Abstract:Aiming at the complex fluctuation characteristics of line loss rate in new power systems and the insufficient prediction accuracy of traditional random forest algorithms in handling weakly correlated and low-autoregressive data, this study proposes an improved random forest algorithm based on residual-driven and gradient boosting mechanisms. By introducing a residual threshold triggering mechanism and a dynamic weight allocation function, a collaborative optimization framework is constructed for real-time detection of base model deviations and gradient boosting compensation. Combined with data reconstruction strategies including cross-cycle sampling and mutation event labeling, the algorithm effectively captures nonlinear variation patterns of line loss rate. Experimental results demonstrate significant improvements in prediction accuracy: the coefficient of determination (R2) shows notable enhancement compared to traditional models, the mean absolute error (MAE) decreases substantially, and the recognition accuracy of abnormal operational scenarios improves markedly. This method overcomes the limitations of traditional models that rely on linear correlations and time-series dependencies, providing an intelligent analytical tool for line loss anomaly localization and loss-reduction strategy formulation in power systems. Furthermore, it offers a transferable technical pathway for predictive modeling of multi-source weakly correlated energy data, demonstrating substantial engineering application value.