Volume 63

Framework for the Development and Sim2Real Transfer of Control Strategies for Industrial Energy Supply Systems Tobias Lademann, Andreas Clement, Arthur Stobert, Matthias Weigold

Abstract

The energy- and cost-efficient operation of industrial energy supply systems is challenging for conventional rule-based control strategies. Operational control algorithms, such as deep reinforcement learning or model predictive control, can improve the operation strategy, but are often benchmarked against simulation without the transfer to the real world. This work presents a Python framework for the development, benchmarking and simulation-to-real transfer of optimized control strategies. The framework application is demonstrated on the heating network of the ETA Research Factory, a representative industrial use case. The implementation is based on the eta_utility library to provide standardized environments and connectivity. The simulation model is based on the ThermalSystemsControlLibrary, which is a Modelica library and models the physical components, automation data model and conventional control strategy. Parameter identification is used to fine-tune selected simulation model parameters and minimize the simulation-to-real gap. Validation experiments are performed with the conventional control strategy for both the real system and the simulation model. During a 24-hour period, the simulation model shows a relative error of -13.3 % for electrical energy consumption and -11.6 % for gas consumption compared to the real system. This framework enables future research to develop and evaluate control strategies in a real-world industrial energy supply system, and it can be used for comprehensive experiments. In future work, various optimization algorithms should be implemented, benchmarked, and compared.

Keywords control optimization, simulation, benchmarking, parameter identification