PM2.5 Concentration Prediction Based on Coupling Multivariable Parameters LSTM

As one of the major air pollutants, PM2.5 poses serious threats to human health and daily life. Improving the accuracy of PM2.5 predictions is of significant importance for pollution control, as it can effectively reduce its harmful impact. Long Short-Term Memory (LSTM) networks are capable of handling long-term dependencies and are commonly used for modeling and remembering long sequence data, making them a key tool for PM2.5 prediction. Existing LSTM models often use a single parameter for PM2.5 prediction, without considering the coupling effects of multiple meteorological factors with different pollutants and their concentrations. Additionally, using a single optimization algorithm may lead to optimization parameters getting stuck in local optimal solutions, disrupting the optimization results. In this study, multiple meteorological factors highly correlated with PM2.5 data using Pearson correlation coefficients (PCCs) and pollutants with concentrations were selected to construct a coupled correlation model. Furthermore, to improve the reliability of optimization results, the Sparrow Search Algorithm (SSA) with strong local search capability and the Particle Swarm Optimization algorithm (PSO) with strong global search capability were employed to simultaneously optimize LSTM hyperparameters and obtain the optimal solution. Using hourly meteorological data from multiple stations in Beijing from March 1, 2013, to February 28, 2017, as the data source, the SSA-PSO algorithm was used to obtain the optimal solution for network hyperparameters. The coupled correlation model of multiple meteorological parameters with multiple pollutants and concentrations was established. The research results indicate that considering the coupled correlation model of meteorological parameters with pollutants and concentrations can effectively improve the prediction accuracy of PM2.5, and the more types of input parameters, the higher the model prediction accuracy. The R² of the coupled correlation model with seven parameters increased to 0.9556, and the RMSE decreased by 7.68% compared to the single-parameter model.