Integrating Reliable Machine Learning Forecasting with Uncertainty Modeling for Agricultural Yield and Electric Mobility Transition Applications

Reliable national-scale crop yield prediction is critical for agricultural planning under increasing climatic variability. Existing studies commonly rely on random data partitioning and point-based predictions, which may overestimate predictive performance and provide limited insight into model reliability under future climatic conditions. This study proposes an uncertainty-aware machine learning framework combined with time-based validation for reliable national-scale crop yield forecasting. National crop yield data were integrated with climatic variables (annual rainfall and mean temperature), agronomic indicators, and temporal information at the country–year level for the period 1990–2013. Random Forest and XG-Boost models were trained and evaluated using both conventional random train–test splitting and a temporally consistent validation strategy designed to reflect realistic forecasting scenarios. Prediction uncertainty was quantified using ensemble-derived prediction intervals, and a rainfall-based climate stress analysis was conducted to evaluate model behaviors under contrasting climatic conditions. The results showed that random data partitioning substantially overestimated predictive performance (R2= 0.96), whereas time-based validation provided a more realistic assessment of model generalization. Under time-based evaluation, XG-Boost achieved superior predictive performance (R2= 0.89, RMSE = 0.38, MAE = 0.27) compared to Random Forest. Prediction uncertainty increased during low- and high-rainfall conditions, indicating reduced predictability under climatic stress.

Hali tarjima qilinmagan