Prediction of Solar Panel Voltage Using LSTM on a PV System with a 5 W Lamp Load

Abstract

This study aims to develop and evaluate a predictive model for the DC voltage output of photovoltaic (PV) panels used to power a 5-watt lamp, leveraging Long Short-Term Memory (LSTM) networks. The dataset, collected from a real PV system at UNESA Ketintang campus over a year, was preprocessed through resampling at one-minute intervals and normalized using Min-Max Scaling to stabilize input data and enhance model training. The LSTM architecture consisted of three stacked layers with dropout regularization to prevent overfitting, trained using the Adam optimizer and mean squared error loss function. Model performance was assessed using MAE, MSE, and MAPE. Results demonstrated that the LSTM model accurately predicted voltage fluctuations with an MSE of 0.02, MAE of 0.03, and MAPE of 0.17%, effectively capturing both short-term variations and long-term trends. A comparative analysis with the Gated Recurrent Unit (GRU) model showed that while GRU offered lower computational complexity, LSTM achieved higher prediction accuracy. These findings highlight the suitability of LSTM for time-series forecasting in PV systems, providing a valuable tool for real-time monitoring and energy management. The study contributes to improving the reliability and efficiency of renewable energy systems by enabling better prediction and control of PV output under variable environmental conditions