A double-gated recurrent unit neural network for contact force prediction in pantograph-catenary systems

The numerical simulation of pantograph-catenary systems is of paramount importance and normally conducted using finite element analysis (FEA), which are computationally intensive due to the complexity of temporal-spatial coupling nature. To address this limitation and facilitate faster simulation, this paper introduces a surrogate modeling approach based on a time-series neural network to predict the contact forces of different pantograph-catenary systems. The proposed machine learning model uses FEA simulated results as the training dataset, which are labelled appropriately. Specifically, a double-gated recurrent unit (D-GRU) neural network is constructed by extending the traditional neural network structure, enabling it to effectively process non-sequential input data and accurately predict the system’s sequential physical response. This design leverages the gated recurrent unit’s (GRU) efficiency in handling sequential data, enhancing the model’s performance. In comparison with the best-performing GRU models from other networks, D-GRU demonstrates a 2.2% increase in R 2 , along with 3.9% and 3.0% decreases in R M S E and M A E , respectively. The paper concludes with an evaluation of the surrogate model’s performance, focusing on its prediction accuracy and potential for integration into pantograph-catenary optimization design workflows.