A method and device for rapid prediction of air quality for large-scale sparse sampling scenarios

Abstract

The application provides a large-scale sparse sampling scene-oriented air quality rapid prediction method and device, and relates to the field of air quality prediction. The method comprises the following steps: acquiring air quality data through an air quality sensor group, wherein the air quality sensor group comprises a plurality of air quality sensors; acquiring a graph structure corresponding to the air quality sensor group according to the positional relationship of the plurality of air quality sensors in a study area, wherein the graph structure comprises a graph node connection relationship between the plurality of air quality sensors; constructing a data structure conversion component, wherein the data structure conversion component is used to convert the graph structure into a compact grid structure according to the graph node connection relationship between the plurality of air quality sensors, and construct a mapping relationship between the graph structure and the compact grid structure; constructing a lightweight parameter sharing spatio-temporal dilated convolution network component, wherein the lightweight parameter sharing spatio-temporal dilated convolution network component is used to quickly capture the spatio-temporal dependence relationship in the air quality data; constructing an integrated predictor through the data structure conversion component and the lightweight parameter sharing spatio-temporal dilated convolution network component according to the mapping relationship and the spatio-temporal dependence relationship; and acquiring an air quality prediction result through the integrated predictor according to the air quality data. The application solves the balance problem that the existing model is difficult to balance large-scale coverage and real-time response, and alleviates the problem that the calculation resources and the prediction accuracy are difficult to balance.

Description

Technical Field

[0001] This application relates to the field of air quality prediction, and in particular to a method and apparatus for rapid air quality prediction in large-scale sparse sampling scenarios. Background Technology

[0002] With the development of industrialization and urbanization, PM2.5... 2.5 Air pollution problems, including sulfur dioxide (SO2) and nitrogen dioxide (NO2), are becoming increasingly serious, posing a severe challenge to the sustainable development of human society. Studies show that long-term exposure to highly polluted environments may increase the incidence of respiratory diseases, cancer, and cardiovascular diseases. Against this backdrop, real-time and accurate air quality prediction models have become an important means of reducing human exposure to air pollution. These technologies can predict changes in pollutant concentrations and fluctuations in the air quality index in advance, providing guidance and reference for the public and governments to formulate targeted protective measures.

[0003] The key to air quality forecasting lies in uncovering the complex spatiotemporal relationships within historical data to accurately and efficiently infer future air quality data. In recent years, deep learning models have become the mainstream air quality forecasting models due to their excellent ability to capture complex spatiotemporal relationships. Although existing models have achieved satisfactory results in most air quality forecasting scenarios, they still face challenges in large-scale sparse sampling scenarios with high real-time requirements. Specifically, air quality sensors are typically sparsely distributed in large-scale sparse sampling scenarios. Common compact grid structure models struggle to effectively capture the spatiotemporal dependencies in sparse images, severely impacting the model's prediction accuracy. While common graph structure models can achieve better prediction accuracy in sparse scenarios, their complexity often increases exponentially with the number of nodes, significantly affecting computational efficiency. Therefore, in large-scale sparse sampling scenarios, existing models struggle to balance large-scale coverage and real-time response, resulting in a trade-off between computational resources and prediction accuracy.

[0004] Therefore, there is an urgent need for a method and device for rapid air quality prediction in large-scale sparse sampling scenarios. Summary of the Invention

[0005] This application provides a fast air quality prediction method for large-scale sparse sampling scenarios, which solves the problem that existing models cannot balance large-scale coverage and real-time response, and that there is a trade-off between computational resources and prediction accuracy.

[0006] The first aspect of this application provides a method for rapid air quality prediction in large-scale sparse sampling scenarios. The method includes: acquiring air quality data using an air quality sensor group, which comprises multiple air quality sensors; obtaining a graph structure corresponding to the air quality sensor group based on the positional relationships of the multiple air quality sensors in a study area, the graph structure including the connection relationships between graph nodes of the multiple air quality sensors; constructing a data structure transformation component, which converts the graph structure into a compact grid structure based on the connection relationships between the graph nodes of the multiple air quality sensors, and constructing a mapping relationship between the graph structure and the compact grid structure; constructing a lightweight parameter-shared spatiotemporal dilated convolutional network component, which captures the spatiotemporal dependencies in the air quality data; constructing an ensemble predictor based on the mapping relationship and the spatiotemporal dependencies using the data structure transformation component and the lightweight parameter-shared spatiotemporal dilated convolutional network component; and obtaining air quality prediction results based on the air quality data using the ensemble predictor.

[0007] Optionally, based on the graph node connection relationship between multiple air quality sensors, the graph structure is converted into a compact grid structure. Specifically, this includes: obtaining a first starting graph node corresponding to the target air quality sensor in the graph structure, where the target air quality sensor is any one of the multiple air quality sensors; obtaining the center pixel of the compact grid structure and constructing a first mapping relationship between the first starting graph node and the center pixel; based on the graph node connection relationship, if it is confirmed that a first-order adjacent graph node corresponding to the first starting graph node has been obtained, then obtaining the first-order neighboring pixel corresponding to the center pixel and constructing a second mapping relationship between the first-order adjacent graph node and the first-order neighboring pixel; obtaining a second starting graph node, where the second starting graph node is any one of the first-order adjacent graph nodes corresponding to the first starting graph node; based on the graph node connection relationship, if it is confirmed that a first-order adjacent graph node corresponding to the second starting graph node has not been obtained, then constructing a compact grid structure based on the center pixel and the first-order neighboring pixel.

[0008] Optionally, constructing a mapping relationship between the graph structure and the compact grid structure specifically includes: constructing a mapping relationship between the graph structure and the compact grid structure based on the first mapping relationship and the second mapping relationship.

[0009] Optionally, capturing the spatiotemporal dependencies in air quality data specifically includes: obtaining air quality data corresponding to multiple time windows; constructing a three-dimensional tensor corresponding to the air quality data through a data structure transformation component; and capturing the spatiotemporal dependencies in air quality data through a lightweight parameter-sharing spatiotemporal dilation convolutional network component using the three-dimensional tensor.

