NILM-Flex: quantifying residential demand-side flexibility potential via non-intrusive load monitoring and behavior-aware appliance modeling

Integrating accurate sensing and quantification of residential flexible loads is key to alleviating supply–demand mismatches in power systems. However, most existing studies treat non-intrusive appliance sensing and flexibility quantification as separate problems, and scalable solutions that address data granularity, deployment cost, and privacy remain limited. To address this gap, this study proposes NILM-Flex, a scalable, robust end-to-end framework for quantifying residential demand-side flexibility from aggregated household electricity data. A BiTCN-TransUNet++ load disaggregation model, which integrates bidirectional temporal convolution and a Transformer module, is developed. By introducing a user discomfort cost (ε), a device-level, physics–behavior-coupled flexibility characterization method is established. The NILM model is benchmarked on two public datasets, UK-DALE and REDD, while the complete NILM-Flex framework is further validated using long-horizon field-measured real-world data. The disaggregation model achieves average MAE/F1 values of 7.090 W/0.890 and 8.794 W/0.738, respectively. Compared with the best-performing baselines, the MAE is reduced by 25.8% and 8.2%, and the F1 score is improved by 0.064 and 0.065, respectively. The model effectively reproduces user behavior in field-measured scenarios. Residential flexibility exhibits pronounced heterogeneity and asymmetry across temporal and power dimensions, and instantaneous power flexibility demonstrates stronger upward than downward regulation. In addition, flexibility potential is highly sensitive to ε: as ε increases, flexibility release does not grow linearly but presents time-of-day and stage-dependent patterns, and the marginal benefit of regulation at the expense of user comfort shows a saturation effect. Notably, the framework exhibits conservative estimation under non-intrusive disaggregation errors, ensuring operational decision reliability. Overall, this study provides theoretical support and a practical reference tool for the low-cost, scalable extraction of residential flexibility resources.