BatteryLife: A Comprehensive Dataset and Benchmark for Battery Life Prediction (2502.18807v6)

Abstract: Battery Life Prediction (BLP), which relies on time series data produced by battery degradation tests, is crucial for battery utilization, optimization, and production. Despite impressive advancements, this research area faces three key challenges. Firstly, the limited size of existing datasets impedes insights into modern battery life data. Secondly, most datasets are restricted to small-capacity lithium-ion batteries tested under a narrow range of diversity in labs, raising concerns about the generalizability of findings. Thirdly, inconsistent and limited benchmarks across studies obscure the effectiveness of baselines and leave it unclear if models popular in other time series fields are effective for BLP. To address these challenges, we propose BatteryLife, a comprehensive dataset and benchmark for BLP. BatteryLife integrates 16 datasets, offering a 2.5 times sample size compared to the previous largest dataset, and provides the most diverse battery life resource with batteries from 8 formats, 59 chemical systems, 9 operating temperatures, and 421 charge/discharge protocols, including both laboratory and industrial tests. Notably, BatteryLife is the first to release battery life datasets of zinc-ion batteries, sodium-ion batteries, and industry-tested large-capacity lithium-ion batteries. With the comprehensive dataset, we revisit the effectiveness of baselines popular in this and other time series fields. Furthermore, we propose CyclePatch, a plug-in technique that can be employed in various neural networks. Extensive benchmarking of 18 methods reveals that models popular in other time series fields can be unsuitable for BLP, and CyclePatch consistently improves model performance establishing state-of-the-art benchmarks. Moreover, BatteryLife evaluates model performance across aging conditions and domains. BatteryLife is available at https://github.com/Ruifeng-Tan/BatteryLife.

• The paper presents BatteryLife, the largest and most diverse dataset and benchmark for battery life prediction, and introduces CyclePatch, a novel modeling technique.
• The BatteryLife dataset comprises over 90,000 samples from 998 batteries, offering significantly increased size and diversity across various aging conditions.
• The novel CyclePatch technique models degradation patterns across battery cycles and achieves state-of-the-art prediction performance on the BatteryLife benchmark.

The paper introduces BatteryLife, a comprehensive dataset and benchmark for battery life prediction (BLP), addressing limitations in existing datasets regarding size, diversity, and inconsistent benchmarks. The authors integrate 16 datasets, increasing the sample size by 2.4 times compared to the previous largest dataset. BatteryLife encompasses batteries from 8 formats, 80 chemical systems, 12 operating temperatures, and 646 charge/discharge protocols, including zinc-ion, sodium-ion, and industry-tested large-capacity lithium-ion batteries.

The paper identifies three key challenges in the field:

• Limited dataset size, hindering comprehensive insights into modern battery life data.
• Restricted data diversity in existing datasets, raising concerns about the generalizability of findings.
• Inconsistent and limited benchmarks, obscuring the effectiveness of baselines.

To address these challenges, the authors propose BatteryLife, a dataset that is 2.4 times larger than BatteryML, with more than 90,000 samples from 998 batteries. It offers unparalleled diversity, delivering 4 times the format, 16 times the chemical system, 2.4 times the operating temperature, and 3.4 times the charge/discharge protocol compared to BatteryML.

The paper introduces CyclePatch, a plug-in technique to model degradation test data. CyclePatch addresses the observation that voltage and current time series exhibit similar patterns across cycles within a protocol. It treats each cycle as a token, capturing recurring patterns in degradation tests. The cycling data from batteries with different life labels show discriminative characteristics, both in terms of cycles of the same number and variance across cycles. CyclePatch employs an intra-cycle encoder to model the interactions among variables within each cycle and generates informed representations for each cycle, on which an inter-cycle encoder is applied to learn patterns across cycle tokens.

The main contributions of this work are:

• BatteryLife is the largest battery life dataset, offering more than 90,000 samples from 998 batteries.
• BatteryLife is the most diverse battery life dataset, containing lab-tested Li-ion, Na-ion, and Zn-ion batteries, as well as industry-tested large-capacity Li-ion batteries.
• BatteryLife provides a comprehensive benchmark for BLP, offering fair comparisons of popular baselines, and introduces CyclePatch as a plug-in technique for BLP.

The paper defines battery life as the cycle number at which the state of health (SOH) becomes no larger than 80%, where SOH is defined as:

$SOH = \frac{Q_i}{Q_0}$

where:

• $SOH$ is the state of health
• $Q_i$ is the capacity of $i^{th}$ cycle
• $Q_0$ is the initial capacity

The cycling data pattern is affected by aging conditions. The aging factors considered in this work are battery format, anode, cathode, electrolyte, charge protocols, discharge protocols, operation temperature, nominal capacity, and manufacturer.

