Weather-Informed Short-Term Solar Power Forecasting-Oriented Assessment Using LSTM, CNN-LSTM, and XGBoost Models

Abstract

Accurate short-term assessment of photovoltaic (PV) output is important for energy scheduling, reserve planning, and grid operation. This paper presents a corrected and clearly scoped comparative study of long short-term memory (LSTM), convolutional long short-term memory (CNN-LSTM), and extreme gradient boosting (XGBoost) for hourly solar power estimation using satellite-derived weather variables. The study uses a 10-year hourly dataset for a 25 MW solar plant from 2015 to 2024. The input variables include global horizontal irradiance (GHI), ambient temperature, wind speed, wind direction, relative humidity, cloud cover, cyclic time variables, lagged plant output, rolling statistics, and interaction terms. To avoid ambiguity between operational forecasting and target-time weather-informed estimation, the experimental scope is explicitly defined as a weather-informed nowcasting/forecasting-oriented benchmark in which satellite-derived meteorological variables are available at the target timestamp. In a strict h-ahead operational forecast, these target-time weather variables would have to be replaced by exogenous weather forecasts or satellite nowcasts issued before the target time. The dataset is divided chronologically into training (2015–2021), validation (2022), and testing (2023–2024) subsets. All three models are evaluated using the same engineered feature pool, with LSTM and CNN-LSTM receiving 48-hour feature sequences and XGBoost receiving the corresponding tabular feature representation. The LSTM model achieves a test RMSE of 0.7039 MW, while CNN-LSTM reduces the RMSE to 0.6269 MW. XGBoost obtains the lowest error with RMSE of 0.0719 MW, MAE of 0.0496 MW, R² of 0.9998, and nRMSE of 1.1671%. These high XGBoost values are interpreted as weather-informed target-time estimation performance rather than proof of a closed-loop past-only solar forecast. Feature-importance analysis shows that GHI and the GHI-temperature interaction dominate the XGBoost model, while lagged features contribute only marginally. The revised study therefore provides a transparent benchmark and identifies the additional validation required before claiming strict operational forecasting performance.