Predicting the Remaining Useful Life of Supercapacitors under Different Operating Conditions

Abstract

With the rapid development of the new energy industry, supercapacitors have become key devices in the field of energy storage. To forecast the remaining useful life (RUL) of supercapacitors, we introduce a new technology that integrates variational mode decomposition (VMD) with a bidirectional long short-term memory (BiLSTM) neural network. Firstly, the aging experiments of supercapacitors under various temperatures and voltages were carried out to obtain aging data. Then, VMD was implemented to decompose the aging data, which helped to eliminate disturbances, including capacity recovery and test errors. Then, the hyperparameters of BiLSTM were adjusted, employing the sparrow search algorithm (SSA) to improve the consistency between the input data and the network structure. After obtaining the optimal hyperparameters of BiLSTM, the decomposed aging data were input into BiLSTM for prediction. The experimental results showed that the VMD-SSA-BiLSTM model proposed in this paper has high prediction accuracy and high robustness under different temperatures and voltages, with an average RMSE of 0.112519, a decrease of 44.3% compared to BiLSTM, and a minimum of 0.031426.

1. Introduction

Supercapacitors constitute an emerging class of energy storage device, characterized by their high energy density and power density, as well as high charging and discharging efficiency, wide operating temperature range, and long useful life [1,2,3]. Therefore, they are widely used in various applications. For example, in the field of transportation, supercapacitors can improve the acceleration performance of electric vehicles that frequently start and stop due to their high instantaneous power [4]. In wind power generation, the unstable current generated by wind turbines can extremely easily damage batteries, but supercapacitors have good power characteristics and weak current collection capabilities, which can be used to handle the impulse current generated by wind turbines [5,6]. In distributed generation, supercapacitors can be used to maintain the stability of busbar voltage and reduce the impact of distributed generation on the power grid during grid connection [7,8]. Although the initial investment of supercapacitors may be higher than that of traditional batteries, their overall cost of ownership is more advantageous in many applications due to their long lifespan and low maintenance costs [9,10,11]. Especially in situations where frequent starting, stopping, or rapid responses are required, predicting the RUL of supercapacitors reduces the frequency of maintenance and replacement with high reliability, thereby saving long-term operating costs [12]. As the core component of energy storage systems, the health status of supercapacitors plays a significant role in the safe operation of the entire energy storage system. Therefore the accurate prediction of the lifespan of supercapacitors has become a top priority [13,14,15].

Methods for predicting supercapacitor life can generally be divided into two categories: model-based methods and data-driven methods [16,17]. Model-based methods need to create an equivalent circuit model that mirrors the aging process of the supercapacitor, which can be used to predict the lifespan through parameter identification [18]. Based on data-driven methods, researchers have used various techniques, including traditional statistical methods, machine learning, and deep-learning techniques [19,20,21]. Statistical methods are one of the earliest methods used for estimating the lifespan of supercapacitors [22]. These methods include linear regression, support vector machine (SVM), principal component analysis (PCA), etc. [23,24,25]. Statistical methods typically rely on linear relationships of data and have limited modeling capabilities for complex relationships. Nevertheless, they are still effective when dealing with small datasets and simple models. Machine learning methods, such as decision trees and random forests, provide more powerful nonlinear modeling capabilities than statistical methods [26,27,28]. The application of machine learning methods in estimating the lifespan of supercapacitors is gradually increasing, but their performance may be affected due to the lack of sufficient annotated data [29,30]. Deep-learning methods, such as convolutional neural networks (CNNs), recurrent neural networks (RNNs), long short-term memory networks (LSTMs), etc., can automatically learn the hierarchical structure of data through multi-layer neural networks, thereby capturing complex nonlinear relationships [31,32,33]. The data-driven methods, on the other hand, continuously learn from a large amount of aging data and predict the lifespan accordingly [34]. In contrast to model-based techniques, data-driven methods eliminate the need for complex mathematical frameworks to emulate the internal aging processes of supercapacitors and mainly rely on a large amount of data [35,36]. Zhou et al. proposed a life prediction method based on the LSTM neural network [37]. The data of the supercapacitor were segmented into a training dataset and a predictive dataset to serve as the input for the neural network [38]. The overall results showed that the method exhibited great performance in the prediction of the RUL of the supercapacitor. Li et al. introduced an optimized forecasting model, which was a combination of the extreme learning machine (ELM) and the heuristic Kalman filter (HKF) algorithm, designed for predicting the capacity of supercapacitors [39]. Furthermore, a mode of the RNN using a stacked BiLSTM was proposed by Liu et al. [40].

