Enabling real-time optimization of dynamic processes of proton exchange membrane fuel cell: Data-driven approach with semi-recurrent sliding window method

Abstract

The reliability of proton exchange membrane fuel cell (PEMFC) tightly depends on the suitable operating conditions during dynamic operations. This study proposes an optimization framework to determine the optimal control strategy for PEMFC cold starts underpinned by a novel artificial intelligence method, to improve cold-start capacity and shorten the start-up time. The effects of constant and dynamic currents on PEMFC cold starts under various initial temperatures are studied. The numerical results from a developed PEMFC dynamic model show that the constant current slope strategy (CCSS) is more efficient than the constant current strategy (CCS) in respect of the cold-start time. In the CCSS study, a too-large current slope can lead to a voltage undershoot and then cause a failed cold start, but a too-small current slope can result in a long start-up process in the investigated range of the operating conditions. A data-driven model is developed for dynamic prediction and real-time optimization during the cold start by a semi-recurrent sliding window (SW) method coupled with artificial neural networks (NN) with the simulation data. Based on this NN-SW model, the specific safety–critical operating condition curve under the CCSS has been identified. A real-time adaptive control strategy (RACS) is further proposed to optimize the operating current during the PEMFC cold starts with various initial temperatures. Compared to the optimal CCSS, RACS proves to be more robust and efficient for PEMFC cold-start startups. Based on RACS, the start-up time for an initial temperature of −20 °C can be cut down by 26.7%. Furthermore, the ice predictions by the NN-SW model are also tested and the results are satisfying with an average R2 = 0.9773.