[0010] Optionally, based on air quality data, an integrated predictor obtains air quality prediction results, specifically including: obtaining the current time window among multiple time windows; obtaining multiple historical time windows, with the multiple historical time windows located before the current time window; obtaining historical air quality data corresponding to each of the multiple historical time windows; and the integrated predictor obtaining air quality prediction results based on the historical air quality data.

[0011] Alternatively, the air quality forecast result can be obtained according to the following formula:

[0012]

[0013] Where M represents the compact grid structure, LiEnPred←G represents the construction process of the compact grid structure, G represents the graph structure, and v i Let A be the i-th air quality sensor in the air quality sensor group, and let A represent the graph node connection relationship between air quality sensors, where A∈R. N×N R is a real number, and N is the total number of air quality sensors. To construct a three-dimensional tensor based on a compact grid structure, t is the t-th time window. Historical air quality data, and Where P and Q represent the length and width of the compact grid structure, respectively, B is the history-dependent step size, and b is any time window in [1,B]. This is the air quality forecast result, and F is the prediction step size, and f is any time window in [1,F].

[0014] Optionally, obtain the dilation factor; based on the dilation factor, reduce the depth of the neural network in the lightweight parameter-sharing spatiotemporally dilated convolutional network component according to the following formula:

[0015]

[0016] In the above formula,

[0017]

[0018] STPDCN represents the forward propagation process of a shared spatiotemporally dilated convolutional network component. Input historical air quality data; For predicted air quality data; These represent the expansion factors in the Q, P, and T dimensions of the l-th hidden layer, respectively. The convolutional kernels are shared in the hidden layers, where K T ,K P ,KQ These represent the convolution kernel sizes for T-dimensional, P-dimensional, and Q-dimensional dimensions, respectively. Let E be the hidden state of the air quality data in the l-th hidden layer, where E is the dimension of the hidden layer, which also corresponds to the number of convolutional kernels. Let e ​​be the three-dimensional tensor obtained from the e-th convolutional kernel; For three-dimensional tensors The elements in the array, p, q, and t, are three-dimensional tensors. Coordinates in P-dimensional, Q-dimensional, and T-dimensional coordinates; For three-dimensional tensors The elements in the array, where i, j, and s are three-dimensional tensors. The coordinates are P-dimensional, Q-dimensional, and T-dimensional; Conv is the convolution operation used for dimension alignment.

[0019] A second aspect of this application provides a rapid air quality prediction device for large-scale sparse sampling scenarios. The device includes an acquisition module, an integrated predictor construction module, and a prediction module, wherein...

[0020] The acquisition module is used to acquire air quality data through an air quality sensor group, which includes multiple air quality sensors; and to acquire the graph structure corresponding to the air quality sensor group based on the positional relationship of the multiple air quality sensors in the study area. The graph structure includes the graph node connection relationship between the multiple air quality sensors.

[0021] An integrated predictor building module is used to construct a data structure transformation component, which converts a graph structure into a compact grid structure based on the graph node connections between multiple air quality sensors, and establishes a mapping relationship between the graph structure and the compact grid structure. A lightweight parameter-shared spatiotemporally dilated convolutional network component is also constructed to capture the spatiotemporal dependencies in air quality data. Based on the mapping relationship and the spatiotemporal dependencies, an integrated predictor is built using the data structure transformation component and the lightweight parameter-shared spatiotemporally dilated convolutional network component.

[0022] The prediction module is used to obtain air quality prediction results based on air quality data through an integrated predictor.

[0023] A third aspect of this application provides an electronic device including a processor, a memory, a user interface, and a network interface. The memory is used to store instructions, the user interface and the network interface are used to communicate with other devices, and the processor is used to execute the instructions stored in the memory to cause the electronic device to perform the method as described above.

[0024] A fourth aspect of this application provides a computer-readable storage medium storing a computer program, which is executed by a processor using the method described in any of the foregoing descriptions.

[0025] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages:

[0026] 1. This application constructs a graph structure corresponding to a large number of air quality sensors in a large-scale sparse scene, and constructs a data structure conversion component. The data structure conversion component converts the graph structure into a compact grid structure based on the graph node connection relationship between the multiple air quality sensors. At the same time, a lightweight parameter-shared spatiotemporal dilated convolutional network component is constructed. The lightweight parameter-shared spatiotemporal dilated convolutional network component captures the spatiotemporal dependencies in the air quality data. Based on the mapping relationship and spatiotemporal dependency relationship, and through the data structure conversion component and the lightweight parameter-shared spatiotemporal dilated convolutional network component, an integrated predictor is constructed. By combining the compact grid structure and the graph structure, the prediction accuracy of the integrated predictor in a large-scale sparse scene is significantly improved.

[0027] 2. This application introduces an inflation factor and shares convolutional kernels between different layers, thereby enabling the convolutional network to significantly reduce the depth of the neural network without sacrificing prediction accuracy, and ensuring that the parameter size of the integrated predictor does not increase with the increase of the number of neural network layers.

[0028] 3. This application constructs a compact grid structure transformed from a graph structure and infers the air quality prediction results for future time windows based on historical air quality data corresponding to past time windows, thereby making the prediction accuracy more accurate and further improving the prediction accuracy of the integrated predictor in large-scale sparse scenes. Attached Figure Description

[0029] Figure 1 This is a distribution map of air quality sensors in a large-scale scene provided in an embodiment of this application;

[0030] Figure 2 This is a flowchart illustrating a method for rapid air quality prediction in large-scale sparse sampling scenarios provided in an embodiment of this application.

