Data-driven fuel cell performance prediction by transfer learning and dynamic time warping

The increase of world energy demand contributes to the development of renewable energy sources. Fuel cells, which use hydrogen to deliver energy, have seen a promising future. Fuel cells have been investigated since the last 50 years but they are still relatively absent for the commercial use due to the large exploitation cost and poor durability. One of the reasons is that the health state of a running fuel cell is hard to evaluate due to the system complexity and insufficient data. To tackle this problem, this paper proposes a data-driven fuel cell performance prediction method based on transfer learning and dynamic time warping. Historical data is trained in the source domain to build a model that can be transferred to the target domain and can be fine-tuned with only a small volume of online measurement. To improve the model recognition, dynamic time warping technology is applied to quantify the similarity between two different time series so that similar instances can be selected to build the neural network prediction model. Results show that, compared to traditional prediction method, the proposed prediction method has improved the prediction accuracy by almost 50%.