An artificial intelligence-based multi-dimensional integral air quality assessment calculation system

Abstract

The application discloses a kind of multi-dimensional integral air quality evaluation calculation systems based on artificial intelligence, it is related to artificial intelligence environmental monitoring technical field.The system includes: multi-dimensional data acquisition terminal, obtains multi-source data;Multi-scale spatiotemporal feature extraction module, by constructing spatiotemporal collaborative tensor, and using diffusion-graph convolution network and path integral attention mechanism, generates dynamic pollution field representation;Multi-dimensional integral calculation module, using explainable neural integrator carries out spatiotemporal volume integration, outputs dynamic air quality comprehensive index and spatiotemporal cumulative flux;Air quality contribution decomposition module, based on contribution tensor decomposition technology generates multi-dimensional contribution atlas;Visual and interactive output module is used for result rendering and release.The application realizes the dynamic, collaborative, explainable accurate evaluation and pollution contribution traceability of regional air quality.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence environmental monitoring technology, and in particular to a multi-dimensional integral air quality assessment and calculation system based on artificial intelligence. Background Technology

[0002] Current air quality assessment systems primarily rely on single-index concentration measurements from discrete monitoring stations and linear weighted calculations using fixed formulas. This approach treats pollutants as independent static variables, neglecting the spatiotemporal synergistic effects, nonlinear superposition, and dynamic migration processes of multiple pollutants under complex meteorological and geographical conditions. While deep learning-based assessment models can capture some nonlinear relationships, they are mostly black-box models, lacking explicit physical mechanism embedding and interpretable contribution levels, making it difficult to dynamically integrate and trace source contributions of regional pollution accumulation processes. Traditional systems cannot continuously integrate and assess pollution fields across three spatial and temporal dimensions, nor can they quantify the specific contributions of different pollution sources and transport paths to the air quality of the target area.

[0003] Therefore, it is necessary to design a multidimensional integral air quality assessment and calculation system that integrates atmospheric physical mechanisms and artificial intelligence technology to achieve dynamic, coordinated, interpretable, and accurate assessment and contribution decomposition of regional air quality. Summary of the Invention

[0004] This invention proposes a multidimensional integral air quality assessment and calculation system based on artificial intelligence. To address the problem that traditional air quality assessment methods neglect the spatiotemporal coordination and dynamic processes of pollutants, a multidimensional integral calculation architecture that integrates atmospheric physical mechanisms and interpretable artificial intelligence is constructed. This system achieves dynamic quantitative assessment of regional air quality by integrating modules for multi-source data acquisition, multi-scale spatiotemporal feature extraction, multidimensional integral calculation, quality contribution decomposition, and visualization output.

[0005] This invention provides a multi-dimensional integral air quality assessment calculation system based on artificial intelligence, applicable to dynamic air quality assessment at the urban, industrial park, and regional scales. The system includes:

[0006] Multi-dimensional data acquisition terminal to acquire multi-source monitoring data within the target area;

[0007] The multi-scale spatiotemporal feature extraction module is connected to the multi-dimensional data acquisition terminal to receive and process multi-source monitoring data, construct a spatiotemporal co-tenent, and extract multi-scale spatiotemporal features through a diffusion-graph convolutional network and a path integral attention mechanism to generate a dynamic pollution field characterization.

[0008] The multidimensional integral calculation module is connected to the multi-scale spatiotemporal feature extraction module. Taking the dynamic pollution field characterization as input, it uses an interpretable neural integrator to perform spatiotemporal volume integral calculation on the target assessment area within a specified time window, and outputs the dynamic air quality comprehensive index and spatiotemporal cumulative flux.

[0009] The air quality contribution decomposition module is connected to the multidimensional integral calculation module and the multi-scale spatiotemporal feature extraction module. Based on the contribution tensor decomposition technology, it decomposes the dynamic air quality comprehensive index and spatiotemporal cumulative flux into quantitative results of contributions from different pollution source types, different transmission paths and different precursors, and generates a multidimensional contribution decomposition map.

[0010] The visualization and interactive output module, connected to the multidimensional integral calculation module and the air quality contribution decomposition module, integrates and visualizes the dynamic comprehensive air quality index, spatiotemporal cumulative flux, and multidimensional contribution decomposition map, and publishes it to the outside world through a standard data interface.

