Hybrid ILM-NILM Smart Plug System

TL;DR

The paper tackles the cost and practicality gap between intrusive load monitoring (ILM) and non-intrusive NILM by proposing a hybrid approach: extending smart plugs to detect multiple loads connected to a single outlet, including through extension cords. It introduces a high-frequency, dimmer-based measurement method that yields Real Power, Apparent Power, and voltage/current matrices, enabling a CNN-based multi-label classifier to identify multiple concurrent loads and their count $N$. The results show high distribution accuracy (~98%) and strict accuracy (~97.4%) for up to three loads, even with reduced multi-load training data, demonstrating strong potential for real-world deployment with lower hardware costs. This work advances practical energy management by enabling appliance-level insight and control in plug-based systems without requiring a plug per device.

Abstract

Electrical load classification is generally divided into intrusive and non-intrusive approaches, both having their limitations and advantages. With the non-intrusive approach, controlling appliances is not possible, but the installation cost of a single measurement device is cheap. In comparison, intrusive, smart plug-based solutions offer individual appliance control, but the installation cost is much higher. There have been very few approaches aiming to combine these methods. In this paper we show that extending a smart plug-based solution to multiple loads per plug can reduce control granularity in favor of lowering the system's installation costs. Connecting various loads to a Smart Plug through an extension cord is seldom considered in the literature, even though it is common in households. This scenario is also handled by the hybrid load classification solution presented in this paper.

Figures (9)

• Figure 1: Real Power measurement matrix [W] of an incandescent light bulb. Each row represents a different cutoff ratio, while each cell shows the Real Power measurement recorded for a single AC period.
• Figure 2: All possible two-load category combinations. $\checkmark$ marks the combinations for which at least 100 measurements were recorded.
• Figure 3: Traditional classification accuracy rates (average of 100 runs) selecting the top prediction of the CNN. For a multi-load sample, the classification was considered correct if the top output returned was one of the several appliances connected to the plug. The training data did not contain any multi-load samples. The training data consisted of 220 single-load measurements from each load category. The first row contains the accuracy rates for the 30 test samples per load class, the table in the middle contains the accuracy rates for the two-load combinations, and the bottom two accuracy rates belong to the three-load combinations.
• Figure 4: The mean of all measured Real Power measurement matrices for the laptop and ledspotlight loads (left and right figures) and the mean Real Power measurement matrix of the combined laptop + ledspotlight measurements (middle figure).
• Figure 5: Summary of class detection accuracy results using 70 training samples from each load combination category and 30 test samples. The accuracy rates are averages of 100 runs. The empty white cells show the combinations for which no measurements were recorded.