Research on air quality prediction method based on spatio-temporal fusion and hyperparameter self-optimization

Accurate prediction of atmospheric pollutant concentrations is an important issue in environmental protection and health management. Aiming at the problem that traditional prediction methods cannot fully utilize the spatio-temporal relationship and complex inter-feature dependencies, this paper proposes a prediction model that integrates spatio-temporal features and hyper-parametric self-optimization (BOT-ConvLSTM). Specifically, the model first uses the (Convolutional Long Short-Term Memory) ConvLSTM network to capture both spatial and temporal dependencies within the data across different stations, and generates a series of feature representations containing spatial and temporal dynamics, and then strengthens the model's sensitivity to temporal patterns through the Temporal Pattern Attention Mechanism (TPA) to accurately capture the changing trends and cyclical characteristics of pollutants. Finally, Bayesian optimization is employed to fine-tune the model's hyperparameters, minimizing the time and errors associated with manual parameter adjustments. In addition, the experimental validation using the PM2.5 concentration data from the monitoring stations in Dalian shows that, compared with the traditional model, the method presented in this paper enhances both prediction accuracy and robustness, which provides a new solution for the prediction of pollutant concentration in the complex spatial and temporal environments.