• Funding : ANRT, Kapsdata
  • Start year :
  • 2025

Joint work with the company Kapsdata

Faced with the heavy consequences of rising energy costs for industrial companies, it has become urgent for the industry to implement solutions that can improve the energy performance of its infrastructures. This challenge involves three main objectives: better controlling the consumption of water, steam, electricity, and gas in complex processes (such as data centers, large technical or commercial buildings, and the agri-food sector), reducing environmental impact, and coping with the constant fluctuations in energy and raw material prices.

To address these issues, this CIFRE PhD project aims to optimize data centers in real time, through edge computing, based on a limited or even non-existent historical dataset. The proposed approach will rely on three main axes:

  • Anomaly Detection: Development of unsupervised learning models, designed to be trained incrementally, in order to detect subtle or emerging anomalies (which would escape static rules) while progressively adapting to normal variations in the system related, for instance, to configuration changes, workload evolution, or environmental conditions.

  • Prescriptive Models: Design of a multi-objective agent based on reinforcement learning enriched with physical knowledge (PIRL), combined with domain expertise. The goal is to achieve a resilient control policy that approaches optimality.

  • Explainability: Without clear explanations, operators may lack confidence in the system and fail to follow the proposed actions. It is therefore essential to develop tailored explainable artificial intelligence (XAI) methods, capable of justifying each alert or recommendation with precise arguments, in particular by identifying the elements or conditions that led to the decision.