Ultra short‐term PV power forecasting based on ELM segmentation model

Abstract

With the rapid increase of the installed capacity of photovoltaic (PV) power plants, it is important to forecast PV power output in order to meet the requirements of power systems. Based on extreme learning machine (ELM), this study proposes a methodology of ultra short-term PV power forecasting by using historical power output data, which can predict power output for a prediction horizon of 30 min with 10 min time step. Compared with auto-regressive and moving average (ARMA) time series model and BPAQ3 neural network, ELM provides better prediction accuracy in each time horizons. To further enhance the model performance, the authors analyse the forecasting error distribution in different time intervals during each day, and according to the error expectation, divide the data set into two subsets, then train each subset separately to establish segmentation model. The result shows that ELM segmentation model based on error distribution provides a better performance than using ELM alone.