SDG-L: A Semiparametric Deep Gaussian Process based Framework for Battery Capacity Prediction

TL;DR

SDG-L tackles battery capacity prediction by leveraging per-cycle state information through an LSTM-based feature extractor and a semiparametric deep Gaussian process (DGPR) that incorporates an explicit degradation prior via an EMF. The framework trains the LSTM and DGPR jointly, modeling residual capacity trends after fitting the EMF, and achieves superior accuracy on the NASA Ames battery dataset (average test MSE around 1.2%), outperforming traditional baselines. Ablation studies confirm that the EMF prior, DGPR expressivity, and LSTM-derived features all contribute to the gains, highlighting the value of combining parametric degradation knowledge with nonparametric deep GP modeling. This approach enhances predictive performance and uncertainty quantification for battery health management, supporting more efficient and reliable energy storage systems.

Abstract

Lithium-ion batteries are becoming increasingly omnipresent in energy supply. However, the durability of energy storage using lithium-ion batteries is threatened by their dropping capacity with the growing number of charging/discharging cycles. An accurate capacity prediction is the key to ensure system efficiency and reliability, where the exploitation of battery state information in each cycle has been largely undervalued. In this paper, we propose a semiparametric deep Gaussian process regression framework named SDG-L to give predictions based on the modeling of time series battery state data. By introducing an LSTM feature extractor, the SDG-L is specially designed to better utilize the auxiliary profiling information during charging/discharging process. In experimental studies based on NASA dataset, our proposed method obtains an average test MSE error of 1.2%. We also show that SDG-L achieves better performance compared to existing works and validate the framework using ablation studies.

Figures (3)

• Figure 1: Method Framework. Part b is the pipeline of our method and the detail of our framework components are in part a. In our framework, X data (input exogenous information) and Y data (battery capacity) are respectively processed to features (F) and residuals (R) before being fed into the DGPR module. X is put through an LSTM feature extractor (illustrated in part a) to get features (F), while residuals (R) are derived from the EMF curve fitting on Y. Predictions can be calculated by adding the DGPR output and EMF results for the test set using the fitted curve parameters. Specifically, the time series of voltage, current, and temperature observed at each cycle is embedded into a $20\times3$ matrix and fed to an LSTM extractor, whose output feature dimension is set to be 2. The LSTM extractor and the DGPR module will be trained simultaneously.
• Figure 2: Visualizations of Estimated Value, Ground Truth, and CI. Orange curves are plotted for the ground truth. Green, red, and purple curves for model prediction of different methods. A 2-sigma CI is marked by translucent regions. The rows from top to bottom (a-c) are the prediction results of GPR, SDG-L w/o curve fitting, and SDG-L; the columns from left to right show LIB cells with ID B0005, B0006, B0007, and B0018. The training number for cells B0005, B0006, and B0007 is set as 125 and B0018 as 110, marked by the vertical dashed line.
• Figure 3: Feature Series Extracted from LSTM Network of Different LIB Cells. (Upper Left: B0018, Upper Right: B0005, Lower Left: B0006, Lower Right: B0007) We have $x$,$y$ axis for the two-dimensional features and $z$ axis for cycle numbers.