Health-Constrained Explicit Model Predictive Control Based on Deep-Neural Networks Applied to Real-Time Charging of Batteries

The application of Model Predictive Control (MPC) for optimal real-time battery charging is attracting growing interest due to its advantages over empirical charging protocols. However, the high complexity and nonlinearity of physics-based models of batteries can hinder MPC applications due to the large computational resources required online. To overcome this challenge, this paper proposes a machine learning (ML) based explicit MPC method for battery charging subjected to health constraints. The method uses deep artificial neural networks (DANNs) to construct offline control laws that describe the optimal charging current as a function of the state of the battery. These DANN-based control laws are developed using data generated by solving the MPC problem using the physics-based model of the battery and considering different expectations for the initial state of the battery. These control laws can then be used online to control the charging process by calculating the optimal closed-loop current via simple and inexpensive predictions. The method is applied to a case study of battery charging based on MPC, and the results prove the capabilities of the DANN-based control laws in terms of i) very high prediction accuracy of the closed-loop profiles of the charging current, ii) good ability to learn the constraints imposed on the MPC problem from the data, and iii) significant reduction in the required computation time compared to traditional MPC.