A Spatial Forecasting Method for Photovoltaic Power Generation Combined of Improved Similar Historical Days and Dynamic Weights Allocation

摘要

This paper proposes a combined method for short-term forecasting of photovoltaic (PV) power generation, especially for PV power forecasting in micro-grids. To decrease the computational complexity and uncertainty of the traditional forecasting method which has to consider many complex external environment factors, the paper focuses only on the direct relationships among the micro-sources. The paper combines three forecasting models so as to improve the precision. In the first model, historical data is quantized for clustering with Self-organizing Map and Least Squares Fitting method. In the second model, Back Propagation Neural Network with the optimization of Mind Evolutionary Computation is used for forecasting. The third model employs Extreme Learning Machine based on the Single-hidden Layer Forward Feed Neural Network. Three models work separately and combine finally with the weights. Variance-covariance method is used for allocating weights between the second and third models. Then the first model has the equal weight of the summary of the other two. The simulation and experiment prove that this combined method is feasible, with better flexibility and higher accuracy.