Neural Network-Based Modelling of the Thermal Dynamics for SBB Trains and their HVAC Systems

Abstract

Over the course of the last year, the Swiss Federal Railways displayed an annual energy consumption of 2300 Gigawatt-hours, with this number, as it stands, destined to grow since the increasing demand for public transportation in Switzerland. Between 15 to 20% of the power demand of the SBB fleet trains is due to Heating, Ventilation and Air Conditioning systems which is responsible for the well-being and comfort of the passengers and has to comply with strict norms. In order to achieve monetary and emission savings, one of the option is a more precise and smart deployment of the HVAC Systems: the implementation of a Model Predictive Control (MPC) techniques could lead to these goals. This study aims to investigate the feasibility of modelling the thermal dynamics of a wagon of the Regio-Dosto train through Neural Networks with the future possibility to use it within a MPC framework. Data from 8 trains during the Winter period and 5 trains during the Summer period are used in a simulation environment to train and test different Neural Networks architecture, taking into account the non-linearities affecting the thermal models of the wagons. In particular, a novel type of Physics informed Neural Networks (PiNNs), named Physically Consistent Neural Networks (PCNN), are used to simulate the complexity of this model using not only the amount of past data retrieved by SBB over the course of past months, but also the basic physics equations behind these processes. This novel architecture can lead to a convex input-output relationship that can be utilized in a future optimization step. Results show the effectiveness of this model architecture: different architecture typologies of the same kind can be exploited in order to make progress towards a potential implementation of MPC in the SBB fleets.