BATTERY STATE OF CHARGE ESTIMATION BASED ON DATA-DRIVEN MODELS WITH MOVING WINDOW FILTERS AND PHYSICS-BASED MODELS WITH EFFICIENT SOLID-PHASE DIFFUSION PDES SOLVED BY THE OPTIMIZED PROJECTION METHOD

Abstract

State of charge (SOC) estimation is one of the most important functions of battery management systems (BMSs), which is defined as the percentage of the remaining charge inside the battery to its maximum capacity. SOC indicates when the battery needs to be recharged. It is necessary for many battery management applications, for example, charge/discharge control, remaining useful time/ driving range predictions, and battery power capability estimations. Inaccurate SOC estimations can lead to user dissatisfaction, mission failures, and premature battery failures. This thesis focuses on the development of advanced battery models and algorithms for SOC estimations. Two SOC estimation approaches are investigated, including electrochemical models and data-driven models.

Electrochemical models have intrinsic advantages for SOC estimation since it can relate battery internal physical parameters, e.g. lithium concentrations, to SOC. However, the computational complexity of the electrochemical model is the major obstacle for its application in a real-time BMS. To address this problem, an efficient solution for the solid phase diffusion equations in the electrochemical model is developed based on projection with optimized basis functions. The developed method generates 20 times fewer equations compared with finite difference-based methods, without losing accuracy. The results also show that the developed method is three times more efficient compared with the conventional projection-based method. Then, a novel moving window filter (MWF) algorithm is developed to infer SOC based on the electrochemical model. MWF converges to true values nearly 15 times faster compared with unscented Kalman filter in experimental test cases.

This work also develops a data-driven SOC estimation approach. Traditional data-driven approaches, e.g. neural network, have generalization problems. For example, the model over-fits to training data and generate erroneous results in the testing data. This thesis investigates algorithms to improve the generalization capability of the data-driven model. An algorithm is developed to select optimal neural network structure and training data inputs. Then, a hybrid approach is developed by combining the neural network and MWF to provide stable SOC estimations. The results show that the SOC estimation error can be reduced from 8% to less 4% compared with the original neural network approach.