Research On Prediction Model Of Industrial Pollution Gas Concentration Based On Deep Learning And Time Series

With the acceleration of industrialization and modernization,the phenomenon of air pollution has become more grievous.It is indispensable for the prevention and monitoring of air pollutants that researchers study the composition content of various gas pollutants emitted by industry and predict the change of future concentration.With the development of deep learning technology,it is a very challenging task to construct a prediction model for spatiotemporal prediction tasks and extract the relevant features among Multi-dimensional input data.This thesis takes the concentration of industrial pollutants as the research object,analyzes and excavates the information rule of concentration of pollutant components in the current period.In view of the characteristic mapping relationship between NOx and other related factors of the current high emissions of air pollution,to achieve the prediction of NOx concentration in the future.The main contributions of this thesis are as follows:(1).This thesis proposes a Multi-Att-LSTM prediction model based on Multi-head attention mechanism and Multi-feature variables combined with LSTM network,which is used to train and learn Multi-dimensional time series features,and predict future pollutant concentrations according to historical pollutant concentration data.Variety.The model uses the transformer model structure as the basic baseline,and extends the traditional encoder-decoder structure.The encoder adopts a locally processed Multi-layer LSTM unit structure to extract relevant features of the input data and input them into the decoder.The Decoder also obtains the prediction output through a Multi-layer LSTM module combined with a Multi-head selfattention layer that can be used for long-term dependencies.(2).In this thesis,a local multivariate attention mechanism is proposed to be used before the LSTM module in the encoder-decoder to capture the dynamic correlation and different feature influences between the target sequence and other correlation factors.According to the hidden layer state and output result in the LSTM network module,it is spliced and passed into the Decoder to obtain more comprehensive hidden state information at the current moment,and extract effective features in a targeted manner.(3).This thesis presents an anomaly detection method based on threshold segmentation.The method combined with time series prediction model to calculate outliers and reconstruct errors,update thresholds and adjust weights.Finally,the algorithm improves the prediction accuracy of Multi-variable and high-dimensional data.By analyzing the correlation between NOx concentration of pollutants and other related factors,the factors with greater correlation are selected as multivariate input features for data processing,and a dataset of NOx concentration of industrial pollutants is constructed,and experiments are carried out on multiple datasets.,and compared with other models,the results verify that the Multi-Att-LSTM prediction model has better prediction effect,and the evaluation indicators RMSE,MAE and MAPE have the smallest results and better performance.