Data-based modeling for prediction of supercapacitor capacity: Integrated machine learning and metaheuristic algorithms

Abstract

Background

Accurately predicting the specific capacity of supercapacitors (SCs) is essential for improving their energy efficiency and performance. This requires robust methods to model the complex, nonlinear relationships among variables.

Methods

In this study, the dataset was divided into three optimal clusters using k-means, based on supercapacitor capacity, each displaying distinct features. Additionally, the unclustered dataset was also analyzed. The training of Multi-Layer Perceptron (MLP) neural networks was examined using six metaheuristic algorithms. Neural network hyperparameters were optimized via grid search, and metaheuristic algorithms via random search. Performance, convergence, and adaptability were evaluated for clustered and unclustered datasets, focusing on accuracy, speed, and generalization.

Significant findings

The cluster-based MLP models demonstrated exceptional predictive accuracy, outperforming unclustered models. Notably, the MLP integrated with Invasive Weed Optimization (MLP-IWO) in cluster 2, with a population size (Np) of 40, achieved the highest coefficient of determination (R²=0.9998), representing a 105.53 % improvement compared to the best unclustered model (R² = 0.4864). Similarly, the MLP integrated with the Firefly Algorithm (MLP-FA) in clusters 1 and 3 (Np = 30) achieved R² values of 0.9983 and 0.9927, respectively. These findings highlight the effectiveness of integrating clustering with metaheuristic optimization for enhancing prediction accuracy in SCs capacity modeling.