Performance Analysis and Implementation of LSTM and GRU based on Synthetic Data

The goal of this work is to develop artificial dataset through the use of autoencoders, a solution to the problem of expensive and scarce real-world data for predictive model training in the field of battery health monitoring. An autoencoder neural network is used to generate synthetic datasets that capture the complex patterns and features of battery behavior by utilizing a large dataset with 81,000 records. The scarcity of battery measurement data poses a challenge for battery researchers, as it can hinder the development of reliable estimation models. This paper addresses this issue by proposing machine learning-based techniques for the synthesis of high-fidelity battery datasets. The generated synthetic datasets can be employed to augment the existing datasets, thereby enhancing the effectiveness of the research endeavors. The presented work demonstrates that synthetic dataset can be used as the replica or as a substitute of the work done with larger dataset. The strategies presented in this paper offer a practical solution to overcome the limitations posed by small battery datasets. Moreover, a comparison has been provided between the ML method proving that LSTM emerges to be a better method based on the performance metrics with one drawback of having longer training time than GRU.