CN114626567A - A Prediction Method of Air Quality Index - Google Patents

Abstract

The invention relates to the technical field of air pollution prediction, in particular to a prediction method of an air quality index, which comprises the following steps: constructing network data; constructing an air quality index prediction model, wherein the air quality index prediction model comprises an attention-based graph convolution network and a long-term and short-term memory neural network; training an air quality index prediction model; and predicting the air quality index by using the trained air quality index prediction model. The air quality index prediction method provided by the invention solves the problem of low accuracy of the existing prediction method.

Description

A Prediction Method of Air Quality Index

technical field

The invention relates to the technical field of air pollution prediction, in particular to a prediction method of air quality index.

Background technique

The Air Quality Index (AQI) is a numerical value used to quantitatively describe the level of air quality. Air pollution can be directly expressed through the air quality index. Air pollution is affected by both regional and temporal factors, and generally speaking, it has a temporal and spatial nature.

Existing forecasting techniques for atmospheric pollutants only focus on temporal and spatial characteristics, and do not consider the effects of both. The use of time features is a traditional method that uses the time-sensitive data characteristics of air quality index to predict air quality index. However, the processing of time series by existing models has reached the bottleneck stage, and the prediction accuracy cannot be further improved.

Only using spatial features is to consider the interaction between the current site and other sites, while ignoring the time series of atmospheric data, and the use of spatial features at this stage is often limited to the learning of neighbor nodes, rather than learning thinking like humans. It has the feature of learning attention, allowing the model to learn the weights of neighbor nodes.

SUMMARY OF THE INVENTION

The invention proposes a prediction method of air quality index to solve the problem of low accuracy of the existing prediction method.

The technical scheme of the present invention to solve the above problem is: a prediction method of air quality index, comprising the following steps:

Construct network data, wherein the network data includes an adjacency matrix between air monitoring sites and a feature matrix of each air monitoring site;

constructing an air quality index prediction model, the air quality index prediction model including an attention-based graph convolutional network and a long short-term memory neural network;

Training an Air Quality Index Prediction Model

inputting the network data into the air quality index prediction model, and training the air quality index prediction model based on the loss function until the air quality index prediction model converges;

The air quality index is predicted using the trained air quality index prediction model.

Preferably, the step of constructing network data specifically includes:

Obtain atmospheric historical data of multiple air monitoring sites and the distance between each air monitoring site, wherein the atmospheric historical data includes historical data of air quality sub-indexes of various pollutants and historical meteorological data;

Construct an adjacency matrix between air monitoring stations according to the distance between each air monitoring station;

The characteristic matrix of each air monitoring station is obtained by preprocessing the atmospheric historical data;

The atmospheric historical data is divided into training set, validation set and test set.

Preferably, the preprocessing method includes mean filling null values, removing abnormal data, and smoothing data.

Preferably, the proportions of the training set, validation set and test set are 70%, 15% and 15% respectively.

Preferably, the step of constructing an adjacency matrix between the air monitoring stations according to the distances between the air monitoring stations includes: setting a threshold value k=1500km, setting a threshold value greater than k to 0, and setting a value less than k to 1 to construct an adjacency matrix .

Preferably, the training training air quality index prediction model specifically includes:

Input the adjacency matrix between air monitoring stations and the feature matrix of the training set into the attention-based graph convolution network to obtain the atmospheric feature matrix;

The atmospheric characteristic matrix is reconstructed by the multi-step multivariate method, and the constructed atmospheric characteristic data is obtained;

Input the constructed atmospheric feature data into the long short-term memory neural network layer for training;

Design loss function;

The air quality index prediction model is trained based on the loss function until the model converges.

Preferably, the loss function is expressed as

Among them, O _i is the actual value of the i-th atmospheric characteristic data, P _i is the predicted value of the i-th atmospheric characteristic data, and N is the total number of data.

Preferably, the step of using the trained air quality index prediction model to predict the air quality index specifically includes inputting the test set into the trained air quality index prediction model to obtain the air quality index.

Compared with the prior art, the beneficial effects of the present invention are: the prediction model of the present invention includes a graph convolution network and a long short-term memory network based on an attention mechanism, and a graph attention convolution network layer is used for the spatial characteristics of each monitoring site. It uses the long short-term memory network layer to process the temporal features of each site, builds an end-to-end network model that integrates the two network layers, and integrates the two data features for training, which can not only provide more abundant feature information for model prediction , which can improve the generalization ability of the model, and ultimately greatly improve the prediction accuracy of the current site. In order to circulate the data of the two neural network layers, the present invention designs a data combination mode of the middle layer, so that each model layer can be propagated by gradient.

Description of drawings

FIG. 1 is a schematic flowchart of the prediction method of the present invention.

Figure 2 is a schematic diagram of the structure of the air quality index prediction model.

Detailed ways

In order to make the purposes, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments These are some embodiments of the present invention, but not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention. Accordingly, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the invention as claimed, but is merely representative of selected embodiments of the invention.

Embodiment 1: a prediction method of air quality index as shown in Figure 1, the specific steps include:

S1: Construct network data

Calculate the Euclidean distance of each station according to the longitude and latitude of the station (air monitoring station), set the threshold k=1500km, and construct the adjacency matrix A between the stations.

