Enhancement of LSTM models based on data pre-processing and optimization of Bayesian hyperparameters for day-ahead photovoltaic generation prediction

In this paper, a forecasting method for day-ahead photovoltaic (PV) generation using long short-term memory (LSTM) neural network is presented. To improve the quality of the predictions, the most relevant hyperparameters of the proposed model are adjusted using a Bayesian optimization algorithm. The study uses a database obtained from a grid-connected solar plant containing historical data of the PV power generated, solar irradiance, ambient temperature, PV panel temperature and wind speed. First, a preprocessing algorithm is applied to enhance the quality of the data and the accuracy of the forecasts. Subsequently, the proposed model is used to predict the power generated at the facility for the following day. The obtained results are compared with other types of models used in the related literature: a gated recurrent unit (GRU) neural network and a multilayer perceptron (MLP) neural network. The tests are performed on days with different meteorological behavior, and it is observed that the proposed model outperforms the comparison models in all cases analyzed in terms of accuracy and quality of predictions.