Abstract: Using machine learning models and methods for short-term load forecasting can improve the overall accuracy of load forecasting, but in small sample forecasting scenarios, such as extreme weather conditions and holidays, there is no obvious advantage over manual forecasting based on expert experience. In order to fully combine the expert knowledge of forecasting operators with the reasoning generalization ability of machine learning, an interpretable load forecasting framework based on Bayesian time-varying coefficient（BTVC） and Cat Boost model is proposed. First, combined with data and expert knowledge, BTVC model is built to forecast, and load components of various influencing factors, trends and seasonality factors are obtained. Secondly, combine the above results with conventional features as the input of Cat Boost regression model for final forecasting. Then, Shapley additive explanation（SHAP） is used for attribution analysis. The quantitative relationship of the framework output can be used as a reference for load forecasting practitioners to create more effective features to further improve the forecasting results. Finally, the proposed load forecasting and result analysis framework are validated by using real load data in power grid of a region as an example.