Multi-source PM2.5 prediction method combining graph attention network and multi-time period characteristics

Abstract

The invention relates to a multi-source PM2.5 (Particulate Matter 2.5) prediction method based on a GAT-MULCYCLE (Graphics Amplified Transform-MULCYCLE) fusion space and multi-time period The method comprises the following steps: firstly, acquiring time sequence data of pollutants (PM2.5, sulfur dioxide and the like) and weather (wind speed, air pressure and the like) and spatial data such as longitude and latitude from a monitoring station, then preprocessing the data and constructing a data set, and then carrying out spatial feature extraction and multi-cycle time feature extraction on the preprocessed data; meanwhile, an original sequence is input into an autoregression layer, and finally output results of the two parts are fused to obtain final PM2.5 concentration value prediction. According to the method, the influence of the related spatial features in the PM2.5 prediction task and the multi-cycle features contained in the time sequence is considered at the same time, so that the model structure and the prediction method are more reasonable.

Description

Technical Field

[0001] The present invention belongs to the field of meteorological forecasting, and specifically is a PM2.5 concentration prediction method based on GAT-MULCYCLE that extracts spatial and multi-time period features. Background Art

[0002] In recent years, the serious problem of air pollution has attracted increasing attention, and the prediction of atmospheric pollutant concentrations is a hot research topic. PM2.5 is the primary pollutant affecting air quality and the main culprit for smog. PM2.5 concentration is not only affected by the concentrations at neighboring stations and meteorological factors, but also exhibits cyclical variations. For example, PM2.5 concentrations at surrounding stations, along with wind speed, air pressure, and other factors, affect the PM2.5 concentration at a given station. Furthermore, concentration trends during the same time period between two days (morning peak, evening peak, etc.) are similar. Therefore, PM2.5 prediction models based on spatial and multi-temporal periods can better learn the diffusion and cyclical variations of PM2.5 and predict its concentration, providing important reference for environmental regulators to address air pollution.

[0003] In the PM2.5 prediction task, existing time series prediction methods mainly fall into three categories: prediction models based on traditional statistical methods, such as the autoregressive integrated moving average model (ARIMA) and the vector autoregressive model (VAR). Prediction models based on machine learning, such as the support vector machine model (SVR) and the random forest model (RF). Prediction models based on deep learning, such as the long short-term memory network (LSTM) and the gated neural unit (GRU). Although these classic models perform well in simple tasks, they are not very effective in multi-source PM2.5 prediction tasks. They have difficulty capturing nonlinear relationships in the data and can cause gradient explosion. Existing methods either only target a single data source and do not consider the mutual influence between data sources, or they cannot effectively consider the changing characteristics of the cycle and do not distinguish between long and short cycles. Therefore, the prediction results for such problems are naturally inaccurate. Summary of the Invention

[0004] In order to better support the air pollution monitoring of the ecological environment department and conduct accurate and rapid PM2.5 concentration prediction, the present invention provides a PM2.5 concentration prediction method based on GAT-MULCYCLE, which can more accurately predict the PM2.5 concentration in the city, thereby providing scientific protection for air quality safety.

[0005] The technical solutions adopted by the present invention for the above-mentioned purpose are as follows:

[0006] A multi-source PM2.5 prediction method combining graph attention network and multi-time period features includes the following steps:

[0007] 1) Obtain environmental time series data and latitude and longitude spatial data from multi-purpose environmental monitoring sites;

[0008] 2) performing data cleaning on the time series data and spatial data of the monitoring site and performing one-hot processing on the seasonal and heating period characteristics to obtain an initial feature matrix;

[0009] 3) constructing an adjacency matrix of the destination site, and using it together with the initial feature matrix as input parameters of the GAT network;

[0010] 4) constructing a GAT network, feeding the input parameters into the GAT network for training, so that the network learns the spatial correlation of each site and outputs the time series data of the next moment of the multi-destination site that integrates the influence of neighboring sites;

[0011] 5) Taking the time series data of a single destination site as input;

[0012] 6) Setting a period for the time series data, performing multi-period time feature extraction on the time series data of the single destination site, and obtaining an environmental prediction result for PM2.5;

[0013] 7) Sending the original environmental time series data in the initial matrix into the autoregressive network to obtain an initial prediction result;

[0014] 8) Fusing and splicing the output results of the autoregressive network and the environmental prediction feature data as the final output prediction result, which is used to reflect multi-source input and prediction output.

[0015] The environmental time series data includes monitoring time, PM2.5, AQI, carbon dioxide, sulfur dioxide, carbon monoxide, wind speed, wind direction, temperature, and air pressure parameter detection values.