[0031] Figure 3 This is a schematic diagram of the transformation from a graph structure to a compact grid structure provided in an embodiment of this application;

[0032] Figure 4 This is a schematic diagram of another graph structure to compact grid structure transformation provided in an embodiment of this application;

[0033] Figure 5This is a schematic diagram of a 3D convolutional network operation provided in an embodiment of this application;

[0034] Figure 6 This is a scatter plot illustrating computational efficiency and prediction accuracy provided in an embodiment of this application;

[0035] Figure 7 This is a scatter plot diagram illustrating the parameter scale and prediction accuracy provided in an embodiment of this application;

[0036] Figure 8 This is a schematic diagram of a module for a rapid air quality prediction device for large-scale sparse sampling scenarios provided in an embodiment of this application;

[0037] Figure 9 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0038] Explanation of reference numerals in the attached figures: 81. Acquisition module; 82. Integrated predictor building module; 83. Prediction module; 901. Processor; 902. Communication bus; 903. User interface; 904. Network interface; 905. Memory. Detailed Implementation

[0039] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0040] The terminology used in the following embodiments of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. As used in the specification of this application, the singular expressions “a,” “an,” “the,” “the,” “the,” and “this” are intended to include the plural expressions as well, unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this application refers to and includes any or all possible combinations of one or more of the listed items.

[0041] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as implying or suggesting relative importance or implicitly indicating the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature, and in the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more.

[0042] To enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings.

[0043] Research has revealed that existing models face challenges in large-scale sparse sampling scenarios with high real-time requirements. Please refer to [reference needed]. Figure 1 The document presents a distribution map of air quality sensors in a large-scale scene, as provided in an embodiment of this application. Air quality sensors in large-scale scenes are typically sparsely distributed. Common grid models used for large-scale prediction struggle to effectively capture the spatiotemporal dependencies in sparse images, severely impacting prediction accuracy. While graph-based models can achieve better prediction accuracy in sparse scenes, their complexity often increases exponentially with the number of nodes, significantly affecting computational efficiency. In real-world scenarios, real-time large-scale sparse sampling prediction is crucial for public health protection and environmental sustainability. However, existing models struggle to balance large-scale coverage and real-time response, facing challenges in terms of computational resources and accuracy.

[0044] Therefore, this application proposes a method and apparatus for rapid air quality prediction in large-scale sparse sampling scenarios. It should be noted that, in the embodiments of this application, air quality data from over 2000 stations in China over the past three years were collected, and the prediction performance of the method in this application was verified using PM2.5 concentration as an example.

[0045] Please refer to Figure 2 The flowchart illustrates a method for rapid air quality prediction in large-scale sparse sampling scenarios provided by an embodiment of this application. The method is applied to a server, and the flowchart mainly includes the following steps: S201 to S206.

[0046] Step S201: Obtain air quality data through an air quality sensor group, which includes multiple air quality sensors.

[0047] Specifically, the air quality sensor group is deployed according to the specific needs of the monitoring area, typically involving the deployment of multiple air quality sensors. In this embodiment, the air quality sensor group is distributed across different geographical locations throughout China, specifically among the air quality sensors at over 2,000 stations over the past three years. The types of air quality sensors include, but are not limited to, monitoring devices for various air quality indicators such as PM2.5, PM10, nitrogen dioxide (NO2), carbon monoxide (CO), and ozone (O3). Each air quality sensor continuously collects air quality data according to a preset sampling frequency and data transmission strategy. In this embodiment, PM2.5 concentration is used as an example for illustration. The aforementioned air quality data is typically stored in time-series format, with each air quality sensor's collected data bearing a specific timestamp to ensure accurate alignment of data collected by different air quality sensors during subsequent analysis.

[0048] Step S202: Based on the positional relationship of multiple air quality sensors in the study area, obtain the graph structure corresponding to the air quality sensor group. The graph structure includes the graph node connection relationship between multiple air quality sensors.

[0049] Specifically, the composition diagram of all air quality sensors in the study area. Where v i ∈V represents the i-th air quality sensor in the graph structure, A∈R N×N This represents the graph node connections or dependencies between air quality sensors, where N represents the total number of all air quality sensors.

[0050] Step S203: Construct a data structure conversion component. The data structure conversion component is used to convert the graph structure into a compact grid structure based on the graph node connection relationship between multiple air quality sensors, and to construct the mapping relationship between the graph structure and the compact grid structure.

[0051] For details, please refer to Figure 3 This illustration shows a graph-to-compact grid structure transformation provided in an embodiment of this application. The Data Structure Transformation (DST) component generates a compact grid structure N∈R using the graph structure between air quality sensors. P×Q , where m pq >0 indicates that the pixel in row p and column q corresponds to a certain air quality sensor (m pq (equal to the index of the air quality sensor in the graph structure), m pq =0 indicates that the pixel in row p and column q of the compact grid structure does not correspond to an air quality sensor. P and Q represent the total length and width of the compact graph, respectively. During the conversion process, the data structure conversion component establishes a mapping relationship between the graph structure and the compact grid structure.

[0052] In one possible implementation, step 203 further includes: obtaining a first starting graph node corresponding to the target air quality sensor in the graph structure, wherein the target air quality sensor is any one of a plurality of air quality sensors; obtaining the center pixel of the compact grid structure, and constructing a first mapping relationship between the first starting graph node and the center pixel; according to the graph node connection relationship, if it is confirmed that a first-order adjacent graph node corresponding to the first starting graph node has been obtained, then obtaining the first-order neighboring pixel corresponding to the center pixel, and constructing a second mapping relationship between the first-order adjacent graph node and the first-order neighboring pixel; obtaining a second starting graph node, wherein the second starting graph node is any one of the first-order adjacent graph nodes corresponding to the first starting graph node, and according to the graph node connection relationship, if it is confirmed that a first-order adjacent graph node corresponding to the second starting graph node has not been obtained, then constructing a compact grid structure based on the center pixel and the first-order neighboring pixel.

