Neural Networks Battery Applications: A Review

Neural network battery applications have drawn tremendous attention. However, recent review papers fail to reflect the popularity of research activities in this area. In addition, neural networks like many other machine learning techniques are data dependent. One neural network architecture may have a much better performance than another architecture in terms of prediction accuracy for a specific application. Therefore, it is crucial to select the neural network architecture based on the available data and the purpose of the application. A review reporting the latest activities regarding neural network battery application is in demand. We selected representative publications from three popular areas: SOC estimation, SOH prediction, and parameter identification. We examined these publications from the aspects such as neural network architectures, inputs, outputs, data requirements, and cell chemistries. We also compared advantages and disadvantages among numerous neural network architectures. Three research trends were found in our study. First, the neural network architecture is getting more and more complex. The complexity comes from either structuring same or different neural networks together or combining the neural network with other techniques. Secondly, more research attention is moving toward the architectures that are suitable for time-dependent applications. For non-RNN architectures, averaged data is used as inputs. Averaging the data allows the past information to be learned by the non-RNN architectures. For RNNs, the past information is carried over by feeding back as either an input or a state. Last, although researchers always try to explore properties other than voltage, current, and temperature, the final selected inputs always have voltage, current and temperature. In addition, voltage is the most important one among these three inputs.