摘要 : With the deployment and real-time monitoring of a large number of micro air quality monitoring stations, new application scenarios have been provided for the research of air quality prediction methods based on artificial intellige... 展开 With the deployment and real-time monitoring of a large number of micro air quality monitoring stations, new application scenarios have been provided for the research of air quality prediction methods based on artificial intelligence. Integrating deep learning with multi-task learning, this paper proposes a hybrid model for air quality prediction to leverage data from intensive air quality monitoring stations. The proposed model consists of a shared layer, a task-specific layer, and a multi-loss joint optimization module. It is tested on three monitoring stations located in three different districts of Lanzhou City, China, for PM2.5 concentration prediction. The results show that: (1) When the number of convolutional layers of convolutional neural network in the shared layer and the number of gated recurrent unit layers in the task-specific layer exist in two layers, model performs the best, and its predictability of the optimization algorithm with early-stopping will be significantly improved. (2) Using the proposed model to predict PM2.5 concentration on horizon t + 1, the mean absolute error and root mean square error are 4.54 and 7.96, respectively, indicating better performance in intensive air quality prediction than previous models based on simple hybridization. (3) The predictive performance on different stations is different, and the proposed model performs better than other models when there are large fluctuations and sudden changes in the data. Overall, the proposed model has good temporal stability and generalization ability and provides a new method for air quality prediction in intensive air quality monitoring scenarios. (C) 2020 Elsevier Ltd. All rights reserved. 收起