PM2.5 concentration prediction method and system

Abstract

The invention relates to a PM2.5 concentration prediction method and system, and the method employs a geometric algebraic space-time diagram recurrent neural network model to predict the PM2.5 concentration of a target region, and the geometric algebraic space-time diagram recurrent neural network model comprises a geometric algebraic-graph convolutional neural network and a geometric algebraic-long and short-term memory network. Comprising the following steps of collecting multi-source data and performing preprocessing; extracting spatial features from the preprocessed data by using a geometric algebraic-graph convolutional neural network; on the basis of the spatial characteristics, utilizing a geometric algebraic-long and short-term memory network to obtain spatial and temporal characteristics of PM2.5; the PM2.5 concentration is predicted according to the spatial and temporal characteristics; the system is used for implementing the method. Compared with the prior art, the method has the advantages of high PM2.5 concentration prediction precision and the like.

Description

Technical Field

[0001] The present invention relates to the technical field of air pollution prediction, and in particular to a PM2.5 concentration prediction method and system. Background Art

[0002] With the construction of industrialization and urbanization, air pollution has become an unavoidable problem, which not only threatens people's health, but also overdraws a country's development prospects. For every 10μg / m3 increase in particulate matter concentration, the incidence rate of the population will increase by 0.38%. Therefore, it is crucial to develop an effective algorithm to predict PM2.5 concentration in the long term.

[0003] At present, there are two main methods for predicting atmospheric pollution concentrations: chemical models based on formation mechanisms and statistical models based on data. Chemical models mainly study the formation mechanism of PM 2.5 and consider the interactions between multiple pollutants, such as the community multi-scale air quality modeling system and the chemical weather research and forecasting model. The advantages of chemical models based on the formation mechanism of chemical models are their consideration of actual conditions and their ability to explain the transformation process between atmospheric pollutants. However, chemical models have limitations in dealing with geographical factors, especially in high-altitude areas and long-distance transportation. In addition, these models require a lot of computing resources, making them unsuitable for predicting air pollution in complex scenarios. Compared with chemical models, statistical models based on data are more efficient and simple. These models usually use historical data for mathematical modeling to improve prediction accuracy. Statistical models mainly include machine learning models and deep learning models. In machine learning models, historical data of air pollutants are frequently used, and a variety of regression analysis methods are used for prediction, such as artificial neural networks (ANN), linear regression (LR), random forests (RF), autoregressive moving average (ARMA), and support vector regression (SVR). However, due to the influence of PM 2.5 There are so many factors that influence concentration, and the diffusion and deposition processes of air pollutants are highly nonlinear in nature, so simple machine learning models may not work well.

[0004] In recent years, the popularity of deep learning models has surged because they are easy to implement and have a strong ability to handle complex data. Compared with traditional machine learning models, deep learning models show strong generalization ability and are more suitable for tasks with nonlinear features, mainly because they usually have more neural network layers and neurons. Representative models include convolutional neural networks (CNNs) and recurrent neural networks (RNNs). Although deep learning has made significant progress in PM2.5 prediction, a single deep learning model still has limitations in predicting PM2.5 concentration. For example, CNN is good at processing spatial features, but it does not work well when processing time series data, while RNN, although it can process time series data, is prone to the dilemma of gradient vanishing when dealing with long-term dependency problems. In addition, the use of a single data source also limits the model's ability to capture the complex spatiotemporal changes in PM2.5 concentration. The existing GNN model may cause information loss when the network layer is deepened, and the problem of information forgetting still exists when LSTM processes long time series data. In addition, the prediction model of a single data source is difficult to fully capture the complex spatiotemporal changes in PM2.5 concentration. Summary of the invention

[0005] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and to provide a PM2.5 concentration prediction method and system based on a geometric algebraic space-time graph recurrent neural network.

