L. E et al.

Green Energy and Intelligent Transportation 4 (2025) 100291

In contrast, data-driven methods do not require understanding in-

Abbreviations and notations ternal physicochemical processes or prior model assumptions, typically
resulting in black-box models. However, they heavily rely on large, high-
ANN Artificial Neural Network quality training datasets. For example, Wang et al. proposed a CNN-based
BO Bayesian Optimization RUL prediction method, mapping five consecutive charge–discharge cy-
CNN Convolutional Neural Network cles to RUL with an RMSE of 501 cycles, but requiring 85% of data for
GRU Gated Recurrent Unit training [9]. Ren et al. used an artificial neural network (ANN) model,
LSTM Long Short-Term Memory achieving an RMSE of 599 cycles using the first 16% of cycles from 88
MAE Mean Absolute Error supercapacitors, highlighting dataset dependency [10]. Zhou et al.
MAPE Mean Absolute Percentage Error applied a residual convolutional neural network (Res-CNN) for lithium
NARX Nonlinear Autoregressive Network with Exogenous iron phosphate (LFP) battery cycle life prediction, with mean absolute
Inputs errors (MAEs) of 43.366 cycles and mean absolute percentage errors
NSGAIII Non-dominated Sorting Genetic Algorithm III (MAPEs) below 4.557%, yet needing data from the first 30% of cycles
OLS Ordinary Least Squares [11]. Recent advances in sensing technologies, such as acoustic sensing
PINN Physics-Informed Neural Network (e.g., ultrasonic testing and acoustic emission) and internal temperature
RMSE Root Mean Square Error monitoring (e.g., NTC sensors, fiber optic sensors), offer insights for
RUL Remaining Useful Life supercapacitor state prediction, particularly in capturing nonlinear aging
SOH State of Health characteristics, despite focusing primarily on lithium-ion batteries [12,
13]. These methods, while achieving high accuracy, share a common
limitation: reliance on data quality and quantity. A summary of relevant
studies is summarized in Table 1.
Both model-based and data-driven methods have inherent limita-
1. Introduction
tions: model-based approaches often oversimplify processes, reducing
flexibility [20], while data-driven methods lack interpretability and
Supercapacitors, as a novel energy storage device, offer higher
require extensive data [21,22].
capacitance, greater specific capacity, a broader operating temperature
Although the application of PINN in predicting supercapacitor degra-
range, and a longer lifespan compared to traditional capacitors [1]. They
dation trajectories and RUL remains limited, their success in other fields
are currently widely utilized in hybrid vehicles, playing a crucial role in
highlights their potential to address the shortcomings of both model-based
power output coordination [2]. To ensure safe supercapacitor usage,
and data-driven approaches. For instance, Gong et al. [23] developed an
optimizing management strategies through accurate prediction of
energy-based PINN for low-frequency electromagnetic prediction, where
degradation trajectories and RUL is essential [3]. This approach ad-
physical constraints were embedded into the loss function, ensuring out-
dresses key challenges in electric vehicle battery monitoring, safety, and
puts aligned with the underlying physical model. Their results demon-
energy management, meeting the demand for high-precision perfor-
strated that physics-informed loss functions significantly outperformed
mance evaluation across the vehicle's lifecycle [4].
conventional deep-learning approaches in predicting magnetic field dis-
Current methods for predicting supercapacitor degradation trajec-
tributions. Similarly, Cho et al. [24] applied PINNs to predict lithium-ion
tories and RUL are mainly divided into model-based and data-driven
battery cell temperature by integrating an empirical thermal model into
approaches [5]. Model-based methods are developed by studying
the loss function, achieving a prediction error of only 0.05 C.
supercapacitor failure mechanisms and aging process parameters, aiming
A key challenge in PINNs is balancing the weight between physical
to capture the underlying physical and chemical processes of aging.
constraints and data fitting. Traditional methods rely on fixed weights,
However, their accuracy is limited by assumptions about degradation
such as those used by Colby et al. for solving the Allen–Cahn equation,
mechanisms and parameter derivation. For instance, Xu et al. developed
where a large fixed weight was assigned to the physical loss term to
a capacity degradation model based on the Arrhenius equation, using
ensure accuracy near initial conditions [25]. However, such static ap-
temperature, current, and cycle numbers as variables, achieving a root
proaches often fail to adapt to diverse datasets or dynamic optimization
mean square error (RMSE) of approximately 10.5 F over the lifetime, but
processes, leading to suboptimal performance, overfitting, or slow
assuming exponential decay under varying conditions, which may
convergence. Manual weight tuning, while straightforward, is
oversimplify the process [6]. German et al. proposed an SEI-based aging
labor-intensive, error-prone, and heavily dependent on expert knowl-
law for capacitance loss during floating aging, tested on 81 commercial
edge, particularly in high-dimensional problems.
supercapacitors, with an RMSE of 3% for state of health (SOH) estima-
To address these, this paper proposes a PINN method for predicting
tion, yet it assumes irreversible aging, potentially overlooking reversible
degradation trajectories and the remaining useful life of supercapacitors,
factors like gas adsorption [7]. Wu et al. introduced a life prediction
combining LSTM with an empirical aging equation in the loss function,
method based on the “ten-degree rule,” assuming reaction rates double
balanced by a parameter optimized via Bayesian optimization for
every 10 C, achieving an RMSE of 1562.5 cycles, but this may over-
enhanced adaptability and performance. The main contributions of this
simplify temperature-degradation relationships, limiting accuracy in
paper are shown as follows:
extreme conditions [8].

Table 1
Summary of literature on data-driven models.
Reference Method Accuracy Training data requirements

[9] CNN-based method RMSE of 501 cycles 85% of the total number of supercapacitor units
[10] ANN-based method RMSE of 599 cycles 16% of full life cycle data
[11] Res–CNN–based method MAEs of 43.366 cycles 30% of full life cycle data
[14] Gaussian process regression-based method RMSE of less than 2% 60% of the total number of supercapacitor units
[15] LSTM-based method RMSE of 0.026,1 on online data and 0.033,8 on offline data 80% of the total number of supercapacitor units
[16] Bidirectional-LSTM-based method RMSE of 0.027,5 60% of full life cycle data
[17] LSTM-based model with hybrid genetic algorithm RMSE ranging from 0.016,1 to 0.026,4 50% of full life cycle data
[18] Bayesian optimization-based hyperband method MAPE of 1.05% 6% of full life cycle data
[19] NARX neural networks method RMSE of 0.032,8 60% of full life cycle data