The Impact of Training Algorithms and Transfer Functions on the Accuracy of Neural Network-Based Temperature Prediction of 3S4P Battery Module

The accurate prediction of lithium-ion battery temperature is of utmost importance for efficient battery thermal management systems. Artificial neural network (ANN) offers a powerful tool to assess battery module behavior under varying operating conditions. In this study, experimental data is generated by charging a 3S4P battery module with 0.5C-rates at ambient temperatures of 30 ℃ and 35 ℃. This data is then utilized to train, validate, and test 54 ANN models with different transfer functions and training algorithms. The input parameters to the ANN models include State of Charge (SOC), current, ambient temperature, and voltage, while the battery module’s maximum temperature serves as the output parameter. Our findings reveal that the Feed Forward Back Propagation ANN structure, employing Levenberg–Marquardt (LM) as the training algorithm, with 40 neurons in the first hidden layer using logistic sigmoid (LOG) function and 1 neuron in the outer layer using the linear function, yields the lowest mean absolute relative deviation (MARD) value (0.027002%) and the highest regression coefficient (R2 = 0.99998). The ANN-LM structure with LOG transfer function demonstrates accuracy in predicting experimentally collected data for battery module under charging and varying ambient temperature.