[0006] The purpose of the present invention can be achieved by the following technical solutions:

[0007] According to a first aspect of the present invention, a PM2.5 concentration prediction method is provided, which uses a geometric algebraic space-time graph recurrent neural network model to predict the PM2.5 concentration in a target area, wherein the geometric algebraic space-time graph recurrent neural network model includes a geometric algebra-graph convolutional neural network and a geometric algebra-long short-term memory network; the method comprises the following steps:

[0008] Collecting multi-source data in the target area and integrating and preprocessing them to obtain preprocessed data; the multi-source data includes PM2.5 concentration, meteorological data, and geographical location information of monitoring stations;

[0009] Extracting spatial features from the preprocessed data using a geometric algebra-graph convolutional neural network; the spatial features are highly expressed and calculated in a geometric algebraic space through geometric algebraic products;

[0010] Based on the spatial characteristics, the spatiotemporal characteristics of PM2.5 are obtained using geometric algebra-long short-term memory network;

[0011] The PM2.5 concentration is predicted based on the spatial characteristics.

[0012] As a preferred technical solution, the method for extracting spatial features comprises the following steps:

[0013] The preprocessed data is input into the geometric algebra-graph convolutional neural network, and the graph convolutional neural network therein is used to extract the node features and edge features of the preprocessed data; the node features include the PM2.5 concentration and meteorological data of the monitoring station; the edge features include the physical distance between the monitoring stations and the wind direction relationship;

[0014] The node features are mapped to the geometric algebraic space by the graph convolutional neural network to obtain the geometric algebraic feature vector, and the geometric algebraic feature vector is transferred in the multi-layer graph convolutional neural network to output the spatial features.

[0015] As a preferred technical solution, the calculation formula for the characteristic transfer is:

[0016]

[0017] in, is the geometric algebraic feature vector of the l+1th layer of monitoring station v; σ(·) is the activation function; θ is the trainable network parameter; (GA) represents the geometric algebraic calculation output; B is the adjacency matrix between monitoring stations, which is the re-normalized Laplace matrix expression: and is the degree matrix is the inverse matrix of , and A is the adjacency matrix.

[0018] As a preferred technical solution, the expression output by the geometric algebra calculation is:

[0019]

[0020] Among them, e 12...k is the orthogonal basis element of the geometric algebra space G; n represents the matrix order; l represents the number of layers of the geometric algebra-graph convolutional neural network; represents geometric algebra weight; W (l)，G represents the geometric algebra weight matrix; represents the geometric algebraic eigenvector of the monitoring station v; ⊙ represents the dot product operation.

[0021] As a preferred technical solution, the adjacency matrix satisfies any of the following requirements: if the Euclidean distance between two nodes is greater than the maximum adjacent distance, then there is no edge between the nodes; if there are three or more mountains between the two nodes whose heights exceed the specified height, then there is no edge between the nodes.

[0022] As a preferred technical solution, the method for obtaining the spatiotemporal characteristics of PM2.5 includes:

[0023] The spatial features extracted by the geometric algebra-graph convolutional neural network and the historical PM2.5 concentration data are fused to obtain the fused features. The calculation formula is:

[0024] H fusion =Concat(H GNN , X hist ),

[0025] Among them, H GNN Represents spatial features; X hist Represents historical PM2.5 concentration data; H fusion Indicates fusion features;

[0026] The fusion features are input into the geometric algebra-long short-term memory network, and the spatiotemporal features are output through the long short-term memory network processing.

[0027] As a preferred technical solution, the processing process of the long short-term memory network is:

[0028] Calculate the forget gate value:

[0029]

[0030] Among them, U and W represent weight matrices; b represents the bias term; σ represents the sigmoid activation function; h t-1 The hidden state at the previous moment; represents the fusion feature at time t; f t is the forget gate value; Geometric product operation; calculates the input gate value and generates a candidate value vector:

[0031]

[0032] Among them, i t is the input gate value; represents the candidate value vector; α represents the tanh activation function;

[0033] Update unit status:

[0034]

[0035] Among them, C t is the current unit state; C t-1 is the unit state at the previous moment;

[0036] Calculate the output gate value and get the final hidden state:

[0037]

[0038] h t =o t ⊙α(C t)),

[0039] Among them, t Indicates the output gate value; h t Represents the final hidden state value, i.e., the spatiotemporal feature.

[0040] As a preferred technical solution, the geometric algebraic space-time graph recurrent neural network model is optimized by the White Shark optimization algorithm.