[0053] For details, please refer to Figure 4 This diagram illustrates another graph structure-compact grid structure conversion provided in this embodiment. The data structure conversion component establishes a mapping relationship between graph nodes and grid pixels. Inspired by the breadth-first search algorithm, the data structure conversion component first determines the first starting graph node and its coordinates in the compact grid structure, i.e., the first starting graph node corresponding to the target air quality sensor in the graph structure. The target air quality sensor can be any one of multiple air quality sensors. In this embodiment, for ease of explanation, we can take the node at the very center of the graph structure and its corresponding air quality sensor as the first starting graph node and the target air quality sensor, and take the first starting graph node as the center pixel for constructing the compact grid structure. Then, based on the graph node connection relationship, the remaining graph nodes are determined layer by layer outward from the center pixel, sequentially determining their coordinates in the compact grid. Figure 4 As shown in Figure a, firstly, we map the first starting graph node v4 to the center pixel of the compact grid structure, and then determine whether the first starting graph node has a first-order adjacent graph node. A first-order adjacent graph node is a graph node directly connected to the first starting graph node, i.e. Figure 4 In the graphs v2, v5, v6, and v8, the first-order adjacent nodes of v4 are mapped clockwise to the first-order neighbor pixels of the center pixel, to form a graph like v4. Figure 4 The compact grid structure shown in Figure a establishes the first mapping relationship between the mapped first-order neighboring pixels and the first-order adjacent graph nodes. Then, we sequentially consider each first-order adjacent graph node of v4 as the center node, meaning we can use any first-order adjacent graph node corresponding to the first starting graph node as the second starting graph node, and determine the second-order neighboring graph nodes of v4. Second-order adjacent graph nodes are graph nodes directly connected to first-order adjacent graph nodes. Taking v2 as the second starting graph node as an example, the second-order adjacent graph nodes are v1 and v3 (for v2, v1 and v3 are its first-order adjacent graph nodes). Clockwise, we map the first-order adjacent graph nodes of v2 to the first-order neighboring pixels of the center pixel (which is v2 at this time), forming a structure as shown in Figure a. Figure 4 In the compact grid structure shown in Figure b, a second mapping relationship is constructed between the mapped second-order neighboring pixels and their second-order adjacent graph nodes. When the eight-neighbor pixels of the target node in the compact grid structure cannot accommodate a graph node, we expand outward from the target node until we find a pixel where a graph node can be placed, such as... Figure 4As shown in c and d. Repeat the above operations until all graph nodes are mapped to the compact grid structure. Furthermore, the data structure conversion component constructs the mapping relationships between adjacent graph nodes of order 1-n in the compact grid structure, namely, the first mapping relationship, the second mapping relationship, ..., the nth mapping relationship, and uses all the first mapping relationships, second mapping relationships, ..., the nth mapping relationship as the mapping relationship between the graph structure and the compact grid structure. In the embodiments of this application, the first mapping relationship and the second mapping relationship represent all mapping relationships between adjacent graph nodes in the compact grid structure, namely, the 1-nth mapping relationships.

[0054] Step S204: Construct a lightweight parameter-shared spatiotemporal dilated convolutional network component. This component is used to capture the spatiotemporal dependencies in air quality data.

[0055] Specifically, through the data structure transformation component, the air quality data from T time windows will be combined into a three-dimensional tensor. In real-time air quality prediction scenarios, rapid and accurate extraction from tensors is crucial. Uncovering the spatiotemporal dependencies in air quality data is crucial. However, most existing prediction models struggle to balance prediction accuracy and computational efficiency. High-accuracy models often suffer from complex design processes and low computational efficiency, making them unsuitable for real-time prediction. To address these challenges, we propose a lightweight parameter-shared spatio-temporal dilation convolution network (STP-DCN).

[0056] In one possible implementation, step 204 further includes: acquiring air quality data corresponding to multiple time windows; constructing a three-dimensional tensor corresponding to the air quality data through a data structure conversion component; and capturing the spatiotemporal dependencies in the air quality data through a lightweight parameter-sharing spatiotemporal dilation convolutional network component using the three-dimensional tensor.

[0057] Specifically, in a compact grid image, air pollutants (e.g., PM2.5) collected over T time windows... 2.5 (Concentrations of nitrogen, nitrogen, and carbon) form a three-dimensional tensor. The lightweight parameter-sharing spatiotemporally dilated convolutional network component captures the spatiotemporal dependencies in air quality data using three-dimensional tensors.

[0058] Step S205: Based on the mapping relationship and spatiotemporal dependency, an integrated predictor is constructed using a data structure transformation component and a lightweight parameter-sharing spatiotemporally dilated convolutional network component.

[0059] Specifically, the ensemble predictor consists of a data structure transformation component and a lightweight parameter-shared spatiotemporally dilated convolutional network component. The data structure transformation component is used to establish the function mapping relationship between graph nodes and compact grid pixels; the lightweight parameter-shared spatiotemporally dilated convolutional network component is used to mine the spatiotemporal dependencies in the compact grid pixels.

[0060] Step S206: Based on the air quality data, obtain the air quality prediction results through the integrated predictor.

[0061] Specifically, the air pollutant data collected in T time windows represent the air quality data corresponding to each of the multiple time windows, where X t ∈R P×Q x represents the air pollutants collected by the air quality sensor group in the t-th time window. pqt ∈R 1×1 Indicates air quality sensor m pq Air pollutants collected during the t-th time window. For any t∈[1,T], the data needs to be collected using historical three-dimensional tensors based on a compact grid structure M. Quickly and efficiently predict future air quality The specific formula is shown below.

[0062]

[0063] Where M represents the compact grid structure, LiEnPred←G represents the construction process of the compact grid structure, G represents the graph structure, and v i Let A be the i-th air quality sensor in the air quality sensor group, and let A represent the graph node connection relationship between air quality sensors, where A∈R. N×N R is a real number, and N is the total number of air quality sensors. To construct a three-dimensional tensor based on a compact grid structure, t is the t-th time window. Historical air quality data, and Where P and Q represent the length and width of the compact grid structure, respectively, B is the history-dependent step size, and b is any time window in [1,B]. This is the air quality forecast result, and F represents the prediction step size, and f is any time window in [1, F]. Using the above method, the air quality prediction results for future time windows are inferred based on historical air quality data corresponding to past time windows, thus making the prediction more accurate and further improving the prediction accuracy of the integrated predictor in large-scale sparse scenes.

