Dual-Channel Ensemble Learning for Dynamic Risk Assessment in Electricity Fee Recovery under Small-Sample Scenarios

This study proposes a dual-channel ensemble learning-based dynamic early warning model to address the challenges of sample imbalance and tail risk quantification in electricity fee recovery risk assessment under small-sample scenarios. The methodology combines heterogeneous ensemble classification models (Gradient Boosting Decision Tree and Light Gradient Boosting Machine) for predicting default probabilities with quantile random forest regression for estimating potential default amounts, enabling dual-dimensional risk evaluation. A hybrid sampling technique (Synthetic Minority Over-sampling Technique and Edited Nearest Neighbours, SMOTE-ENN)) optimises data distribution, while a dynamic fusion mechanism adaptively adjusts classification and regression weights. Experiments on real-world electricity user data demonstrate that the model outperforms traditional methods in key metrics, achieving an AUC value of 0.88 and a large default coverage rate of 95%. This research provides theoretical and practical tools for optimising resource allocation in fee recovery management.