Research On PM_2.5 Concentration Prediction Based On Empirical Mode Decomposition And Deep Learning Method

In recent years,many areas in our country have been seriously affected by the haze climate,which has harmed the ecological environment and people’s health.Excessive PM_2.5concentration is one of the main causes of haze.Exploring the law of PM_2.5concentration changes has become the main content of air quality prediction research.Accurate and effective air pollutant concentration prediction methods help prevent the impact of haze climate,and at the same time provide reference value for formulating air pollution control strategies.The prediction of air pollutants in the time dimension can be regarded as a problem of multivariate time series prediction,but previous studies of the prediction of air pollutants had not taken the non-stationarity of air pollutants as time series,and researchers have often overlooked the impact of non-stationarity in time series predictions.This made the prediction accuracy of the model lower due to the limited prediction performance.Therefore,this paper proposes a PM_2.5concentration prediction method based on Empirical Mode Decomposition（EMD）and Gated Unit Recurrent Neural Network（GRU）.Firstly,in order to identify the sequence characteristics of the forecast target,this paper used the construction of statistical test quantities and unit root test methods to test the PM_2.5concentration series for time series stationarity and non-linearity.Secondly,in view of the impact of the unevenness and nonlinearity of the PM_2.5concentration sequence in the prediction process,this paper proposed a PM_2.5concentration prediction model based on the EMD-GRU model.The model used the EMD decomposition method to decompose the PM_2.5concentration sequence into multiple stationary sub-sequences,and then in turn the sub-sequences were combined with meteorological features into the built multi-step of two-layer GRU neural network for training and prediction.The sub-sequence values of the prediction output were summed to obtain the final prediction results.The prediction model proposed in this paper can train the marked historical data features,and the PM_2.5concentration data of the next hour can be output after being tested by the trained model.Finally,this paper used the PM_2.5concentration prediction of the air quality monitoring station of the Capital Airport as an example to conduct experimental verification based on the deep learning method and the hybrid model based on empirical mode decomposition and GRU neural network（EMD-GRU）.