Convolutional neural networks for non-intrusive load monitoring: a critical analysis and future directions

Convolutional Neural Networks (CNNs) have become a leading deep learning approach for Non-Intrusive Load Monitoring (NILM) due to their ability to learn appliance features directly from aggregate signals. Despite rapid progress in CNN-based NILM, existing reviews typically treat CNNs as one of the components within broader deep-learning overviews or cover only a narrow subset of CNN designs; consequently, an up-to-date, CNN-based comparison of architectures, datasets, sampling settings, and performance trends remains essential for guiding both researchers and industry experts in model selection and implementation decisions. This critical review consolidates findings from 165 CNN-based NILM studies published between 2018 and 2025, grouped into pure CNNs, signal-transformation (image-based) CNNs, hybrid CNN architectures, and advanced optimized CNN frameworks. Across the reviewed studies, UK-DALE and REDD are the most frequently used datasets (∼37.8% and 31.4% of studies, respectively), and there is a strong preference for high-frequency sensing, with 10–100 kHz sampling used in 49.2% of experiments (nearly half of the reported studies). Across categories, hybrid CNN models consistently deliver the strongest disaggregation outcomes by coupling convolutional feature extraction with temporal sequence modeling, improving appliance-level F1-scores by roughly 20–30 percentage points for low-power and multi-state appliances with overlapping signatures in representative studies. Therefore, this review provides practical guidance for developing more generalizable and deployment-ready NILM systems.