Research On Estimation For Lithium Battery Remaining Capacity Of Electric Vehicles

In the face of bad climate change, electric vehicles (EVs) will become the main direction of future development. But the battery is the important factor to EVs. In order to guarantee security and stable operation of EVs, it is necessary to control and manage the battery to improve its lifetime. So it is important to best use battery.However, battery remaining capacity is the main subject in managing the battery, which directly reflects the state of the batteries. From this, we could calculate the driving range of EVs, avoid over-charge or over-discharge, and the inconsistency between performances for guarantee the stable running of EVs. So it is quite important and significance to estimate the battery remaining capacity online in real-time.Due to complicated nonlinear relations between battery remaining capacity and kinds of factors, it's difficult to create accurate mathematical model. So the thesis uses Radial Basis Function Neural Network (RBFNN) to estimate battery remaining capacity. Firstly, the battery data exported to tables by the remote management system of the battery. This thesis designs the algorithm of data interpolation and processing in order to obtain experiment data. Based on the data, the thesis analyzes the charge-discharge principles of battery. The model of RBFNN to battery remaining capacity is established with relative factors as input variable and battery remaining capacity as output variable, and is trained by K-mean clustering algorithm on Matlab Platform.In order to apply battery remaining capacity estimation methods into the embedded system with lower costs, this thesis collects the input variable having much effect on battery remaining capacity to simplify the structure of the Neural Network and retraining for network through Neural Network effect analysis method to get prediction model of battery remaining capacity.In the end, this thesis research genetic algorithm to improve RBFNN algorithm for solve dependence of K-mean clustering algorithm to initial value, and global optimization search for the center of RBF in hidden layer and width in order to obtain the optimal resolution. Experiment results indicate that this method is not sensitive for choice of initial value, and is much better than K-mean clustering algorithm.