Modeling of a solid oxide fuel cell power plant using an ensemble of neural networks based on a combination of the adaptive particle swarm optimization and Levenberg–Marquardt algorithms

In this paper, a dynamic nonlinear model of a Solid Oxide Fuel Cell (SOFC) is determined using an ensemble of Neural Networks (NNs). The Component NNs (CNNs) of the ensemble are trained using the Levenberg–Marquardt (LM) algorithm. The ensemble of NNs is optimized using the hybrid adaptive Particle Swarm Optimization (PSO) with a newly developed Negative Correlation Learning (NCL) method. In this work, PSO has been used to pre-train CNNs before the LM algorithm is employed. In this way, we will get rid of the shortcomings of both of these two methods; PSO has a global vision but it cannot find the global minimum precisely while LM can generally find this minimum but needs a good starting point near the global minimum. An ensemble of NNs shows greater generalization performance compared to a single neural network. In NCL, all the CNNs in the ensemble are trained simultaneously and interactively using the correlation penalty terms in their error functions. Instead of constructing unbiased individual networks whose errors are uncorrelated, NCL can generate negatively correlated networks to boost specialization and cooperation among the individual networks. NNs are excellent tools to model complex nonlinear dynamic systems and can be useful for simulation of SOFC. The results indicated that the SOFC outputs can be accurately estimated using the trained model in this work.