Improvement of electricity consumption model using variables constructed by zero-shot labelling on social media data

The local authorities and utilities need a deeper understanding of the drivers of electricity consumption in order to not only maintain energy balance, operational planning and grid resilience but also to design effective flexibility mechanisms and demand-side management actions. Most existing models rely on calendar-derived variables and meteorological data and they exhibit limitations in discerning granular local variations, hence, consequently fail to capture details in local shifts. Capturing local trends and accurately modelling meteorological impacts during extreme weather events remains challenging at a local scale. Additionally, diffused and unexpected factors (e.g. sports events, strike etc.) may influence electricity consumption in ways that are not anticipated. Building on earlier successes in refining well-known effects, i.e., Christmas or sports events, this research work leverages Natural Language Processing (NLP) and zero-shot labeling using a pretrained, open-source, natural language inference (NLI) model to classify textual data of social media. Our method generates new variables in the form of time series based on the content of these posts. These context-specific variables reflect local characteristics. In this paper, we focus on evaluating the impact of adding a single variable to the baseline prediction model of three French cities, without examining their combined effects altogether. The results are promising. For instance, the results show that the information about the public demonstration in a city improves the prediction error by up to 9%, therefore, improving the performance of model.