Internal Water States Estimation of Dead-end PEMFC Based on Multi-modal Data Fusion

Dead-end operation of proton-exchange-membrane fuel cells (PEMFCs) causes significant water accumulation, and the resulting water states strongly modulate catalytic activity, heat transfer, and mass transport. Accurate, real-time knowledge of these internal water states is therefore indispensable for effective PEMFC health management. Nevertheless, current experimental and modeling techniques cannot yet deliver dynamic, online predictions. To address this gap, we propose the LSTM-Inception-Transformer, a multi-modal data-fusion network tailored for water-state estimation. The network is trained exclusively on data produced by a validated, three-dimensional, non-isothermal, two-phase, single-channel PEMFC model. By combining long short-term memory (LSTM), Inception, and Transformer blocks, the architecture achieves cross-modal feature fusion and yields reliable water-state estimates under arbitrary load voltages and times. In contrast to conventional water-management strategies that depend only on output voltage/current signals or empirical rules, the proposed approach supplies direct, dynamic decision support. Numerical experiments demonstrate a 51.8% accuracy gain relative to a single-modal LSTM baseline.