[0016] The data cleaning is to fill in missing values ​​and remove outliers to ensure data integrity and avoid abnormal fluctuations.

[0017] The adjacency matrix is ​​defined according to the distance value between stations and is used to indicate whether the actual distance between any two stations satisfies the distance value.

[0018] It constructs a GAT network based on the distance of neighbor nodes, weights all neighbors, and dynamically assigns different weights to each node, capturing the relationship between nodes so that the network finally outputs the feature matrix of the current site at the next moment.

[0019] The training includes setting network parameters and a loss function, and using the loss function as an evaluation indicator to determine the training end point.

[0020] A multi-source PM2.5 prediction system combining a graph attention network and multi-time period features includes a data acquisition sensor group, a cloud server, and a management terminal. The data acquisition sensor group is a number of environmental monitoring sites deployed at the monitored environment, which collect environmental time series data and latitude and longitude spatial data and upload them to the cloud server; the management terminal is provided with a front-end interface and a management backend, the front-end interface collects user instructions and displays real-time monitoring data and prediction calculation process data, and the backend is used to convert user instructions into program call commands and send them to the cloud server for execution; the cloud server is provided with a prediction program, which is used to receive instructions from the management terminal to load the prediction program and execute the method steps as described above to realize multi-source PM2.5 prediction combining a graph attention network and multi-time period features.

[0021] The present invention has the following beneficial effects and advantages:

[0022] 1. High algorithm accuracy. This method simultaneously considers the impact of neighboring nodes on the PM2.5 concentration of the site and the multiple periodic features contained in time series data. Using GAT and GRU as basic models, it is a more reasonable new PM2.5 concentration prediction method. Compared with previous prediction methods that only consider time series features and do not consider periodic features, this method has higher accuracy.

[0023] 2. The model is highly automated. For a specific monitoring site, the present invention requires only a one-time GMC model training. Afterwards, PM2.5 concentrations can be automatically predicted without human intervention, significantly saving labor costs. This feature makes the present invention more convenient and economical in practical applications, providing an efficient and automated solution for environmental monitoring and prediction.

[0024] 3. Highly versatile algorithm. The core concept of this invention is highly portable and universal, requiring no additional adaptation for different time and geographical scenarios. This allows the invention to be easily extended to PM2.5 prediction tasks in a variety of different scenarios. This characteristic enables the invention to perform well in different environments and conditions, providing a flexible and effective solution for PM2.5 prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 This is a structural diagram of the PM2.5 concentration prediction model based on the fusion spatial features of GAT-MULCYCLE in the present invention; DETAILED DESCRIPTION

[0026] To make the above-mentioned objects, features, and advantages of the present invention more clearly understood, the specific implementation methods of the present invention are further described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways than those described herein, and those skilled in the art can make similar improvements without violating the scope of the invention. Therefore, the present invention is not limited to the specific implementation disclosed below.

[0027] The method first obtains time series data of pollutants (PM2.5, sulfur dioxide, etc.) and meteorological (wind speed, air pressure, etc.) from monitoring sites, as well as spatial data such as longitude and latitude. The data is then preprocessed and a dataset is constructed. The preprocessed data is then subjected to spatial feature extraction and multi-period temporal feature extraction. The original sequence is then fed into an autoregressive layer. Finally, the outputs of the two parts are fused to obtain the final prediction result. In the spatial feature extraction part, the present invention first feeds the time series data of multiple sites into the GAT network and dynamically calculates the attention coefficient between nodes based on the hyperparameter (DISTANCE). Different weights are assigned to different nodes, automatically learning the degree of mutual influence between nodes and more flexibly capturing information between different nodes. In the temporal feature extraction part, the present invention first obtains the target city sequence after spatial feature extraction. Based on the specific data situation, multiple skip connection layers are dynamically set and fed into the GRU network to capture the multi-period features of the time series. In the model fusion part, the original sequence is fed into the AR network and fused with the sequence obtained in the previous step to obtain the PM2.5 concentration value prediction. The present invention considers the influence of relevant spatial features and multi-period features contained in the time series in the PM2.5 prediction task, making the model structure and prediction method more reasonable. Figure 1 The specific steps are as follows:

[0028] Step 1: Acquire Data. Environmental monitoring sites are equipped with a variety of environmental data monitoring devices, and the data is stored in a central database. First, retrieve the time series data recorded by all environmental monitoring stations from this database: including PM2.5, AQI, carbon dioxide, sulfur dioxide, carbon monoxide, monitoring time, wind speed, wind direction, temperature, air pressure, and other test values. Also obtain the latitude and longitude of all environmental monitoring stations.

