Solar Irradiance Forecasting With Deep Learning and Ensemble Models: LSTM, Random Forest and Extra Trees With Multivariate Meteorological Data

Abstract

The penetration of solar energy within power systems is imperative to ensure the enduring sustainability of power systems, especially in cases of isolated systems such as microgrids. However, the variability in solar energy generation, which can impact grid stability and supply-demand balance, highlights the necessity for reliable forecasting methods and advanced technological tools in addressing power system challenges. This paper presents an in-depth exploration of solar irradiance forecasting, utilising a combination of advanced techniques, including deep learning and ensemble models. In particular, the utilisation of long short-term memory (LSTM), random forest and extra trees in conjunction with multivariate meteorological data is investigated to enhance the accuracy and reliability of solar irradiance predictions. A comprehensive technique for modelling multivariate time series is employed to predict solar irradiance by incorporating various meteorological factors, such as temperature, relative humidity, and barometric pressure, among other relevant variables. A case study in Cuenca, Ecuador, was chosen based on real data obtained from a meteorological station, ensuring the accuracy and reliability of the data. The proposed method significantly enhances performance when compared to a baseline model, with the LSTM model notably excelling in predicting maximum and minimum solar irradiance, offering valuable insights for extended forecasting applications.