Uncertainty quantification of proton-exchange-membrane fuel cells degradation prediction based on Bayesian-Gated Recurrent Unit

Machine learning is very important in predicting the degraded performance of fuel cell systems for advanced diagnosis and control. Unfortunately, existing machine-learning-based schemes are usually designed with point estimation, making it difficult to quantify the uncertainty of the prediction result. In this paper, we propose a Bayesian-Gated Recurrent Unit model (B-GRU) that combines the Bayesian Theory and GRU to predict the phenomenon of fuel cell voltage decay. Fuel cell data are preprocessed by the random forest, and the key feature data are then imported into the B-GRU. Variational inference and adaptive moment estimation is used to obtain the optimal parameters in the B-GRU. Probability density distributions are calculated by replacing the param-eters in GRU with random variables to quantify the uncertainty in the model. In addition to providing point estimates, the B-GRU also gives interval estimates for uncertainty quantification. With small training data, the point estimation result of B-GRU is more accurate than traditional neural networks. Furthermore, compared to the Bayesian neural networks, the proposed B-GRU also exhibits superior performance both in point and interval estimation results based on the IEEE PHM 2014 DATA Challenge dataset. With its excellent ability for noise immunity and uncertainty quantification, the proposed prediction method can provide more useful decision -making recommendations for hydrogen energy devices.

Automated summary

In this paper, a B-GRU model combining Bayesian Theory and GRU is proposed for predicting the voltage decay phenomenon in fuel cells. The B-GRU outperforms traditional neural networks in point and interval estimation results, and exhibits excellent ability for noise immunity and uncertainty quantification.