[0041] According to a second aspect of the present invention, a PM2.5 concentration prediction system is provided, the system comprising:

[0042] Data collection and processing module: used to collect relevant data of the target area and perform preprocessing to obtain preprocessed data;

[0043] Feature extraction and fusion module: used to receive and process data, first extract spatial features through geometric algebra-graph convolutional neural network, extract spatial features at a certain moment and corresponding neighborhood comprehensive information and perform feature fusion, and pass the fused features through geometric algebra-long short-term memory network to obtain the corresponding spatiotemporal features of PM2.5;

[0044] Prediction module: Predict PM2.5 concentration based on spatiotemporal characteristics.

[0045] According to a third aspect of the present invention, there is provided a computer-readable storage medium having a computer program stored thereon, wherein the program implements the method described above when executed by a processor.

[0046] Compared with the prior art, the present invention constructs a high-dimensional geometric algebraic space in a non-Euclidean domain for better learning vector embedding between nodes, and constructs a recurrent neural network model based on a geometric algebraic space-time graph. Based on the collected multi-source data sets, the spatial features of PM2.5 are extracted through the geometric algebra-graph convolutional neural network in the model, and then the spatial features are fused with historical data through geometric algebra-long short-term memory network processing to predict PM2.5 concentration, which effectively solves the problem of feature information distortion in the traditional PM2.5 prediction process and performs well in prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 is a flow chart of the method of the present invention;

[0048] Figure 2 It is a structural diagram of a geometric algebraic space-time graph recurrent neural network of the present invention;

[0049] Figure 3 This is the optimization flow chart of the White Shark optimization algorithm of the present invention. DETAILED DESCRIPTION

[0050] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0051] Details of one or more embodiments of the present application are set forth in the following drawings and description to make other features, objects, and advantages of the present application more readily apparent.

[0052] Reference to "embodiments" in this application means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present application. The appearance of the phrase in various locations in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those of ordinary skill in the art that the embodiments described in this application may be combined with other embodiments without conflict.

[0053] A hybrid deep learning model refers to the integration of two or more independent models, and allows the hybrid model to leverage the strengths of each model to extract different aspects of data features. Compared with a single model, the hybrid model has higher prediction accuracy and can more skillfully extract spatial and temporal features; it has the advantages of a single model and makes up for its weaknesses.

[0054] Therefore, a geometric algebraic space-time graph recurrent neural network model is constructed in the present invention. The network model includes a geometric algebra-graph convolutional neural network and a geometric algebra-long short-term memory network, which avoids the problem of information loss caused by a single graph convolutional neural network as the number of network layers increases and the problem of information forgetting in the long short-term memory network when dealing with long-term dependencies, and reduces computational complexity by expressing and calculating features in geometric algebraic space.

[0055] Embodiment 1:

[0056] This embodiment provides a PM2.5 concentration prediction method, the flow chart of which is as follows Figure 1 As shown, the following steps are included:

[0057] S1, collect multi-source data in the target area and integrate and pre-process them to obtain pre-processed data;

[0058] S11. Collect multi-source data: Obtain multi-source data from multiple data sources such as air quality monitoring stations, meteorological stations and geographic information systems, including PM2.5 concentration, air temperature data (including temperature, humidity, wind speed and direction) and geographical location information of monitoring stations;

[0059] S12. Integrate and process multi-source data: Use interpolation to fill in missing data to ensure data integrity, and perform data normalization to unify data scales to ensure the stability and consistency of model inputs, providing high-quality data inputs for subsequent model training and predictions.

[0060] S2. Use geometric algebra-graph convolutional neural network to extract spatial features from preprocessed data. The processing process in the network is as follows: Figure 2 As shown in the GA-GNN box;

[0061] S21. Input the preprocessed data into the geometric algebra-graph convolutional neural network, and use the graph convolutional neural network therein to extract the node features and edge features of the preprocessed data; use the geometric algebra enhanced graph neural network to extract spatial features, which is suitable for environmental monitoring data with complex spatial dependencies;

[0062] In the graph structure, nodes represent monitoring stations, edges represent spatial relationships between monitoring stations, PM2.5 concentration and meteorological data of each node are used as node features, and physical distance and wind direction relationship between nodes are used as edge features;

[0063] S22. The node features are mapped to the geometric algebraic space by the graph convolutional neural network to obtain the geometric algebraic feature vector, and the geometric algebraic feature vector is transferred in the multi-layer graph convolutional neural network to output the spatial features.