[0029] Step 2: Data cleaning. Preprocess the time series and spatial data from the monitoring stations, primarily including handling and normalizing missing values ​​and outliers. One-hot processing is also performed for seasonal and heating period characteristics.

[0030] Step 3: Construct the input parameters for the GAT network. The GAT-required data structures are constructed on the cleaned data, including the adjacency matrix and initial feature matrix for the city nodes. The adjacency matrix is ​​determined by the parameter DSITANCE, and the initial feature matrix is ​​the result of processing in Step 2.

[0031] Step 4: The data obtained in step 3 is fed into the GAT network to extract spatial correlation, and time series data that incorporates the influence of neighboring nodes is obtained.

[0032] Step 5: Select the target city time series data from the time series data obtained in step 4.

[0033] Step 6: Select a suitable period based on the periodic characteristics of the time series data, and extract multi-period time features from the data obtained in step 5 based on the selected period set.

[0034] Step 7: Feed the original time series data obtained in step 2 into the autoregressive network.

[0035] Step 8: Fuse the data obtained in step 6 and step 7 to obtain the final prediction result.

[0036] Specifically, data cleaning involves using the k-nearest neighbor algorithm for outlier detection. For each data point, the algorithm determines whether it is an outlier by evaluating the distance to its nearest neighbors. To ensure data continuity, we employ multiple methods to address short-term missing data, including forward filling, backward filling, and mean filling. The combined use of these filling methods improves filling coverage and ensures data integrity. Furthermore, we employ smoothing techniques, such as moving averages or exponential smoothing, to stabilize the data and reduce the impact of noise on the analysis.

[0037] The one-hot processing of the season and heating period features is specifically as follows: the categorical information of the season and heating period features is converted into a numerical form that the model can understand. First, we map this information into integer form, for example, spring is represented as 1, summer is represented as 2, and so on. Then, through one-hot encoding, we convert each season into a binary vector, where the length of the vector is equal to the number of seasons. In this vector, each position represents a season, the value of the corresponding position is 1, and the values ​​of other positions are 0. The same method is used to binary encode the heating period of the destination site to obtain a one-hot encoding vector representing the heating period.

[0038] The normalization operation for seasonal and heating period characteristics is to use min-max standardization to map the data to [0, 1].

[0039] The data structure construction process is as follows: for the construction of the adjacency matrix, it is constructed based on the parameter DISTANCE and the actual distance between the two stations. If the distance is greater than DISTANCE, the value is 0, otherwise it is 1. The rows and columns of the adjacency matrix are the sequence numbers of each station; for the initial feature matrix, it is set to the time series data processed by step 2.

[0040] The entire model of the method of the present invention is GMC, and its training process for time series prediction is:

[0041] 1) Input the time step, first send it to the GAT network to obtain spatial correlation features, then send it to the multi-period extraction layer, and finally fuse it with the original time series data after the autoregressive layer to obtain the final prediction result.

[0042] 2) After grid search, it can be seen that the model parameters are selected as 24*2*7 time steps, 300 kilometers distance, K=2 (number of attention heads), 128 hidden layer sizes, and the SGD method is used for model optimization to achieve the best prediction effect.

[0043] 3) Definition of loss function: We use MAE (Mean Absolute Error) and MSE (Mean Squared Error) as evaluation indicators.

[0044] The specific process of feature extraction of the model of the method of the present invention is:

[0045] 1) In the spatial feature extraction stage: a GAT network is constructed based on the distance between neighbor nodes and all neighbors are weighted. GAT can dynamically assign different weights to each node, more effectively capturing the relationship between nodes and ultimately obtaining a new feature matrix for the node.

[0046] 2) Multi-head attention mechanism: In order to improve the spatial expression and generalization capabilities of the model, this paper uses a multi-head attention mechanism to further extract the spatial features of the node graph. We repeat the previous step times to obtain K calculation results, where K represents the number of attention heads, and take the average to obtain the final feature.

[0047] 3) Multi-cycle temporal feature extraction: We introduce the GRU, whose gated unit design can better capture and preserve long-term dependencies in the input sequence while mitigating gradient issues. Unlike the traditional GRU, we define T to represent a selected set of p cycles. Because we focus only on specific cycles, skipping many hidden layer states, we can significantly mitigate the gradient explosion problem even for long time series. The calculation process for a specific cycle is as follows:

[0048]

[0049]

[0050]

[0051]

[0052] Where σ represents the sigmoid function and tanh represents the bitangent function Indicates concatenation of hidden state and input, ⊙ indicates element-wise multiplication, and W and b are both training parameters. t This is the time series of the target city we initially selected. The h in the formula represents the hidden layer calculation result, and the remaining parameters are the training parameters required by the GRU. Since p periods are selected, p outputs will ultimately be generated, representing the multi-period time features extracted for the selected node.

