Short-Term Photovoltaic Power Forecasting Using an LSTM Neural Network

In this paper, two algorithms are proposed for short-term PV power forecasting using a long short-term memory (LSTM) neural network (NN). The first algorithm is designed to predict a single step ahead PV power, whereas the latter is capable of forecasting time horizons with variable starting points, which makes it very useful for rolling horizon based energy management algorithms. The effect of the input sequence length on the performance of the single-step model is investigated. The prediction accuracy of the multi-step model is examined with different lengths of rolling prediction horizons. It is shown that in the case of intraday rolling horizons, adding certain new predictors can effectively improve the machine performance. Hourly and half hourly data from different seasons are used to train and test the performance of the forecasting machine. Moreover, to demonstrate the superiority of the proposed LSTM based algorithms, the performance of other neural networks, namely the generalized recurrent neural network (GRNN) and the nonlinear autoregressive exogenous (NARX) neural network, is also explored.