Unified Univariate-Neural Network Models for Lithium-Ion Battery State-of-Charge Forecasting Using Minimized Akaike Information Criterion Algorithm

Lithium-Ion batteries require step-ahead information to apply contingency plans prevent them from operating beyond their safe operation thresholds in grid storage and electric vehicle applications. Recently, machine learning techniques have been increasingly applied forecast one such battery metric, State-of-Charge % ( SOC ). Conventional standalone recent works suffer an accuracy standpoint thus replaced by high-fidelity hybrid techniques. Existing on either perform in-sample predictions, provide limited of the underlying model, or do not consider varying charging-discharging rate (C-rate) dynamics battery. To address this issue, article presents unified forecasting models using a Minimized Akaike Information Criterion ${m}$ -AIC) algorithm. Initially, -AIC algorithm is used manner precisely tune search ARIMA (Autoregressive Integrated Moving Average) terms’ order automatically accurately battery’s current, voltage, parameters univariate at lower C-rate given C-rates. The based models’ xmlns:xlink="http://www.w3.org/1999/xlink">m-AIC ) further enhanced first modeling unifying with Multilayer Perceptron (MLP) then Nonlinear Autoregressive Neural Network external input (NARX) neural networks respectively previously forecasted (from This results unified-MLP xmlns:xlink="http://www.w3.org/1999/xlink">u- MLP) unified-NARX NARX) which higher out-of-sample for different optimizer variants. Results show that proposed MLP NARX reduce mean squared error 0.1048% 0.0175% comparison counterparts lowest corresponding values 0.271% 0.0236% respectively. Furthermore, additional reference Holt-Winters Exponential Smoothing (HWES), analyzed (by replacing comparative performance evaluation both manners. Recommendations usage preferred respective are also presented epochs values.