Advancement in Supercapacitors for IoT Applications by Using Machine Learning: Current Trends and Future Technology

Carbon materials with a substantial specific surface area, pores, electrical conductivity, and thermal resistance are often considered as suitable choices for hybrid supercapacitors. The efficacy of SCs is affected by non-linear effects resulting from variations in both their structural and operational characteristics. In the past decade, there has been a growing use of ML algorithms to gain insights into the correlation among the performance of carbon-derived supercapacitors and various structural and operational characteristics. Liu et al. [92] suggested a method for predicting the capacitive behavior of carbon-based materials by investigating the correlation among their structural characteristics and capacitance. A dataset of 105 diverse carbon compounds, each defined by 11 structural factors, was obtained with the aim of predicting capacitance. This work used several models, namely RFs, Gradient Boosting Machines (GBMs), and Extreme Gradient Boosting (XGB), for the purpose of predicting capacitance. Furthermore, the use of supervised learning models, including ANNs, SVMs, and MLR, have been employed for the same objective. However, the complexity of ensemble models may hinder the interpretability of the results. Zhou et al. [93] used four regression models, including ANNs, RFs, SVMs, and Generalized Linear Regression (GLR), to forecast the essential characteristics of activated carbon that provide the highest amount of energy and power density. The input parameters that were utilized for model training consisted of the scan rate and the surface area of micropores and mesopores. The output parameters that were implemented were capacitance and power density. The use of an ANN model exhibits superior efficacy in determining the value of capacitance for activated carbon, hence establishing a greater correlation between the predicted and measured power density values. In addition, an ANN model predicts that the highest energy density may be attained by activated carbon molecules with a surface area of 920 m²/g for micropores and 770 m²/g for mesopores, respectively. The ANN model excels at recognizing the complex patterns and relationships within large datasets, enabling the extraction of valuable insights of supercapacitor performances.

Tawfik et al. [94] investigated the machine learning methods that provide direct, efficient, and accurate forecasting skills in the context of designing materials for porous carbon supercapacitor electrodes. A total of 260 distinct carbon-based electrodes were obtained, each exhibiting unique morphologies. Various machine learning methods were examined to forecast the capacity of the porous carbon supercapacitors, including artificial neural networks with diverse architectures, Lasso, and support vector machine models. Their findings demonstrated that the artificial neural network with two hidden layers has superior performance in the context of SCs, as shown by the RMSE, MAE, and R values of 28.67, 37.59, and 0.895, respectively, while the XGB shows the second-highest performance after the ANN with R value of 0.892, as shown in Figure 7a,b. The performance and relative contribution of the predictive features are shown in Figure 7c. It shows the importance of the SSA, which affects more in terms of the output capacitance. Furthermore, Saad et al. [95] used several ML models, including k-nearest neighbors’ regression, Bayesian ridge regression, decision trees, and artificial neural networks, to achieve precise predictions of graphene’s capacitance. These models were trained using diverse electrochemical and physiochemical characteristics of graphene. The ANN model has superior performance compared to the other three models. The ANN suggested by Lu et al. [96] was utilized to predict the capacitance of the carbonized metal–organic framework. This study involved a comparison between the experimental capacitances and those predicted by the ANN. The results demonstrated that the suggested model surpasses in its ability to forecast the capacitance of supercapacitors, exhibiting the lowest error range of 0.02% to 1.05%.

Chemical doping using heteroatoms, such as boron, nitrogen, sulfur, phosphorus, and other elements, may greatly enhance the capacitive abilities of carbon-based materials [97,98,99,100]. Consequently, there has been a shift in focus towards using ML models to forecast the effectiveness of SCs, considering the impact of doping materials as a percentage, as well as other structural and operational characteristics. Mishra et al. [101] evaluated the impact of heteroatom doping composition and structural characteristics of carbon materials on the effectiveness of capacitance using ML models. A comprehensive dataset of 147 carbon-based supercapacitor sets was compiled from the existing literature. This dataset encompasses several input parameters, such as current density, pore volume, pore size, presence of defects, potential window, specific surface area, oxygen, and nitrogen content, of the carbon-based electrode material. The regression analysis of the target specific capacitance from the physicochemical features of the SCs involved the implementation of five distinct approaches. These approaches encompassed the Ordinary Least Square Regression (OLS) method, as well as several data-driven techniques, namely SVMs, DTs, RFs, and XGB. The XGB and RF performed better, which is evident by the R² values of 0.79 and 0.75, respectively. They used different carbon electrodes to produce the proportional impact of any given input parameter on the capacitance using the trained XGB model. Moreover, Zhu et al. [102] used ML methods, including ANNs, linear regression, and Lasso, to determine the capacitance of carbon materials. A comprehensive data collection consisting of 681 supercapacitors based on carbon materials has been compiled from over 300 scholarly articles. This ML model was trained using five input features, specifically the specific surface area, PS, ID/IG (intensity ratio of D and G bands), N-doping level, and voltage window. The results indicated that an ANN model outperforms both the linear regression and Lasso models, as shown by its higher R² value of 0.91.

