An evolutionary method for transaction fee prediction based on NBEATSx optimized by DCNo-GWO

Transaction fees, mandated by blockchain protocols for the sender of a transaction, provide a sensitive measure of congestion within blockchain networks and the quality of services offered. Predicting these fees is a challenging task in time series modeling, and recurrent neural network-based methods commonly used in this domain often suffer from performance limitations due to insufficient incorporation of effective exogenous variables and underexplored hyperparameter optimization. To address these issues, we propose a novel heuristic algorithm-based approach for transaction fee prediction. Specifically, we develop an architecture based on Neural Basis Expansion Analysis with exogenous variables (NBEATSx) to model transaction fee time series, integrating ten relevant external variables. A distributed evolutionary algorithm, named the Distributed Chaotic Nonlinear Grey Wolf Optimizer (DCNo-GWO), is introduced to simultaneously optimize thirteen hyperparameters of NBEATSx and select informative exogenous variables. Experiments are conducted on two Ethereum transaction datasets, spanning 8/7/2015–11/29/2023 and 5/29/2024–11/29/2024, with 3037 and 185 daily samples, respectively. Results show that DCNo-GWO achieves effective convergence, demonstrating scalability in hyperparameter optimization. The combined DCNo-GWO and NBEATSx method outperforms six baseline models, reducing MAPE by 16.86% and 2.07% on the two datasets. These findings highlight the practical significance of our approach, confirming its ability to meaningfully improve prediction accuracy in real-world blockchain transaction scenarios.