An Accurate Practical Technique for Real-Time State-of-Charge Estimation of Li-Ion Batteries Using Neural Networks
Lithium-ion (Li-ion) batteries have attracted significant attention in terms of technical features, but to use them within their specific operating region and prevent undue degradation, it is crucial to constantly monitor their state of charge (SOC). However, most available techniques are either too complex, too computationally demanding, or require a large measurement window. This paper proposes a general framework for the data-driven SOC estimation and demonstrates its effectiveness via a very simple Neural Network (NN) classifier. Accordingly, a nonlinear input-output neural network (NIONN) is developed based on the proposed network and then further optimized. Additionally, we present an input engineering process using available data to provide a good accuracy while considering the practical aspects such as complexity as well as the update speed for a low-cost online application. The numerical results from the measurements validate the performance of the proposed general framework. Per extensive optimizations, it is proven that a shallow network structure with 13 neurons in the first hidden layer and 12 neurons in the second layer can achieve <98.3% accuracy and a mean squared error of <10-5 which is on par with more complex networks with higher neurons and/or hidden layers. Additionally, the optimum observation window was determined to be 5 seconds.
