Accurate solar radiation forecasting is essential for optimising photovoltaic (PV) energy generation and grid integration. This paper presents an LSTM-based forecasting framework enhanced with physically meaningful features and rigorous evaluation protocols. Unlike prior approaches, we prevent temporal data leakage by applying a strictly chronological data split (last 20% reserved for testing) and rolling-window cross-validation. The dataset comprises three years (2020–2023) of high-resolution solar radiation and meteorological data, supplemented by derived features such as solar zenith angle, clear-sky index, and cloud cover. Benchmark models—including persistence, clear-sky, and linear regression baselines—are evaluated alongside the proposed LSTM using standard metrics (MAE, RMSE, MAPE, and skill score) with statistical significance testing. Results confirm that the enhanced LSTM outperforms benchmarks, reducing RMSE by 14% and achieving a skill score of 0.68. Furthermore, nighttime irradiance predictions are eliminated through day/night masking. These findings demonstrate that a carefully validated, feature-rich LSTM framework can significantly improve solar forecasting reliability while ensuring reproducibility and transparency.
Parameswari et al. (2026) studied this question.