Isomorphic structured pruning of temporal CNNs for scalable NILM on edge devices

This work introduces isomorphic structured pruning (ISP) within the Non-Intrusive Load Monitoring (NILM) domain to support the deployment of optimized Deep Learning (DL) models in resource-constrained edge environments. By grouping functionally equivalent substructures and ranking them independently, it increases the tolerance to neural network units removal thereby offering improved pruning-computational performance trade-off, along with consistency and reliability in the pruning results. Experimental results using a temporal Convolutional Neural Network (CNN) architecture on two datasets, the Plegma, containing Mediterranean-based appliances, and UK Dale, containing widely-used domestic appliances, demonstrate up to 97% model size reduction and computational efficiency gains of up to 42 × in terms of Multiply-and-Accumulate (MAC) operations, all with negligible classification performance losses. Comparisons with popular pruning techniques highlight ISP’s ability in maximizing the trade-off between pruning threshold values and classification performance with the results showing 85% pruning of the temporal CNN, resulting in a value of approximately 0.5 MB in size, combined with up to 99% in accuracy and up 81.4% in F1-score across all devices considered. To showcase its real-world applicability, ISP was deployed on two industrial edge devices with diverse computational resources, an EI-52 and an LEC-7230M, achieving over 95% faster inference times, along with reductions exceeding 85% in both CPU usage and energy consumption compared to the baseline model. These findings underscore the advantages of the ISP as a practical and scalable solution for deploying high-performance NILM models in real-world edge computing scenarios. • Introduce an isomorphic structured pruning method for NILM. • Compare the proposed method against unstructured and conventional structured pruning. • Validate real-world feasibility through deployment on two edge devices. • Provide evaluation on two distinct datasets, the UK Dale and the Plegma dataset.