Targeting energy conservation

C04 considers the energy transition towards sustainability from the perspective of households and companies. The project leverages machine learning (ML) for domain and out-of-distribution (OOD) generalizations to identify characteristics of individuals that favor energy conservation. The aim is to determine relevant target groups that are susceptible for such interventions. In the long run, the project aims for scalable methods for OOD generalization which can be used for energy conservation campaigns at larger scales.

The transition towards sustainability requires effective policies that trigger changes in economic decisions by households and companies. This project addresses several key questions concerning policies targeting household energy use, energy use in firms, demand flexibility, and mobility behavior. It focuses on experimental methods and deep learning algorithms to demonstrate how machine learning (ML) on spatio-temporal data can be utilized to increase the effectiveness of policies through targeting. Targeting policies is typically understood as targeting those individuals who are expected to have the greatest behavioral response to the policy. We complement this view by targeting policies at the right time, such as peak hours on the power grid or rush hours on the road, and at the right place, such as in certain congested power distribution networks or on certain highways or city districts.

In particular, we conduct large-scale surveys and field experiments to investigate the effectiveness of promising innovative interventions aimed at changing individual behavior. Besides energy consumption, we also aim to study the effects on the well-being of the people affected. For these analyses, we leverage ML models such as deep neural networks to uncover the sensitivity of intervention effects with respect to personal characteristics as well as spatial and temporal circumstances. Further, we develop methods that tackle the problem of domain and out-of-distribution (OOD) generalization as well as test-time model adaptation for better generalization performance of deep learning models.

As a further step, we investigate how these findings can help to identify target groups in relevant populations outside the experimental setting, which leads to improved cost effectiveness of targeted interventions. ML methods trained on the basis of the field experiment are used to make predictions about the impact of such targeted interventions at the individual level based on readily available spatio-temporal data such as electricity load profiles, locations or traffic behavior. However, such predictions require generalizations of the patterns found in the training data. From an ML perspective, this poses a major challenge, as the populations to be generalized over may differ greatly from the training data. A particular focus of the project therefore lies on the development and the application of deep learning methods that allow for such OOD generalizations to succeed with as few additional data requirements as possible.