[0064] In one possible implementation, step S204 further includes: obtaining the current time window among multiple time windows; obtaining multiple historical time windows, wherein the multiple historical time windows are located before the current time window; obtaining historical air quality data corresponding to the multiple historical time windows respectively; and integrating the predictor to obtain air quality prediction results based on the historical air quality data.

[0065] For details, please refer to Figure 5 The document presents a schematic diagram of a 3D convolutional network operation provided in an embodiment of this application. The lightweight parameter-shared spatiotemporal dilated convolutional network component extends from the classic 3D convolutional network. Compared to the classic 3D convolutional network, the lightweight parameter-shared spatiotemporal dilated convolutional network component constrains the convolutional operation to meet the time-dimensional dependency principle, ensuring that the convolutional kernel can only operate on time windows prior to the current time, i.e., multiple historical time windows are located before the current time window. Secondly, a dilation factor is introduced on the basis of the classic 3D convolutional network. Based on the dilation factor, the depth of the neural network in the lightweight parameter-shared spatiotemporal dilated convolutional network component is reduced, allowing the convolutional network to significantly reduce its depth without sacrificing prediction accuracy. Finally, the lightweight parameter-shared spatiotemporal dilated convolutional network component includes multiple neurons that share the same set of convolutional kernels, ensuring that the parameter size of the model does not increase with the increase of the number of neural network layers. Compared to the classic 3D convolutional network, the lightweight parameter-shared spatiotemporal dilated convolutional network component module has three main advantages. First, the lightweight parameter-shared spatiotemporally dilated convolutional network component satisfies the fundamental temporal constraints, making its convolution operations more efficient than those in classic 3D convolutional networks. Second, the lightweight parameter-shared spatiotemporally dilated convolutional network module controls the depth of the neural network through a dilation factor, resulting in higher computational efficiency than classic 3D convolutional networks. Finally, by sharing convolutional kernels across different layers, the lightweight parameter-shared spatiotemporally dilated convolutional network component has a smaller parameter size than classic 3D convolutional networks. (Using three-dimensional tensors) For example, the following equation illustrates the forward propagation process of a lightweight parameter-sharing spatiotemporally dilated convolutional network component:

[0066]

[0067] STPDCN represents the forward propagation process of a shared spatiotemporally dilated convolutional network component. Input historical air quality data; For predicted air quality data; These represent the expansion factors in the Q, P, and T dimensions of the l-th hidden layer, respectively. The convolutional kernels are shared in the hidden layers, where K T ,K P ,K Q These represent the convolution kernel sizes for T-dimensional, P-dimensional, and Q-dimensional dimensions, respectively. Let E be the hidden state of the air quality data in the l-th hidden layer, where E is the dimension of the hidden layer, which also corresponds to the number of convolutional kernels. Let e ​​be the three-dimensional tensor obtained from the e-th convolutional kernel; For three-dimensional tensors The elements in the array, p, q, and t, are three-dimensional tensors. Coordinates in P-dimensional, Q-dimensional, and T-dimensional coordinates; For three-dimensional tensors The elements in the array, where i, j, and s are three-dimensional tensors. The coordinates are P-dimensional, Q-dimensional, and T-dimensional; Conv is the convolution operation used for dimension alignment.

[0068] Furthermore, during the forward propagation of the lightweight parameter-sharing spatiotemporally dilated convolutional network component in step S204, the Lightweight Ensemble Predictor (LiEnPred) uses the air quality data from the first B time windows. Predict air quality data for the next F time windows Generally speaking, directly optimize the truth value. and predicted value The trained prediction model can be obtained by squared loss. However, considering that the compact grid structure matrix M may not be completely filled by graph nodes, we only include the loss at some locations in the loss function, as shown in the following formula:

[0069]

[0070] in, Indicates air quality sensor m pq The observations (ground truth) in the tb-th time window; Indicates air quality sensor m pq The predicted value in the tb-th time window; It means only m pq We will only calculate the loss if the value is greater than 0. This represents the learnable weights in the ensemble predictor.

[0071] It should be noted that in this application, we use real, large-scale sparse sampling data to evaluate the predictive performance of the LiEnPred method. The air quality data comes from 2014 monitoring sensors in China, including seven indicators: PM2.5, CO, NO2, PM10, SO2, O3, and AQI. The sampling interval for all seven indicators is 60 minutes, spanning from January 1, 2021 to December 31, 2023. Considering that the ensemble predictor is a universal air quality prediction model, we only used PM2.5 concentration data to verify the predictive performance of the ensemble predictor.

[0072] The baseline models in this application can be broadly classified into two categories: the former are spatiotemporal prediction models based on compact grid structures, including ConvLSTM, ST-3DNet, and ST-3DMDDN models; the latter are spatiotemporal prediction models based on graph structures, including DSTAGNN and STGDN models.

[0073] To fairly compare the prediction accuracy of the compact grid structure model and the graph structure model, this application addresses the correspondence between pixels and sensors in the sparse grid. Specifically, we use a 200×200 grid to divide the study area and ensure that one pixel corresponds to one sensor (if one pixel corresponds to multiple sensors, we assign the sensor to a neighboring pixel). The prediction accuracy results of the integrated predictor and the baseline model are presented together in Table 1.

[0074] [Table 1]

[0075] Evaluation metrics(RMSE / MAPE) of LiEnPred and baselines.

[0076]

[0077] The results show that the prediction accuracy of the second type of model is better than that of the first type of model, meaning that the graph structure model performs better than the compact grid structure model. This is because the graph structure model explicitly defines the spatial dependencies between sensors, thereby improving prediction capabilities. The compact grid structure model, however, struggles to capture these spatial dependencies in a sparse grid, reducing prediction performance. Furthermore, although the ensemble predictor is a compact grid structure model, its prediction accuracy is better than the first type of model, and in multi-step predictions, it approaches or surpasses the accuracy of the second type of model. This indicates that, without considering computational efficiency and parameter size, the ensemble predictor has achieved prediction accuracy superior to or comparable to the benchmark model. Further analysis of the computational efficiency of the ensemble predictor is presented in Table 2. The inference time primarily reflects the online prediction speed of the ensemble predictor, while the optimization time primarily reflects its offline training speed. The results show that the computational efficiency of the ensemble predictor is not only superior to that of the graph structure model but also superior to that of the classic compact grid structure model.

