Direct multi-step forecasting of uranium concentration in in-situ recovery wellfields based on domain-aware feature selection and physics-guided ensemble learning

Abstract

Accurate long-horizon forecasting of uranium concentration is essential for production scheduling, stage identification, and risk-aware operation in in-situ recovery (ISR) uranium wellfields. However, field-scale ISR data are often characterized by limited sample size, strong fluctuations, and unstable generalization of purely data-driven models. To address these challenges, this study proposes a direct multi-step forecasting framework integrating domain-aware feature selection and a physics-guided ensemble. The method uses a 30-day historical observation window and future injection plans to forecast uranium concentration over the next 60 days. A 12-dimensional physics-guided feature set is constructed using grouped feature screening based on Shapley additive explanations (SHAP), incorporating remaining uranium state, dynamic operational control, hydrogeochemical conditions, and geological properties. Based on this feature space, support vector regression (SVR), random forest (RF), and Light Gradient Boosting Machine (LightGBM) are combined using a static heterogeneous ensemble. Validation using real production data from more than 30 production units in an ISR mine shows that the proposed model achieves a global coefficient of determination (

$$\:{R}^{2}$$

) of 0.8902 and a root mean squared error (RMSE) of 1.7042, outperforming individual machine-learning models and a long short-term memory (LSTM) baseline. Ablation experiments indicate that physics-guided features and static equal-weight ensemble learning improve long-horizon forecasting stability. Prediction intervals are further constructed from tree-level dispersion in the RF branch, providing risk-aware boundaries for engineering scheduling.