Interpretable Deep Learning Approach for Long Horizon Health Prognosis of a Li-ion Battery

The safety of a wide range of devices depends on the accurate long-term estimation of the State-of-health (SoH) of Lithium-ion battery. However, the existing techniques have several limitations in terms of prediction accuracy, computational complexity and applicability. To mitigate these drawbacks, a novel Deep learning-based method with low computational complexity has been proposed that can predict battery’s future performance upto 100 cycles, thereby providing an early warning of battery failure. A unique feature called Interval of time for spectrum of same charging voltages (ITSSCV) along with maximum and minimum voltages for each cycle are used to predict the SoH of the battery, which greatly reduces the size of the training data and the storage requirements. A Long Short Term Memory (LSTM) neural network has been utilized to predict the SoH with 0.005 RMSE. Further, the proposed solution can perform the long horizon SoH estimation upto the next 100 cycles with an RMSE of 0.0029. Moreover, Explainable AI is applied to the model to interpret the prediction outcome and improve trustworthiness. The model can quickly and accurately predict the long horizon SoH of battery using low-end computation resources, thereby proving its usefulness in real-life critical battery-operated equipment.

I. Introduction

Due to the high energy density, high efficiency, and decreasing manufacturing costs, Energy Storage Systems based on lithium-ion cell technologies, Lithium-ion batteries are quickly taking over a wide range of storage uses. Despite their robustness, the motors may malfunctioning due to harsh working conditions, high temperature, high humidity, and overloading, leading to unexpected downtime [1], [2]. Eventually, LIB’s efficiency would decline because of ageing and operating environment factors. The unexpected failure may result in emergent maintenance and unexpected system shutdown, which can be catastrophic [3]. To guarantee a satisfactory performance, it is critical to precisely determine the state of health (SoH) of LIBs.