Search

CN-121995007-A - PM based on STL-VMD-transducer hybrid deep learning2.5With O3Collaborative prediction method

CN121995007ACN 121995007 ACN121995007 ACN 121995007ACN-121995007-A

Abstract

The invention discloses a PM 2.5 and O 3 collaborative prediction method based on STL-VMD-transform mixed deep learning, which comprises the steps of collecting historical data of target predicted pollutant data, historical data of contemporaneous other pollutant data and meteorological data, preprocessing the historical data to obtain input independent variables, enabling the target predicted pollutant data to comprise PM 2.5 concentration data and O 3 concentration data, enabling time series data to be subjected to STL decomposition into trend items, season items and residual items, enabling the residual items decomposed by the STL to be subjected to iteration decomposition by means of a variation mode decomposition algorithm VMD to obtain residual sets with set number modes, training a transform decoder to output converged transform decoder, collecting new other pollutant data and gas image data, and inputting the new other pollutant data and gas image data into the converged transform decoder to output predicted PM 2.5 concentration data and O 3 concentration data by means of a full-connection layer.

Inventors

  • LIU YULU
  • LI RUI
  • CEN WANGLAI

Assignees

  • 成都信息工程大学

Dates

Publication Date
20260508
Application Date
20260127

Claims (5)

  1. 1. The PM 2.5 and O 3 collaborative prediction method based on STL-VMD-transducer mixed deep learning is characterized by comprising the following steps: S1, collecting historical data of target predicted pollutant data, historical data of other synchronous pollutant data and meteorological data, preprocessing, calculating correlation between the target predicted pollutant data and other pollutant data and meteorological data, and screening out a key variable set with obvious influence on the target predicted pollutant data to obtain an input independent variable, wherein the target predicted pollutant data comprises PM 2.5 concentration data and O 3 concentration data; S2, constructing time series data of historical data of target predicted pollutant data, and performing STL decomposition on the time series data into trend items, season items and residual items; s3, continuing to carry out iterative decomposition on residual items decomposed by the STL by using a variation mode decomposition algorithm VMD to obtain residual sets of set number of modes; S4, taking the residual error set, the trend item and the season item as output dependent variables, combining the input independent variables and the output dependent variables to form a training data set, constructing a transducer decoder, inputting the training data set into the transducer decoder, training the transducer decoder, and outputting a converged transducer decoder; And S5, collecting new other pollutant data and gas image data, inputting the new other pollutant data and gas image data into a converged converter decoder, and outputting predicted PM 2.5 concentration data and O 3 concentration data by using a full connection layer.
  2. 2. The method for collaborative prediction of PM 2.5 and O 3 based on STL-VMD-transducer mixed deep learning according to claim 1, wherein step S1 comprises: s11, collecting historical data of target predicted pollutant data, historical data of other pollutant data and meteorological data in the same period; s12, carrying out missing value and abnormal value processing on the historical data, and then carrying out data normalization processing to obtain a normalized value of historical data of target predicted pollutant data Normalized values of historical data of contemporaneous other contaminant data and meteorological data ; S13, calculating the Speermann correlation coefficient between target predicted pollutant data and contemporaneous other pollutant data and gas image data by using a normalization value to obtain a symmetrical correlation coefficient matrix, and intuitively displaying a correlation coefficient matrix by using a thermodynamic diagram; ; Wherein, the As a correlation coefficient of PM 2.5 concentration data and O 3 concentration data, As a correlation coefficient of PM 2.5 concentration data and NO 2 concentration data, As a correlation coefficient of the O 3 concentration data and the PM 2.5 concentration data, As a correlation coefficient of the O 3 concentration data and the NO 2 concentration data, For the NO 2 concentration data and the PM 2.5 concentration data, Correlation coefficients of NO 2 concentration data and O 3 concentration data; S14, screening out correlation coefficients associated with target predicted pollutant data in the correlation number matrix And calculate each correlation coefficient P value of (2); ; Wherein, the For the test statistic corresponding to the j-th other contaminant data and the meteorological data, df is the degree of freedom, P-value corresponding to j-th other pollutant data and gas image data; S15, setting a significance threshold of statistical test If (if) Judging that the j-th other pollutant data and the gas image data have obvious influence on the target predicted pollutant data, otherwise, judging that the j-th other pollutant data and the gas image data have no obvious influence on the target predicted pollutant data; and S16, screening out other pollutant data and meteorological data which have obvious influence on the target predicted pollutant data, taking the other pollutant data and meteorological data as a key variable set of the target predicted pollutant data, and acquiring normalized values of historical data of other pollutant data and meteorological data in the key variable set, and taking the normalized values as input independent variables.
