Source-linked AI summary
Sera: Semantic Representation Aggregation for Reliable and Interpretable Battery Health Forecasting
Jiawei Li, Fang Liu, Wei Zhang, Zuming Liu, Man-Fai Ng, Zhi Wei Seh
TL;DR
SoH forecasting must address nonlinear degradation and variation across batteries, while existing data-driven temporal models often do not explicitly represent higher-level degradation characteristics that could guide reliable and interpretable forecasts. Sera independently encodes rule-based and LLM-based degradation semantics alongside temporal features, then integrates the representations through gated aggregation. Across the reported evaluations, Sera reduces forecasting errors by up to 37.3% over the temporal baseline and improves generalization to unseen batteries.
Problem
SoH forecasting must address nonlinear degradation and variation across batteries, while existing data-driven temporal models often do not explicitly represent higher-level degradation characteristics that could guide reliable and interpretable forecasts.
Source: §1 · pp. 1–2Method
Sera independently encodes rule-based and LLM-based degradation semantics alongside temporal features, then integrates the representations through gated aggregation.
Source: §1 · p. 2Results
Across the reported evaluations, Sera reduces forecasting errors by up to 37.3% over the temporal baseline and improves generalization to unseen batteries.
Source: §5 · p. 8Takeaways & Limitations
Counterfactual analysis defines semantic consistency as agreement between a model’s response to a rule intervention and the expectation from that rule’s degradation meaning.
Source: §4.3.1 · p. 7
Abstract
from arXiv · showhide
Battery state of health (SoH) forecasting is important for battery management, but remains challenging due to nonlinear degradation and heterogeneity across batteries. Existing data-driven approaches primarily use temporal models to learn from numerical battery time series, and higher-level degradation characteristics are often not explicitly represented. These characteristics, however, can provide degradation guidance to support reliable forecasting and make the influence of degradation more interpretable. In this paper, we propose \textsc{Sera}, a \underline{se}mantic \underline{r}epresentation \underline{a}ggregation framework that complements temporal modelling with degradation semantics. Guided by battery domain expertise, \textsc{Sera} extracts degradation semantics from time series and constructs two complementary representations using rule-based knowledge and LLM-based interpretation. The representations are independently encoded and integrated with the representation learned by temporal models through gated aggregations. Experiments on the mainstream benchmark across multiple prediction horizons and different temporal models show that \textsc{Sera} consistently improves forecasting performance, achieving up to a 37.3\% reduction in prediction error over the temporal baseline and enhanced generalizability. Counterfactual analysis examines how forecasts respond to changes in degradation semantics to assess interpretability. The results show that prediction responses are consistent with the meanings of key degradation descriptors across tested horizons. Together, these findings demonstrate that structured degradation semantics and effective aggregation can improve forecasting accuracy and support reliable and interpretable battery health forecasting for advanced battery management.
1 Introduction
Battery SoH forecasting must handle nonlinear, heterogeneous degradation while supporting reliable predictions across batteries and interpretable behavior. Sera adds rule- and LLM-based semantic representations to temporal modelling and reports lower forecasting error and improved generalization.
1 Introduction
- Forecasting is difficult because degradation trajectories are nonlinear and vary even among batteries of the same model, with longer horizons amplifying the challenge.Forecasts are intended to be accurate while representing degradation state to support reliability and interpretability.Source: §1 · p. 1
- Data-driven methods often model numerical time series without explicitly representing higher-level degradation characteristics such as trends, severity, and possible mechanisms.Battery-domain knowledge can also be encoded through equations, parameters, or constraints.Source: §1 · p. 2
- Sera reduces MAE by up to 37.3% over the temporal baseline and improves generalization to unseen batteries.The evaluation also uses counterfactual analysis to examine whether forecast responses align with the meanings of key degradation descriptors.Source: §1 · p. 2
- Sera constructs complementary rule-based and LLM-based semantic representations, encodes them separately from temporal features, and combines them through gated aggregation.The rule-based representation applies predefined rules, while the LLM-based representation interprets relationships among multiple rules.Source: §1 · p. 2
- The evaluation spans a real-world dataset, multiple prediction horizons and temporal encoders, and counterfactual semantic interventions.The interventions assess whether forecast responses are consistent with the meanings of degradation semantics.Source: §1 · p. 2
2 Battery SoH Forecasting Problem Formulation
The forecasting problem is formulated as predicting a future SoH trajectory from recent multivariate battery-cycle measurements. The model maps the observed window to the future trajectory.
