Données textuelles pour la modélisation avancée des dynamiques de la demande et des prix d'électricité

Despite the increasing maturity of current power systems and markets, they still face numerous uncertain risks, such as sudden changes in electricity demand due to special events. In this context, leveraging machine learning and natural lan-guage processing (NLP) technologies for integrated modelling to address uncertainties in power systems holds significant promise.This thesis revisits traditional power forecasting models, exploring the potential of using information embedded in social texts and investigates multimodal forecasting methods to enhance the robustness and decision-making intelligence of power systems. It designs a novel power forecasting framework that enables the acquisition, modelling, and interpretation of social text knowledge to improve power forecasting.To extract knowledge from vast amounts of text, various text embedding, feature extraction, and feature selection tech-niques have been developed. For modelling this knowledge, the thesis enhances machine learning methods by consider-ing both deterministic and probabilistic forecasting. To explain the impact of this knowledge on power systems, different approaches based on correlation and causality are designed, providing a comprehensive picture of the power-social-economic relationship.The proposed methods have been validated through experiments across different countries, languages, levels, and ob-jectives. In summary, the research findings of this thesis contribute to strengthening the capability of power systems to cope with uncertainties and promote the intelligent development of power systems.