Machine learning methods are also used to forecast efficiency and optimizing the design of pseudocapacitive SCs, which include oxides and composite materials. Numerous scholars have conducted investigations on the utilization of artificial intelligence models for the purpose of effective prediction, pertaining to various types of pseudocapacitive supercapacitors. Mathew et al. [103] used a ML model to investigate the impact of the electrode area on the efficiency of the hybrid supercapacitor. The design of hybrid supercapacitor electrodes involves the use of both MnO₂ and activated carbon materials. These electrodes are configured in rectangular and square geometries, with their areas being subjected to variation. The findings suggest that electrodes with a rectangular form exhibit improved qualities in comparison to electrodes with a square shape. The model that has been presented demonstrates reliability in its predictive capabilities, as shown by its mean squared error value of 0.02. The ML model is well-suited for handling large and diverse datasets of the SC, whereas the conventional methods struggle to extract meaningful information. Similarly, Lokhande and Chavan [104] conducted an investigation of the cyclic behavior of the Ni(OH)₂ supercapacitor electrode that was created using the ML model. The electrochemical test of the constructed electrode was conducted at various scan rates, and the resulting experimental data were used to train the model. The calculation of the specific capacitance involves the consideration of several factors, including instantaneous current, active mass, scan rate, and potential window. The ANN model has superior performance, as seen by its low percentage error of 0.14%. Additionally, Alimi et al. [105] investigated an ANN model to achieve precise forecasting of the CV characteristics of trijunction supercapacitor electrodes composed of MnO₂, NiO, and ZnO. The obtained R² value of 0.999 indicates a high level of accuracy in their predictions. Table 2 presents a summary of the literature studies on the machine learning applications for predicting the capacitance of supercapacitors.

The degradation of supercapacitors is affected by several variables, including temperature, voltage, and the materials used for the electrode and separator [125]. To effectively manage the intricate aging state system of supercapacitors, it is essential to use indirect monitoring and prediction techniques. Two crucial indicators, state of health (SOH) [126] and remaining usable life (RUL) [127], are identified among them. Supercapacitors possess a comparatively extended operational lifespan in comparison to other energy storage technologies. However, it is important to note that their lifespan is subjected to limitations imposed by external stress factors encountered during actual use. The primary elements that contribute to the aging of supercapacitors are electrical stress, namely voltage and current, and thermal stress, specifically temperature. SCs consist of electrodes, electrolytes, diaphragms, and fluid collectors. Therefore, the aging characteristics of supercapacitors often include shell damage, electrolyte breakdown, and electrode deterioration [128]. The accurate and timely monitoring of the SOH and RUL of supercapacitors is essential for the precise assessment of their aging process [129,130].

Machine learning prediction models have been developed for the purpose of monitoring the state of health of supercapacitors in the respective area. The aging process of supercapacitors is influenced by a multitude of variables. The hybrid genetic algorithm (HGA) was presented by Zhou et al. [131] to enhance the long short-term memory (LSTM) model. In this approach, the genetic algorithm incorporates sequence quadratic programming as its local search operator, leading to the development of a novel recurrent neural network (RNN) model [91]. The experimental findings demonstrated that the HGA-LSTM model exhibited superior prediction accuracy, with a prediction error of less than 2%, as well as enhanced resilience compared to both the separate LSTM and HGA models. However, HGA-LSTM lacks interpretability, making the results challenging to understand the reasoning behind the prediction accuracy. In addition, a prediction approach was developed by Wang et al. [132], which utilizes a novel variant of LSTM in conjunction with the Adam and Dropout algorithms. This study aims to estimate the status of the cycle aging of supercapacitors in various operating conditions. The experimental findings demonstrated that the newly developed LSTM-RNN exhibited superior prediction accuracy, as shown by an RMSE of 0.0261. Furthermore, Weigert et al. [133] employed a neural network with full connectivity to forecast the cycle life of hybrid electric automobiles that utilize battery–supercapacitor technology. In addition, they examined the main factors that contribute to the aging of SCs. The RUL of the SC is determined by the examination of a brief charge–discharge curve, resulting in a strong correlation coefficient of 0.95 for the predicted results. The literature review on machine learning applications for predicting the RUL and SOH of SCs is shown in Table 3.

Deploying ML models requires significant upfront expenditures in technology, infrastructure, and proficient personnel. The collecting and processing of big datasets of energy storage devices, especially supercapacitors, the development of models, and the upkeep of computing resources, incur significant expenses. Continual costs include the ongoing surveillance, upkeep, and improvements required to ensure that the ML models remain current and efficient. Operational expenses are influenced by regular updates, data quality assurance, and system monitoring. The economic importance of storing and processing vast amounts of data on a large scale cannot be emphasized enough, necessitating strong expenditures on infrastructure [142]. Furthermore, the issue of energy consumption becomes relevant when dealing with intensive processing requirements, especially in the case of deep learning models. This highlights the need to implement efficient strategies to control energy costs in large-scale installations.