Abstract:
Smart grids generate extensive data, yet much analysis remains association-based and cannot reliably inform decisions about the effects of interventions in planning and operations. Randomized trials provide strong causal evidence but are costly and limited in scope, predictive models cannot isolate intervention effects from confounding factors, and existing causal studies in smart grids each focus on a single aspect, such as market allocation, network evolution, or temporal dependencies, without covering the complete workflow from problem formulation to decision-making. This paper introduces a domain-specific causal inference framework for smart grids that integrates established causal modeling, identification, and validation techniques into a structured workflow tailored to energy-system analysis. The framework is organized into seven phases: causal problem framing, causal model construction, data collection and preparation, causal discovery, causal inference, refutation tests and validation, and intervention evaluation and recommendation.
To support adaptation across settings, the paper categorizes common causal problems in smart grids into three types (confounder-intensive, temporally structured, and real-time decision problems) and shows how the framework maps to each, with the confounder-intensive case demonstrated in detail.
The framework is demonstrated through a demand response case study assessing the causal impact of pricing schemes on household electricity consumption. Using real-world data, we apply three representative causal inference methods (Backdoor Adjustment, Difference in Differences, and Double Machine Learning), benchmarked against an association-based baseline. The results show that factors influential under association can become negligible or even reverse under causal estimation, while seemingly minor factors emerge as plausible levers for decision-making under the stated identification assumptions.