The LSTM effectively solves the problem of gradient explosion and vanishing in RNNs [41,42]. Although the LSTM can obtain long-distance feature information, the information obtained is all from before the output time and does not utilize reverse information [43,44]. The BiLSTM network has the capability to explore deeper into the connections among the current data points and their corresponding past and future time data, thus enhancing the predictive accuracy of the model and making more comprehensive and detailed decisions using past and future information [45,46,47].

The SSA represents a new approach to swarm-based optimization, drawing inspiration from sparrows’ foraging and anti-predator strategies [48,49]. It is superior to other algorithms, like grey wolf optimization (GWO) and particle swarm optimization (PSO) in terms of the local search strength and speed of convergence [50,51].

Because the actual capacity of a fully charged supercapacitor can directly reflect the attenuation of the supercapacitor RUL, it can serve as a crucial indicator of supercapacitor health in predicting the remaining lifespan [52]. However, there are noise and capacity recovery phenomena in the degradation curve of supercapacitors, which makes it difficult to accurately track the degradation trend of supercapacitors using capacity prediction directly [53].

Following the aforementioned analysis, this article proposes the VMD-SSA-BiLSTM model. VMD has been proven to have good performance in handling nonstationary signals, as well as having good decomposition ability and resistance to noise interference [54,55]. Therefore, in this article, VMD was used to process the aging data of supercapacitors, obtaining a sequence of intrinsic mode functions (IMFs) characterizing local features in order to remove noise interference and better extract the capacity degradation features. Then, a multi-layer BiLSTM model was used to predict multiple IMF components separately. To solve the issue of the model’s unstable prediction, the SSA was used to improve the count of concealed units in BiLSTM, as well as to adjust the number of epochs and the starting learning rate. Eventually, the forecasts from each component were merged to achieve the ultimate prediction outcome. To ensure the generalization ability of the model, the RULs of the supercapacitors under various temperatures and voltages were predicted. The results of this experiment indicate that the model exhibits strong generalization ability under various working conditions.

The rest of the paper is organized as follows: In Section 2, the VMD-SSA-BiLSTM method is presented. In Section 3, the aging state is introduced. In Section 4, the introduced model is used to forecast the RULs of supercapacitors under various temperature and voltage conditions. The conclusion is presented in Section 5.

The main research content of the article is shown in Figure 1. Firstly, obtain aging data under different voltage, current, and temperature conditions. The aging data obtained from the battery experiments were decomposed to five IMF components using the VMD method, and then the RULs of supercapacitors under different operating conditions were predicted using the SSA-BiLSTM method.

2. VMD-SSA-BiLSTM Method

2.1. VMD

VMD was introduced by Dragomiretskiy et al. in 2014 [54]. Constructing a constrained variational model is the essence of the VMD algorithm, and the adaptive decomposition signal is achieved through the quest for the best solution within this model [56,57].

Firstly, original signal $f(t)$ of the supercapacitor is decomposed into K components $u$, with the following expression:

$$\{\begin{array}{c}\underset{\left\{u_k\right\},\left\{{\omega }_k\right\}}{\min }\left\{{\displaystyle \sum _{k=1}^K{\Vert {\partial }_t\left[(\delta (t)+\frac{j}{\pi t})\ast u_k(t)\right]e^{-j{\omega }_{k}t}\Vert }_2^2}\right\}\\ s.t.{\displaystyle \sum _{k=1}^K{u}_k=f(t)}\end{array}$$

(1)

