A Novel Online Learning-Based Optimal Control Algorithm for Enhancing Solid Oxide Fuel Cells Performance in Islanded DC Microgrids

Solid oxide fuel cells (SOFCs) offer a promising solution for enhancing reliability and sustainability in microgrid power supply with the growing penetration of renewable energy sources. The proposed method addresses key challenges in existing SOFC control approaches, including model dependence, usage of nonoptimal control policy, reliance on an offline-trained neural network (NN), and complex design. Compared with model-based methods, this method uses NN and policy iteration technology to learn system dynamics and approximate optimal control policy, thereby eliminating model dependence. Compared with offline learning-based methods, this method achieves online policy evaluation and NN updating to eliminate tedious offline training and data acquisition processes. Compared with the online learning-based SOFC control approaches, this method employs a fixed-weight recurrent NN to avoid slow or even no convergence caused by recursive least squares-based NN weights updating process, reducing design complexity without sacrificing control performance. The superiority of the proposed method is validated through hardware-in-the-loop tests.