[0064] The expression of feature transfer in a multi-layer graph convolutional neural network is:

[0065]

[0066] is the geometric algebraic feature vector of the l+1th layer of monitoring station v; σ(·) is the activation function; θ is the trainable network parameter; (GA) represents the geometric algebraic calculation output; B is the adjacency matrix between monitoring stations, which is the re-normalized Laplace matrix expression: and is the degree matrix The inverse matrix of , A is the adjacency matrix, and the adjacency matrix satisfies that if the Euclidean distance between two nodes is greater than 300 kilometers or there are three or more mountains with a height of more than 1200 meters between the two nodes, then there is no edge between the nodes;

[0067] Among them, the expression of the geometric algebra calculation output in the above formula is:

[0068]

[0069] Among them, e 12...k is the geometric algebraic space G kOrthogonal basis elements; n represents the matrix order; l represents the number of geometric algebra-graph convolutional neural network layers; represents geometric algebra weight; W (l)，G represents the geometric algebra weight matrix; represents the geometric algebraic eigenvector of the monitoring station v; ⊙ represents the dot product operation;

[0070] In the above geometric algebraic space, G k is an orthogonal basis 2 k dimensional GA space, expressed as: {{e i},{e ij}, …, {e 12...k}}, and a multivector can be represented as a scalar and 2 k -1 basis linear combination. Taking G3 geometric algebraic space as an example, multivectors are the extension of vectors to higher dimensions. For any multivector β∈G3, it can be described as

[0071] β=λ0+λ1e1+λ2e2+λ3e3+λ 12 e 12 +λ 13 e 13 +λ 23 e 23 +λ 123 e 123 ,λ∈R,e∈E

[0072] Among them, R represents the real number field, E represents the basic elements of G3, and scalars and vectors constitute different subspaces in geometric algebra.

[0073] According to the above formula, the geometric algebraic weight matrix W of node v can be (l)，G and the geometric algebraic eigenvector of node v It is expressed in the form of multi-vectors, as follows:

[0074]

[0075] Among them, in the geometric algebraic product process, is the shared geometric algebraic weight, e 12...k Denotes the geometric algebraic space G k are the orthogonal basis elements of , and n represents the matrix order.

[0076] S3, based on spatial features, use geometric algebra-long short-term memory network to obtain the spatiotemporal characteristics of PM2.5. The processing process is as follows Figure 2 As shown in the GA-LSTM box;

[0077] S31. The spatial features extracted by the geometric algebra-graph convolutional neural network and the historical PM2.5 concentration data are fused to obtain fused features. The calculation formula is:

[0078] H fusion =Concat(H GNN , X hist ),

[0079] Among them, H GNN Represents spatial features; X hist Represents historical PM2.5 concentration data; H fusion Indicates fusion features;

[0080] S32, input the fusion features into the geometric algebra-long short-term memory network and process them through steps S321-S324 to output the spatiotemporal features;

[0081] S321, calculate the forget gate value:

[0082]

[0083] Among them, U and W represent weight matrices; b represents the bias term; σ represents the sigmoid activation function; h t-1 The hidden state at the previous moment; represents the fusion feature at time t; f t is the forget gate value; Geometric product operation;

[0084] S322, calculate the input gate value and generate a candidate value vector:

[0085]

[0086] Among them, i t is the input gate value; represents the candidate value vector; α represents the tanh activation function;

[0087] S323, update unit status:

[0088]

[0089] Among them, C t is the current unit state; C t-1 is the unit state at the previous moment;

[0090] S324, calculate the output gate value and obtain the final hidden state:

[0091]

[0092] h t =o t ⊙α(C t )),

[0093] Among them, tIndicates the output gate value; h t Represents the final hidden state value, i.e., the spatiotemporal feature.

[0094] S4. Predict PM2.5 concentration based on temporal and spatial characteristics.

