A PM2.5 concentration prediction model based on multi-task deep learning for intensive air quality monitoring stations

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