Zhaoming Liu

Fast online access to fuel cell impedance is important for health diagnosis and fine-grained management, yet existing onboard methods still struggle to balance update rate and measurement fidelity. This paper presents a hybrid monitoring framework that combines sparse physical calibration from a production DC-DC converter with continuous virtual estimation. A six-phase interleaved DC-DC converter injects dual-frequency excitation at 10 Hz and 1 kHz and, with synchronous Fast Fourier Transform (FFT) extraction under integer-period sampling, provides physically grounded impedance references while keeping current ripple below 1%. Between calibration instants, a compact recurrent estimator, FR-LSTM-lite, infers impedance continuously from electrochemistry-guided inputs. Structured pruning and 8-bit quantization reduce the model to a form suitable for automotive microcontrollers. Hardware-in-the-Loop (HIL) and 140-kW bench tests show inference latency below 1 ms and impedance error no greater than 1.5%, delivering milliohm-level monitoring without extra hardware or sustained excitation. Together, these capabilities make milliohm-resolution impedance available in real time on production hardware, opening a practical path toward impedance-feedback fuel cell control.

This paper proposes a novel co-control framework that integrates TD3-based deep reinforcement learning (DRL) with an long short-term memory (LSTM) risk prediction model for optimal purge management in proton exchange membrane fuel cell (PEMFC) systems. The framework is developed using a dynamic model identified and validated on a 60 kW fuel cell system. The DRL agent learns reward-optimized purge strategies, while the auxiliary LSTM module predict the risk of nitrogen-water accumulation to enable protective control actions. Comparative evaluations of control strategies demonstrate the proposed DRL framework’s effectiveness in enhancing anode circulation adaptability. The LSTM component successfully triggers preventive purges upon detecting anode circulation risk states. This hybrid approach mitigates risks inherent to the trial-and-error limitations of pure DRL methods while improving system robustness. The results demonstrate a promising application of DRL and risk protection for fuel cell systems, which achieves safe comprehensive improvements by optimally balancing the system’s output response and fuel efficiency.

Effective thermal management is essential for the performance of high-power fuel cell systems, especially for vehicular applications. This study presents a novel self-adaptive predictive control (SAPC) strategy to address challenges posed by large time delay effects and time-variant characteristics in fuel cell systems. By integrating prediction and current information, SAPC reduces the time-domain signal mismatch induced by delay effects in the basic proportional–integral–derivative (PID) controller. Adaptive weight factors for the prediction component are tuned online for various scenarios, and online/offline model adjustments automatically adapt the dynamic transfer-function model to varying operational points and time-variant characteristics. Experimental validation on an 80 kW fuel cell system test bench demonstrates that SAPC significantly reduces temperature control error by 59% and 64.5%, overshoot by 74.6% and 70%, and mean settling time by 87.9% and 77.6% compared with two other conventional PID frameworks. This highlights SAPC as a promising solution for enhancing precise temperature control in high-power fuel cell systems within dynamic load conditions.

Maintaining an appropriate permeated nitrogen concentration and stable anode pressure are crucial for fuel cell performance. However, direct detection of nitrogen concentration encounters challenges due to measurement limitations, and an abrupt pressure drop occurs during purging. In this paper, an extended Kalman filter observer is devised based on the lumped model and utilized for the anode side to filter the pressure signal, estimate the nitrogen fraction, and calculate the purge flow. The filtered pressure is utilized as feedback for a linear active disturbance rejection controller (LADRC) to achieve precise anode inlet pressure control. The estimated purge flow serves as the variable feedforward (VF) to compensate pressure fluctuation during purging, triggered by the nitrogen fraction threshold. Analysis between the traditional feedforward proportional-integral controller and the LADRC-VF underscores better performance of the latter, showcasing substantial reduction in anode inlet pressure. Finally, the LADRC-VF is validated on a 60 kW fuel cell system, demonstrating an average pressure fluctuation maintained within ±2 kPa and a 13% decrease in root mean square error. Meanwhile, the observer-based purge strategy proves advantageous over traditional fixed interval purging, showcasing efficiency in total purge time savings while ensuring a low nitrogen concentration under dynamic load conditions.

Advanced sensors based on electrochemical impedance spectrum (EIS) are essential for fuel cell system state estimation, fault diagnosis, and health management. However, traditional EIS methods are unsuitable for real-time state management due to high costs, large sizes, and long acquisition times, limiting their online industrial application. To enhance onboard EIS application, this study proposes a novel multistage and multisine injection method to improve wideband frequency sweeping, balancing EIS quality, and measurement time using a four-phase floating-interleaving parallel dc–dc converter prototype with EIS injection functions. In addition, a fuel cell coupled with a four-phase floating-interleaving parallel dc–dc equivalent circuit model is established for comprehensive performance analysis in impedance calculation with different excitation methods. Experimental validation showed an impedance magnitude error within 3.2% from 1 kHz to 0.1 Hz, and the implementation reduces EIS measurement time from 900 s to approximately 150 s. Furthermore, a case study on fuel cell EIS sensitivity under dynamic conditions indicates that the proposed prototype can capture dynamic EIS variations using local multifrequency points.

The prediction of the fuel cell voltage and impedance facilitates the development of advanced energy and health management strategies. In this paper, a novel joint prediction framework for the voltage and characteristic frequency impedance that can reflect internal dynamics loss based on the residual network (ResNet) with the spatial and channel squeeze & excitation (scSE) attention mechanism (scSEResNet) is proposed, where seven measurable signals in the vehicle fuel cell system are selected as input signals. The proposed joint prediction framework is validated through a series of test sequences, with the results displaying the mean absolute percentage error of 0.9512% and 0.9286% for the voltage and 2500 Hz impedance, respectively, and 0.7881% and 2.1069% for the voltage and 10 Hz impedance, respectively. Besides, the proposed scSEResNet is compared with other typical deep learning networks, demonstrating the highest accuracy without causing additional computational burden even with the inclusion of the attention mechanism. Moreover, the proposed framework is further validated and compared under dissimilar noise levels, the removal of the specific input signals, and different types and degrees of faults, respectively.