Distribution-based residual analysis of load forecasting for enhanced anomaly detection

Anomalous behavior, spanning from faulty sensors to energy theft, introduces significant uncertainty in energy data, contributing to resource misallocation, profitability loss and the development of inefficient tariff structures within the energy sector. We propose a novel detection framework that leverages the utility sector’s existing energy forecasting models by differentiating between the residuals on potentially anomalous days and those from “known” benign days by quantifying the dissimilarity between their underlying distributions. Our unsupervised approach is rigorously tested on two distinct datasets utilizing LSTM and ARIMA forecasting models. The results underscore our method’s enhanced detection capabilities leading to substantially lower false-positive rates compared to state-of-the-art unsupervised techniques that leverage forecasting errors. Specifically, our approach achieves accuracy improvements of up to 20.7% over the state-of-the-art and reduces false-positive rates to as low as 6.6%, compared to the 11% reported by these techniques. Further analysis reveals that higher sampling rates markedly improve detection performance, positioning our method as a superior alternative within existing literature when higher-sampling rate data is available. • Novel unsupervised method: residual analysis detects anomalies using distribution dissimilarity. • Outperforms state-of-the-art: up to 20.7% more accurate; false positives reduced to 6.6%. • Robust across models (LSTM, ARIMA) and horizons (6h, 12h, 24h) for broad applicability.