Estimation of Rolling Resistance of Heavy‐Duty Battery Electrical Vehicle with Physics Informed Neural Networks

Sammanfattning

Precise estimations of energy consumption are important for range prediction in Heavy-Duty Battery Electric Vehicles (HDBEVs), which enables more efficient route planning. One significant contributor to energy consumption is the rolling resistance, commonly modeled as proportional to the vehicle load through the Rolling Resistance Coefficient (RRC). In this thesis, the RRC is instead estimated dynamically using eight different Physics-Informed Neural Network (PINN) approaches. The models were trained on data from real-world trips and evaluated against two experimental benchmark setups from an earlier study, where the RRC is measured under controlled conditions. Starting from a baseline PINN from different prior work, the impact of incorporating tyre temperature and tyre temperature relations were investigated, along with the addition of a time-series architecture based on Long Short-Term Memory (LSTM) networks. Using tyre temperature as an output feature, together with additional relations in the loss function, did not improve the RRC estimation. In contrast, including tyre temperature as an input feature improved the results, achieving Mean Absolute Percentage Error (MAPE) values of 8.89 % and 12.3 % for the two benchmarks. Incorporating an LSTM architecture into the original method stabilized the RRC estimations but resulted in static outputs for the RRC. Combining tyre temperature as an input with an LSTM architecture further reduced the MAPE to 8.33 % and 26.9 %, but introduced irregular behavior during sudden velocity changes, which significantly reduced the correlation in one of the benchmarks.