An on-Line Electrochemical Parameter Estimation Study on Lithium-Ion Batteries Using Neural Network (NN)

Several studies have been devoted to the modeling of electrochemical lithium-ion (Li-ion) batteries. The success of all these models relies, among other things, on the precise knowledge of the electrochemical properties of the battery. Direct measurement of these properties is, however, a tedious task. It typically requires the dismantling of the battery. Also, the measured properties are dependent on the battery’s age and may vary according to the measurement technique. To overcome the difficulties of estimating the battery properties, inverse methods are proposed. These methods are based on optimization algorithms that aim at minimizing the discrepency between the predictions of a direct model (including estimated parameters) and experimental data. The major drawback of inverse methods is that they are computationally demanding. As a result, they cannot be retained for on-line control, monitoring or Battery Management Systems (BMSs). In this paper, for the first time, an inverse method resting on a trained neural network is presented for the on-line estimation of the following five electrochemical properties of a Li-ion battery: The diffusion coefficients of the electrodes (Dn & Dp), the intercalation/deintercalation reaction-rate constant for both electrodes (kn & kp) and the electrolyte resistance (Rcell). The black box model employs the 1C discharge curve of a Li-ion battery with a LiCoO2cathode material. Due to the complexity of Li-ion batteries, four different layers were chosen for the neural network: one input layer, two hidden layers and one output layer. The 1C discharge curve is fed to the model via a time domain matrix for the input layer. In order to reduce the training error, two different hidden layers comprising 50 and 75 neurons were employed. The output layer was composed of five signals which characterize the electrochemical properties. The data needed for the training of the neural network was generated with an improved Single Particle Model (SPM). The 1C discharge curves were first calculated for a reasonable range of the expected parameters. These data were then used to determine the values of the weights and functions of the neural network. The process was next implemented into two steps. First, the optimum number of training iterations was calculated. Seventy percent (70 %) of the generated data was randomly employed to train the network. The remaining data (30%) was used to test the performance of the network. Second, once the optimum value of the iteration number had been determined, the model was trained again by introducing the available data for the optimum iteration number. Finally, the trained neural network was used to estimate the electrochemical properties for different 1C discharge curves. The model was successfully validated with experimental data. Moreover, due to the matrix structure of the neural network, the parameter estimation process is adaptable and easy to implement. As a result, the proposed neural network model is suitable for real-time control and monitoring applications. Acknowledgements: The financial supports of Natural Sciences and Engineering Research Council of Canada (NSERC) and Hydro-Quebec are gratefully acknowledged. Figure 1

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.