Forecasting energy output of a solar power plant in curtailment condition based on LSTM using P/GHI coefficient and validation in training process, a case study in Vietnam

Abstract

This study presents how to improve the short-term forecast of photovoltaic plant's output power by applying the Long Short-Term Memory, LSTM, neural networks for industrial-scale solar power plants in Vietnam under possible curtailment operation. Since the actual output power does not correspond to the available power, new techniques (Global Horizontal Irradiance - GHI interval division, P/GHI factor addition (P - Power)) have been designed and applied for processing errors and missing data. The prediction model (LSTM network, structure of hidden layers, number of nodes) has been developed by the authors in a previous work. In this new version of the model, the training technique is improved by using validation and experiments to determine the appropriate relevant parameters. The forecast results show that the proposed new method is more efficient than the old method, as the MAPE (Mean Absolute Percentage Error) forecast error is reduced by 6.059% and the RMSE (Root Mean Square Error) is reduced by 6.710%.