[0011] Furthermore, the multi-dimensional data acquisition terminal acquires multi-source monitoring data within the target area, specifically including the following steps:

[0012] Step S1: Air quality monitoring sensors deployed at fixed sites and mobile platforms in the target area collect time-series data on pollutant concentrations;

[0013] Step S2: Connect to the regional meteorological monitoring network and obtain multi-dimensional meteorological field data;

[0014] Step S3: Access the geographic information system and obtain static geographic information data. The pollutant concentration time series data and multi-dimensional meteorological field data obtained in steps S1 and S2, together with the static geographic information data obtained in this step, constitute multi-source monitoring data.

[0015] Step S4: After preprocessing, the multi-source monitoring data is combined into a spatiotemporally aligned multidimensional input data stream and sent to the multi-scale spatiotemporal feature extraction module.

[0016] Furthermore, the multi-scale spatiotemporal feature extraction module receives and processes multi-dimensional input data streams to generate a dynamic pollution field characterization, specifically including the following:

[0017] Spatiotemporal Co-operation Tensor Construction Unit: The preprocessed pollutant concentration time series data, multidimensional meteorological field data and static geographic information data are fused and upgraded according to a unified spatiotemporal grid to construct a spatiotemporal co-operation tensor;

[0018] Diffusion-Graph Convolutional Network Unit: Taking the spatiotemporal co-current tensor as input, it uses a graph convolution operator embedded with the unsteady advection-diffusion equation to perform message passing and feature aggregation on a dynamic graph composed of monitoring stations and virtual grid nodes, explicitly modeling the spatial transport process of pollutants under the action of non-uniform wind field and turbulent diffusion, and outputting node feature maps.

[0019] The path integral attention mechanism unit receives the node feature map, uses the pollution state of the previous moment as the initial distribution, uses the wind speed and direction sequence as the driving field, calculates the pollution quality transport contribution along multiple virtual trajectories through learnable attention weights, and integrates it to the grid state at the current moment, thereby generating a dynamic pollution field representation that can reflect the historical transport impact.

[0020] Furthermore, the multidimensional integral calculation module takes the dynamic pollution field characterization as input and outputs the dynamic comprehensive air quality index and spatiotemporal cumulative flux, specifically including the following:

[0021] Interpretable neural integrator unit: The dynamic pollution field is characterized in the three-dimensional spatial domain of the target area and discretized in the specified time window. A physically constrained integral kernel function is designed. The integral kernel function is used to perform weighted summation on each spatiotemporal unit to realize the spatiotemporal volume integral calculation under physical guidance and output the dynamic comprehensive air quality index.

[0022] Spatiotemporal cumulative flux calculation unit: During the neural integration process, the pollutant flux components at each boundary of the evaluation area are recorded synchronously and accumulated along the time dimension to calculate the net mass of pollutants input or output to the area from different directions during the evaluation time, i.e., the spatiotemporal cumulative flux.

[0023] Furthermore, the air quality contribution decomposition module generates a multidimensional contribution decomposition map, specifically including the following:

[0024] Contribution Tensor Construction Unit: By associating the dynamic pollution field characterization, pollution source inventory data, and transport path labels simulated by the reverse trajectory model, a high-order contribution tensor is constructed, whose dimensions correspond to the spatial grid, time step, pollution source type, and transport path category, respectively.

[0025] Tensor decomposition unit: The nonnegative tensor decomposition algorithm is used to reduce the dimensionality and decompose the high-order contribution tensor to obtain a set of low-rank spatiotemporal mode factors, source contribution factors and path contribution factors.

[0026] Contribution quantification and map generation unit: The various factors obtained from the decomposition are back-mapped and proportionally allocated with the dynamic air quality comprehensive index and spatiotemporal cumulative flux output by the multidimensional integral calculation module. The contribution percentage of each type of pollution source and each transport path to the final integral result is quantitatively calculated, and multidimensional contribution decomposition map is generated by merging the data in the form of heat map, streamline map and stacked chart.

[0027] By adopting the above solution, the beneficial effects achieved by the present invention are as follows:

[0028] This invention achieves high-fidelity modeling of the spatiotemporal synergistic effects and dynamic migration accumulation processes of multi-pollutant fields by constructing a spatiotemporal cooperative tensor and introducing a path integral attention mechanism. This mechanism can explicitly integrate atmospheric dynamic processes and simulate the historical transport impact of pollutants under complex wind fields, overcoming the limitations of traditional static or instantaneous assessment methods. This invention enhances the physical interpretability of model behavior and significantly improves the spatiotemporal accuracy and process realism of regional air quality dynamic assessment in scenarios of non-uniform distribution and rapid changes of pollutants.

