Explainable Multi-Class Electricity Theft Detection on the TDD2022 Benchmark

This work presents the first explainable, per-attack-type analysis of electricity theft detection on the TDD2022 benchmark. Ten classifiers are evaluated under a frozen, temporally ordered, per-building-type split that prevents leakage between training and test sets. Rather than reporting aggregate accuracy alone, the study reveals that detection difficulty varies dramatically by attack type from perfect detection to near-complete evasion and uses SHAP to attribute why each attack is caught or missed to specific consumption features. A McNemar’s test across ten models, including modern gradient-boosting methods, demonstrates that performance is bounded by the feature representation rather than model choice, motivating sequence-aware detection as future work.