The atmospheric historical data of each station is preprocessed, including filling empty values with the mean, removing abnormal data, smoothing data, etc., to obtain the characteristic matrix X of each station.

Divide the site atmospheric historical data into training set, validation set and test set.

S2: Build an air quality index prediction model (GA-LSTM), input the data provided by S1, and get the trained model after network model training.

S21: Put the site adjacency matrix A and the atmospheric feature matrix X of the training set into the attention graph convolution network for forward propagation, learn the characteristics of the neighboring sites through the site adjacency matrix relationship, correct the feature data, and obtain the atmosphere The feature matrix is equivalent to utilizing the spatial information of the site. Calculated as follows

Where: i is the current node, j is the currently calculated adjacent node, α _ij ^k is the linearized attention correlation value calculated by k attention mechanisms, W ^k is the weight matrix, and K is the graph attention of multiple heads layer, σ is the nonlinear layer.

S22: According to the characteristics of the long-term and short-term memory neural network layer, the atmospheric characteristic matrix is reconstructed according to the multi-step multi-variable method, and the data matrix is reconstructed to obtain the constructed atmospheric characteristic data in the form of using multi-day multi-characteristics to predict the next few days. This constitutes the data format of the long short-term memory network characteristics.

Multi-step multi-variable method: The air quality index of the previous n days of atmospheric data is used to form the characteristics of a training data, and the air quality index value of n+t days constitutes the current training data mark. For example, using the air data and air quality index from January 1st to 14th as features to predict the air quality index from January 15th to 21st, that is, using the data of the previous 14 days to predict the target data of the next 7 days.

n days of data as features Data for n+t days as markers

S23: Put the constructed atmospheric characteristic data into the long-term and short-term memory network for training, learn the time characteristics of the data, and use the basic structure of the long-term and short-term memory neural network.

S24: Calculate the loss using the mean square loss method according to the difference between the predicted value and the real value.

Among them: O _i is the actual value of the i-th atmospheric characteristic data, P _i is the predicted value of the i-th atmospheric characteristic data, and N is the total number of data.

S3: Backpropagation is performed according to the result of forward propagation.

Perform backpropagation of the network according to Loss, and calculate all learnable model parameters in the gradient update model;

The S2 and S3 processes are iteratively and alternately performed until the model converges, that is, the trained air quality index prediction model is obtained.

S4: Use the trained model to predict the air quality index for unlabeled nodes.

The above descriptions are only the embodiments of the present invention, and are not intended to limit the protection scope of the present invention. Any equivalent structure or equivalent process transformation made by using the contents of the description and drawings of the present invention, or directly or indirectly applied to other related The system field is similarly included in the protection scope of the present invention.

Claims (8)

1. A method for predicting an air quality index is characterized by comprising the following steps:

constructing network data, wherein the network data comprises an adjacency matrix among air monitoring stations and a characteristic matrix of each air monitoring station;

constructing an air quality index prediction model, wherein the air quality index prediction model comprises an attention-based graph convolution network and a long-short term memory neural network;

training air quality index prediction model

Inputting the network data into an air quality index prediction model, and training the air quality index prediction model based on a loss function until the air quality index prediction model converges;

and predicting the air quality index by using the trained air quality index prediction model.

2. The method for predicting an air quality index according to claim 1, wherein the step of constructing network data specifically comprises:

acquiring atmosphere historical data of a plurality of air monitoring stations and distances among the air monitoring stations, wherein the atmosphere historical data comprises historical data of air quality index of each pollutant and historical meteorological data;

constructing an adjacent matrix among the air monitoring stations according to the distance among the air monitoring stations;

preprocessing the historical atmospheric data to obtain a feature matrix of each air monitoring station;

and dividing the atmosphere historical data into a training set, a verification set and a test set.

3. The method of claim 2, wherein the preprocessing method comprises mean filling null values, removing abnormal data, and smoothing data.

4. The method of claim 3, wherein the training set, the validation set, and the test set are 70%, 15%, and 15% respectively.

5. The method for predicting the air quality index as claimed in claim 4, wherein the step of constructing the adjacency matrix between the air monitoring stations according to the distance between each air monitoring station comprises the following steps: setting a threshold k to 1500km, setting k larger than k to 0, and setting k smaller than k to 1 to construct an adjacency matrix of each station.

6. The method according to claim 1 or 2, wherein the training of the air quality index prediction model specifically comprises:

inputting an adjacency matrix between air monitoring stations and a feature matrix of a training set into a graph convolution network based on attention to obtain an atmospheric feature matrix;

reconstructing an atmospheric characteristic matrix by adopting a multi-step multi-variable method to obtain constructed atmospheric characteristic data;

inputting the constructed atmospheric characteristic data into a long-term and short-term memory neural network layer for training;

designing a loss function;

and training the air quality index prediction model based on a loss function until the model converges.

7. The method of claim 6, wherein the loss function is expressed as

Wherein, O_iIs the true value, P, of the ith atmosphere characteristic data_iAnd N is the predicted value of the ith atmosphere characteristic data and the total number of the data.

8. The method according to claim 7, wherein the step of predicting the air quality index by using the trained air quality index prediction model specifically comprises inputting a test set into the trained air quality index prediction model to obtain the air quality index.

Images