Hybrid battery state estimation framework and advanced algorithm development

Lithium-ion batteries are playing a key role in shifting our society to a low-carbon future, they are being increasingly used in automotive and power industries to enable the intake of more renewable energy and to reduce greenhouse gas emissions. To ensure the safe, reliable, and efficient operation of batteries, a battery management system (BMS) is indispensable. Among all functionalities of the BMS, real-time battery state estimation provides important information for achieving high-fidelity and high-performance operations. This thesis focus on the development of novel techniques and frameworks to provide a coherent body of work on the estimation of battery states of interest, that is, state of charge (SOC) and state of health (SOH). The nonlinear variants of the Kalman filter (KF) framework have proven to be powerful and elegant solutions for the real-time state estimation of battery systems. As a state-space model is fundamental for using filtering algorithms and their estimation results highly depend on the model accuracy, an accurate battery model that can well capture the battery dynamics is established with suitable model structure and parameters. Then the nonlinear version of the KF, namely Extended Kalman Filter (EKF) and Unscented Kalman Filter (UKF), as well as the Particle Filter (PF) are used to estimate the battery SOC. As it is not a trivial task to obtain a precise model that can well describe the battery degradation trend, an intelligent estimation technique is proposed based on the convolutional neural network (CNN) to achieve fast and accurate online battery capacity estimation, which integrates health feature extraction, parameters identification, and capacity estimation into one framework. Taking into account the limitation of CNN on small degradation datasets, the transfer learning technique is incorporated into the CNN-based framework to improve the estimation performance on small datasets by taking advantage of the knowledge learned from large datasets. Further, in view of the limited computational capability of the current BMSs, a new network pruning technique is proposed to reduce the size and computation cost of the final model. Considering the intrinsic coupling relationship between SOC and SOH, a co-estimation framework is proposed to estimate the SOC and capacity simultaneously. Due to the estimators can be mutually optimized in the co-estimation framework, the estimation results of SOC and capacity are both more accurate than the separated estimation methods. Further, to acquire more informative measurements, the fiber optic sensors are attached to the cell surface and their measurements are utilized for battery SOC estimation to further improve the accuracy. Experimental data are collected from the lithium iron phosphate batteries to analyze and evaluate the efficacy of the methods and frameworks proposed in this thesis.