Day-Ahead Photovoltaic Power Forecasting Using Deep Learning with an Autoencoder-Based Correction Strategy

Accurate forecasting is crucial for successfully integrating photovoltaic (PV) power plants into electrical grids and microgrids. Accordingly, this work presents a hybrid methodology for day-ahead PV power forecasting (PPF). It begins by examining three deep learning (DL) techniques, long short-term memory (LSTM), gated recurrent unit (GRU), and multilayer perceptron (MLP), as potential forecasting models. To find a robust forecasting model, feature selection is employed to select the most relevant input features, and additionally, hyperparameter optimization is performed using the Chu–Beasley genetic algorithm to automatically set the hyperparameters for each technique. An initial day-ahead PPF is computed recursively after selecting the optimal forecasting model. Subsequently, this initial forecast is refined using a long-short-term memory autoencoder (LSTM-AE) that corrects the initial PPF. To further enhance the interpretability of the final forecast, the k-means algorithm, incorporating a soft-dynamic time warping (DTW) metric, is utilized. The efficacy of the methodology is validated using real data from a solar farm at the State University of Campinas (UNICAMP) in Brazil. Empirical results demonstrate that the proposed methodology improves the forecast accuracy by more than 3.5% when LSTM-AE is applied for correction compared to state-of-the-art models.