Machine Learning–Driven Rainfall Prediction: A Case Study at Mashhad Weather Station, Iran

In regions like Iran, where the climate is highly variable and precipitation plays a critical role in water resource management and agriculture, reliable short-term rainfall forecasting is essential. This study focuses on ۲۴ hourly precipitation prediction (one-day lag) at the Mashhad meteorological station using daily data from ۲۰۰۰ to ۲۰۲۳. Several ensemble-based machine learning algorithms were evaluated, including Random Forest, AdaBoost, CatBoost, LightGBM, and XGBoost regressors. Ten key one-day lagged meteorological variables, including precipitation, temperature, humidity, wind speed, sunshine hours, and sea level pressure, were used as input features. The results showed that the XGBoost regression model achieved the best balance between predictive accuracy and generalization, with a test MAE of ۰.۸۴ mm, RMSE of ۲.۳۱ mm, and R² of ۰.۱۷۴. While the model demonstrated strong capability in distinguishing dry and wet days, its performance in capturing high-intensity rainfall events remained limited. This study highlights the potential of ensemble machine learning methods for data-driven rainfall forecasting in semi-arid climates. It also emphasizes the need for future improvements through deep learning architectures (e.g., GRU, TCN, Transformers) and integration with upper-atmospheric features or numerical weather prediction (NWP) systems for enhanced modeling of extreme rainfall events.

کلیدواژه ها:

Rainfall Forecast ، XGBoost ، Ensemble Learning ، Mashhad