A Study On Power Battery SOC Estimation Considering Data Uncertainty

With high energy density and power density, lithium-ion batteries have been widely used in electric vehicles (EV). Battery management system (EMS) is essential for electric vehicles,related to the efficient use of energy and the safe operation of the battery, which is dependent on the accurate estimation of battery state of charge (SOC). However, SOC cannot be measured directly, traditional estimation methods suffer the disadvantages of parameter selection difficulty,low accuracy and poor versatility. As the representative of data-driven modeling methods, machine learning technologies make great progress in recent years. Such methods have excellent nonlinear approximation ability and portability. The large amount of state data generated by electric vehicles provide favorable condition for data-driven SOC estimation methods. However, the data-driven modeling methods highly depend on the data reliability. Due to the limited precision of on-board data acquisition equipments and complex working environment, the noise and outliers are usually unavoidable. And because of the"black box" assumption, it is difficult to analyze the cause of the error. We try to solve these problems from two aspects, first, we try to reduce the negative effects of unreliable data, on the other hand, the estimation posterior probability is obtained to help avoid the risk of BMS system malfunction. The main work of this paper can be summarized as follows:1. Least squares support vector machine (LSSVM) has excellent generalization performance, a SOC estimation method based on LSSVM is proposed. Particle swarm optimization method is used to optimize this model. Experiments are organized based on the lithium iron phosphate (LFP) batteries, the performance of different kernel functions are compared and the effectiveness of the proposed PSO-LSSVM SOC estimation method is verified through the experiments.2. To reduce the adverse impact of data noise on the SOC estimation model, a weighted support vector machine regression method is proposed. The reliability of this approach is improved by reducing the weight of samples with high noise level. The noise distribution for the different data attributes is relatively independent, a T-S fuzzy based weight function is proposed to comprehensively evaluate the influence of the data noise for individual feature attributes. In addition,a nonlinear correlation measure is proposed to identify the data contribution. The robustness of the proposed methods to the data noise is verified by using Toyota COMS EV test platform.3. Aiming at the heteroskedasticity and non-stationary characteristics of SOC regression problem, a flexible Gaussian mixture regression (GMR) SOC estimation method is proposed.Gaussian mixture model (GMM) is used to capsule the information of original dataset.Gaussian process regression (GPR) is used to estimate the result for each Gaussian component respectively. In the end, the estimation result of the subsystem is summed. An evolutionary expectation maximization (EEM) is proposed to optimize the GMM model. A Pearson nonlinear correlation feature selection algorithm is proposed to reduce the data dimension and avoid overfitting.4. We study the effect of outliers on mechine leaning methods such as support vector machines, neural networks and Gaussian process regression. Weighted Gaussian process regression (WGPR) method is proposed to deal with the adverse effects of outliers. In addition,a weighted density based local outlier factor is proposed to identify the outlilers. This method is based on the local density outlier detection, and the weighted Euclidean distance is used to improve the model performance for high dimensional dataset. The validity of the proposed methods is validated in case studies of SOC estimation and short-term photovoltaic power forecasting.