Most existing electricity theft detection methods rely on a single model to analyze user electricity usage patterns.These approaches suffer from limited generalization capabilities and are prone to overfitting.Moreover

the highly covert nature of electricity theft makes it difficult to capture effective user behavior features.Therefore

the paper proposes an electricity theft detection method based on cascaded distillation.User electricity consumption data is first collec-ted

corrected

and subjected to dimensionality reduction.A shared backbone network enhanced by cascaded distillation is then applied to extract temporal features from the processed user behavior data.A dynamic anchor box distillation module is introduced to augment the data based on its temporal characteristics

while a dynamic anchor box refinement module is used to mine deep features from user behavior data.This process enables the model to capture diverse electricity usage patterns

improve generalization

and mitigate overfitting.Based on the extracted behavior features

an electricity theft index is computed to classify the type of electricity theft.Experimental results show that the designed method achieves high accuracy in identifying electricity theft behaviors and effectively distinguishes different types of violations.