Hybrid LSTM-XGBoost Model for Lightning Strike Frequency Forecasting Using Spatio-Temporal Features

Frama Andika (Sriwijaya State Polytechnic, Palembang, Indonesia)
Muhammad Nawawi (Sriwijaya State Polytechnic, Palembang, Indonesia)
Masayu Anisah (Sriwijaya State Polytechnic, Palembang, Indonesia)

Indonesian Journal of Artificial Intelligence and Data Mining
Vol. 9 No. 2 (2026): July 2026

Lightning strikes pose significant risks to human safety, infrastructure, and aviation systems, making accurate frequency forecasting essential for early warning and disaster mitigation. However, existing approaches often fail to capture both the temporal dependencies and nonlinear spatial patterns inherent in real-time lightning data. This study proposes a hybrid model integrating Long Short-Term Memory (LSTM) and Extreme Gradient Boosting (XGBoost) for lightning strike frequency forecasting using spatio-temporal features. The LSTM component was employed to extract sequential temporal dependencies from time-series lightning data, while XGBoost was utilized to model complex nonlinear relationships among spatial features, including latitude, longitude, and region. The dataset comprised 742,543 real-time lightning strike records with seven features: timestamp, coordinates, region, multi-point distance score (MDS), multi-channel gradient (MCG), and status. Feature engineering was performed to construct temporal aggregations and spatial encodings as model inputs. The proposed hybrid model was evaluated using Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and the coefficient of determination (R²). The experimental results demonstrate that the hybrid LSTM-XGBoost model outperformed the standalone LSTM and XGBoost baselines across all evaluation metrics, confirming the effectiveness of integrating deep learning and ensemble methods for spatio-temporal lightning forecasting. The novelty of this study lies in the application of a sequential hybrid LSTM-XGBoost pipeline specifically designed for real-time spatio-temporal lightning strike frequency forecasting, an area that remains underexplored in the existing literature. The findings suggest that the proposed model holds strong practical potential for integration into operational lightning early warning systems, thereby contributing to enhanced disaster mitigation and risk assessment across aviation, infrastructure, and public safety sectors.