[0078] [Table 2]

[0079] Running time(milliseconds)of LiEnPred and baselines with batch size being 8.

[0080]

[0081] In addition, please refer to Figure 6 The document presents a scatter plot illustrating the computational efficiency and prediction accuracy provided in an embodiment of this application. Figure 6 In the table, the closer the scatter points are to the coordinate center, the better the computational efficiency and prediction accuracy of the model. The results demonstrate that the ensemble predictor balances prediction accuracy and computational efficiency to a certain extent, especially in large-scale sparse scenes. Similar to computational efficiency, we analyzed the parameter scale of the learnable weights in the ensemble predictor, and the results are shown in Table 3. The results show that the ensemble predictor significantly outperforms the ConvLSTM model, ST-3DNet model, ST-3DMDDN model, and DSTAGNN model in terms of parameter scale, and approaches the lightweight STGDN model.

[0082] [Table 3]

[0083] Parameter scale(kilobyte)of learnable weights for LiEnPred andbaselines.

[0084]

[0085]

[0086] In addition, please refer to Figure 7 It presents a scatter plot illustrating the parameter scale and prediction accuracy provided in an embodiment of this application. Figure 7 In the diagram, the closer the scatter points are to the coordinate center, the better the prediction accuracy and the lower the parameter size of the model. The results show that the ensemble predictor achieves acceptable prediction accuracy without significantly increasing the model parameter size.

[0087] This application employs the aforementioned method to construct a graph structure corresponding to a large number of air quality sensors in a large-scale sparse scene, and constructs a data structure conversion component. This component converts the graph structure into a compact grid structure based on the connection relationships between the graph nodes of the multiple air quality sensors. Simultaneously, a lightweight parameter-shared spatiotemporal dilated convolutional network component is constructed. This component captures the spatiotemporal dependencies in the air quality data. Compared to classic prediction models based on compact grid structures, this component, based on mapping relationships and spatiotemporal dependencies, constructs an ensemble predictor through the data structure conversion component and the lightweight parameter-shared spatiotemporal dilated convolutional network component. By combining the compact grid structure and the graph structure, the prediction accuracy of the ensemble predictor in a large-scale sparse scene is significantly improved.

[0088] Please refer to Figure 8 This illustration shows a schematic diagram of a rapid air quality prediction device for large-scale sparse sampling scenarios provided in an embodiment of this application. The device includes an acquisition module 81, an integrated predictor construction module 82, and a prediction module 83.

[0089] The acquisition module 81 is used to acquire air quality data through an air quality sensor group, which includes multiple air quality sensors; and to acquire the graph structure corresponding to the air quality sensor group, which includes the graph node connection relationships between the multiple air quality sensors.

[0090] The integrated predictor building module 82 is used to acquire a data structure transformation component, which is used to convert a graph structure into a compact grid structure based on the graph node connection relationship between multiple air quality sensors, and to build a mapping relationship between the graph structure and the compact grid structure; acquire a lightweight parameter-shared spatiotemporal dilated convolutional network component, which is used to capture the spatiotemporal dependencies in air quality data; and build an integrated predictor based on the mapping relationship and spatiotemporal dependencies through the data structure transformation component and the lightweight parameter-shared spatiotemporal dilated convolutional network component.

[0091] The prediction module 83 is used to obtain air quality prediction results based on air quality data through an integrated predictor.

[0092] In one possible implementation, the integrated predictor building module 82 is used to convert a graph structure into a compact grid structure based on the graph node connection relationship between multiple air quality sensors. Specifically, this includes: obtaining a first starting graph node corresponding to a target air quality sensor in the graph structure, where the target air quality sensor is any one of the multiple air quality sensors; obtaining the center pixel of the compact grid structure and constructing a first mapping relationship between the first starting graph node and the center pixel; based on the graph node connection relationship, if it is confirmed that a first-order adjacent graph node corresponding to the first starting graph node has been obtained, then obtaining the first-order neighboring pixel corresponding to the center pixel and constructing a second mapping relationship between the first-order adjacent graph node and the first-order neighboring pixel; obtaining a second starting graph node, where the second starting graph node is any one of the first-order adjacent graph nodes corresponding to the first starting graph node; based on the graph node connection relationship, if it is confirmed that a first-order adjacent graph node corresponding to the second starting graph node has not been obtained, then constructing a compact grid structure based on the center pixel and the first-order neighboring pixel.

[0093] In one possible implementation, the integrated predictor building module 82 is used to build a mapping relationship between the graph structure and the compact grid structure, specifically including: building a mapping relationship between the graph structure and the compact grid structure based on a first mapping relationship and a second mapping relationship.

[0094] In one possible implementation, the integrated predictor building module 82 is used to capture the spatiotemporal dependencies in air quality data, specifically including: acquiring air quality data corresponding to multiple time windows respectively; constructing a three-dimensional tensor corresponding to the air quality data through a data structure transformation component; and capturing the spatiotemporal dependencies in air quality data through a lightweight parameter-sharing spatiotemporal dilated convolutional network component using the three-dimensional tensor.

[0095] In one possible implementation, the prediction module 83 is used to obtain air quality prediction results through an integrated predictor based on air quality data. Specifically, this includes: obtaining the current time window among multiple time windows; obtaining multiple historical time windows, wherein the multiple historical time windows are located before the current time window; obtaining historical air quality data corresponding to the multiple historical time windows respectively; and the integrated predictor obtaining air quality prediction results based on the historical air quality data.

