Mixed data-driven decision-making in demand response management: An empirical evidence from dynamic time-warping based nonparametric-matching DID

As an important approach for demand-side management in the power sector, demand responses (DRs) are increasingly important in guiding scientific energy consumption behaviour. However, most related prior studies are based on small-scale experimental or survey data with a rule-based optimization algorithm; scientific DR management and strategy formulation studies driven by large-scale, hybrid frequency data are rare. This paper integrates a large-scale controlled trial, 15 min high-frequency power consumption data, and individual residents’ monthly low-frequency power consumption data on a micro-scale. The data-driven and causal analysis methods are combined and a machine-learning algorithm have been adopted to propose a dynamic time-warping (DTW) clustering-based difference-in-differences (DID) method. This non-parametric matching method successfully results in an intra-group randomized experiment. Empirical results reveal that cash-incentive-based DR can effectively stimulate electricity-saving behaviour, and families from the treatment groups save an average of 27.3% of their total electricity consumption in the experimental period. Further, a dynamic response process analysis indicates that a substantial discrepancy exists in the degree of demand response and the response modes of residents with different power consumption patterns. More importantly, prior empirical studies proved this method's effectiveness and feasibility: based on the DTW non-parametric matching method, the control and treatment groups can well support the parallel trend hypothesis. This work provides important implications for the accurate, efficient implementation and scientific decision-making of subsequent DR programs.