taken into account; 2) based on the accuracy and stability of each adopted forecasting model, the corresponding weight in the combined forecasting model is to be adjusted dynamically. Finally, the training samples and testing samples are established. According to the comparison between the performance of single forecasting model and combined forecasting model, the feasibility and stability of our proposed combined forecasting model can be verified.