[0029] This invention achieves multidimensional spatiotemporal volume integral calculation of dynamic pollution fields by designing a physically constrained interpretable neural integrator. This integrator embeds physical laws such as mass conservation into the learning process in the form of soft constraints, ensuring the physical rationality of the integration results. At the same time, it adaptively adjusts the weight of the integrator kernel through data-driven methods. This mechanism can accurately quantify and assess the overall pollution load and net flux of a region over a period of time, solving the problem that traditional point measurement or surface averaging methods cannot characterize the three-dimensional spatial cumulative effect. It provides an innovative technical tool for the assessment and quantification of regional environmental capacity and control effectiveness.

[0030] Furthermore, this invention innovatively proposes a contribution tensor decomposition method, which enables multi-dimensional contribution tracing of air quality assessment results. This method can quantitatively decompose the comprehensive index and cumulative flux to specific pollution source types and transport paths, generating an intuitive multi-dimensional contribution map. This mechanism significantly improves the precision and scientific level of diagnosis and source tracing analysis of air quality deterioration, and solves the problems of low spatiotemporal resolution and difficulty in directly correlating with dynamic assessment results in traditional source apportionment methods. It provides core data support and decision-making basis for formulating precise and differentiated regional joint prevention and control strategies. Attached Figure Description

[0031] Figure 1 This is a flowchart illustrating the operation of a multi-dimensional integral air quality assessment and calculation system based on artificial intelligence proposed in this invention.

[0032] Figure 2 This is a flowchart illustrating the generation of dynamic pollution field characterization by the multi-scale spatiotemporal feature extraction module proposed in this embodiment of the invention. Detailed Implementation

[0033] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0034] Example 1

[0035] according to Figure 1 This invention provides a multi-dimensional integral air quality assessment calculation system based on artificial intelligence, applicable to dynamic air quality assessment at the urban, industrial park, and regional scales. The system includes:

[0036] The multi-dimensional data acquisition terminal, deployed in fixed monitoring stations, mobile monitoring vehicles, and UAV platforms in the target area, includes optical particle counters, gas electrochemical sensors, meteorological six-element instruments, and miniature vertical detectors. It collects data on the concentration of various pollutants, three-dimensional wind field, temperature, humidity, pressure, and boundary layer height. Simultaneously, it connects to the pollution source online monitoring system, traffic flow database, and high-resolution geographic information database through data interfaces to obtain data on pollution source intensity, traffic activity intensity, and underlying surface type. All data undergoes timestamp calibration, outlier removal, and format standardization via edge computing nodes to generate a spatiotemporally aligned multi-dimensional input data stream.

[0037] The multi-scale spatiotemporal feature extraction module is deployed on a high-performance edge computing server. It receives multi-dimensional input data streams and fuses them into a five-dimensional tensor through a spatiotemporal co-construction tensor unit. Then, it uses a diffusion-graph convolutional network unit to simulate the spatial diffusion of pollutants and traces the historical transport impact through a path integral attention mechanism unit. Finally, it outputs a three-dimensional spatiotemporal dynamic pollution field characterization that can represent the dynamic distribution and evolution of pollutants.

[0038] The multidimensional integral calculation module, deployed in collaboration with the multi-scale spatiotemporal feature extraction module, receives dynamic pollution field characterizations from its interpretable neural integrator unit. It performs spatiotemporal volume integral calculations under physical constraints on the user-specified polygonal region and time interval (past 24 hours), and outputs the dynamic air quality comprehensive index (a dimensionless scalar) and the spatiotemporal cumulative flux of pollutants passing through each boundary of the region (a set of directional vector values) during that period.

[0039] The air quality contribution decomposition module calls the pollution source list and backward trajectory simulation data to construct a contribution tensor and perform non-negative tensor decomposition. It calculates the contribution ratio of different types of sources such as industrial sources, mobile sources, and dust sources, as well as different transport channels such as the northwest path and the southeast path, to the current dynamic air quality comprehensive index and spatiotemporal cumulative flux, and generates a multi-dimensional contribution decomposition map that includes spatial distribution and statistical charts.

[0040] The visualization and interactive output module uses WebGL technology for 3D rendering, dynamically displays the dynamic pollution field characterization in the form of isosurface cloud maps, and comprehensively presents the integral results and contribution decomposition results in the form of dashboards, Sankey diagrams and layered color maps. It also provides standardized data services for environmental management platforms and public information service applications through RESTful API and MQTT protocol.

