MULTIVARIATE INTRA-MINUTE SOLAR FORECASTING FOR OPTIMIZED GRID INTEGRATION: A COMPARATIVE STUDY OF XGBOOST AND LSTM IN BAUCHI, NIGERIA

Abstract

Accurate intra-minute solar forecasting is critical for real-time grid stabilization and the optimization of Energy Management Systems (EMS), particularly in volatile tropical regions. This study evaluates the performance of Extreme Gradient Boosting (XGBoost) and Long Short-Term Memory (LSTM) architectures for 1-minute ahead Global Horizontal Irradiance (GHI) forecasting using high-resolution ground-measured data from the World Bank WAPP campaign in Bauchi, Nigeria (2021–2023). To ensure meteorological relevance, a multivariate approach was adopted, integrating physical context from DNI, DHI, and module temperature sensors alongside engineered autoregressive lags and rolling smoothing kernels. Performance was benchmarked against a persistence baseline across 58,024 operational daylight observations. The experimental results reveal a definitive "Deep Learning Deficit" at high frequencies; while the persistence baseline proved difficult to beat at 60-second intervals (MAE 12.26 ), the proposed XGBoost model significantly outperformed the LSTM, achieving an MAE of 13.75 and of 0.9909, compared to an MAE of 21.97 for the LSTM. Feature importance analysis confirmed that at intra-minute timescales, immediate physical momentum ( ) and auxiliary sensor correlations dominate long-term temporal sequences, which likely introduced smoothing-noise in the LSTM. Furthermore, XGBoost achieved an inference speed of 0.001 ms, approximately 62 times faster than the LSTM. These findings prove that lightweight tree-based ensemble models are the optimal choice for deployment in microgrid control and smart inverters, providing the statistical precision and low latency required for real-time ramp-rate mitigation and industrial grid security in Sub-Saharan Africa.