Section: Application Domains
Energy management, our prioritary application field, involves sequential decision making with:
Significant challenges also include:
spatial distribution of problems; due to capacity limits we can not consider a power grid like Europe + North Africa as a single “production = demand” constraint; with extra connections we can equilibrate excess production by renewables for remote areas, but not in an unlimited manner.
We have had several related projects in the past, many of them together with the SME Artelys, working on optimization in general, and in particular on energy management. In particular, we had with them an Inria ILAB (Metis, ended in end 2014), and are currently working on POST, an ADEME BIA project about investments in power systems that will end in July 2017. Another project has been submitted to ADEME about the optimization of the local grids (at the city level) depending on the demand and the prediction of the market prices.
In 2016, we started to work with RTE, the company that is managing the global electric network in France. They fund Benjamin Donnot's CIFRE PhD thesis about learning the parries to prevent the loss off security of the network in case of material failures or unexpected consumption peaks. This collaboration had several follow-up, including the organization of a large scale challenge funded by the EU http://see4c.eu/, which will be endowed with 2 million euros in prizes (Isabelle Guyon co-organizer). The participants will be asked to predict the power flow on the entire French territory over several years. This challenge will eventually be followed by a challenge in reinforcement learning (RL), in the context of the PhD thesis of Lisheng Sun who just started working on the problem of RL and Automatic Machine Learning (reducing to the largest possible extend thuman intervention in reinforcement learning). Another direction being explored are uses of causal models to improve explainability of predictive models in decision support systems (Inria-funded post-doc Berna Batu). This should allow making more intelligible suggestions of corrective actions to operators to bring network operations back to safety when incidents or stress occur.
Technical challenges: Our work with Artelys focuses on the combination of reinforcement learning tools, with their anytime behavior and asymptotic guarantees, with existing fast approximate algorithms. Our goal is to extend the state of the art by taking into account non-linearities which are often neglected in power systems due to the huge computational cost. We study various modelling errors, such as biases due to finite samples, linearization, and we propose corrections. The work with RTE involves modeling the network itself from archives, because the numerical simulation is both too expensive and not robust, and modeling the client demand in order to be able to predict possible outlier consumptions.