When solving this model, it is necessary to introduce $\alpha$ and $\lambda (t)$ to transform a constrained variational issue into an unconstrained variational issue. The expression is as follows:

$$\begin{array}{c}L(\left\{u_k\right\},\left\{{\omega }_k\right\},\lambda )=\alpha {{\displaystyle \sum _{k=1}^K\Vert {\partial }_t\left[(\delta (t)+\frac{j}{\pi t})\ast u_k(t)\right]e^{-j{\omega }_{k}t}\Vert }}_2^2+\\ {\Vert f(t)-{\displaystyle \sum _{k=1}^K{u}_k(t)}\Vert }_2^2+\langle \lambda (t),f(t)-{\displaystyle \sum _{k=1}^K{u}_k(t)}\rangle \end{array}$$

(2)

Initialize $u_k^1$, ${\omega }_k^1$, and ${\lambda }^1$, and use the alternating direction algorithm to adjust the values of $u_k^{N+1}$, $w_k^{N+1}$, and ${\lambda }^{N+1}$ to search for the saddle point of the Lagrangian expression (2), which corresponds to the optimal solution of the problem. The updated expressions are as follows:

$${\widehat{u}}_k^{n+1}\left(\omega \right)=\frac{\widehat{f}(w)-{\displaystyle \sum _{i\ne k}{\widehat{u}}_i^{n+1}(\omega )}+\frac{{\widehat{\lambda }}^n(\omega )}{2}}{1+2\alpha {(\omega -{\omega }_k^n)}^2}$$

(3)

$${\omega }_k^{n+1}=\frac{{\displaystyle \underset{0}{\overset{\infty }{\int }}\omega {\left|{\widehat{u}}_k^{n+1}(\omega )\right|}^{2}d\omega }}{{\displaystyle \underset{0}{\overset{\infty }{\int }}{\left|{\widehat{u}}_k^{n+1}(\omega )\right|}^{2}d\omega }}$$

(4)

$${\widehat{\lambda }}^{n+1}(\omega )={\widehat{\lambda }}^n(\omega )+\tau \left[\widehat{f}(\omega )-{\displaystyle \sum _{k=1}^K\left|{\widehat{u}}_k^{n+1}(\omega )\right|}\right]$$

(5)

where ${\widehat{u}}_k^{n+1}\left(\omega \right)$, $\widehat{f}(w)$, and $\widehat{\lambda }(\omega )$ are the corresponding Fourier transforms of $u_k^{n+1}(t)$, $f(t)$, and $\lambda (t)$, respectively, and $n$ is the number of iterations.

Expression (6) tests whether each component of the IMF can reach the required error accuracy and stop updating after meeting the requirements, which is as follows:

$${\displaystyle \sum _{k=1}^K\frac{{\Vert {\widehat{u}}_k^{n+1}-{\widehat{u}}_k^n\Vert }_2^2}{{\Vert {\widehat{u}}_k^n\Vert }_2^2}}<\epsilon$$

(6)

2.2. BiLSTM

In the BiLSTM architecture, $y_t$ at each layer is composed of a combination of three components: $h_{t+1}$, $h_{i-1}$, and $x_t$. The combination process of each hidden layer state is represented by Equation (7).

$$\{\begin{array}{c}{h}_t=LSTM(x_t,h_{t-1})\\ h_i=LSTM(x_t,h_{i-1})\\ y_t=a_t{h}_t+b_t{h}_i+c_t\end{array}$$

(7)

In the equation, LSTM stands for the computation procedures of the traditional LSTM network.

From Figure 2, it can be observed that BiLSTM incorporates a data stream moving from the future toward the past, which is in addition to the unidirectional flow from the past to the future in standard LSTM models. The hidden layers from previous and subsequent time steps are not connected, which allows BiLSTM to better explore the temporal features of the data. However, the number of units in the hidden layer, the epochs required for the iterative optimization of the loss function, and the starting learning rate can all affect the accuracy of the predicted output data. The manual tuning of these parameters can be time-consuming and may not result in optimal parameters. Therefore, a sparrow swarm particle dimension was constructed based on the number of hyperparameters requiring optimization, using the SSA to adjust those parameters.

