Leveraging Neural Networks in a Hybrid Model Predictive control framework for district heating networks
Résumé
District heating networks are urban-scale heat energy supply systems. These networks offer a flexible manner
to supply heat energy by leveraging the thermal inertia of the network pipes and the presence of distributed heat
sources. To exploit such flexibility and reach energy supply efficiency, optimal operational control strategies of
the networks are essential. However, existing models developed for an optimal control strategy are hindered
by their computational costs, rendering them impractical for large networks. The computational burden stems
mostly from the iterative simulation of the networks required for each optimization resolution step, until optimal
control parameters are attained. Indeed, each simulation of the network has computational costs proportional to
the size of the networks. In this study, we propose a hybrid Model Predictive Control (MPC), a control
framework using a hybrid model to simulate the network and optimize the mass flow rates and supply
temperatures injected by the heat sources to fulfill substation energy demands with the lowest surplus and
deficits of delivered heat. The proposed approach substitutes clusters of consumer substations with artificial
Neural Networks (NN) models, trained to replicate real-time consumption patterns and leaving waters’
temperatures evolution of the replaced clusters. This hybrid approach reduces computational costs while
maintaining prediction errors below 0.52 %. Results demonstrate that replacing one-third of the network leads
to a reduction of 9% of the computational time required by a physics-based MPC.
Origine | Fichiers produits par l'(les) auteur(s) |
---|---|
licence |