Tuning of Rule-Based Control Parameters: A Bayesian Optimization Approach

Many hierarchical control systems consist of a rule-based controller on the top-level. Since the parameters of rule-based controllers may not be properly tuned, the corresponding closed-loop systems are not operating efficiently in terms of energy consumption, monetary costs, or other performance measures. However, often an advanced control strategy like MPC is not economically feasible due to the large effort associated with its commissioning. This fact highlights the importance of tuning rule-based controllers, preferably in an automated and model-free way so that they can be easily tuned specifically when the access and modification to the plant is very limited and minimal. Recently in the literature and in practice, data-driven approaches based on black-box and Bayesian optimization are utilized for toy examples such as inverse pendulum and applications like robotics, combustion engines, PID parameter optimization for heat pumps, linear axial systems and grinding machines.

Many hierarchical control systems consist of a rule-based controller on the top-level. Since the parameters of rule-based controllers may not be properly tuned, the corresponding closed-loop systems are not operating efficiently in terms of energy consumption, monetary costs, or other performance measures. However, often an advanced control strategy like MPC is not economically feasible due to the large effort associated with its commissioning. This fact highlights the importance of tuning rule-based controllers, preferably in an automated and model-free way so that they can be easily tuned specifically when the access and modification to the plant is very limited and minimal.
Recently in the literature and in practice, data-driven approaches based on black-box and Bayesian optimization are utilized for toy examples such as inverse pendulum and applications like robotics, combustion engines, PID parameter optimization for heat pumps, linear axial systems and grinding machines.

Following the same lines, we will investigate tuning of a rule-based controller implemented for an electrical and/or heat storage system installed in the ehub at Empa, Dübendorf with unsatisfactory performance. We will utilized an approach based on safe Bayesian optimization with Gaussian processes which is well in agreement with the given operational constraints and available resources. By applying problem-relevant performance metrics and safety constraints for the safe Bayesian optimization, we will show that the performance in terms of flexible capacity, peak-power reduction, self-consumption can be improved. A very important aspect is that the safety of the system is maintained during the optimization while no model for the system is considered or exploited.

Following the same lines, we will investigate tuning of a rule-based controller implemented for an electrical and/or heat storage system installed in the ehub at Empa, Dübendorf with unsatisfactory performance. We will utilized an approach based on safe Bayesian optimization with Gaussian processes which is well in agreement with the given operational constraints and available resources. By applying problem-relevant performance metrics and safety constraints for the safe Bayesian optimization, we will show that the performance in terms of flexible capacity, peak-power reduction, self-consumption can be improved. A very important aspect is that the safety of the system is maintained during the optimization while no model for the system is considered or exploited.

Joint project between ETHZ & Empa: Mohammad Khosravi - khosravm@control.ee.ethz.ch, Benjamin Huber - benjamin.huber@empa.ch

Joint project between ETHZ & Empa:
Mohammad Khosravi - khosravm@control.ee.ethz.ch,
Benjamin Huber - benjamin.huber@empa.ch