Sample-Efficient Hyperparameter Optimization of an Aim Point Controller for Solar Power Tower Plants by Bayesian Optimization

This is the latest version of this item.

Abstract

In this work, sample-efficient algorithms for a controller hyperparameter optimization of an arbitrary aim point controller for solar power tower plants are introduced. The objective is to find controller parameters, which optimize the performance of the aim point controller, and thus increase the efficiency of the plant. This should be accomplished within a minimum number of optimization steps, which implies the need of a sample-efficient optimization strategy. The algorithms, proposed in this work, are based on the Bayesian Optimization (BO) approach and enhance the algorithm's sample efficiency by leveraging simulation data as prior information. It is assumed that the utilized simulation data is possibly corrupted by mismatches to the system's real behavior and thus does not contain information about the optimal controller parameter configurations. Therefore, it is not possible to choose them directly from the simulation data, however it can still contain helpful information to accelerate the optimization. The controller parameters, selected by an optimization algorithm, have to be evaluated on the plant, after every optimization iteration. Testing the controller parameters on the real system is a time-consuming procedure, which explains the need to reduce the optimization iterations to a minimum. The algorithms, proposed for this purpose, are mostly based on the methods for leveraging prior information in BO of Antonova and Rai et al. , and extended to the use of multiple sets of simulation data, which was not sufficiently covered in literature so far. Moreover, a novel approach for utilizing simulation data in BO is introduced in this work, named Prior-Guided Expected Improvement. The algorithms were tested on a six-dimensional test function, which imitates the performance of an aim point controller, dependent on six controller hyperparameters. Several sets of simulation data were deployed, that partly resemble the function and do not contain the function's global optimum. As a reference, the standard BO algorithm was used. Two of the proposed approaches outperformed the reference by reaching close to optimal controller hyperparameters within 33 % less optimization steps, than the standard BO. In addition, the prior-informed algorithms seemed to be less prone to get stuck in local optima, than the standard BO. Moreover, in case of high simulation to reality mismatches or unsuitable simulation data, the prior-informed algorithms still yielded results similar to the reference. In a second test case, the proposed approaches were used to optimize a simulated Vant-Hull aim point controller with two hyperparameters, where they needed 23 % less optimization iterations than the standard BO. However, to test the prior-informed aim point controller optimization on a real solar power tower plant, further development has to be done to guarantee save controller behavior during the hyperparameter optimization. Thereby, damages to the receiver, caused by overheating, can be prevented.

Available Versions of this Item

• Sample-Efficient Hyperparameter Optimization of an Aim Point Controller for Solar Power Tower Plants by Bayesian Optimization. (deposited 28 Oct 2022 10:04) [Currently Displayed]