Performance Evaluation of S2P-CNN for Multi-Load Disaggregation of EV and PV Generation under Varying Data Sampling Rates

Abstract

This study evaluates the performance of a multi-output regression Sequence-to-Point Convolutional Neural Network (S2P-CNN) for Non-Intrusive Load Monitoring (NILM), with a specific focus on disaggregating Electric Vehicle (EV) charging load and Photovoltaic (PV) generation under varying data sampling rates. While the growing integration of Distributed Energy Resources (DERs) has increased the complexity of household energy patterns, most utility smart meters still record data at low resolutions (e.g., 15 minutes), presenting a significant challenge for traditional algorithms. To address this issue, experiments were conducted utilizing the real-world UK-DALE dataset solely as a background load, onto which synthetic EV and PV profiles were superimposed at resolutions of 1, 5, 10, and 15 minutes. To ensure statistical robustness, all evaluations were validated across multiple random seeds and compared against a standard Long Short-Term Memory (LSTM) sequence-modeling baseline. The quantitative results reveal a clear engineering trade-off: while lowering the sampling rate degrades point-wise signal shape accuracy due to the loss of high-frequency transients, the model successfully preserves the total energy mass. Crucially, multi-seed validation highlights a specific sensitivity at certain resolutions (e.g., 10 minutes), causing feature extraction to be highly dependent on initial weights. However, at the primary target resolution of 15 minutes, the model regains stability and demonstrates highly consistent behavior, maintaining Signal Aggregate Error (SAE) values safely below the 10% acceptable threshold. Furthermore, baseline evaluations indicate that the S2P-CNN architecture offers greater robustness and computational efficiency under coarse sampling conditions compared to the LSTM. These findings demonstrate that the S2P-CNN is a viable solution for aggregate-level energy estimation using existing Advanced Metering Infrastructure (AMI), offering a practical and cost-effective deployment pathway for smart grid management without necessitating immediate hardware upgrades.