Aiming at the online control problem of microbial fuel cells, this article presents a class of explicit model-predictive control methods based on the machine learning data model. The proposed method is divided into two stages: off-line design and on-line control. In the off-line design stage, (1) a feasible data set is collected by sampling the admissible state in the feasible region and solving the optimal model predictive control law for each sampling data point off-line, (2) a feasible sample discriminator is constructed based on the support vector machine–based binary classification in order to judge the whether the real sampling state is feasible, and (3) according to the feasible samples and the corresponding optimal control law, the control surface of explicit model predictive controller is constructed based on the machine learning methods. In the on-line control stage, the process data are collected in real time and the feasible control output is calculated by using the trained explicit predictive control surface. Extensive testing and comparison among the different machine learning algorithms, such as artificial neural network, extreme learning machine, Gaussian process regression, and relevance vector machine, are performed on the benchmark model of a class of microbial desalination fuel cells. These results demonstrate that the proposed explicit model predictive control method can avoid the exhausting optimization computing and is easy to realize on-line with good control performance.
Learning a feasible and stabilizing explicit model predictive control law by robust optimization