2.3. SSA

The SSA is inspired by the foraging and anti-predator strategies of sparrows [50]. The location of the sparrows is represented as follows:

$$X=\left[\begin{array}{cccc}{X}_{1,1}& X_{1,2}& \cdots & X_{1,d}\\ X_{2,1}& X_{2,2}& \cdots & X_{2,d}\\ ⋮& ⋮& ⋮& ⋮\\ X_{n,1}& X_{n,2}& \cdots & X_{n,d}\end{array}\right]$$

(8)

The fitness values of all the sparrows can be represented as the following vector:

$$F_x=\left[\begin{array}{c}f(\left[\begin{array}{cccc}{X}_{1,1}& X_{1,2}& \cdots & X_{1,d}\end{array}\right])\\ f(\left[\begin{array}{cccc}{X}_{2,1}& X_{2,2}& \cdots & X_{2,d}\end{array}\right])\\ ⋮\\ f(\left[\begin{array}{cccc}{X}_{n,1}& X_{n,2}& \cdots & X_{n,d}\end{array}\right])\end{array}\right]$$

(9)

The values in each row of $F_x$ represent individual fitness values.

In the SSA, it is the responsibility of the producers to forage for sustenance and direct the travel path of the group. The positional information of the producers is as follows:

$$x_{i,j}^{t+1}=\{\begin{array}{c}{x}_{i,j}^t\cdot \exp (\frac{-i}{\alpha \cdot ite{r}_{\max }}),R_2<ST\\ x_{i,j}^t+Q\cdot L,R_2\ge ST\end{array}$$

(10)

In the formula, $\alpha$ $\in$ (0,1], $R_2$ $\in$ [0,1], $ST$ $\in$ [0.5,1], and $L$ is a 1 $\times$ d matrix with all elements equal to 1.

Scroungers persistently keep an eye on the producers. Once the producers stumble upon better food, the scroungers promptly leave their existing position to contend for the food, thus updating their location as follows:

$$x_{i,j}^{t+1}=\{\begin{array}{c}Q\cdot \exp (\frac{x_{worst}^t-x_{i,j}^t}{i^2}),i>\frac{N}{2}\\ x_e^{t+1}+\left|x_{i,j}^t-x_e^{t+1}\right|\cdot A^{+}\cdot L,i\le \frac{N}{2}\end{array}$$

(11)

where $x_e^{t+1}$ is the best position in generation t + 1. A is a 1 $\times$ d matrix, where each element is randomly assigned either 1 or −1, and $A^{+}=A^T{(A{A}^T)}^{-1}$.

The algorithm presupposes that between 10% and 20% of individuals within the population will become aware of the threat. The starting location for these individuals is randomly generated within the population as follows:

$$x_{i,j}^{t+1}=\{\begin{array}{c}{x}_{best}^t+\beta \cdot \left|x_{i,j}^t-x_{best}^t\right|,f_i>f_b\\ x_{i,j}^t+k\cdot (\frac{x_{i,j}^t-x_{worst}^t}{(f_i-f_w)+\epsilon }),f_i=f_b\end{array}$$

(12)

where $k$ $\in$ [−1,1] and $\epsilon$ are a constant.

The flowchart of the supercapacitor RUL prediction method based on VMD-SSA-BiLSTM is shown in Figure 3. VMD decomposition can effectively reduce the noise interference in the data through modular processing, such as capacity recovery and testing errors; the BiLSTM network can capture complex features in time series data; the SSA, with its powerful optimization ability and fast convergence speed, is used to optimize the number of hidden layer units, epochs, and initial learning rate of BiLSTM networks, thereby improving the matching degree between the data and network structure.

3. Supercapacitor Aging State Test

3.1. Test Platform and Aging Data

The aging status testing platform for the supercapacitors is mainly divided into three parts: the tester, the temperature chamber, and the computer. The model of the tester was NEWARE CT-4008-5V100A. The temperature chamber was employed to furnish a particular temperature environment to test the aging status of the supercapacitors, and the model was MENTEK MHP-150-AA. Figure 4 displays the aging supercapacitor test platform.