The paper defines the problem as: given the input $X_{1:S}$ with $\forall S\leq 100$, predict the battery life denoted by $y\in\mathbb{R}^{1}$. $$X_{i:N}=[X_i,X_{i+1},\cdots,X_N]\in\mathbb{R}^{3\times T}$$ represents the voltage, current, and capacity variables across $T$ time steps starting from the $i^{th}$ cycle to the $N^{th}$ cycle, where $X_i\in\mathbb{R}^{3\times T_i}$ is the cycling data of the $i^{th}$ cycle with $T_i$ time steps.

The paper splits BatteryLife into four parts: Li-ion, Zn-ion, Na-ion, and CALB. The data statistics of each part are summarized in Table 2 of the paper.

The popular benchmark methods that the paper considers are: Transformer encoder, LSTM (Long short-term memory), BiLSTM (bidirectional long short-term memory), GRU (gate recurrent unit), BiGRU (bidirectional gate recurrent unit), CNN (convolutional neural network), MLP (multilayer perceptron), DLinear, PatchTST, Autoformer, iTransformer and MICN.

CyclePatch segments the cycling time series into basic units that have recurring patterns throughout degradation tests. The cycle token is computed by:

$$\left[X_1, X_2, X_3, \cdots, X_S \right] = {\rm Segment}(X_{i:S})$$

$\hat{X}_i = W {\rm flatten}(X_i) + b$

where:

• $X_i$ is the $i^{th}$ cycle
• $W \in \mathbb{R}^{D_1 \times 900}$
• $b \in \mathbb{R}^{D_1}$
• ${\rm flatten}(X_i) \in \mathbb{R}^{900}$

The computation at the $l^{th}$ layer of the intra-cycle encoder is given by:

$$\hat{z}^l_i = W_2^l \sigma(W_1^l z^{l-1}_i + b_1^l) + b_2^l$$

$z^l_i = LN\left(\hat{z}^l_i + z^{l-1}_i\right)$

where:

• $W_1^l \in \mathbb{R}^{D_2 \times D_1}$
• $W_2^l \in \mathbb{R}^{D_1 \times D_2}$
• $b_1^l \in \mathbb{R}^{D_1}$
• $b_2^l \in \mathbb{R}^{D_2}$
• $LN$ denotes layer normalization

An inter-cycle encoder is then applied to extract key patterns across cycle token embeddings:

$v = f(H)$

$\hat{y} = {\rm Projection}(v)$

where:

• $f(\cdot)$ represents the inter-cycle encoder
• $v$ captures both intra-cycle and inter-cycle information
• $\hat{y}$ is the final prediction

The paper conducts experiments to answer four research questions:

• RQ1: How do the benchmark methods perform on different domains?
• RQ2: How do the main components of CyclePatch framework affect the performance?
• RQ3: How adaptable are benchmark methods when applied across aging conditions in each domain?
• RQ4: How transferrable is the model pretrained on the Li-ion domain for other domains?

The paper employs two metrics to evaluate model performance: MAPE (mean absolute percentage error) and $\alpha$-accuracy.

The MAPE is computed as:

$${\rm MAPE}=\frac{1}{N}\sum_{i}^{N}\frac{\left|y_i-\hat{y}_i\right|}{y_i}$$

where:

• $y_i$ is the ground truth battery life of the $i^{th}$ sample
• $\hat{y}_i$ is the predicted battery life of the $i^{th}$ sample
• $N$ is the number of samples in the testing set

The $\alpha$-accuracy is computed as:

$$\alpha{\rm -accuracy}=\frac{1}{N}\sum_{i=1}^{N}1_{\left|y_i-\hat{y}_i\right|\leq\alpha y_i}(\hat{y}_i)$$

The paper finds that CyclePatch methods achieve the best performance across all domains. The best MAPE are 0.179, 0.515, 0.255, and 0.141 for the Li-ion, Zn-ion, Na-ion, and CALB datasets, respectively. Techniques successful in other time series fields cannot be naively applied to BLP. The model performance improves with the increase in the number of usable cycles initially and then plateaus. All components in CyclePatch significantly contribute to its performance. Models generally perform worse on unseen aging conditions compared to those seen ones. Domain adaptation significantly improves model performance on Zn-ion and Na-ion datasets.

In future studies, the authors plan to incorporate more datasets into BatteryLife and focus on developing more transferrable models for BLP.