[0041] Furthermore, all the control and calculation logic of the system of the present invention can also be implemented by a computer program. A computer-readable storage medium containing the program (such as a hard disk deployed in an edge server or cloud platform) can control the corresponding hardware devices (such as data acquisition terminals and computing servers) to work together when the program stored in it is loaded and run by the processor, thereby realizing all the functions described in this embodiment and constituting a complete integrated software and hardware evaluation system.

[0042] Example 2

[0043] according to Figure 2 This embodiment is based on Embodiment 1. In this embodiment, the multi-scale spatiotemporal feature extraction module generates a dynamic pollution field characterization, and its internal workflow specifically includes:

[0044] Spatiotemporal Coordination Tensor Construction Unit: A basic spatiotemporal grid is constructed using a 1 km × 1 km horizontal grid, 50 m vertical layers, and a 1 hour time step. The hourly pollutant concentrations, wind speed and direction vectors, temperature and humidity, and land type index features are mapped to this grid system through three-dimensional kriging interpolation and feature encoding methods, forming a five-dimensional spatiotemporal coordination tensor with dimensions of [number of longitude grids, number of latitude grids, number of altitude layers, number of time steps, and number of feature channels].

[0045] The diffusion-graph convolutional network unit treats each spatiotemporal grid node as a graph node, with the node feature being the multidimensional feature vector of the grid at the corresponding time. A dynamic directed weighted adjacency matrix is ​​constructed based on the spatial adjacency relationship between nodes and wind speed and direction information. The weights reflect the transmission probability between nodes under the influence of the wind field. A graph convolutional kernel embedded with the unsteady advection-diffusion equation is used to aggregate the features of the spatiotemporal cooperative tensor. Its message passing function simulates the transmission and diffusion process of pollutants from upstream nodes to downstream nodes. After multi-layer stacking, an updated node feature map is output, which already contains physically guided spatial correlation features.

[0046] The path integral attention mechanism unit receives node feature maps and, for each target grid node at the current time, backtracks the set of source regions it could potentially reach in the previous time (past 12 hours) driven by historical wind fields. Using a learnable multi-head attention network, it calculates the attention weights from source nodes at each historical time to the current target node. These weights simulate the contribution ratio of pollutants transported along the virtual trajectory. The dynamic pollution field representation of the current node is jointly determined by the weighted integral of its own features updated by the diffusion-graph convolutional network and the features of all historical source nodes. Finally, a dynamic pollution field representation containing historical transport memories is generated.

[0047] The core calculation formula for the above path integral attention mechanism unit is as follows:

[0048] ;

[0049] in, Is At any given moment, the dynamic pollution field characterization of the target grid node; It is a fusion function that can combine two parts: one part is the feature of the current node after being updated by the graph convolutional network, and the other part is the feature of the historical source nodes aggregated through the attention mechanism; This is a node. At any moment The node feature vector updated via a diffusion-graph convolutional network unit; It is a time index variable that iterates through historical moments, from... arrive Looking back One time step; It is the length of the time window for path backtracking; At a historical moment All possible influences on the current target node that may be transported through the wind farm The set of source nodes; It is a source node and belongs to the set. ; These are learnable attention weights, calculated by a multi-head attention network; It is the source node At a historical moment Node feature vectors; It is a double summation, over all historical moments and all possible source nodes at each historical moment. Perform a weighted summation.

[0050] Example 3

[0051] This embodiment is based on Embodiment 2. In this embodiment, the working process of the multidimensional integral calculation module specifically includes the following steps:

[0052] Step S1: Receive the dynamic pollution field characterization, which is essentially a four-dimensional data volume (longitude, latitude, altitude, and features).

[0053] Step S2: The user selects the three-dimensional geographic region and time interval to be evaluated through the interactive interface;

[0054] Step S3: The interpretable neural integrator unit discretizes the spatiotemporal region into tiny units (spatial volume elements) and (time steps). For each unit, an integral kernel function parameterized by a small neural network is defined. The neural network takes the center point coordinates, time, and local meteorological features of the unit as input and outputs a weight value. A regularization term is added to the training loss function of the neural network, and the output conforms to the principle of mass conservation.

