Improve Vehicle Efficiency Accuracy Through Virtual Sensors - A Comparative Study of MLP, LSTM, and Transformer Architectures with and without Physics-Informed Constraints for Fuel Consumption Prediction

Abstract

In modern heavy-duty vehicles, measuring fuel consumption relies on advanced flow meters, which are expensive and challenging to install. Volvo Group currently employs an empirical calculation model based on predefined coefficients, but this thesis explores the potential of replacing the flow meter with machine learning models. The primary objective is to develop predictive models that outperform the existing baseline. To this end, Multi-Layer Perceptron (MLP), Long Short-Term Memory (LSTM), and Transformer architectures are evaluated to determine the most suitable model for this scenario, while the impact of incorporating Physics-informed Neural Networks (PINNs) on prediction performance is also investigated. Using historical field test data, six models were trained and evaluated. All trained models demonstrated superior performance compared to the baseline. The outcomes of this thesis work have the potential to be embedded in the Electronic Control Unit (ECU) system for future field tests, providing a practical and cost-effective alternative to the physical flow meter.