Research on AQI prediction of Chengdu-Chongqing economic circle based on CNN-BiLSTM-Selfattention model

¹ College of Culture and Mathematics, Chengdu College of University of Electronic Science And Technology of China, Chengdu, China
² College of Mathematics and Physics, Chengdu University of Technology, Chengdu, China

^* Correspondence: changhd1998@163.com

Abstract

Air pollution has emerged as a significant environmental challenge worldwide. The Chengdu- Chongqing economic circle is central to regional development in China. Research into predicting air quality aims to support sustainable development efforts in China and across the globe. Due to the chaotic, disordered, and non-stationary nature of the Air Quality Index (AQI) data, traditional statistical forecasting models are inadequate for AQI predictions. Therefore, this study focuses on the AQI of 16 cities at or above the prefecture level within the Chengdu-Chongqing economic circle and identifies six major pollutants, including PM2.5, PM10, carbon monoxide (CO), and sulfur dioxide (SO2), as key contributors to AQI levels. To analyze AQI data characteristics, the K-Shape clustering method is initially employed to categorize the 16 cities in the Chengdu-Chongqing economic circle. Following this, a CNN-BiLSTM-Selfattention prediction model is developed, integrating the CNN, BiLSTM, and Selfattention models to forecast the AQI for both high- representative and low-representative cities in the region. Additionally, the performance of the CNN- BiLSTM-Selfattention model is compared with that of the BiLSTM model, CNN-LSTM model, and CNN- BiLSTM model to validate its accuracy. Finally, the CNN-BiLSTM-Selfattention model is utilized to project the AQI for the 16 cities within the Chengdu-Chongqing economic circle over an eight-day period from November 12, 2023, to November 19, 2023. The findings indicate that: (1) Utilizing the K-Shape clustering technique, Chengdu and Neijiang emerge as the cities with high AQI representation in the Chengdu- Chongqing economic region, whereas Yibin and Luzhou are identified as cities with low representation. (2) A comparison of the RMSE, MSE, MAPE, MAE, and R2 values across the four models reveals that the CNN- BiLSTM-Selfattention model demonstrates superior prediction accuracy and enhanced stability. (3) The forecast analysis suggests that while certain days experience significant air quality pollution in the Chengdu- Chongqing economic circle, the overall air quality exhibits a trend towards improvement, with pollution indices across most areas remaining below level 3.