[0055] Step S4: For the characteristic channels (fine particulate matter concentration field) representing the pollution level in the dynamic pollution field characterization, perform weighted summation and integration on discrete spatiotemporal units to calculate the dynamic air quality comprehensive index:

[0056] ;

[0057] in, It is a dynamic comprehensive air quality index; and These are the start and end times of the time interval selected by the user. It is a three-dimensional geographic space region; A learnable integral kernel function is a function consisting of parameters... The weight function is defined for a small neural network, and the input is the coordinates of the center point of the spatiotemporal unit. ,time In addition to local meteorological characteristics, the output is a weighted value; It is a pollutant concentration field; Represents a spatial volume element; Represents the time infinitesimal element; It is a summation of time; It is a spatial summation; These are discrete weight values; These are discrete concentration values; It is a discrete volume; It is the discrete time step; It is a spatiotemporal quadruple integral, which represents the continuous summation of the weighted pollutant concentration at every point in a specified three-dimensional spatial region from the start time to the end time.

[0058] Step S5: During the integration process, the pollutant flux passing through each boundary surface (east, west, south, north, top, bottom) of the three-dimensional geographic space region at each time step is calculated simultaneously. The flux is determined by the average concentration of the boundary surface, the wind speed component perpendicular to the interface, and the weight of the integration kernel. The fluxes over the entire time interval are accumulated to obtain the net pollutant mass flowing into the region from each direction, i.e., the spatiotemporal cumulative flux vector.

[0059] Step S6: Output the dynamic air quality composite index and spatiotemporal cumulative flux vector to the downstream module.

[0060] Example 4

[0061] This embodiment is based on Embodiment 1. In this embodiment, the air quality contribution decomposition module generates a multidimensional contribution decomposition map. The specific implementation process includes:

[0062] Step S1: Contribution tensor construction unit prepares input data, dynamic pollution field characterization provides pollution intensity for each grid at each time, pollution source inventory data provides spatial distribution intensity of different source categories, and trajectory model simulation obtains the main air mass arrival path for each grid at each time and classifies it into a finite number of path categories;

[0063] Step S2: Construct a fourth-order contribution tensor with four dimensions: spatial grid index, time step index, pollution source type index, and transport path category index. The initial value of each element in this tensor is determined by multiplying the pollution intensity of that grid at that moment by its classification into the [number]th [grid]. Class source and first Estimate by the membership degree of the class path;

[0064] Step S3: The tensor decomposition unit performs nonnegative CP decomposition on the fourth-order contribution tensor, approximating it as the sum of several rank tensors, expressed by the formula:

[0065] ;

[0066] in, It is a spatial factor vector, characterizing the first Spatial distribution of the patterns; It is a time factor vector, characterizing the first The temporal evolution of this pattern It is the source contribution factor vector, which characterizes the contribution ratio of each pollution source in the r-th mode. It is a path contribution factor vector, characterizing the first The contribution ratio of each transport path in the various modes, Represents the cross product of vectors; Let be the decomposition rank, be the number of potential patterns, and r be the pattern index, which is the r-th potential pollution pattern. For a fourth-order contribution tensor; This indicates that starting from the first potential contamination mode (r=1) and ending at the Rth potential contamination mode (r=R), the rank tensor corresponding to each mode is calculated sequentially. ), and then this By superimposing these rank tensors, we obtain a final tensor that can approximate the original fourth-order contribution tensor. The composite tensor;

[0067] Step S4: The contribution quantification and map generation unit quantifies and maps the factor vectors obtained from the decomposition, normalizes them, allocates the dynamic air quality comprehensive index to each decomposition mode proportionally, and further allocates the contribution to specific pollution source types and transport paths according to the source contribution factors and path contribution factors in each mode.

[0068] Step S5: Based on the above calculation results, generate a series of maps: a) Spatial contribution heat map: map the contributions of different sources back to geographic space for display; b) Contribution flow Sankey map: show the proportion of contributions of different source types to the evaluation area through different transport paths; c) Temporal contribution sequence map: show the changing trend of contributions from different sources over time; these maps together constitute a multidimensional contribution decomposition map and are published through the visualization and interactive output module.

[0069] The present invention and its embodiments have been described above. This description is not restrictive. The accompanying drawings are only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if a person skilled in the art is inspired by this description and designs a similar structure and embodiment without departing from the spirit of the present invention, such design should fall within the protection scope of the present invention.