[0096] In one possible implementation, the prediction module 83 is used to obtain the air quality prediction result according to the following formula:

[0097]

[0098] Where M represents the compact grid structure, LiEnPred←G represents the construction process of the compact grid structure, G represents the graph structure, and v iLet A be the i-th air quality sensor in the air quality sensor group, and let A represent the graph node connection relationship between air quality sensors, where A∈R. N×N R is a real number, and N is the total number of air quality sensors. To construct a three-dimensional tensor based on a compact grid structure, t is the t-th time window. Historical air quality data, and Where P and Q represent the length and width of the compact grid structure, respectively, B is the history-dependent step size, and b is any time window in [1,B]. This is the air quality forecast result, and F is the prediction step size, and f is any time window from 1 to F.

[0099] In one possible implementation, prediction module 83 is used to obtain the dilation factor; based on the dilation factor, the depth of the neural network in the lightweight parameter-sharing spatiotemporally dilated convolutional network component is reduced according to the following formula:

[0100]

[0101] In the above formula,

[0102]

[0103] STPDCN represents the forward propagation process of a shared spatiotemporally dilated convolutional network component. Input historical air quality data; For predicted air quality data; These represent the expansion factors in the Q, P, and T dimensions of the l-th hidden layer, respectively. The convolutional kernels are shared in the hidden layers, where K T ,K P ,K Q These represent the convolution kernel sizes for T-dimensional, P-dimensional, and Q-dimensional dimensions, respectively. Let E be the hidden state of the air quality data in the l-th hidden layer, where E is the dimension of the hidden layer, which also corresponds to the number of convolutional kernels. Let e ​​be the three-dimensional tensor obtained from the e-th convolutional kernel; For three-dimensional tensors The elements in the array, p, q, and t, are three-dimensional tensors. Coordinates in P-dimensional, Q-dimensional, and T-dimensional coordinates; For three-dimensional tensors The elements in the array, where i, j, and s are three-dimensional tensors. The coordinates are P-dimensional, Q-dimensional, and T-dimensional; Conv is the convolution operation used for dimension alignment.

[0104] It should be noted that the above embodiments of the apparatus are only illustrated by the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the apparatus and method embodiments provided in the above embodiments belong to the same concept, and the specific implementation process can be found in the method embodiments, which will not be repeated here.

[0105] This application also provides an electronic device. (See reference...) Figure 9 , Figure 9 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. The electronic device may include: at least one processor 901, at least one communication bus 902, a user interface 903, at least one network interface 904, and a memory 905.

[0106] The communication bus 902 is used to enable communication between these components.

[0107] The user interface 903 may include a display screen and a camera. Optionally, the user interface 903 may also include a standard wired interface and a wireless interface.

[0108] The network interface 904 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface).

[0109] The processor 901 may include one or more processing cores. The processor 901 connects to various parts of the server using various interfaces and lines, and performs various server functions and processes data by running or executing instructions, programs, code sets, or instruction sets stored in the memory 905, and by calling data stored in the memory 905. Optionally, the processor 901 may be implemented using at least one hardware form of Digital Signal Processing (DSP), Field-Programmable Gate Array (FPGA), or Programmable Logic Array (PLA). The processor 901 may integrate one or a combination of several of the following: Central Processing Unit (CPU), Graphics Processing Unit (GPU), and modem. The CPU primarily handles the operating system, user interface, and applications; the GPU is responsible for rendering and drawing the content to be displayed on the screen; and the modem handles wireless communication. It is understood that the modem may also not be integrated into the processor 901 and may be implemented as a separate chip.

[0110] The memory 905 may include random access memory (RAM) or read-only memory. Optionally, the memory 905 may include a non-transitory computer-readable storage medium. The memory 905 may be used to store instructions, programs, code, code sets, or instruction sets. The memory 905 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for at least one function (such as touch function, sound playback function, image playback function, etc.), instructions for implementing the above-described method embodiments, etc.; the data storage area may store data involved in the above-described method embodiments, etc. Optionally, the memory 905 may also be at least one storage device located remotely from the aforementioned processor 901. (Refer to...) Figure 9 The memory 905, which serves as a computer storage medium, may include an operating system, a network communication module, a user interface module, and an application for rapid air quality prediction in large-scale sparse sampling scenarios.

[0111] exist Figure 9In the illustrated electronic device, the user interface 903 is primarily used to provide an input interface for the user and acquire user input data; while the processor 901 can be used to call the air quality rapid prediction application stored in the memory 905 for large-scale sparse sampling scenarios. When executed by one or more processors 901, the electronic device performs one or more of the methods described in the above embodiments. It should be noted that, for the foregoing method embodiments, for the sake of simplicity, they are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, because according to this application, some steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also understand that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0112] This application also provides a computer-readable storage medium storing instructions. When executed by one or more processors, these instructions cause an electronic device to perform one or more of the methods described in the above embodiments.

[0113] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0114] In the various embodiments provided in this application, it should be understood that the disclosed apparatus can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some service interface; the indirect coupling or communication connection between apparatuses or units may be electrical or other forms.

[0115] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0116] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0117] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage device (CMD). Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned memory includes various media capable of storing program code, such as USB flash drives, portable hard drives, magnetic disks, or optical disks.

[0118] The above description is merely an exemplary embodiment disclosed in this application and should not be construed as limiting the scope of this application. Any equivalent changes and modifications made in accordance with the teachings of this application shall still fall within the scope of this application. Those skilled in the art will readily conceive of other embodiments disclosed in this application upon considering the specification and practical truths presented.

[0119] This application is intended to cover any variations, uses, or adaptations disclosed herein that follow the general principles disclosed herein and include common knowledge or customary technical means in the technical field not described herein.

Claims

1. A method for rapid air quality prediction in large-scale sparse sampling scenarios, characterized in that, The method includes: Air quality data is acquired through an air quality sensor group, which includes multiple air quality sensors. Based on the positional relationship of the multiple air quality sensors in the study area, a graph structure corresponding to the air quality sensor group is obtained, and the graph structure includes the graph node connection relationship between the multiple air quality sensors. A data structure conversion component is constructed, which is used to convert the graph structure into a compact grid structure according to the graph node connection relationship between the multiple air quality sensors, and to construct the mapping relationship between the graph structure and the compact grid structure; A lightweight parameter-shared spatiotemporal dilated convolutional network component is constructed to capture the spatiotemporal dependencies in the air quality data. Based on the mapping relationship and the spatiotemporal dependency relationship, an integrated predictor is constructed through the data structure transformation component and the lightweight parameter-sharing spatiotemporally dilated convolutional network component; Based on the air quality data, air quality prediction results are obtained through the integrated predictor.

