Method for predicting SOC of power battery by using extreme learning machine under manifold regularization framework

Abstract

The invention discloses a method for predicting an SOC of a power battery by using an extreme learning machine under a manifold regularization framework, and belongs to the technical field of batterymanagement systems. The method comprises the following steps: collecting laboratory sample data; establishing a model to obtain sample data normalization; using a z-score normalization method, and selecting a Gaussian kernel function to calculate the similarity between each point and xi; reflecting the manifold of a sample space, and setting a relevant parameter and a function form according toan extreme learning machine prediction SOC model; and under the manifold regularization framework, creating an extreme learning machine SOC prediction model, and using a differential evolution methodto optimize regularization parameters. According to the method for predicting the SOC of the power battery by using the extreme learning machine under the manifold regularization framework provided bythe invention, a manifold regularization theory is introduced to optimize the extreme learning machine, thus the generalization performance of the extreme learning machine is improved; a differentialevolution optimization algorithm is introduced to improve the prediction performance of the extreme learning machine; and a power battery SOC prediction model established by using the extreme learning machine can improve the prediction accuracy, the prediction efficiency and the prediction stability.

Description

technical field [0001] The invention relates to a method for predicting the SOC of a power battery, in particular to a method for predicting the SOC of a power battery by an extreme learning machine under a manifold regularization framework, and belongs to the technical field of battery management systems. Background technique [0002] Predicting the SOC of the car battery is one of the core functions of the power battery management system. It mainly estimates the cruising range based on the battery SOC, in case the car breaks down during driving or the battery itself is damaged due to excessive discharge of the battery and causes danger. Therefore, how to accurately predict the SOC of the vehicle battery has become one of the key issues of the vehicle battery management system. [0003] At present, the estimation methods of SOC have their own advantages, but there are also certain application limitations. The SOC prediction methods of vehicle power batteries mainly include ...

AI Technical Summary

Problems solved by technology

[0003] At present, the estimation methods of SOC have their own advantages, but there are also certain application limitations. The SOC prediction methods of vehicle power batteries mainly include discharge experiment method, ampere-hour measurement method, open circuit voltage method, internal resistance method, Kalman filter method and neural network method. Among them, the discharge experiment and the open circuit voltage method have reliable measurement data, but it is not easy to achieve real-time measurement and it is difficult to directly apply to the vehicle; the measurement error of the ampere-hour measurement method is cumulative, and it is more affected by high temperature. It is difficult to guarantee the prediction accuracy; the open circuit voltage method requires the battery to stand still for a long time, which is only suitable for electric vehicles in the parking state; the internal resistance method can better reflect the SOC of the battery, but the internal resistance of the battery is very small, and the measurement equipment is inconvenient And the cost is high; the Kalman filter method can be applied to batteries of various types and at different aging stages, but the amount of calculation is too large, and it is difficult to accurately express the nonlinear factors in the battery during the modeling process. In the neural network method, the weights, There is no complete theoretical guidance for the selection of the threshold, so the training of the samples may fall into a local optimum

Images