Voltage prognosis of PEMFC estimated using Multi-Reservoir Bidirectional Echo State Network

Authors:

Abstract
One of the key points in the development of fuel cell technologies is to increase their lifetime. For this, it is necessary to be able to detect a degraded state in order to correct it, but also to be able to estimate the future state, which can facilitate predictive tasks such as control and maintenance. Echo State Networks are good candidates to estimate the future performance of proton exchange membrane fuel cells. They are part of the reservoir computing family and process time series using a neural reservoir. This paper proposes a method to predict the future voltage of a low-temperature proton exchange membrane operating in a steady state. For this purpose, several bidirectional reservoirs are used in parallel which improves the captured dynamics but also reduces the reservoir optimization effort which is currently one of the limitations in the development of Echo State Networks. The forecasting performance is quantified in three different ways: First, to simulate the temporal evolution of a system, the network is trained by varying the size of the training base. This allowed observing that the network can correctly capture the degradation dynamics while having an adaptation capacity when new data are added. Secondly, a comparison with a unidirectional ESN shows that bidirectional reservoirs will more easily capture non-periodic recovery and noise phenomena. Lastly, a comparison with 3 stacked bidirectional long-term memories shows that the proposed approach provided similar results while facilitating the optimization due to its low number of parameters to be set.
Funding Information
  • Fuel Cells and Hydrogen 2 Joint Undertaking (875047)