Optimized ANN for LiFePO4 battery charge estimation using principal components based feature generation

Electric vehicles (EVs) have gained prominence in the present energy transition scenario. Widespread adoption of EVs necessitates an accurate state of charge estimation (SoC) algorithm. Integrating predictive SoC estimations with smart charging strategies not only optimizes charging efficiency and grid reliability but also extends battery lifespan while continuously enhancing the accuracy of SoC predictions, marking a crucial milestone in sustainable electric vehicle technology. In this research study, machine learning methods, particularly Artificial Neural Networks (ANN), are employed for SoC estimation of LiFePO4 batteries, resulting in efficient and accurate estimation algorithms. The investigation first focuses on developing a custom-designed battery pack with 12V, 4Ah capacity with a facility for real-time data collection through a dedicated hardware setup. The voltage, current and open-circuit voltage of the battery are monitored with computerized battery analyzer. The battery temperature is sensed with a DHT22 temperature sensor interfaced with Raspberry Pi. Principal components are derived for the collected battery data set and analyzed for feature engineering. Three principal components were generated as input parameters for the developed ANN. Early Stopping for the ANN was also implemented to achieve faster convergence of the ANN. While considering eleven combinations for ten different optimizers loss function is minimized. Comparative analysis of hyperparameter tuning and optimizer selection revealed that the Adafactor optimizer with specific settings produced the best results with an RMSE value of 0.4083 and an R2 Score of 0.9998. The proposed algorithm was also implemented for two different types of datasets, a UDDS drive cycle and a standard cell-level dataset. The results obtained were in line with the results obtained with the ANN model developed based on the data collected from the developed experimental setup.