[0095] In this embodiment, a network model optimization algorithm based on the White Shark algorithm is provided, and is used to optimize the geometric algebraic space-time graph recurrent neural network model. The optimization process is as follows: Figure 3 As shown, the following steps are included:

[0096] A1. Use mean square error as the loss function, update the hyperparameters through the Adam optimizer, and use the updated hyperparameters as the candidate hyperparameter solutions;

[0097] A2. Obtain candidate solutions for the hyperparameters of the network model, input the White Shark optimization algorithm model for preliminary fitness evaluation to obtain the fitness values ​​of each candidate solution, and use the prediction error of the model on the training data as the fitness function. The formula is as follows:

[0098]

[0099] Among them, x represents the candidate solution, y i represents the true value, represents the predicted value, and n represents the number of samples;

[0100] A3. Select several candidate solutions of corresponding hyperparameters with excellent fitness values ​​as the initial individual positions of white sharks, and the optimal solution as the prey. Perform hunting simulation, calculate the distance between each white shark individual and the prey to find the best hunting white shark. The calculation formula is as follows:

[0101]

[0102] Among them, X(t) represents the current solution, X best represents the current optimal solution η and Represent the step size coefficient and perturbation coefficient respectively, Z represents the random vector;

[0103] A4. Taking each of the fitness values ​​as a white shark individual and the optimal solution as prey, a hunting simulation is performed, and the distance between the white shark individual and the prey is calculated by an iterative formula to seek the best white shark individual position, i.e., the candidate solution of the current best hyperparameter, and update the white shark individual position, and repeat the position update until the global optimal hyperparameter candidate solution is found;

[0104] A5. Update the hyperparameter values ​​of the network model based on the global optimal hyperparameter candidate solution;

[0105] A6. Return the hyperparameter values ​​obtained in step S5 to the network model to obtain a trained network model, and evaluate the model performance by calculating the accuracy, recall, and F1-Score indicators, including:

[0106] A61. Use the trained network model to predict the future PM2.5 concentration and output the PM2.5 concentration at a future time point;

[0107] A62. Compare the predicted concentration with the actual observed value, calculate the prediction error at each time point, and construct a confusion matrix based on the error value;

[0108] A63. According to the confusion matrix, the accuracy, recall and F1-Score indicators at each time point are further calculated. The calculation formulas of each indicator are as follows:

[0109] Accuracy:

[0110]

[0111] Among them, TP represents the number of samples predicted to be positive and actually positive, TN represents the number of samples predicted to be negative and actually negative, FP represents the number of samples predicted to be positive but actually negative, and FN represents the number of samples predicted to be negative but actually positive.

[0112] Recall:

[0113]

[0114] F1-Score:

[0115]

[0116] Among them, Precision is defined as:

[0117]

[0118] A63. If the evaluation result is excellent, the trained network model is selected as the final geometric algebraic space-time graph recurrent neural network model.

[0119] In order to improve the reliability of the evaluation results, the cross-validation method can be used to divide the data set into a training set and a test set, repeat the training and testing process multiple times, calculate the evaluation index each time, and take the average evaluation value as the final evaluation result. This evaluation method can comprehensively and objectively reflect the prediction ability of the model and its stability under different conditions.

[0120] In order to verify the performance of the geometric algebraic space-time graph recurrent neural network model after training in steps A1 to A6, this embodiment was experimented on the KnowAir dataset and compared with GRU and GC-LSTM. GRU (model 1), GC-LSTM (model 2) and the geometric algebraic space-time graph recurrent neural network model (model 3) provided by the present invention were used to predict PM2.5 concentrations in 184 cities respectively. The test time is 72 hours, and predictions are made every three hours. The average value of each 24 indicators in the sample city is calculated. The experiment is repeated ten times to obtain the root mean square error (RMSE) and critical success index (CSI) of the indicators and perform statistics. The root mean square error (RMSE) is used to directly test the prediction accuracy, and the critical success index (CSI) is used to measure the performance near the pollution threshold of 75μg / m3, as shown in the following table:

[0121] Table 1 Comparison of model experimental data

[0122]

[0123] According to Table 1, the geometric algebraic space-time graph recurrent network model of the present invention has high prediction accuracy and excellent performance in measuring the pollution threshold of 75 μg / m3.