  3. 3. The method for collaborative prediction of PM 2.5 and O 3 based on STL-VMD-transducer mixed deep learning according to claim 2, wherein step S2 comprises: S21, setting the seasonal period length P and the sliding window width of trend fitting And seasonal fit sliding window width ; S22, using weighted regression smoothing algorithm LOESS to time series data Smoothing to obtain initial trend item ; S23, calculating time series data With initial trend terms The differences between the sequences to obtain trended sequences The trended sequence is removed Grouping according to the seasonal period length P, extracting trending sequences of all time points belonging to the corresponding seasonal position of each group j ; S24, utilizing a weighted regression smoothing algorithm LOESS to remove the sequence after trend Smoothing to obtain a smoothed sequence of the season position And will smooth the sequence Expanding to full-length sequence data to obtain periodic preliminary season term And meet the following ; S25, preliminary season items of different season position points Considered as a time sequence, and the weighted regression smoothing algorithm LOESS is utilized to carry out smoothing again in the time direction so as to obtain stable periodic season items ; S26, calculating time series data And periodic seasonal term The difference between them to obtain the seasonal sequence Then the weighted regression smoothing algorithm LOESS is utilized to remove the seasonal sequence Smoothing to obtain new trend item ; S27, utilizing new trend items And periodic seasonal items Calculating the current residual Defining temporary residual error scale And calculate the standard deviation corresponding to the current residual error ; ; Wherein, the Is a median function; s28, utilizing standard deviation Calculating the current Lu Bangquan weight W t ; ; Wherein, the Is Bisquare weight functions; S29, updating a weighted regression smoothing algorithm LOESS based on the robust weight W t , returning to step S23, and using trend terms Replacement of initial trend terms And steps S23-S26 are executed to output new trend items And periodic seasonal items ; S210, cycling steps S23-S29 until the trend term and the periodical season term converge or reach the set cycling times, and outputting the final trend term Season term Sum and residual terms 。
  4. 4. The method for collaborative prediction of PM 2.5 and O 3 based on STL-VMD-transducer mixed deep learning according to claim 3, wherein the variational modal decomposition algorithm VMD decomposes residual terms The method of (a) comprises the following steps: S31, residual error item Input as a variational modal decomposition algorithm VMD Initializing VMD parameters of a variation modal decomposition algorithm: Initial modality : Initial center frequency of mode : Initial Lagrangian multiplier K is the modal number; s32, input Fourier transforming to obtain signal components of different modes Using signal components of different modalities Updating the mode, and updating the center frequency and Lagrange multiplier of the mode; ; where w is the center frequency of the mode, i is the mode number of the current iteration process, Modality i updated for the n +1 iteration process, For the mode i of the nth iteration process, Is the lagrangian multiplier for the nth iteration process, For the center frequency of the kth modality of the nth iteration process, In order to be a bandwidth constraint strength, The center frequency of the kth modality is updated for the n +1 th iteration process, The kth modality updated for the nth iteration process, The lagrangian multiplier updated for the nth iteration process, Is noise tolerance; s33, judging whether the updated mode of the iterative process meets the convergence condition; if it meets Judging convergence of the iterative decomposition process, outputting signal components of each mode to obtain residual error sets of K modes Otherwise, judging that the iterative decomposition converges, and continuing the iterative decomposition process; To converge tolerances.
  5. 5. The method for collaborative prediction of PM 2.5 and O 3 based on STL-VMD-transducer mixed deep learning according to claim 4, wherein step S4 comprises: S41, collecting residual errors Trend item And season items As output dependent variables, the input independent variables and the output dependent variables are combined to form a training data set; S42, constructing a transducer decoder, wherein an N-layer decoder stack is constructed in the transducer decoder, each layer decoder comprises a multi-head self-attention mechanism and a feedforward neural network, and the multi-head self-attention mechanism comprises attention heads focusing on residual mutation and synergistic effect in residual error collection and trend terms Attention head and attention season term of middle day change period An attention head for the duration of the contamination process; s43, inputting the training data set into a transducer decoder, training the transducer decoder, acquiring front-back time information of target predicted pollutant data and other pollutant data and meteorological data in the key variable set through a multi-head self-attention mechanism, inputting the front-back time information into a feedforward neural network, and outputting a predicted output module vector of the transducer decoder ; ; Where u is the number of predicted output variables of the transducer decoder, A predicted output variable for the transducer decoder; S44, outputting the prediction module vector Input/output layer for outputting predicted value of PM 2.5 concentration data and O 3 concentration data respectively ; ; Wherein, the As an activation function of the fully connected layer, As a function of the regularization, Activating a function for a Gaussian error linear unit; s45, calculating a predicted value Mean square error with measured PM 2.5 concentration data and O 3 concentration data And carrying out weighted summation to obtain the predicted total loss ; ; Wherein, the Predictive weights for PM 2.5 concentration data and O 3 concentration data; s46, setting a predicted error loss threshold If (if) And judging that the training process is converged, outputting a converged converter decoder, otherwise, not converging the training process, and continuously training the converter decoder until the training process is converged, and outputting the converged converter decoder.

