Bayesian-optimized LSTM networks for accurate day-ahead photovoltaic power prediction

Abstract

Accurate day-ahead photovoltaic (PV) power forecasting is essential for effective energy management and grid balancing. This study proposes a Bayesian-optimized long short-term memory (LSTM) network for day-ahead PV power prediction. The model was evaluated using a PV-meteorological time-series dataset collected from a 500-kWp grid-connected PV installation in Mosul, Iraq, between January 1, 2021 and December 31, 2023 at an hourly sampling interval. After data-quality screening, 25,842 valid synchronized observations were retained from 26,280 timestamps. Each forecasting sample used the previous 24 hours of PV power, solar irradiance, ambient temperature, relative humidity, wind speed, and temporal indicators to predict the subsequent 24-hour PV power profile. The data were divided chronologically into training, validation, and independent test subsets. Preprocessing included duplicate removal, missing-value treatment, IQR-based outlier handling, temporal alignment, and min-max normalization fitted only on the training subset. Bayesian optimization tuned the LSTM architecture and training hyperparameters, while a hybrid MSE-MAE loss balanced large deviations and overall error. The proposed model achieved an MAE of 15.2 kW, an RMSE of 20.5 kW, a MAPE of 8.3%, and an R² of 0.93, outperforming linear regression and autoregressive integrated moving average (ARIMA) under the evaluated conditions.