Interpretable Deep Learning using Kolmogorov–Arnold Networks for Energy Theft Detection

Abstract

The increasing availability of high-resolution metering data has positioned deep learning as a powerful tool for energy theft detection; however, most existing approaches rely on black-box models that lack interpretability and require large, labeled datasets, limiting their applicability in regulated and safety-critical environments. This paper proposes an interpretable deep learn-ing approach for energy theft detection based on Kolmogorov–Arnold Net-works (KAN), explicitly designed to combine non-linear modeling capability with feature-level transparency. The novelty of the proposed method lies in embedding explainability directly into the learning architecture, rather than relying on post hoc explanation techniques. By representing the detection function as a composition of learnable univariate spline functions, the KAN-based model enables direct visualization and quantitative interpretation of each input feature’s contribution to the detection outcome. This property al-lows energy theft detection to be formulated as a transparent decision-support process, suitable for operational deployment and regulatory auditing. The proposed approach integrates engineered consumption features with KAN-based inference and supports deployment within a distributed edge in-telligence architecture, enabling low-latency detection and privacy-aware processing. Experimental evaluation on representative low-voltage network data demonstrates that the proposed approach achieves competitive detection accuracy while significantly improving interpretability and robustness under prosumer-induced variability. The results confirm that inherently interpreta-ble deep learning models represent a viable and effective alternative to con-ventional black-box techniques for energy theft detection.