Two-stage Photovoltaic Power Forecasting based on Extreme Learning Machine and Improved Pointwise Mutual Information

Abstract

Photovoltaic power forecasting plays an important role in the power system planning and contributes to the development of renewable energy. This paper proposed a two-stage forecasting method based on Extreme Learning Machine (ELM) and Improved Pointwise Mutual Information (IPMI), which is responsible for short-term forecasting of the small-scale PV station. The method like a hybrid method requires past measured Numerical Weather Prediction (NWP) data and time series of PV power output as input of the system. During the first stage, PMI algorithm is applied to solve the coupling problem and to determine the weather features that contribute to the PV power output. Locally weighted regression (LOESS) smoothing is used as a nonparametric technique to fit a smooth curve to calculate regression error. Then, historical data will be classified into several groups by affinity propagation (AP) clustering method, combing with principal component analysis (PCA). For the second stage, training models and prediction models based on ELM network will be set up for each group obtained by AP clustering respectively. Numerical results for a small-scale PV plant in Beijing present improvements in forecasting accuracy and computation efficiency when compared to other forecasting methods.