[0124] Embodiment 2:

[0125] The above is an introduction to the method embodiment. The following is a further explanation of the solution of the present invention through a system embodiment.

[0126] This embodiment provides a PM2.5 concentration prediction system, which includes:

[0127] Data collection and processing module: used to collect relevant data of the target area and perform preprocessing to obtain preprocessed data;

[0128] Feature extraction and fusion module: used to receive and process data, first extract spatial features through geometric algebra-graph convolutional neural network, extract spatial features at a certain moment and corresponding neighborhood comprehensive information and perform feature fusion, and pass the fused features through geometric algebra-long short-term memory network to obtain the corresponding spatiotemporal features of PM2.5;

[0129] Prediction module: Predict PM2.5 concentration based on spatiotemporal characteristics.

[0130] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the described module can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0131] Embodiment 3:

[0132] The present embodiment provides that the electronic device includes a central processing unit (CPU), which can perform various appropriate actions and processes according to computer program instructions stored in a read-only memory (ROM) or loaded from a storage unit into a random access memory (RAM). In the RAM, various programs and data required for the operation of the device can also be stored. The CPU, ROM and RAM are connected to each other via a bus. An input / output (I / O) interface is also connected to the bus.

[0133] Multiple components in the device are connected to the I / O interface, including: input units, such as keyboards, mice, etc.; output units, such as various types of displays, speakers, etc.; storage units, such as disks, optical disks, etc.; and communication units, such as network cards, modems, wireless communication transceivers, etc. The communication unit allows the device to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunication networks.

[0134] The processing unit performs the various methods and processes described above, such as methods S1 to S4 and A1 to A6 in the embodiments. For example, in a certain embodiment, methods S1 to S4 and A1 to A6 may be implemented as a computer software program, which is tangibly contained in a machine-readable medium, such as a storage unit. In some embodiments, part or all of the computer program can be loaded and / or installed on the device via a ROM and / or a communication unit. When the computer program is loaded into the RAM and executed by the CPU, one or more steps of methods S1 to S4 and A1 to A6 in the embodiments described above may be executed. Alternatively, in other embodiments, the CPU may be configured to execute methods S1 to S4 and A1 to A6 in the embodiments by any other appropriate means (e.g., by means of firmware).

[0135] The functions described above herein may be performed at least in part by one or more hardware logic components. For example, without limitation, exemplary types of hardware logic components that may be used include: field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chips (SOCs), complex programmable logic devices (CPLDs), and the like.

[0136] The program code for implementing the method of the present invention can be written in any combination of one or more programming languages. These program codes can be provided to a processor or controller of a general-purpose computer, a special-purpose computer or other programmable data processing device, so that the program code, when executed by the processor or controller, enables the functions / operations specified in the flow chart and / or block diagram to be implemented. The program code can be executed entirely on the machine, partially on the machine, partially on the machine as a stand-alone software package and partially on a remote machine, or entirely on a remote machine or server.

[0137] In the context of the present invention, a machine-readable medium may be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, device, or equipment. A machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or equipment, or any suitable combination of the foregoing. A more specific example of a machine-readable storage medium may include an electrical connection based on one or more lines, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0138] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed by the present invention, and these modifications or replacements should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A PM2.5 concentration prediction method, characterized in that: The method uses a geometric algebraic space-time graph recurrent neural network model to predict the PM2.5 concentration in the target area. The geometric algebraic space-time graph recurrent neural network model includes a geometric algebra-graph convolutional neural network and a geometric algebra-long short-term memory network. The following steps are involved: Collect multi-source data in the target area and integrate and pre-process it to obtain pre-processed data; the multi-source data includes PM2.5 concentration, meteorological data and geographical location information of monitoring stations; Extracting spatial features from the preprocessed data using a geometric algebra-graph convolutional neural network; the spatial features are highly expressed and calculated in a geometric algebraic space through geometric algebraic products; Based on the spatial characteristics, the spatiotemporal characteristics of PM2.5 are obtained using geometric algebra-long short-term memory network; The PM2.5 concentration is predicted based on the described spatiotemporal characteristics.