In this study, we used the Maxwell BCAP0010 P270 T01 capacitor. This supercapacitor comes with the following characteristics: a rated capacity of 10 F, a minimum capacity of 8 F, a maximum equivalent series resistance (ESR) of 75 mΩ, a rated voltage of 2.7 V, a maximum voltage not exceeding 2.85 V, a maximum current not exceeding 7.2 A, a leakage current of 0.030 mA, an operating temperature range of −40 °C to 85 °C, and an energy storage temperature range of −40 °C to 70 °C. This model of supercapacitor has been widely used in consumer electronics, UPS uninterruptible power supplies, intelligent instruments, car recorders, toys, program-controlled switches, etc.

The aging of the supercapacitors shows that the aging factors are mainly the voltage, temperature, and discharge/charge current [58,59]. In order to confirm the adaptability of the proposed model, aging experiments on the supercapacitors under various temperatures and voltages were carried out, and the aging data were measured and predicted.

Table 1. Data for the supercapacitors.

	Voltage	Current	Temperature
SC1	2.9 V	3 A	25 °C
SC2	2.9 V	3 A	50 °C
SC3	2.9 V	3 A	65 °C
SC4	2.7 V	3 A	25 °C
SC5	3.2 V	3 A	25 °C
SC6	3.7 V	3 A	25 °C

provides detailed information. SC1 represents the first supercapacitor, and so on.

3.2. Simulation Platform

The methodology introduced in this article was implemented using MATLAB 2021b, running on a Windows 10 operating system, and executed on a hardware platform consisting of an Intel Core i5-7300 CPU and NVIDIA GTX 1050 GPU.

4. Predictions and Result Analysis

4.1. Decomposition of Supercapacitor Capacity Sequences Using VMD

In the degradation curves of the supercapacitors, noise and capacity recovery phenomena exist [60,61]. VMD was applied to decompose the supercapacitor capacity sequences into K IMF components, and the predictive results of each component were superimposed as the final output. Gradually increasing K during the training process, this work selected K = 5, corresponding to the optimal tradeoff between training time and prediction error as the final number of IMF components.

Figure 5 shows the decay graph of the capacity of SC5 with the cycle times and the VMD decomposition graph of the capacity decline curve. It shows that the capacity exhibits a degradation trend as the number of cycles increases, and the attenuation of capacity is often accompanied by an increase in the internal resistance of the battery. During the use of supercapacitors, changes in internal resistance can often lead to performance degradation. An increase in the internal resistance will reduce the power output and energy efficiency of supercapacitors. There is a capacity regeneration phenomenon during the capacity drop process. In some cases, supercapacitors may experience partial capacity recovery after a period of rest or appropriate voltage/temperature treatment. IMF1 and IMF2 reflect the overall trend of the original data, IMF3 shows the fluctuation within a smaller time period, and IMF4–IMF5 represent the changes in higher frequency bands.

4.2. Data Processing and Evaluation Index

We found that the cycle number and capacity decay of supercapacitors have a close relationship with a Pearson correlation coefficient of −0.9 or higher, indicating a high degree of negative correlation. Therefore, we included both the cycle number and decomposed capacitance as inputs to our BiLSTM model. Normalization was performed first as follows.

Normalization is a prevalent technique in deep learning for preparing datasets before training models. Its main aim is to neutralize the influence of varying scales of numerical features within the data on the model’s predictive outcomes [62]. For example, the lifespan of supercapacitors is considerable, potentially extending to hundreds of thousands of charge–discharge cycles. In contrast, the capacity of supercapacitors measures around 10 F. The discrepancy in the magnitude between these two data is huge. In addition, normalization can accelerate the speed with which the model processes data [63].

$$y^{\prime }=\frac{y-y_{\min }}{y_{\max }-y_{\min }}$$

(13)

To ensure an accurate evaluation of the model performance, this paper adopted three evaluation indicators, namely the root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE) as the evaluation indicators. The calculation formulas for these indicators are as follows:

$$RMSE=\sqrt{\frac{1}{N}\left({{\displaystyle \sum _{n=1}^N\left|x_{(n)}-{\widehat{x}}_{(n)}\right|}}^2\right)}$$

(14)

$$MAE=\frac{1}{N}\left({\displaystyle \sum _{n=1}^N\left|x_{(n)}-{\widehat{x}}_{(n)}\right|}\right)$$

