Machine learning-based simulation and experiment of a concentration sensor-free control strategy for direct methanol fuel cells

The methanol concentration critically influenced the performance of direct methanol fuel cells (DMFCs); yet conventional sensor-based approaches suffered from limited accuracy, high costs, and progressive sensor degradation. To overcome these limitations, we developed a sensor-less concentration-control strategy grounded in data-driven modeling. A semi-empirical electrochemical model was first calibrated against experimental polarization curves to expand the voltage-performance dataset, whereas the methanol-consumption dataset was densified via linear interpolation and k-nearest-neighbour augmentation. The enriched data enhanced model robustness and enabled a systematic comparison of machine-learning algorithms. A Bayesian-regularised artificial neural network yielded the most accurate voltage prediction (R2=0.99838), while an interaction-effect linear regression captured the methanol consumption rate with high fidelity (RMSE=0.000473,R2=0.99391). We then integrated these surrogate models into a dynamic Matlab/Simulink framework and validated them through two one-hour random-current tests. The integrated model reproduced the transient stack voltage with mean absolute relative errors of 2.1% and 1.6%; the predicted methanol concentrations deviated from measurements by only 0.12wt% and 0.09wt%, respectively. Finally, a 24 h sensor-less control experiment on a 100 W DMFC stack maintained the inlet methanol concentration within 1.9–2.1wt%, with deviations remaining below 0.2wt%. These results validated the effectiveness and practicality of the proposed data-driven strategy, offering a scalable pathway toward reliable, sensor-free methanol-concentration control and performance prediction in high-power DMFC systems. Significantly, this work offers a novel perspective for advancing sensorless methanol concentration control strategies in DMFC systems.