Data-Driven Model Optimization for Battery Electric Vehicles

Abstract

Digital models of real world trucks are beneficial for analysis of truck components, especially in regards to degradation and anomaly detection. This thesis’ objective is to develop methods for optimizing model performance. This was done by using realworld data from logged driving sessions as input to the modeled components, and comparing its output to the corresponding logged signals. One method that was developed optimizes the accuracy of lookup tables. The method consisted of the gradient descend method ADAM combined with Gaussian kernel smoothing. It was developed and tested on one lookup table and its corresponding signal, showing an MAE reduction to 34-38% of the original MAE. The updated values in the table had a generally good shape but lack of data in the extremes led to some inconsistencies. The other method focused on using linear regression to fit a simulated signal to its corresponding logged signal, and preferably overestimating the logged signal. This was done by using a weighted penalty function where the weight was decided by a custom-built algorithm. The linear regression resulted in small changes in MAE, but the maximum error decreased for all vehicles while the fraction of the time the simulated signal overestimated the logged signal increased across all vehicles.