Impact of Solar Peak Energy and Weather Classification on BiLSTM-Based Solar Irradiance Forecasting in Equatorial Regions

Abstract

Accurate solar irradiance forecasting is critical for managing solar energy systems in equatorial regions, where high solar potential is coupled with significant variability. This study investigates the influence of solar peak energy and weather classification features on a Bidirectional Long Short-Term Memory (BiLSTM) model for multi-step Global Horizontal Irradiance (GHI) prediction. The research method involved four phases: data preprocessing (including cyclical time encoding, lag features, and solar peak extraction), weather classification (pseudo-labelling refined by decision trees), BiLSTM-based forecasting with Optuna hyperparameter tuning, and model evaluation using standard error metrics. Three configurations were compared: (A) solar peak + weather classification, (B) weather classification only, and (C) core meteorological and temporal features without additional inputs. The workflow incorporated cyclical time encoding, pseudo-labelling with decision tree refinement, lag feature construction, and Optuna-based hyperparameter tuning. Model performance was assessed using MAE, RMSE, MAPE, R², and MASE. Scenario A achieved the lowest MAPE (28.74%), whereas Scenario C yielded the smallest MAE (103.59 W/m²) and MASE (0.786). Scenario B performed worst with a MAPE of 29.85% and MAE of 105.64 W/m², highlighting the limited standalone value of weather classification. Across all scenarios, RMSE values remained within 148–150 W/m² and R² around 0.68, reflecting minimal differences in variance explanation. These findings suggest that simpler models can perform as well as or even outperform more complex configurations, offering efficiency benefits for operational forecasting. The practical implication of these results is that reliable irradiance forecasts can be achieved with simpler BiLSTM configurations, reducing computational cost and supporting real-time energy management, PV system sizing, and grid stability in equatorial regions. Future work should incorporate satellite imagery and real-time cloud tracking to further enhance prediction accuracy.