Description
Introduction: Accurate weather forecasting is essential for agriculture, disaster risk reduction, aviation, transportation, and climate monitoring. However, predicting meteorological variables such as temperature, humidity, and wind speed remains challenging due to the nonlinear and spatio-temporal nature of atmospheric processes. Conventional Numerical Weather Prediction (NWP) models, including ECMWF, often struggle to capture localized weather variability. Recent advances in artificial intelligence and deep learning offer alternative approaches for learning complex meteorological patterns. This study evaluates the performance of NWP/ECMWF, Recurrent Neural Networks (RNN), Convolutional Neural Networks (CNN), and Graph Neural Networks (GNN) for weather forecasting across diverse climatic regions of Tanzania.
Methodology: Daily observations of temperature, relative humidity, wind speed, and wind direction were collected from five meteorological stations: Zanzibar, Dar es Salaam, Kibaha, Dodoma, and Iringa. Data preprocessing included missing-value treatment, temporal feature engineering, cyclical encoding, lag-feature generation, and standardization. A spatio-temporal graph was constructed by representing stations as graph nodes connected through geographical proximity and statistical similarity. NWP/ECMWF, RNN, CNN, and GNN models were trained and evaluated using identical datasets and forecasting horizons. Performance was assessed using Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Coefficient of Determination (R²).
Results: The proposed GNN demonstrated the strongest overall performance for humidity and wind-speed forecasting across the study regions. For humidity prediction, the GNN achieved an average RMSE of 1.92%, MAE of 1.41%, and R² of 0.95, while wind-speed prediction achieved RMSE of 0.24 kt, MAE of 0.20 kt, and R² of 0.98. For temperature forecasting, the GNN achieved competitive performance (RMSE = 0.66°C, MAE = 0.47°C, R² = 0.42), although NWP/ECMWF performed slightly better at some stations.
Conclusion: Integrating spatial and temporal dependencies through graph-based learning substantially improves meteorological forecasting accuracy. The proposed GNN framework provides a robust and scalable approach for localized weather prediction and operational meteorological applications in Tanzania and similar climatic regions.
| Title | Comparative Evaluation of Spatio-Temporal Graph Neural Networks and Deep Learning Models for meteorological forecasting |
|---|