مدل سازي غلظت آلودگي هوا در محيط ناهمگون و ديناميك شهري بر پايه يادگيري تقويتي

Exposure to air pollutants, particularly PM2.5, is inevitable and poses a significant threat to human health. These fine particles (with an aerodynamic diameter of less than 2.5 micrometers) quickly penetrate the respiratory system and can cause irreversible effects, such as lung cancer and cardiovascular diseases. PM2.5 is recognized as the primary pollutant contributing to Tehranʹs air pollution. Accurate modeling of this pollutant is crucial in reducing exposure and formulating effective policies to safeguard public health. Modeling PM2.5 concentrations in urban environments is challenging due to the dynamic and heterogeneous nature of these settings and the uncertainties in spatial and temporal relationships among PM2.5 monitoring stations. This dissertation aims to improve the accuracy of PM2.5 concentration modeling by addressing missing data, reducing uncertainties in inter-station relationships, and accounting for the dynamics and heterogeneity of urban environments. To manage missing data, various single and multiple imputation methods were examined and eva‎luated to identify the most accurate approach. Ultimately, missing air pollution data were imputed using the XGBoost method, incorporating meteorological variables, pollutant concentrations, and temporal variables (such as hour and day of the week). These data were collected from twelve air pollution monitoring stations and three synoptic stations in Tehran, covering the period from March 21, 2017, to November 11, 2022. Accurately modeling PM2.5 concentrations requires simultaneous extraction of spatial and temporal relationships among stations due to the spatiotemporal nature of its distribution. To reduce uncertainties in spatial relationships among stations, convolutional neural networks (CNN) and graph convolutional networks (GCN) were employed. This research proposes the Dynamic Spatial-Temporal Hybrid Attention Network (DST-HAN) to enhance modeling accuracy over 1-hour to 24-hour intervals. DST-HAN consists of two developed components: Spatial-Temporal Relationship Extraction: This component integrates the Spatial-Temporal Hybrid Attention Network (STHAN) and the Attention Temporal Graph Convolutional Network (A3TGCN). STHAN utilizes CNNs to extract spatial relationships, making it suitable for Euclidean spaces. Conversely, A3TGCN uses GCNs to capture spatial relationships in non-Euclidean spaces, making it ideal for complex and heterogeneous environments. For temporal relationship extraction, a combination of gated recurrent units (GRU) and attention mechanisms was employed. This integration enables the model to identify short-term and long-term temporal dependencies in time-series data across stations, thereby improving modeling accuracy.Dynamic Weight Adjustment: To address the dynamics and heterogeneity of urban environments, the Deep Q-Network (DQN) algorithm was used to dynamically and optimally adjust the weights of the hybrid models. This algorithm analyzes large datasets and continuously interacts with the environment, assigning weights to the hybrid models based on rewards received from the environment. Models with better performance receive higher weights. Unlike traditional methods with fixed weighting, this approach identifies dynamic changes and station-specific features, incorporating them into the modeling process.The results demonstrate that the DST-HAN model predicts PM2.5 concentrations with high accuracy (R² = 0.88 for the next hour and R² = 0.81 for the next 24 hours) and can serve as an effective tool for managing exposure to this pollutant.

5 فصل