Toward Explainable NILM: Real-Time Event-Based NILM Framework for High-Frequency Data

TL;DR

This work tackles the challenge of trustworthy NILM by delivering an explainable, real-time, event-based NILM framework tailored for high-frequency data. The pipeline processes high-frequency signals with FIT-PS, detects on/off events via a z-score detector, estimates robust appliance signatures from activation currents, and constructs a small, interpretable Fourier feature set that feeds an XGBoost classifier, with SHAP providing post hoc explanations. On the PLAID dataset, the method achieves about $90\%$ accuracy while maintaining sub-second latency and low computational burden, demonstrating practical feasibility for edge deployment. The approach emphasizes transparency across preprocessing, feature extraction, and decision-making, aiming to promote adoption of NILM in real-world energy management.

Abstract

Non-Intrusive Load Monitoring (NILM) is an advanced, and cost-effective technique for monitoring appliance-level energy consumption. However, its adaptability is hindered by the lack of transparency and explainability. To address this challenge, this paper presents an explainable, real-time, event-based NILM framework specifically designed for high-frequency datasets. The proposed framework ensures transparency at every stage by integrating a z-score-based event detector, appliance signature estimation, Fourier-based feature extraction, an XG-Boost classifier, and post hoc SHAP analysis. The SHAP analysis further quantifies the contribution of individual features, such as cosine of specific harmonic phases, to appliance classification. The framework is trained and evaluated on the PLAID dataset, and achieved a classification accuracy of 90% while maintaining low computational requirements and a latency of less than one second.

Figures (7)

• Figure 1: Block diagram of an event-based NILM system.
• Figure 2: The flow chart of a proposed approach for high-frequency event-based NILM.
• Figure 3: Activation current processing for a turn-on event. (a) Aggregated current signal with highlighted $N_b = 1$ cycle before and $N_a = 18$ cycles after an event. (b) Activation current $\Delta i$ calculated for 18 cycles. (c) Estimated appliance current signature calculated as in Eq. \ref{['eq:est']}.
• Figure 4: Impact of the number of cycles after an event on the appliance classification accuracy.
• Figure 5: Correlation matrix of Fourier features.