UCLA

This work focuses on developing electric-vehicle battery models that can precisely predict voltage from measurable properties with limited errors using feed-forward neural network models of the backpropagation algorithm. Recently, the neural network has been utilized in a variety of different predictions such as the state of charge prediction or the state of health prediction. Also, electric vehicles like the model X, model 3, and model Y from Tesla have been widespread from 2015 until today. Our model for electric-vehicle battery voltage prediction achieves 25 times reduction in the maximum voltage error and 273 times reduction in the average voltage error comparing to the existing models from Contemporary Amperex Technology (CATL). This is accomplished by using the neural network models in comparison to the equivalent circuit model, which is a way to describe working conditions in a circuit by using the mathematical method, for the lithium-ion battery. Advantages of using a battery model to run the test instead of installing a pack in a vehicle are that our model can reach the tolerant error range. This allows automakers to use our model to design cars at an initial stage and provide guidance to choose the particular specification of battery packs to run the vehicle performance test without much cost.