Residual-Corrected Equivalent-Circuit Model with Universal Differential Equations for Robust Battery Voltage Prediction under Operating-Condition Shift

Technical Summary: Residual-Corrected Equivalent-Circuit Model with Universal Differential Equations

Problem Statement
Accurate terminal-voltage prediction is fundamental to model-based battery management systems (BMS), serving as the primary correction signal for state estimators like the Extended Kalman Filter (EKF). However, existing approaches face a trade-off:

• Low-order Equivalent-Circuit Models (ECM): While computationally efficient and interpretable, first-order Thevenin models (ECM-1RC) lack the expressiveness to capture nonlinear, temperature-dependent, and transient behaviors under realistic operating conditions. This structural mismatch leads to voltage errors that can bias downstream state estimation.
• Purely Data-Driven Models: Deep learning approaches (e.g., LSTMs) offer flexibility and high accuracy under matched conditions but often sacrifice interpretability and suffer from performance degradation when faced with operating-condition shifts (e.g., temperature changes or unseen drive cycles).

The paper addresses the challenge of achieving robust voltage prediction that maintains the interpretability of physical models while compensating for their structural limitations, specifically under distribution shifts such as temperature variations and different drive-cycle profiles.

Methodology
The authors propose a Residual-Corrected Hybrid Formulation implemented as an ECM-based Universal Differential Equation (ECM-UDE). The core architecture integrates a physical model with a learnable neural component:

1. Physical Backbone (ECM-1RC): The model retains a first-order Thevenin circuit structure. The terminal voltage $V(t)$ is governed by the standard relation:
   $$V(t) = OCV(z(t)) - R_0 I(t) - V_1(t)$$
   where $OCV$ is the open-circuit voltage, $R_0$ is ohmic resistance, and $V_1(t)$ is the polarization voltage.
2. Latent Polarization Dynamics: The evolution of the polarization state $V_1(t)$ is modified to include a learnable correction term $f_\theta$:
   $$\dot{V}_1(t) = -\frac{V_1(t)}{R_1 C_1} + \frac{I(t)}{C_1} + f_\theta(V_1(t), I(t), z(t), T_{cell}(t))$$
   Here, $f_\theta$ is a compact neural network (a multilayer perceptron) that corrects only the structured mismatch (residuals) that the reduced-order circuit cannot capture.
3. State Propagation: Unlike standard ECMs, the hybrid model propagates the State of Charge (SOC, $z$) internally as a dynamical state ($\dot{z} = -\eta I(t)/Q_{nom}$), reducing dependence on externally supplied SOC trajectories within the prediction window.
4. Training Protocol:
   • Warm-Start: ECM-1RC parameters ($R_0, R_1, C_1$, and OCV coefficients) are first identified via nonlinear least squares on training data. These parameters initialize the hybrid model, ensuring the neural correction operates in a "low-residual regime" from the first training epoch.
   • Optimization: The hybrid model is trained end-to-end using Adam optimization on mean-squared error loss.
   • Normalization: The pipeline uses a hybrid normalization strategy: current and voltage are normalized via empirical z-scores, while temperature and SOC are normalized using fixed, physics-informed constants to ensure stability during out-of-distribution transfer.

Key Contributions

1. Structurally Localized Residual Correction: Unlike previous hybrid approaches that may apply corrections at the output level, this method embeds the neural component exclusively within the latent polarization dynamics. The physical model retains responsibility for the dominant voltage structure (OCV, ohmic drop, and RC dynamics).
2. Robustness-Centered Evaluation: The study introduces a rigorous evaluation protocol assessing performance across four axes:
   • Matched-condition prediction (UDDS at 25°C).
   • Sensitivity to inference-time SOC uncertainty.
   • Zero-shot temperature transfer (25°C to -20°C).
   • Zero-shot drive-cycle transfer (UDDS to US06, LA92, HWFET).
3. Demonstration of Stability: The work highlights that the hybrid approach not only improves accuracy but also significantly reduces inter-seed variability compared to purely data-driven baselines, indicating greater optimization stability.

Results
Experiments were conducted on a public Panasonic 18650PF dataset. The ECM-UDE was compared against a standalone ECM-1RC and a stacked LSTM baseline.

• Matched Conditions (UDDS, 25°C): ECM-UDE achieved a Mean Absolute Error (MAE) of 39.16 mV, outperforming the LSTM (75.89 mV) and ECM-1RC (127.97 mV). This represents a 48% reduction in MAE relative to the LSTM. Furthermore, the ECM-UDE exhibited an order-of-magnitude lower inter-seed variability (Coefficient of Variation: 0.44% vs. 6.20% for LSTM).
• Inference-Time SOC Perturbation: When Gaussian noise was added to the SOC input, the ECM-UDE maintained the lowest absolute error across all noise levels ($\sigma_z \in \{0.01, 0.02, 0.05\}$), though it showed slightly higher sensitivity than the deterministic ECM-1RC.
• Temperature Transfer (Zero-Shot): As temperature decreased from 25°C to -20°C, all models degraded. However, ECM-UDE consistently remained the most accurate. At -20°C, ECM-UDE achieved an MAE of 173.63 mV, compared to 193.45 mV for the LSTM and 182.92 mV for the ECM-1RC.
• Drive-Cycle Transfer: Under zero-shot transfer to unseen cycles (US06, LA92, HWFET), ECM-UDE achieved the lowest error on all targets. The relative gain was most pronounced on the most aggressive cycle (US06), where ECM-UDE reduced MAE by 45.2% compared to the LSTM.

Significance and Claims
The paper claims that the residual-corrected ECM-UDE offers a lightweight and interpretable enhancement to low-order circuit models. The significance lies in the demonstration that:

• Physical Anchoring: The physical model provides a stable "voltage anchor" that prevents the purely learned model from degrading under operating-condition shifts (temperature and drive-cycle variations).
• Complementarity: The physical backbone and the learned correction are complementary rather than redundant. The circuit handles the dominant structure, while the neural component compensates for structured residuals.
• Practicality: The approach achieves substantial accuracy gains without replacing the ECM with a large black-box model, making it a viable candidate for integration into observer-based SOC estimation schemes in BMS.

The authors note that while the results are promising for the specific NCA chemistry cell tested, generalizability to other chemistries (e.g., LFP, NMC) and cell formats requires further assessment. Future work is suggested to explore observer-level integration and explicit temperature-dependent parameter adaptation.