Development of a Real-time Solar Panel Power Prediction Model Using the Long Short-term Memory

Abstract

Indonesia’s remote islands face significant challenges in electricity access due to the high cost and logistical difficulties of extending the national grid, leading many communities to rely on expensive and polluting diesel generators. Solar-based microgrids offer a sustainable alternative, yet the intermittent nature of solar energy, driven by fluctuating weather conditions, poses major obstacles to a reliable power supply and efficient system sizing. This study addresses these issues by developing a real-time solar panel power prediction model using Long Short-Term Memory (LSTM) networks. A 50 Wp solar panel system equipped with an INA260 current sensor, a voltage sensor, a DHT-22 temperature sensor, and an ESP32 microcontroller was constructed to collect real-time voltage, current, and temperature data at 10-second intervals. The collected data underwent preprocessing, feature engineering, and transformation into supervised learning sequences for training. Three temporal resolutions of the LSTM model were systematically evaluated: 3-minute, 2-minute, and 1-minute, all with 30 output timesteps. Performance was assessed using rolling-window predictions on the held-out test set with metrics including RMSE, MAPE, and R². Results demonstrated that finer temporal resolution significantly improves forecasting accuracy. The 1-minute variation achieved the best performance with the lowest RMSE and highest R², effectively capturing both diurnal patterns and short-term fluctuations in solar power output. The developed LSTM model enables accurate short-term predictions (30–90 minutes ahead), supporting proactive energy management, including optimized battery charging, load scheduling, and reduced grid dependency. Future work will incorporate additional meteorological variables and seasonal data to improve model robustness further. Three temporal variations of the LSTM model were systematically evaluated: 3-minute, 2-minute, and 1-minute resolutions, all with 30 output timesteps. Performance was assessed using rolling-window predictions on the held-out test set with metrics including RMSE, MAPE, and R². Results demonstrated that finer temporal resolution significantly improves forecasting accuracy. The 1-minute variation achieved the best performance with the lowest RMSE and highest R², effectively capturing both diurnal patterns and short-term fluctuations in solar power output. The developed LSTM model enables accurate short-term predictions (30–90 minutes ahead), supporting proactive energy management such as optimized battery charging, load scheduling, and reduced grid dependency. Future work will incorporate additional meteorological variables and seasonal data to further improve model robustness.