2 Battery SoH Forecasting Problem Formulation
- Battery health forecasting is formulated as a multivariate sequence-to-sequence prediction problem over a future horizon.Each cycle is represented by a feature vector of measurements and derived health indicators, including voltage, current, temperature, and internal resistance.Source: §2 · p. 2
- The forecasting target is the battery’s future SoH trajectory over the next H cycles.The input window comprises the most recent L cycles ending at cycle t.Source: §2 · p. 2
- A parameterized forecasting model maps the historical observation window to a predicted future SoH trajectory.The prediction is expected to approximate the ground-truth trajectory.Source: §2 · p. 2
3 Sera Methodology
Sera combines temporal modelling with rule-based and LLM-based degradation semantics, encoding these information sources separately before adaptive gated aggregation for SoH forecasting.
3.2 Structured Semantic Representation
- Seven semantic rules describe degradation trend, risk, mechanism, severity, change point, key metric, and metric value.Trend and change point characterize progression; risk and severity indicate degradation severity, while mechanism, key metric, and metric value connect descriptions to processes and quantitative features.Source: §3.2 · p. 3
- The shared feature vector uses capacity-based SoH slope and acceleration ratio, plus temperature and internal-resistance statistics, to support semantic construction.SoH slope and acceleration ratio distinguish trend categories including stable, mild_decay, linear_decay, and accelerating_fade.Source: §3.2 · p. 3
3.1 Design Overview of Sera
- The rule-based and LLM-based branches construct complementary degradation descriptions, with the LLM interpreting feature relationships more flexibly while retaining the rules’ shared schema.The rule-based branch uses fixed domain-knowledge rules; the LLM-based branch offers greater flexibility in capturing contextual relationships among degradation features.Source: §3.2 · p. 3, §3.1 · p. 3
3.3.1 Temporal Encoding.
- The temporal branch processes a battery time-series window with an encoder to produce a temporal embedding; Sera can use temporal encoders beyond PatchTST.PatchTST groups consecutive observations into patches and uses a CLS token to summarize the sequence.Source: §3.3.1 · pp. 3–4
3.3.2 Semantic Representation Encoding.
- Separate, independently parameterized encoders map the rule-based and LLM-based semantic representations to embeddings matching the temporal embedding’s dimension.Each semantic branch encodes structured text, applies an MLP projection and ReLU activation, then uses a final linear mapping.Source: §3.3.2 · p. 4
3.3.3 Adaptive Gated Aggregation.
- Input-dependent gates weight the temporal, rule-based, and LLM-based embeddings dimension by dimension before their aggregation for SoH prediction.Each gate can suppress or retain embedding dimensions, allowing the contributions of the three representations to adapt to individual input windows.Source: §3.3.3 · p. 4
3.4 SoH Forecasting and Model Training
- The forecasting head uses the aggregated embedding to predict the future SoH trajectory, while training combines prediction loss with temporal-behavior constraints.The horizon-weighted SmoothL1 loss gives larger weights to later predictions; monotonicity discourages increases and smoothness discourages large consecutive changes.Source: §3.4 · pp. 4–5
4 Experiments
The section evaluates Sera’s forecasting performance and previews its experimental setup, main results, and further analyses.
4.1 Experimental Setup
The experiments use cell-identity-separated splits of the MIT Battery Degradation Dataset, fixed 30-cycle histories, several forecast horizons, and comparisons across semantic variants and temporal encoders.
4.1.1 Dataset and Forecasting Sample Construction.
- The MIT Battery Degradation Dataset contains 124 commercial LFP/graphite cells cycled to end-of-life under fast charging at 30°C.The cells are A123 Systems APR18650M1A cells with nominal capacity 1.1 Ah and nominal voltage 3.3 V; measurements include capacity, voltage, current, temperature, and internal resistance.Source: §4.1.1 · p. 5
- Cell identities are split 70%/15%/15% across training, validation, and test sets, with no cell appearing in more than one set.Sliding-window samples pair an L-cycle observation window with its subsequent H-cycle forecasting target; the stride is one cycle.Source: §4.1.1 · p. 5
- The experiments fix the observed history at 30 cycles and test forecast horizons of 30, 50, 70, and 90 cycles.These settings correspond to forecast-to-observation ratios of 1.0, 1.67, 2.33, and 3.0.Source: §4.1.1 · p. 5
4.1.2 Sera Configurations.
- The PatchTST temporal branch uses patch length 5 and dimension 128, while semantic encoders project to dimension 64 before mapping to the shared dimension.Training uses Adam, early stopping, and the same configuration across prediction horizons.Source: §4.1.2 · p. 5
4.1.3 Comparison Models.
- Comparisons include the temporal-only baseline, rule and LLM semantic variants, the full Sera model, and semantic-only models under shared data and training settings.Sera is also evaluated with PatchTST, a standard Transformer, and LSTM; forecasting is measured using MAE and RMSE.Source: §4.1.3 · p. 5
4.2 Forecasting Performance Analysis
Sera improves SoH forecasting over the temporal baseline across prediction horizons, with the combined semantic representations especially beneficial at longer horizons. Results across encoders and representative trajectories show that semantic information complements rather than replaces temporal modelling.