(15)

$$MAPE=\frac{100\%}{N}{\displaystyle \sum _{n=1}^N\left|\frac{x_{(n)}-{\widehat{x}}_{(n)}}{x_{(n)}}\right|}$$

(16)

4.3. SSA Optimization Results

The quantity of hidden layer units is a critical parameter of BiLSTM. When there are too few units in the hidden layer, the neural network does not learn enough about the information in the training set, which will lead to underfitting. When a neural network has too many hidden layer units, the insufficient information within the training dataset leads to overfitting, as it cannot adequately train all the neurons in the hidden layer [64]. Furthermore, even when the training data is rich in information, the excessive number of units in the hidden layer can extend the training duration, which may hinder the achievement of the desired outcomes. The same is true for epochs, where if they are too small, it can lead to underfitting, and if they are too large, it will lead to overfitting. In addition, a suitable initial learning rate enables the objective function to converge to a local minimum at an appropriate time. Obviously, selecting appropriate hyperparameters is crucial [65,66].

The SSA optimization results in different temperature groups are shown in

Table 2. SSA optimization results under different temperatures.

	Hidden Layer Units	Epochs	Initial Learning Rate
SC1	297	297	0.005
SC2	100	100	0.024179
SC3	239	276	0.019278

. The SSA optimization results in different voltage groups are shown in

Table 3. SSA optimization results under different voltages.

	Hidden Layer Units	Epochs	Initial Learning Rate
SC4	100	300	0.031601
SC5	222	292	0.014609
SC6	268	289	0.020300

.

4.4. Experimental Results and Comparative Analysis

In this experiment, the first 70% of the data was selected as the training set, and the remaining 30% was reserved for the test set. To verify the superiority and robustness of the proposed VMD-SSA-BiLSTM model, the BiLSTM and VMD-BiLSTM methods were used as comparisons. The training and prediction results of the proposed model were compared with the measured values, as well as the BiLSTM and VMD-BiLSTM models, as shown in Figure 6. Figure 6A–C shows the prediction results at different temperatures, and Figure 6D–F shows the prediction results at different voltages. As shown in Figure 6, compared with the BiLSTM and VMD-BiLSTM methods, the capacity prediction curve of the VMD-SSA-BiLSTM method was closest to the true degradation trend. In addition, it was observed that the VMD-BiLSTM method had inferior prediction results to the BiLSTM method for some supercapacitors. This may be due to the larger data size resulting from the decomposition of the capacity degradation data, which led to an underfitting caused by mismatched hyperparameters and input data, such as the number of hidden layer units.

The estimation errors under different temperatures are shown in

Table 4. Estimation errors of the supercapacitors under different temperatures.

		Training Set			Testing Set
		RMSE	MAE	MAPE	RMSE	MAE	MAPE
SC1	BiLSTM	0.043893	0.029867	0.28958%	0.079955	0.065333	0.64824%
	VMD-BiLSTM	0.13671	0.069661	0.68393%	0.18063	0.16537	1.6435%
	VMD-SSA-BiLSTM	0.10247	0.047862	0.47019%	0.031426	0.024544	0.24534%
SC2	BiLSTM	0.046512	0.030149	0.32085%	0.20125	0.18135	1.9963%
	VMD-BiLSTM	0.028876	0.019002	0.19854%	0.15693	0.14837	1.6361%
	VMD-SSA-BiLSTM	0.028669	0.021189	0.22016%	0.070354	0.058781	0.64668%
SC3	BiLSTM	0.050802	0.037722	0.38232%	0.20378	0.18068	1.9025%
	VMD-BiLSTM	0.26284	0.14614	1.5108%	0.26251	0.23696	2.4964%
	VMD-SSA-BiLSTM	0.32929	0.18782	1.9430%	0.061636	0.051728	0.54446%

. The estimation errors under different voltages are shown in

Table 5. Estimation errors of the supercapacitors under different voltages.

