Automatic Parameter Scheduling of Equivalent Circuit Battery Models Using Local Linear Model Trees and Amplitude-Modulated Pseudo-Random Excitation Signals

<div class="section abstract"><div class="htmlview paragraph">The automotive industry is moving rapidly to electrification through development of Battery Electric Vehicles (BEV). Development and sizing of the battery and powertrain requires a detailed understanding of battery cell behavior under different conditions. Achieving this is difficult due to the range of cells available and the large range of condition variables of each cell. Equivalent-circuit models are used for BEV development. However, conventional battery cell characterization testing to parameterize these models are time and resource intensive. Characterization can be performed using well-known techniques such as Hybrid Pulse Power Characterization (HPPC) or Galvanostatic Intermittent Titration Technique (GITT) which are used to optimize parameters of an ESC model pertinent to the dynamics of its voltage response. However, the discrete State-of-Charge (SoC) intervals and demand current amplitudes of these experiments are not optimized for a balance of time and model effectiveness. There is scope to develop methods that can excite a range of current amplitudes and SoC points in a shorter timeframe whilst revealing non-linearities in the system’s time response. In this work an excitation signal design is presented that aims to maximize the amount of information gained about the dynamics of the battery across the SOC range within a short timeframe. An automated means of parameter-scheduling an ESC model is also introduced to best utilise the data from such a randomized, unstructured experiment. The excitation signal design takes the form of pulse signals with pseudo-randomly generated amplitudes and duration, constrained to ensure a specified test duration. This data is used to characterize a variant of the ESC model that automatically schedules parameters of the model via self-organizing Locally Linear Model Trees (LoLiMoT). The resulting models have strong predictive capability even in the extreme low SOC condition, resulting in an 18% reduction in Mean Absolute Error when compared to an ESC model without parameter scheduling over a validation cycle. This constitutes a strong step toward rapid, robust battery modelling processes for the purpose of cell selection and simulation for powertrain/vehicle design.</div></div>