EMPOWER

The project EMPOWER aims to improve short-term operation (and thus cost-effectiveness) of power systems by enhancing the Optimal Power Flow (OPF) formulations. The limitations of the current formulations are ever more prominent because of the increasing uncertainty in power systems arising from the growing penetration of renewable energies. In EMPOWER, we will embed new Neural Network (NN) architectures within the OPF problem to model complex, intractable and/or non-convex constraints.

Firstly, static constraints (with no time dependency), such as the power flow equations, will be investigated. Replacing those constraints by NNs not only reduces the accuracy loss compared to conventional relaxations, but also significantly improves computation time. Different types of NNs and implementations (classifier and regressor) will be tested. The activation functions and the loss function of the NN will also be investigated. In addition, we will ensure that each weight matrix associated to each layer of the NN presents a high level of sparsity to quicken the subsequent optimization process. EMPOWER will go further by integrating extremely hard-to-model time dependent processes within OPF using recurrent neural networks. The consideration of these processes (e.g., pumped-hydro energy storage, thermal loads) is rapidly gaining in importance due to the flexibility they offer.

The innovations developed within EMPOWER will be applied to real-life Belgian systems in due course to ensure the practical performances and tractability thereof and to benchmark them in terms of computational time, reliability and cost-effectiveness of the solution with respect to classical methods. From an improved approximation of the OPF problems to the consideration of new time-dependent processes, EMPOWER will have a strong impact on all the power system actors (end-users, system operators, balancing responsible parties…) and facilitate the ongoing crucial energy transition.