2. A PM2.5 concentration prediction method according to claim 1, characterized in that: The method for extracting spatial features comprises the following steps: The preprocessed data is input into a geometric algebra-graph convolutional neural network, and the graph convolutional neural network therein is used to extract node features and edge features of the preprocessed data; the node features include PM2.5 concentration and meteorological data of the monitoring station; the edge features include the physical distance between the monitoring stations and the wind direction relationship; The node features are mapped to the geometric algebraic space by the graph convolutional neural network to obtain the geometric algebraic feature vector, and the geometric algebraic feature vector is transferred in the multi-layer graph convolutional neural network to output the spatial features.

3. A PM2.5 concentration prediction method according to claim 2, characterized in that: The calculation formula of the feature transfer is: in, is the geometric algebraic feature vector of the l+1th layer of monitoring station v; σ(·) is the activation function; θ is the trainable network parameter; (GA) represents the geometric algebraic calculation output; B is the adjacency matrix between monitoring stations, which is the re-normalized Laplace matrix expression: and is the degree matrix is the inverse matrix of , and A is the adjacency matrix.

4. A PM2.5 concentration prediction method according to claim 3, characterized in that: The expression output by the geometric algebra calculation is: Among them, e 12...k is the geometric algebraic space G k Orthogonal basis elements; n represents the matrix order; l represents the number of geometric algebra-graph convolutional neural network layers; represents geometric algebra weight; W (l),G represents the geometric algebra weight matrix; represents the geometric algebraic eigenvector of the monitoring station v; ⊙ represents the dot product operation.

5. A PM2.5 concentration prediction method according to claim 3, characterized in that: The adjacency matrix meets any of the following requirements: If the Euclidean distance between two nodes is greater than the maximum neighbor distance, then there is no edge between the nodes; If there are three or more mountains between two nodes whose height exceeds the specified height, there is no edge between the nodes.

6. A PM2.5 concentration prediction method according to claim 1, characterized in that: The method for obtaining the spatiotemporal characteristics of PM2.5 includes: The spatial features extracted by the geometric algebra-graph convolutional neural network and the historical PM2.5 concentration data are fused to obtain the fused features. The calculation formula is: H fusion =Concat(H GNN ,X hist ), Among them, H GNN Represents spatial features; X hist Represents historical PM2.5 concentration data; H fusion Indicates fusion features; The fusion features are input into the geometric algebra-long short-term memory network, and the spatiotemporal features are output through the long short-term memory network processing.

7. A PM2.5 concentration prediction method according to claim 6, characterized in that: The processing process of long short-term memory network is: Calculate the forget gate value: Among them, U and W represent weight matrices; b represents the bias term; σ represents the sigmoid activation function; h t-1 The hidden state at the previous moment; represents the fusion feature at time t; f t is the forget gate value; Geometric product operation; calculates the input gate value and generates a candidate value vector: Among them, i r is the input gate value; represents the candidate value vector; α represents the tanh activation function; Update unit status: Among them, C t is the current unit state; C t-1 is the unit state at the previous moment; Calculate the output gate value and get the final hidden state: h t =o t ⊙α(C t )), Among them, t Indicates the output gate value; h t Represents the final hidden state value, i.e., the spatiotemporal feature.

8. A PM2.5 concentration prediction method according to claim 1, characterized in that: The geometric algebraic space-time graph recurrent neural network model is optimized by the White Shark optimization algorithm.

9. A PM2.5 concentration prediction system, characterized in that: include: Data collection and processing module: used to collect relevant data of the target area and perform preprocessing to obtain preprocessed data; Feature extraction and fusion module: used to receive and process data, first extract spatial features through geometric algebra-graph convolutional neural network, extract spatial features at a certain moment and corresponding neighborhood comprehensive information and perform feature fusion, and pass the fused features through geometric algebra-long short-term memory network to obtain the corresponding spatiotemporal features of PM2.5; Prediction module: Predict PM2.5 concentration based on spatiotemporal characteristics.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method according to any one of claims 1 to 8 is implemented.

Images