Ice condition forecasting and driving factor analysis based on XGBoost-SHAP model

To address practical constraints, including short ice-condition observation series and complex physical processes during winter water conveyance in the Middle Route of the South-to-North Water Diversion Project, this study integrates the XGBoost and SHAP frameworks. Leveraging prototype observation data collected from the main trunk channel between 2011 and 2022, we developed an ice-condition classification forecasting system to predict floating ice and ice cover with lead times of 7, 14, and 21 days. SHAP is used to quantify the contribution and direction of influence of each driving factor. The results demonstrate that the proposed model achieves higher forecasting accuracy and better key-node prediction performance than Random Forest, Support Vector Machine, and BP Neural Network across all lead-times. Ice-condition evolution is jointly governed by thermal and dynamic processes, with cumulative negative air temperature serving as the dominant thermal factor and water-conveyance discharge as the core dynamic factor. Short-term forecasts are more sensitive to air-temperature fluctuations, whereas medium- and long-term forecasts are more reliant on cumulative low-temperature effects. This proposed model achieves high accuracy and strong interpretability in ice-condition forecasting, even with a relatively short dataset. These capabilities provide critical technical support for ensuring the safety of winter water conveyance and enabling intelligent regulation of the Middle Route of the South-to-North Water Diversion Project.