Short-term Prediction of Photovoltaic Output Based on Improved Extreme Learning Machines

In allusion to the defect of traditional extreme learning machine easily falling into local optimal solutions and the characteristic of environment variation leading to the fluctuation of photovoltaic (abbr. PV) output, based on complete ensemble empirical mode decomposition with adaptive noise (abbr. CEEMDAN) algorithm and combining with extreme learning machine neural network optimized by chimp optimization algorithm a short-term PV output prediction model was constructed. Firstly, by use of CEEMDAN algorithm the key environment factor series impacting PV output was decomposed to obtain the local feature of data signals in different time-scales to reduce the non-stationary of environment factor series. Secondly, taking each decomposed subseries and PV historical data series as the input of extreme learning machine prediction model optimized by the chimp algorithm the prediction was performed. Finally, the data set of DKASC Solar Centre PV station was chosen to conduct the contrast and verifying for different prediction models. Results of simulation example show that the prediction effect of each index of the constructed improved PV output prediction combined model is better and suitable to the prediction of PV generation in different environments.