Solar Power Forecasting of 8 MWp Solar Farm Malacca using LSTM-based Model with Weather Forecast Data: A Case Study of Malaysia

Abstract

Malaysia's tropical climate offers significant potential for photovoltaic (PV) installations due to abundant solar irradiance. However, the variability in solar energy generation due to weather fluctuations presents challenges for grid integration and energy reliability. The study evaluates the performance Univariate LSTM, Multivariate LSTM (with weather sensor data), Multivariate LSTM (with weather sensor and meteorological data), and Bidirectional LSTM (Bi-LSTM) models using input data from 8 MWp solar farm in Ayer Keroh, Malacca, along with weather sensor and meteorological data. Results show that the Univariate LSTM model consistently outperformed others across all forecasting horizons, achieving the lowest error metrics (MAE: 0.0275, MSE: 0.0037, RMSE: 0.0611) and the highest R² value (0.94), making it the most reliable choice for both short- and long-term forecasts. However, weather uncertainty remains a significant challenge, directly impacting solar power production. Thus, Multivariate LSTM models is more practical. From the result, Multivariate LSTM with weather sensor and MET input data demonstrated some advantages as its gives lowest error metrics (MAE: 0.0375, MSE: 0.0044, RMSE: 0.0664) and the highest R² value (0.92), compared to Multivariate LSTM with weather sensor input and Bi-LSTM model. However, at intermediate horizons the accuracy is decreased which might be caused by the increased complexity of meteorological inputs. In contrast, the Bi-LSTM model performed the weakest, with the highest error metrics, suggesting potential overfitting or limited generalization capabilities. This research provides valuable insights into the trade-offs between model simplicity and performance including across different forecasting horizons in renewable energy forecasting.