Interpretable XGBoost Based Cold Load Pickup Behavior Identification Method

In order to provide decision support for load restoration optimization in distribution networks, taking the constant temperature control load of air conditioning as an example, which is the most likely to induce cold load recovery phenomena, we investigate the outage duration dependency characteristics of cold load recovery. We then use the Shapley Additive Explanation (SHAP) method to construct an interpretability mechanism for an XGBoost machine learning model, SHAP defines the average contribution of a feature value to the prediction in different coalitions. We select features that influence load parameters as inputs and cold load recovery power as output, and train a cold load recovery power prediction model. By comparing it with mainstream machine learning models, the interpretable XGBoost algorithm has best behavior, achieving a mean absolute error of 0.380%, a mean squared error of 0.030%, and a mean absolute percentage error of 0.445%. Finally, we uncover the feature correlations that affect cold load pickup power prediction, elucidating the decision-making process of the machine learning model at a global level.