2. The method according to claim 1, characterized in that, The step of converting the graph structure into a compact grid structure based on the graph node connection relationships between the multiple air quality sensors specifically includes: Obtain the first starting graph node corresponding to the target air quality sensor in the graph structure, wherein the target air quality sensor is any one of the plurality of air quality sensors; Obtain the center pixel of the compact grid structure and construct a first mapping relationship between the first starting graph node and the center pixel; Based on the graph node connection relationship, if it is confirmed that the first-order adjacent graph node corresponding to the first starting graph node has been obtained, then the first-order neighboring pixel corresponding to the center pixel is obtained, and a second mapping relationship between the first-order adjacent graph node and the first-order neighboring pixel is constructed. Obtain a second starting graph node, which is any one of the first-order adjacent graph nodes corresponding to the first starting graph node. Based on the graph node connection relationship, if it is confirmed that no first-order adjacent graph node corresponding to the second starting graph node has been obtained, construct the compact grid structure based on the center pixel and the first-order neighboring pixels.

3. The method according to claim 2, characterized in that, The construction of the mapping relationship between the graph structure and the compact grid structure specifically includes: Based on the first mapping relationship and the second mapping relationship, the mapping relationship between the graph structure and the compact grid structure is constructed.

4. The method according to claim 1, characterized in that, The capture of the spatiotemporal dependencies in the air quality data specifically includes: Obtain the air quality data corresponding to multiple time windows; The data structure conversion component is used to construct the three-dimensional tensor corresponding to the air quality data; The lightweight parameter-shared spatiotemporally dilated convolutional network component captures the spatiotemporal dependencies in the air quality data using the three-dimensional tensor.

5. The method according to claim 4, characterized in that, The step of obtaining air quality prediction results through the integrated predictor based on the air quality data specifically includes: Obtain the current time window from the multiple time windows; Multiple historical time windows are obtained, and these multiple historical time windows are located before the current time window; The system acquires historical air quality data corresponding to multiple historical time windows, and the integrated predictor obtains the air quality prediction result based on the historical air quality data.

6. The method according to claim 5, characterized in that, The air quality prediction result is obtained according to the following formula: Where M represents the compact grid structure, LiEnPred←G represents the construction process of the compact grid structure, G represents the graph structure, and v i Let A be the i-th air quality sensor in the air quality sensor group, and let A represent the graph node connection relationship between the air quality sensors, where A ∈ R. N×N R is a real number, and N is the total number of the various air quality sensors. To construct the three-dimensional tensor based on the compact grid structure, t is the t-th time window. The historical air quality data, and Where P and Q represent the length and width of the compact grid structure, respectively, B is the history-dependent step size, and b is any time window in [1,B]. The air quality prediction result is given, and F is the prediction step size, and f is any time window in [1,F].

7. The method according to claim 6, characterized in that, After acquiring the lightweight parameter-shared spatiotemporally dilated convolutional network component, the method further includes: Obtain the expansion factor; Based on the dilation factor, the depth of the neural network in the lightweight parameter-sharing spatiotemporally dilated convolutional network component is reduced according to the following formula: In the above formula, Wherein, STPDCN represents the forward propagation process of the shared spatiotemporal dilated convolutional network component. For the historical air quality data mentioned above; The air quality prediction results are as follows; These are the magnitudes of the expansion factors in the Q, P, and T dimensions of the l-th hidden layer, respectively. The convolutional kernels shared in the hidden layers, where K T ,K P ,K Q These represent the convolution kernel sizes for T-dimensional, P-dimensional, and Q-dimensional dimensions, respectively. E represents the hidden state of the air quality data in the l-th hidden layer, where E is the dimension of the hidden layer, which also corresponds to the number of convolutional kernels. The three-dimensional tensor is obtained by the e-th convolutional kernel; For three-dimensional tensors The elements in the tensor, p, q, and t, are respectively the three-dimensional tensor. Coordinates in P-dimensional, Q-dimensional, and T-dimensional coordinates; For three-dimensional tensors The elements in the tensor, where i, j, and s are the three-dimensional tensor. The coordinates are P-dimensional, Q-dimensional, and T-dimensional; Conv is the convolution operation used for dimension alignment.

8. A rapid air quality prediction device for large-scale sparse sampling scenarios, characterized in that, The device includes an acquisition module, an integrated predictor construction module, and a prediction module, wherein, The acquisition module is used to acquire air quality data through an air quality sensor group, the air quality sensor group including multiple air quality sensors; and to acquire a graph structure corresponding to the air quality sensor group based on the positional relationship of the multiple air quality sensors in the study area, the graph structure including the graph node connection relationship between the multiple air quality sensors; The integrated predictor building module is used to construct a data structure transformation component, which is used to convert the graph structure into a compact grid structure based on the graph node connection relationship between multiple air quality sensors, and to construct a mapping relationship between the graph structure and the compact grid structure; construct a lightweight parameter-shared spatiotemporal dilated convolutional network component, which is used to capture the spatiotemporal dependencies in the air quality data; and construct an integrated predictor based on the mapping relationship and the spatiotemporal dependencies through the data structure transformation component and the lightweight parameter-shared spatiotemporal dilated convolutional network component. The prediction module is used to obtain air quality prediction results based on the air quality data through the integrated predictor.

9. An electronic device, characterized in that, The device includes a processor, a communication bus, a user interface, a network interface, and a memory. The memory is used to store instructions. The user interface and the network interface are used to communicate with other devices. The processor is used to execute the instructions stored in the memory to cause the electronic device to perform the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions that, when executed, perform the method as described in any one of claims 1 to 7.

Images