Research On Deep Learning Prediction Method For Air Pollution Based On K-Line Pattern Matching Algorithm

Recently,as the effective application of deep learning in time series prediction,a series of significant progresses have been made in atmospheric pollution research.But the phenomenon of atmospheric pollution is still extremely flinty which is largely owing to the deepening of human civilization and economic construction.Upgrading deep learning algorithms to improve the precision of time series prediction in atmospheric pollution,which is a powerful means to achieve residents’thoroughly health of physical and mental,steady improvement of the socio-economic,and sustainable development of the natural environment.In the process of time series prediction,the problem of imprecision in sudden changes prediction and the iterative errors in multi-step prediction have always been pain points that have not been effectively solved.Targeting at the time series with strong discreteness and large dimensions of feature space,a range of methods for the design,implementation,and optimization are discussed in the article,including K-line pattern matching model,local to global correlation feature extraction model,and continuous time dependence feature extraction model.And these methods have been applied to the prediction of atmospheric pollution time-series.The specific work is shown below:1.A method based on K-line diagrams to represent the original sequence and analyze the data with the technical indicators is proposed to address the issue of current time series data prediction models being insensitive to mutation signals leading to untimely prediction.The experimental results show that in short-term prediction,compared to the traditional time series data prediction methods,the root mean squared error of the proposed method is reduced by 48%,which can reflect that data K linearization both comprehensively characterizes the trend of time series data changes,and can express the fluctuation amplitude of the data by means of its unique shadow attribute.The design effectively alleviate the problem of inaccurate prediction caused by model lag in sudden change value prediction.2.A historical similar K-line search method based on pattern matching is designed aiming at the problem of error accumulation caused by iterative training in multi-step prediction of current time series data prediction models.Experimental analysis shows that in long-term prediction,compared to the traditional time series data prediction methods,the root mean squared error of the designed method is reduced by 54%,which can reflect that the algorithm of pattern matching can integrate posterior global trend features.To some extent,the method can avoid the problem of severe deviation yet in prediction results due to error accumulation in multi-step prediction.3.A feature extraction model incorporating deep learning methods is constructed to address the issue of poor feature capture capabilities in traditional pattern matching techniques.And a processing of image feature enhancement is carried out in response to the problem of fewer pixels in the K-line map in similar historical temporal patterns after matching,which can result in unclear features.In the meantime,a global average pooling blocks is used to prune the model and perform dimensionality reduction on feature maps.The feature extraction process focuses on the main features,which can extremely reduce the number of model parameters by 71%and greatly accelerate the speed of model training and convergence.In this article,a range of prediction methods with deep learning algorithm are analyzed for time series.Based on the work,a multi-step PM_2.5 prediction combination model is constructed with the introducing of K-line analysis and pattern matching techniques,and combined with the diffusion mechanism of atmospheric pollution.Through comparative experiments,it has been shown that the designed and combined algorithm model in the paper meets the requirements of scientific nature,integrates physical and chemical diffusion mechanisms,overcomes the limitations of traditional deep learning prediction methods,settles the pain spots in prediction for atmospheric pollution time series.Ultimately,it drives the model to achieve higher accuracy.