Enhancing Short-Term Solar Generation Forecasting: The Superiority of LSTM Over SVR

Abstract

The global surge in solar power systems and AI-driven technology has spurred the development of precise solar generation forecasts. These are particularly essential in Malaysia's tropical climate with abundant sunlight. Traditional forecasting methods like ARIMA, SARIMA, and ANN are inadequate in modelling solar energy systems' inherent complexity and non-linearity. This study addresses the challenge of forecasting solar generation in a tropical region. The objective is to develop a 30-minute-ahead solar generation forecasting model using advanced techniques, specifically Long Short-Term Memory (LSTM), and compare its performance with Support Vector Regression (SVR). Utilising data from February 2022 to April 2023 collected from a rooftop solar system in Pulau Pinang, Malaysia, this approach effectively handles sudden changes in solar output, known as "ramping events," caused by cloud movement and unpredictable weather. The results reveal LSTM's superiority, with an nRMSE of 6.75%, outperforming SVR (nRMSE of 7.28%). This pattern recognition capability of LSTM holds promise for larger datasets, offering precise forecasts beneficial for weather prediction and power management. Implementing this technique in more solar PV systems can enhance power reliability and promote sustainable energy practices, showcasing LSTM's potential for optimising solar generation forecasting.