Evaluation of Multi-Scenario Distribution Network Planning Based on Bayesian Optimization of Heterogeneous Stacked Ensemble Learning

[Objective] With the large-scale integration of renewable energy sources， diverse flexible loads， and digital-intelligent equipment， traditional distribution network planning evaluation methods struggle to flexibly adapt to the multifaceted and differentiated characteristics of modern distribution networks. To achieve data-driven， scenario-based， and differentiated planning evaluations， this study constructs an indicator system tailored to new-type distribution networks and proposes a multi-scenario planning evaluation method based on Bayesian optimization and heterogeneous stacked ensemble learning. [Methods] First， a scenario-aware feature space is constructed via one-hot encoding to enable unified ensemble learning across multiple scenarios. Second， a two-layer stacked ensemble strategy is proposed： the first layer establishes a diverse base model pool comprising Random Forest， XGBoost， LightGBM， and CatBoost， balancing the variance reduction of Bagging with the bias reduction of Boosting； the second layer employs an XGBoost meta-learner for decision fusion to integrate complementary algorithmic strengths. Furthermore， Bayesian optimization based on Gaussian Process Regression is utilized to iteratively tune hyperparameters across all layers， enhancing overall model performance and generalization capability. [Results] Case study results demonstrate that the proposed integrated model achieves an accuracy of 91.33%， representing an average improvement of 12.35% over traditional single models. Across various scenarios， accuracy remains stable within a narrow range of 90.61%-91.99%， indicating no overfitting or maladaptation in specific scenarios. [Conclusions] The proposed method effectively integrates distribution network planning scenario characteristics and leverages the complementary advantages of multiple algorithms. It successfully addresses multi-scenario， high-dimensional， and nonlinear decision-making problems in distribution network planning， demonstrating robust overall performance and superior generalization capabilities.