Section: Application Domains

Energy Management

Energy management has been one of our priority application fields since 2012, under the lead of Olivier Teytaud. The first works were concerned with sequential decision making, and were based on TAO experience in games, in particular GO, starting with the Associated Team (EA) with Tainan (Taiwan) and the Inria ILAB Metis, in collaborations with SME Artelys. This collaboration continued to be very fruitful, with the ADEME BIA project POST (2014-2017), about long-term investments in power systems, and the ADEME BIA NEXT, that started in April 2017 for 4 years, about the optimization of local grids (at the city or region level). Another line of research is addressed in collaboration with RTE, the company that manages the global French electric network, through Benjamin Donnot's CIFRE PhD.

The collaboration with Artelys had moved from sequential decision making in the Metis ILAB to reinforcement learning, and the design of the Direct Policy Search approach to handle non-anticipativity, in the POST project. Currently, the NEXT project is concerned with the optimization of local networks to meet customer demand, and hightlights the need for an accurate, robust, and fast simulator (Big Data), and some efficient modeling of the demand (Small Data). This is the topic of Victor Berger's PhD (started Oct. 2017). Another issue is directly related to the network optimization - and the optimal setting (possibly online) of graph optimization algorithms, which this is the topic of Herilalaina Rakotoarison's PhD, started Nov. 2017.

The on-going collaboration with RTE is about learning the parries in reaction to network or demand changes to enforce the "n-1" security constraint: at any time, the failure of any of the 30000 links in the network should preserve the security constraints. Logs of network operations over many years are available, but without any "parry" label. This can be achieved by simulating what would have happened without that particular operations regarding the n-1 constraint. The available network simulator is far too slow and sensitive to noise to be useful here. Modeling the network using Deep Networks is straightforward, for a given topology, though computationally costly. The challenge is to take into account the tolopology so that the n-1 constraint can be quickly checked with a single network. The first results on a small grid (118 nodes) outperform the classical DC approximation while providing a significant speedup in calculations [42]. Further works include scaling up, and incorporating all the intricacies of real data.

Several other energy-related works have been, or will be addressed [20], including the organization of a large scale challenge funded by the EU, which was endowed with 2 million euros in prizes (Isabelle Guyon co-organizer), in the context of the EU project See.4C. The participants are asked to predict the power flow on the entire French territory over several years. This challenge will be followed by a challenge in reinforcement learning (RL), in the context of Lisheng Sun's PhD thesis (started Oct. 2016), who is now working on the problem of RL and Automatic Machine Learning (reducing to the largest possible extend human intervention in reinforcement learning). Another direction being explored is the use of causal models to improve explainability of predictive models in decision support systems (Inria-funded post-doc Berna Batu). This should allow us making more intelligible suggestions of corrective actions of operators to bring network operations back to safety when incidents or stress occur.