LSTM-Based Daily Power Forecasting for a 1 MWp PV System in Tropical Indonesia: Toward Operational Optimization

Abstract

Variations in solar irradiance and module temperature significantly affect the performance and operational efficiency of large-scale photovoltaic (PV) power systems, especially in tropical regions. This study investigates the application of a Long Short-Term Memory (LSTM) network for accurate real-time power prediction in a 1 MWp PV power plant at Institut Teknologi Sumatera (ITERA), Indonesia. Unlike traditional approaches and conventional artificial neural networks (ANN), LSTM networks can effectively capture long-term temporal dependencies and highly nonlinear patterns in PV output data. A five-minute resolution dataset, including actual power output, solar irradiance, and module temperature, was collected throughout March 2025 for model training, with validation performed using independent data from April. The developed LSTM model achieved a mean absolute error (MAE) of 42.8 kW (approximately 4–6% of maximum plant capacity) and a coefficient of determination (R²) of 0.84 during active hours (05:00–19:00) on the validation dataset. These findings indicate that the model performs well not only on the training data, but also maintains strong generalization to unfamiliar data. The proposed approach enables reliable real-time power prediction, supporting applications such as energy forecasting, inverter control, dispatch planning, and anomaly detection in PV systems. This work provides a practical and scalable solution for improving the adaptability and integration of solar power plants in dynamic tropical environments, contributing to the advancement of AI-driven sustainable energy systems.