		Training Set			Testing Set
		RMSE	MAE	MAPE	RMSE	MAE	MAPE
SC4	BiLSTM	0.035153	0.029692	0.27802%	0.1705	0.15799	1.5159%
	VMD-BiLSTM	0.028173	0.02316	0.21595%	0.18888	0.17493	1.6786%
	VMD-SSA-BiLSTM	0.026636	0.02115	0.198%	0.079715	0.068047	0.65223%
SC5	BiLSTM	0.062386	0.046734	0.43529%	0.11715	0.108	1.0397%
	VMD-BiLSTM	0.048143	0.035024	0.327%	0.16045	0.15339	1.4775%
	VMD-SSA-BiLSTM	0.031661	0.022411	0.21008%	0.068823	0.063384	0.61025%
SC6	BiLSTM	0.19471	0.15851	1.6174%	0.44068	0.35786	4.1353%
	VMD-BiLSTM	0.26529	0.20879	2.2027%	0.40365	0.32478	3.7523%
	VMD-SSA-BiLSTM	0.23935	0.1869	1.9557%	0.36316	0.28924	3.3398%

. Among them, the smaller the RMSE, MAE, and MAPE, the higher the prediction accuracy of the prediction model. Whether at different temperatures or voltages, the RUL calculation results of VMD-SSA-BiLSTM displayed the least amount of error. In the testing set, the minimum RMSE achievable was 0.031426, while the minimum MAE was 0.024544, and the minimum MAPE was 0.24534%.

In different temperature groups, the capacity degradation curves of the supercapacitor were relatively smooth without significant fluctuations. Therefore, the predictive results of VMD-SSA-BiLSTM had a good fitting effect with the actual values.

In different voltage groups, the capacitance degradation curves of the supercapacitors exhibited significant fluctuations. This is due to complex physical and chemical reactions and environmental factors, resulting in capacity recovery phenomena and noise in the degradation curves of the supercapacitors. Figure 6D shows that the prediction results of the BiLSTM model did not reflect the trend of the capacity fluctuations, while the VMD-BiLSTM model and VMD-SSA-BiLSTM reflected the trend of the capacity degradation curve fluctuations. This is due to the use of a sequence of inherent mode functions that characterize local features obtained through VMD decomposition. Moreover, after the optimization of the SSA, the predicted results of VMD-SSA-BiLSTM were closer to the actual value than BiLSTM, which shows that the SSA method supports the optimization of the hyperparameter of BiLSTM to have good results.

Compared with different temperature groups, the degradation curves of the supercapacitors in the different voltage groups showed significant fluctuations, which affected the prediction accuracies to some extent. However, VMD-SSA-BiLSTM still reflected the degradation trend and capacity regeneration phenomenon, and compared with the control group, the accuracy was still the highest.

With respect to intuitively assessing the accuracies of the three approaches, Figure 7 and Figure 8 show the statistical errors of their predictions. Observing these figures reveals that the VMD-SSA-BiLSTM model consistently yielded the smallest error and highest precision across various temperatures and voltages. Other models exhibited limited generalization ability to data under diverse conditions and had large errors.

5. Conclusions

With the continuous improvements in the energy density and power performance requirements in energy storage systems, researchers have invested more and more attention in research into advanced electrochemical energy storage devices. As an important member of these energy storage devices, the condition of supercapacitors is closely related to the safe and stable operation of the entire energy storage system. This article proposed a method that combines VMD decomposition, SSA optimization, and the BiLSTM network to accurately predict the RULs of supercapacitors. VMD decomposition effectively reduced the noise interference in the data through modular processing, such as capacity recovery and testing errors; the BiLSTM network could capture complex features in time series data; the SSA, with its powerful optimization ability and fast convergence speed, was used to optimize the number of hidden layer units, epochs, and the initial learning rate of BiLSTM networks, thereby improving the matching degree between the data and network structure.

In this study, we selected temperature and voltage as the input variables, which had the most significant impacts on the degradation of supercapacitor capacity in the aging experiments of supercapacitors. We input the input variables into the VMD-SSA-BiLSTM model to predict the remaining service life. The results of the experiments showed that the model performed well in predicting the RULs of the supercapacitors and could maintain high prediction accuracy and strong robustness under different temperature and voltage conditions.