Air conditioning load identification is an important basis for the participation of massive air conditioning loads on the user side in demand response regulation. Due to the variety of air-conditioning load types and the complexity and variability of operation modes, it is difficult for the existing means to effectively screen out the features that are relatively typical for air-conditioning load identification, which in turn affects the accuracy of identification. On this basis, this paper presents an interpretable model for air conditioning load identification, leveraging mutual information and CatBoost. It explores a method for selecting electrical characteristic indices of air conditioning loads, investigates the correlation between these indices and equipment labels, and uses Shapley additive explanations (SHAP) to assess the impact of electrical characteristics on the accuracy of the air conditioning load identification model. Experimental results demonstrate that five key features significantly influence the load identification outcomes. This approach provides valuable insights for advancing the lightweight design, generalization, and interpretability of air conditioning load identification models.