Towards edge-computed NILM: Insights from a Mediterranean Use Case

An optimized structured pruning methodology within the context of Non-Intrusive Load Monitoring (NILM) is introduced. The proposed methodology exploits unstructured pruning to determine the optimal sparsity ratio for each layer in the deep neural network model. Subsequently, structured pruning is applied to remove entire units from each layer according to the sparsity values guided by the unstructured pruning. By doing so, important feature information is preserved, resulting in improved classification performance as the pruning threshold is not arbitrarily selected based on a random percentage ratio. Experimental results, evaluated on the Plegma dataset—one of the first datasets from the Mediterranean area capturing local devices and consumption patterns—demonstrate that the proposed methodology significantly optimizes inference performance. Specifically, the approach reduces the baseline model’s MFLOPs by up to 48.85% while keeping a satisfactory disaggregation performance, a stark contrast to the widely used unstructured pruning approach, which does not achieve FLOPs reduction. These findings underscore the potential of edge NILM for promoting flexibility and energy transition in the Mediterranean region, facilitating the broader adoption and implementation of NILM solutions in real-world scenarios.