Analysis of battery capacity state estimations using big data and machine learning

Sammanfattning

The battery is a key component that strongly influences both the cost and environmental impact of a battery electric vehicle (those referred to as ’electric vehicles’ in everyday speech). The capacity of a battery is the amount of charge the battery can store and deliver. The state of health is a figure of merit for the battery’s capacity condition compared to the battery’s capacity condition at beginning of life. The state of health decays over time and usage due to chemical and mechanical degradation inside the battery. For battery electric vehicles, a battery is generally considered operational until the state of health decreases to around 70% - 80%. Below this threshold, the reduced driving range and performance make the battery less suitable for vehicle use. Remaining useful life describes the estimated time or number of cycles a battery can continue operating before reaching its end of life. In order to better estimate a battery’s remaining useful lifetime in a certain application and to maximize its utilization, it is necessary to understand which conditions affect the capacity’s state. Factors such as temperature, charging conditions, discharge rates, all influence battery degradation. By analyzing the battery cell capacity and battery cell conditions during charging, one can obtain knowledge about how the capacity is affected by the charging conditions, and learn how to prevent damaging conditions, in order to maximize remaining useful life. This thesis is an analysis of data generated by a capacity estimation algorithm measuring cell capacity of lithium ion battery cells installed in Battery electric vehicles under operation. However, in operating battery electric vehicles, the true cell capacity cannot be measured directly during normal use. Instead, the battery management system estimates capacity from sensor data. These onboard capacity estimates may therefore reflect both the actual battery degradation state and temporary influences from the measurement and estimation conditions. The thesis focuses on analyzing time and usage-detrended capacity samples and its relation to real world influencing features (charging conditions). To do so it uses field data from operating battery electric vehicles, and a Random Forest machine learning model. The results show that the state of charge at the start of sampling, temperature before sample, and relaxation after sample are important features when estimating the capacity’s state. Feature interaction show that the state of charge had a strong influence in interaction with other features, while temperature had a positive linear interaction pattern, meaning the model predicts higher capacity estimates under the higher temperature condition. Relaxation followed a positive logarithmic trend. Further feature interaction analysis indicated that SOC and temperature form the dominant interaction pair. These results may be used to improve onboard capacity estimation algorithms and the accuracy in the capacity estimation.