4.2.1 Overall Forecasting Performance.
- Sera achieves the lowest RMSE at all four horizons and the lowest MAE at H∈{50, 70, 90}, reducing baseline MAE from 1.6 to 1.0, 2.3 to 1.7, 3.0 to 2.1, and 3.6 to 2.2 (all ×10−3).Source: §4.2.1 · p. 5
- Semantic-only variants have MAE values of 0.011–0.014, substantially above models using temporal information, so descriptors work better as complementary guidance than standalone predictors.Source: §4.2.1 · p. 6
4.2.2 Semantic Contribution and Generalization.
- At H=30, TS + Rule reduces MAE by 41.4% versus 34.8% for Sera; at H∈{50, 70, 90}, Sera’s MAE reduction rises from 27.1% to 29.6% to 37.3%.Source: §4.2.2 · p. 6
- As H increases, the TS test–training MAE gap reaches 1.7 ×10−3 at H=90, while Sera maintains a relatively small gap across horizons.Source: §4.2.2 · p. 6, Figure 2 · p. 6
4.2.3 Temporal Encoder Sensitivity.
- Semantic representations improve forecasting with LSTM, TFM, and PatchTST, but gains vary by encoder and horizon; with PatchTST at H=90, Sera lowers MAE from 3.6×10−3 to 2.2×10−3.Source: §4.2.3 · p. 6, Table 3 · p. 6
4.2.4 Case Study.
- For short-life batteries, TS increasingly overestimates SoH after about 300 cycles; for a long-life battery, it underestimates SoH around 800–1,600 cycles, while Sera follows decline more closely.Source: §4.2.4 · pp. 6–7, Figure 3 · p. 7
4.3 Interpretability Analysis
Counterfactual semantic interventions test whether Sera’s forecasts respond in directions consistent with degradation meanings while the observed time series stays fixed. The responses generally align with those meanings, with consistency varying by intervention and horizon.
4.3.1 Counterfactual Semantic Analysis.
- Counterfactual tests hold X fixed, alter rules in both semantic views, and score 11 interventions on 200 sampled windows per horizon by expected-direction consistency and final-cycle SoH shift.Applying each intervention to both semantic representations maintains a consistent hypothetical degradation state.Source: §4.3.1 · p. 7
4.3.2 Semantic Consistency.
- Optimistic bundles raise predicted SoH with 69.0–90.5% semantic consistency, while pessimistic bundles lower it with 100% consistency across prediction horizons.The bundles jointly modify rules toward healthier or worsening degradation states, respectively.Source: §4.3.2 · p. 7, Table 4 · p. 8
- Slope →Steep reaches 100% semantic consistency and lowers mean SoH by 6.9% at H=30 and 25.5% at H=90; Trend →Accelerating yields 87.0–100.0% consistency and 0.25–1.23% decreases.The slope effect grows at longer horizons.Source: §4.3.2 · p. 7
- Under tested settings, quantitative slope interventions influence forecasts more than categorical trend interventions; Li Plating usually has high consistency, while Risk →High reaches only 19.0–58.5%.The authors suggest risk effects may depend more strongly on the battery state represented by the temporal input.Source: §4.3.2 · p. 8
5 Conclusion
Sera combines structured degradation semantics with numerical battery time series for SoH forecasting. It improves forecasting results across temporal encoders and uses counterfactual analysis to examine how semantic interventions affect forecasts.
5 Conclusion
- Sera reduces SoH forecasting errors by up to 37.3% versus the temporal baseline and improves generalization to unseen batteries.Forecasting gains are consistent across different temporal encoders.Source: §5 · p. 8
- Sera independently learns temporal, rule-based semantic, and LLM-based semantic representations, then fuses them through adaptive gated aggregation.Source: §5 · p. 8
- Counterfactual analysis shows more consistent forecast responses to semantic rules that directly characterize degradation progression.The analysis examines the direction and magnitude of prediction changes under controlled interventions.Source: §5 · p. 8