[0053] The specific operation of obtaining the prediction output is: sending the data obtained in step 6 that integrates the spatiotemporal features into the fully connected layer and fusing it with the original sequence sent to the autoregressive layer in step 7 to obtain the final PM2.5 concentration prediction value.

[0054] The present invention also provides a multi-source PM2.5 prediction system combining a graph attention network and multi-time period features, including a data acquisition sensor group, a cloud server, and a management terminal. The data acquisition sensor group is a number of environmental monitoring sites deployed at the monitored environment, which collect environmental time series data and latitude and longitude spatial data and upload them to the cloud server; the management terminal is provided with a front-end interface and a management backend, the front-end interface collects user instructions and displays real-time monitoring data and prediction calculation process data, and the backend is used to convert user instructions into program call commands and send them to the cloud server for execution; the cloud server is provided with a prediction program, which is used to receive instructions from the management terminal to load the prediction program and execute the method steps as described above to realize multi-source PM2.5 prediction combining a graph attention network and multi-time period features.

[0055] Finally, it should be noted that the above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles described in the present invention. These improvements and modifications should be regarded as the scope of protection of the present invention.

Claims

1. A multi-source PM2.5 prediction method combining graph attention network and multi-time period features, characterized by: The following steps are involved: 1) Obtain environmental time series data and latitude and longitude spatial data from multi-purpose environmental monitoring sites; 2) performing data cleaning on the time series data and spatial data of the monitoring site and performing one-hot processing on the seasonal and heating period characteristics to obtain an initial feature matrix; 3) constructing an adjacency matrix of the destination site, and using it together with the initial feature matrix as input parameters of the GAT network; 4) constructing a GAT network, feeding the input parameters into the GAT network for training, so that the network learns the spatial correlation of each site and outputs the time series data of the next moment of the multi-destination site that integrates the influence of neighboring sites; 5) Taking the time series data of a single destination site as input; 6) Setting a period for the time series data, performing multi-period time feature extraction on the time series data of the single destination site, and obtaining an environmental prediction result for PM2.5; 7) Sending the original environmental time series data in the initial matrix into the autoregressive network to obtain an initial prediction result; 8) Fusing and splicing the output results of the autoregressive network and the environmental prediction feature data as the final output prediction result, which is used to reflect multi-source input and prediction output.

2. A multi-source PM2.5 prediction method combining graph attention network and multi-time period features according to claim 1, characterized in that: The environmental time series data includes monitoring time, PM2.5, AQI, carbon dioxide, sulfur dioxide, carbon monoxide, wind speed, wind direction, temperature, and air pressure parameter detection values.

3. The multi-source PM2.5 prediction method combining graph attention network and multi-time period features according to claim 1 is characterized in that: The data cleaning is to fill in missing values ​​and remove outliers to ensure data integrity and avoid abnormal fluctuations.

4. The multi-source PM2.5 prediction method combining graph attention network and multi-time period features according to claim 1 is characterized in that: The adjacency matrix is ​​defined according to the distance value between stations and is used to indicate whether the actual distance between any two stations satisfies the distance value.

5. The multi-source PM2.5 prediction method combining graph attention network and multi-time period features according to claim 1 is characterized in that: It constructs a GAT network based on the distance of neighbor nodes, weights all neighbors, and dynamically assigns different weights to each node, capturing the relationship between nodes so that the network finally outputs the feature matrix of the current site at the next moment.

6. The multi-source PM2.5 prediction method combining graph attention network and multi-time period features according to claim 1 is characterized in that: The training includes setting network parameters and a loss function, and using the loss function as an evaluation indicator to determine the training end point.

7. A multi-source PM2.5 prediction system combining graph attention network and multi-time period features, characterized by: It includes a data acquisition sensor group, a cloud server, and a management terminal. The data acquisition sensor group is a number of environmental monitoring sites deployed at the monitored environment, which collect environmental time series data and latitude and longitude spatial data and upload them to the cloud server; the management terminal is provided with a front-end interface and a management backend. The front-end interface collects user instructions and displays real-time monitoring data and prediction calculation process data. The backend is used to convert user instructions into program call commands and send them to the cloud server for execution; the cloud server is provided with a prediction program, which is used to receive instructions from the management terminal to load the prediction program and execute the method steps as described in any one of claims 1-6 to realize multi-source PM2.5 prediction combining graph attention network and multi-time period features.

Images