Multi-Reservoir Echo State Network for Proton Exchange Membrane Fuel Cell Remaining Useful Life prediction

Abstract

In this paper, a Multi-Reservoir Echo State Network is used to estimate the Fuel Cell degradation, and its remaining useful lifespan. It proposes a methodology for predicting the fuel cell output voltage evolution with time. Echo State Network is a powerful Artificial Intelligence tool for time series predicting which main characteristics is the use of a reservoir of neurons, randomly created, instead of hidden layers such as for Artificial Neural Networks. Only the output layer is optimized by a multilinear regression, resulting in a time reduced training phase. This leads to a possible increase of the reservoir size to preserve, even improve, its accuracy. However, the bottleneck linked to the use of this tool lies in its architecture optimization. This paper proposes a way to overcome the echo state network parameters optimization process by using a Multi-Reservoir Echo State Network. Then a comparison between an Echo State Network optimized algorithm and a Multi-Reservoir Echo State Network for fuel cell RUL prediction is proposed. In order to have a good prediction of the FC lifetime, an innovative approach based on the Multi-Reservoir Echo State Network is developed and validated using experimental data.

An Echo State Network for fuel cell lifetime prediction under a dynamic micro-cogeneration load profile

“…The use of an optimized Echo State Network for PEMFC prognostics under a constant load gives good results with low average errors below 5% for voltage prediction. This method has shown its effectiveness under constant load in a previous work [71]. However, the errors are higher when performing prognostics under a variable load.…”

Scalability, Explainability and Performance of Data-Driven Algorithms in Predicting the Remaining Useful Life: A Comprehensive Review

“…Mezzi et al proposed a multi-reservoir ESN (MR-ESN) to solve the parameter optimization problem in RUL, as this usually involves extensive work that negates the fast training speeds of ESNs. The MR-ESN is trained by feeding the training data into the multiple reservoirs, and the optimized output weights are based on the output weights of all the reservoirs [205], which allows for a problem-independent setup.…”

Scalability, Explainability and Performance of Data-Driven Algorithms in Predicting the Remaining Useful Life: A Comprehensive Review

“…LED Degradation [28] Li-ion Battery Aging [178] [42, 66, 67, 111, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 200] Mikron Milling Machine [201] Proton Exchange Membrane Fuel Cell [202,203,204,205,206] Planetary Gearbox Experiment Platform [108,207] Rail Fastener of Subway Operation Line [208] Steam turbines of Power Generator [60] Switched-Mode Power Supply [209] Super Capacitors [124,210] Transformer Deterioration [30,74,211] UH-60 Helicopter Planetary Gear Plate [212] Vehicle [213] Vertical Roller Mill [214] Wind Turbines [215,216,217] This section evaluates and summarizes the state-of-the-art ML algorithms by dividing them into five categories: neural network-based algorithms, generative models, kernel-based learning, probabilistic models, and fuzzy logic-based methods. Table 1 lists the datasets used by the articles evaluated in the present manuscript.…”