Development of a Neural Network for Electric Vehicle Cabin Modeling towards advanced HVAC Management Strategies

The energy transition is a key challenge and opportunity for the transport sector. In this context, the adoption of electric vehicles (EVs) is emerging as a key solution to reduce environmental impact and mitigate problems related to traditional energy sources. One of the biggest problems related to electric mobility is the limited driving range it offers compared to the time needed for recharging, leading to what’s commonly known as “range anxiety” among users.Significant part of the energy consumption of an electric vehicle is represented by the management of the HVAC system, which aim is to ensure the achievement and maintenance of thermal comfort conditions for the occupants of the vehicle. Currently the HVAC control logics are based on the pursuing of specific cabin setpoint temperature, which does not always guarantee the thermal comfort; more advanced human-based control logics allow to attain the thermal comfort in a zone around the subjects, as known as “heat bubble”, rather than acclimatizing the entire cabin, increasing the system efficiency and often reducing the thermal demand. It is therefore useful to develop a dynamic model that predicts and monitors the evolution of comfort parameters during the vehicle usage.This study proposes to develop a simplified thermal model of the cabin system of a light duty commercial vehicle based on experimental data and numerical simulations, which is able to locally estimate the parameters of thermo-hygrometric comfort, and therefore allows a targeted management of the HVAC system with consequent energy optimization. First the cabin of a commercial BEV has been acquired and processed through reverse engineering techniques (3D scanning) in order to create the 3D CAD model; consequently, a CFD analysis based digital twin has been developed and validated with experimental data in different temperature conditions. Then the cabin system has been modeled with a neural network trained with results of CFD simulations, in order to replicate temperature behavior in the areas of interest.The purpose of this modelling is to provide a starting point for the development of a reduced order model (ROM) that can be the basis of the development of advanced control logics to be integrated into the vehicle’s on-board computer system. Results show a good agreement between the CFD and simplified model (normalized Root Mean Square Error always below 0.29) and fast execution time (0.7 s on an 8 cores Intel i7 - 9700 processor) confirming the suitability of the approach for the proposed application.