Meta classifiers on existing forecasting models to enhance energy theft detection

Proposes meta-classifiers for energy theft detection that directly utilize forecast and residual sequences from existing forecasting models, significantly outperforming traditional methods relying on aggregated residual statistics. Demonstrates the practical advantage of leveraging potentially pre-existing utility forecasting infrastructure (e.g., LSTM, ARIMA), achieving substantial detection improvements with minimal additional training overhead. Validates the approach's robustness through comprehensive experiments across multiple datasets, classifiers, synthetic attack types, and shows that incorporating derived crafted features further enhances performance, particularly across varying data sampling rates.