Description

PM 2.5 and O 3 collaborative prediction method based on STL-VMD-converter mixed deep learning Technical Field The invention relates to an air quality prediction neighborhood, in particular to a PM 2.5 and O 3 collaborative prediction method based on STL-VMD-transducer mixed deep learning. Background With the rapid development of urban and industrialized production, the problems of composite air pollution characterized by fine particles (PM 2.5) and ozone (O 3) are increasingly prominent, PM 2.5 pollution is characterized by 'stable winter', reverse temperature in basin, Under the unfavorable diffusion conditions of high humidity and static wind, the particles discharged at the first time and converted at the second time are extremely easy to accumulate, and haze pollution is formed. O 3 pollution belongs to 'summer photochemistry', and is not directly discharged, but is a product of complex photochemical reaction of precursors (volatile organic compounds VOCs and nitrogen oxides NO x) under high temperature and strong solar radiation. the two pollutants are not only beneficial to human health, The ecological system constitutes a serious threat, the formation mechanism is complicated, and the ecological system presents remarkable synergistic and antagonistic effects under specific meteorological conditions, namely, the generation of O 3 can promote the generation of secondary components in PM 2.5 to exacerbate PM 2.5 pollution, and the high-concentration PM 2.5 can in turn inhibit the generation of O 3, so that the concentration of O 3 is at a lower level. These two effects under different meteorological conditions make pollution of the capital exhibit the characteristic of alternating complex pollution of the typical "summer O 3 type" and "winter PM 2.5 type". Therefore, implementing the cooperative control of PM 2.5 and O 3 is imperative, and accurately predicting the concentration change is a precondition for realizing scientific prevention and control, and is also a research hotspot in the current environmental science and information science intersection field. Although the numerical models (such as WRF-CMAQ and NAQPMS) based on atmospheric chemical transmission can describe the physical and chemical process of atmospheric transmission, the accuracy dependence on an emission list is extremely high, the calculation cost is huge, and the service forecast requirement of high timeliness is difficult to meet. Most of the existing statistical or machine learning prediction methods based on historical data consider an atmospheric environment system as a 'black box', namely only an input end and an output end, and key internal modes (such as strong seasonal periodicity, long-term trends caused by policies and emission reduction and random residual errors caused by sudden weather events) for driving PM 2.5 and O 3 to change cannot be effectively resolved and separated, so that the interpretation is lacking. This results in an insufficient predictive ability of the model to "inflection points" of the contamination process (e.g., sudden generation and dissipation of contamination) and poor predictive stability. Furthermore, most predictive models predict PM 2.5 and O 3 in isolation, ignoring their inherent photochemical relevance. For example, some of the components in PM 2.5 (e.g., nitrates, secondary organic aerosols) share a common precursor with O 3, and the concentration of particulates in the atmosphere also affects the rate of photolysis by the effect of radiation, thereby inhibiting the production of O 3. Such predictions of cracking may lead to decision contradictions, e.g., measures taken to reduce PM 2.5 may exacerbate O 3 contamination. Therefore, a new generation intelligent prediction model capable of deeply understanding and analyzing the inherent structure of the pollution time sequence and capturing the cooperative change rule of various pollutants is developed, and is a scientific bottleneck to be broken through in the current environment monitoring and forecasting field. Disclosure of Invention In view of the above-mentioned shortcomings of the prior art, the present invention provides a method of In order to achieve the aim of the invention, the invention adopts the following technical scheme: The PM 2.5 and O 3 collaborative prediction method based on STL-VMD-transducer mixed deep learning comprises the following steps: S1, collecting historical data of target predicted pollutant data, historical data of other synchronous pollutant data and meteorological data, preprocessing, calculating correlation between the target predicted pollutant data and other pollutant data and meteorological data, and screening out a key variable set with obvious influence on the target predicted pollutant data to obtain an input independent variable, wherein the target predicted pollutant data comprises PM 2.5 concentration data and O 3 concentration data; S2, constructing time series data of historical d