A Bayesian‐Optimized GRU Network With WMD‐EFAST Self‐Organizing Mechanism for Multimodal Atmospheric Pollutant Prediction

Abstract

Atmospheric pollutant concentration prediction is a critical component of environmental governance. Deep learning‐based methods now dominate this prediction field, yet current deep learning models still face several challenges: inefficient hyperparameter optimization leading to high training costs, static network structures struggling to adapt to the nonlinear spatiotemporal coupling between pollutant concentrations and meteorological factors, and these limitations consequently causing performance fluctuations in cross‐timescale predictions. To address these challenges, this paper proposes a Bayesian‐optimized self‐organizing gated recurrent unit (BO‐SOGRU) model, achieving dual synergy between Bayesian optimization and dynamic self‐organizing mechanisms. First, a hyperparameter self‐configuration mechanism is constructed within the Bayesian optimization framework, where a Gaussian process surrogate model performs global optimal search for key parameters, significantly improving hyperparameter optimization efficiency. Second, a dynamic self‐organizing mechanism is designed, employing Weighted Minkowski Distance (WMD) to measure neuron similarity and integrating the Extended Fourier Amplitude Sensitivity Test (EFAST) to dynamically evaluate neuron contributions, thereby enabling adaptive network structure optimization. Convergence analysis verifies the algorithm's stability and feasibility. Experiments were conducted on two benchmark datasets and real‐world pollutant concentration prediction tasks across multiple time scales, and the prediction accuracy was calculated. Compared with other models of the same type, BO‐SOGRU has significant advantages in prediction accuracy and model compactness.