Comparison of Neural Network Topologies for Sensor Virtualisation in BEV Thermal Management Loughborough University

Format:

Book

Conference/Event

Author/Creator:

Wray, Alex, author.

Contributor:

Dutta, Nilabza

Ebrahimi, Kambiz

Pipes, Harry

Conference Name:

WCX SAE World Congress Experience (2024-04-16 : Detroit, Michigan, United States)

Language:

English

Physical Description:

1 online resource cm

Place of Publication:

Warrendale, PA SAE International 2024

Summary:

Energy management of battery electric vehicle (BEV) is a very important and complex multi-system optimisation problem. The thermal energy management of a BEV plays a crucial role in consistent efficiency and performance of vehicle in all weather conditions. But in order to manage the thermal management, it requires a significant number of temperature sensors throughout the car including high voltage batteries, thus increasing the cost, complexity and weight of the car. Virtual sensors can replace physical sensors with a data-driven, physical relation-driven or machine learning-based prediction approach. This paper presents a framework for the development of a neural network virtual sensor using a thermal system hardware-in-the-loop test rig as the target system. The various neural network topologies, including RNN, LSTM, GRU, and CNN, are evaluated to determine the most effective approach. The solution proposed intends to use a combination of the states determined in other sensors and the control inputs made into the system to predict the state of the sensor to be virtualised, with the aim of an average accuracy of 95% and a worst-case accuracy of 80%. Also discussed are the potential methods of nonlinear system identification that can be used to achieve these goals, concluding through a literature review that a Neural Network solution is the most probable method to produce an accurate result. Based on this an analysis is performed of the challenges of neural network development, from collecting and processing data, to actually training the neural network and evaluating the performance outcome. Establishing that depending on the quality and quantity of data collection a range of methods that could be implemented

Notes:

Vendor supplied data

Publisher Number:

2024-01-2005

Access Restriction:

Restricted for use by site license