Attention-based CNN-LSTM and XGBoost Hybrid Model for Air Quality Prediction

Accurately predicting air quality is of significant importance for environmental management, economics, and public health. Over the past few decades, scientific algorithms for understanding and forecasting weather have continuously advanced, leading to the emergence of many excellent algorithms and applications. Due to the complex fluctuations in the environment, Conventional time series models like ARIMA have proven inadequate in capturing the nonlinear characteristics observed in air quality data. Neural networks possess powerful nonlinear generalization capabilities. In this paper, we propose the use of a hybrid model based on the attention mechanism of CNN-LSTM and XGBoost for predicting air quality based on five years of data from Wuhan City. We conduct experimental analysis and compare it with other models. Firstly, ARIMA is used for preprocessing, then a deep learning architecture is formed using a pre-training and fine-tuning framework. Finally, fine-tuning is performed on the XGBoost model. Compared with ARIMA (R2: 0.65758) and ARIMA-BiLSTM (R2: 0.61897), AttCLX achieves an R2 of 0.98887, indicating a significant improvement. By more accurately predicting air quality, we can better protect the environment, promote economic development, and safeguard public health. This has significant implications for government environmental policy-making, business planning of production activities, and individual health management.