This report overviews the opportunities offered by both traditional and novel machine learning techniques to assess the causal outcome of large-scale behavioral interventions affecting power consumption. By applying these methods to the large smart data ensembles collected in natural experimental contexts, like those set up by utilities or even those naturally arisen by the progressive application of lockdown policies during the COVID-19 pandemic, it is possible to explore a variety of response models to different behavioral levers, while trying to unpack the heterogeneity and to guide the policy discussion.
Different feedback mechanisms have been devised with varying level of success in affecting consumption. The causal effect of such interventions has been traditionally estimated with econometric techniques. More recently, thanks to the availability of large data ensemble from smart meters, machine learning brings new opportunities for causal inference analyses.
This work reviews available machine learning techniques that can be used to leverage on smart meter datasets for impact evaluation. Concepts like forecasting, clustering, explainable machine learning, and causal forests are presented for this purpose and their benefits and limits emphasized.
Three case studies corresponding to natural and artificial behavioral interventions monitored via smart meters in Italy and Poland showcase a variety of traditional and novel techniques used to analyze the corresponding smart meter datasets and uncover patterns of household-behavioral changes in power consumption associable with those interventions.
Overall, novel large-scale data-driven assessments of behavioral intervention suggest energy savings of few percentage points. Nonetheless high heterogeneity emerges from the data. Machine learning can help to better understand this heterogeneity. In general, more data and experiments are needed to further refine the match between different classes of households and the most effective behavioral intervention, as well as to scale the insights to other regional contexts.
