Ultra-short-term PV power forecasting based on LSTM with PeepHoles connections

Abstract

With the gradual depletion of global traditional energy sources, solar energy has attracted more and more attentions as a new alternative energy source. According to the research statistics, the total installed capacity of solar stations in the world is expected to reach 124GW in 2019. However, at the same time, due to the volatility and intermittence of photovoltaic (PV) power generation, it is inevitable that the output power of the PV arrays will have an impact on the power system operation. Therefore, PV forecasting technology has a great significance for the application of new energy sources. In this paper, an advanced Long Short-Term Memory (LSTM) model with PeepHoles connections is proposed, and the weather information and current PV power are used as the input of the prediction model. On the basis of the standard LSTM neuron, the state of the previous neuron is added in the forget gate and output gate to prevent the abnormal close of the output gate and the forget gate during training process and actual testing process. At the same time, the new model's ability to extract trend information from the input stream is improved, and the ability to track the change of PV power is enhanced, hence the prediction accuracy of the model is improved. The experiment has been carried out based on the data from 150MW solar station in Yueqing, Zhejiang. It shows that the advanced LSTM model has better weather adaptability and prediction accuracy. The prediction accuracy has been improved by 5.734% compared with standard LSTM in MAPE and the RMSE has been reduced by 0.254MW.