Design And Implementation Of Air Quality Prediction System Based On Seq2seq Model

With the continuous advancement of urbanization,when people enjoy the convenient life brought by science and technology,they are also harmed by the negative effects of science and technology.The problems of increasing number of factories,increasing traffic scale and unreasonable utilization of energy have brought great challenges to the atmospheric environment in China.Air pollution is always a difficult problem to solve.Ground air quality monitoring stations can only monitor the coarse-grained air quality in two-dimensional space,which is not enough to meet the needs of fine-grained air quality monitoring.Today,with a lot of tall buildings,more and more people live in middle and high-rise buildings,so there is a demand for air quality monitoring in the three-dimensional space of the city.In order to understand the future air pollution situation in advance and formulate relevant measures to solve the pollution problem as soon as possible,it is necessary to establish a set of three-dimensional air pollution prediction system.In this thesis,the air quality data in three-dimensional space were obtained by UAV and ground monitoring points,and a prediction model based on neural network was constructed to monitor air quality and PM_2.5 concentration prediction in Zhengzhou.Based on the Seq2seq model,the prediction model uses the long short-term memory network(LSTM)as the cyclic unit.Attention mechanism is added to the model to improve the prediction accuracy.The sample data are grouped according to Pearson correlation coefficient,and four feature groups are obtained according to the grouping strategy.The input and output of the encoder and decoder are constructed using the form of encoder fusion+group interaction.After preprocessing the acquired data,the model is trained and compared with several other baseline methods during testing.The mean absolute error(MAE)and root mean square error(RMSE)are used as the evaluation indicators of model performance.The experimental results show that this model is superior to other baseline methods in terms of performance indexes.The verification of each module in the model shows that the three modules of attention mechanism,feature grouping and encoder fusion+group interaction can reduce the error value of the prediction results.