Claims

1. A multi-dimensional integral air quality assessment and calculation system based on artificial intelligence, characterized in that, The system includes: a multi-dimensional data acquisition terminal, a multi-scale spatiotemporal feature extraction module, a multi-dimensional integral calculation module, an air quality contribution decomposition module, and a visualization and interactive output module; Multi-dimensional data acquisition terminal to acquire multi-source monitoring data within the target area; The multi-scale spatiotemporal feature extraction module is connected to the multi-dimensional data acquisition terminal to receive and process multi-source monitoring data, construct a spatiotemporal co-tenent, and extract multi-scale spatiotemporal features through a diffusion-graph convolutional network and a path integral attention mechanism to generate a dynamic pollution field characterization. The multidimensional integral calculation module is connected to the multi-scale spatiotemporal feature extraction module. Taking the dynamic pollution field characterization as input, it uses an interpretable neural integrator to perform spatiotemporal volume integral calculation on the target assessment area within a specified time window, and outputs the dynamic air quality comprehensive index and spatiotemporal cumulative flux. The air quality contribution decomposition module is connected to the multidimensional integral calculation module and the multi-scale spatiotemporal feature extraction module. Based on the contribution tensor decomposition technology, it decomposes the dynamic air quality comprehensive index and spatiotemporal cumulative flux into quantitative results of contributions from different pollution source types and different transmission paths, generating a multidimensional contribution decomposition map. The visualization and interactive output module is connected to the multidimensional integral calculation module and the air quality contribution decomposition module. It integrates and visualizes the dynamic air quality comprehensive index, spatiotemporal cumulative flux and multidimensional contribution decomposition map, and publishes it to the outside world through the standard data interface. The multi-scale spatiotemporal feature extraction module includes a spatiotemporal co-construction tensor unit, a diffusion-graph convolutional network unit, and a path integral attention mechanism unit. The process by which the multi-scale spatiotemporal feature extraction module generates a dynamic pollution field representation is as follows: The spatiotemporal collaborative tensor construction unit integrates and upgrades preprocessed pollutant concentration time-series data, multidimensional meteorological field data, and static geographic information data according to a unified spatiotemporal grid to construct a spatiotemporal collaborative tensor. The diffusion-graph convolutional network unit takes the spatiotemporal co-current tensor as input and uses graph convolution operators embedded with unsteady diffusion equations to perform message passing and feature aggregation on a dynamic graph composed of monitoring stations and virtual grid nodes. It explicitly models the spatial transport process of pollutants under the action of non-uniform wind field and turbulent diffusion, and outputs node feature maps. The path integral attention mechanism unit receives the node feature map, uses the pollution state of the previous time step as the initial distribution, uses the wind speed and direction sequence as the driving field, calculates the pollution quality transport contribution along multiple virtual trajectories through learnable attention weights, and integrates it to the grid state of the current time step, thereby generating a dynamic pollution field representation.

2. The multidimensional integral air quality assessment and calculation system based on artificial intelligence according to claim 1, characterized in that, The core calculation formula for the path integral attention mechanism unit is: ； in, Is Dynamic pollution field characterization at any given time; It is a fusion function; It is a node At any moment The node feature vector updated via a diffusion-graph convolutional network unit; It is a time index variable that iterates through historical moments, from... arrive ; It is the length of the time window for path backtracking; At a historical moment All possible influences on the current target node that may be transported through the wind farm The set of source nodes; It is a source node and belongs to the set. ; It is a learnable attention weight, calculated by a multi-head attention network; It is the source node At a historical moment Node feature vectors; It is a double summation, over all historical moments and all possible source nodes at each historical moment. Perform a weighted summation.

3. The multidimensional integral air quality assessment and calculation system based on artificial intelligence according to claim 1, characterized in that, The multidimensional integral calculation module includes an interpretable neural integrator unit and a spatiotemporal cumulative flux calculation unit. The process by which the multidimensional integral calculation module outputs the dynamic air quality composite index and spatiotemporal cumulative flux is as follows: The interpretable neural integrator unit represents the dynamic pollution field in the three-dimensional spatial domain of the target area and the specified time window, and uses a physically constrained integral kernel function parameterized by a neural network to perform weighted summation on each spatiotemporal unit, thereby realizing physical-guided spatiotemporal volume integration and outputting a dynamic comprehensive air quality index. During the integration process of the neural integrator unit, the spatiotemporal cumulative flux calculation unit simultaneously records the pollutant flux components at each boundary of the evaluation area and accumulates them along the time dimension. It then calculates the net mass of pollutants input into the area from different directions within the evaluation time, which is taken as the spatiotemporal cumulative flux.

4. The multidimensional integral air quality assessment and calculation system based on artificial intelligence according to claim 1, characterized in that, The air quality contribution decomposition module includes a contribution tensor construction unit, a tensor decomposition unit, and a contribution quantification and graph generation unit. The process by which the air quality contribution decomposition module generates a multidimensional contribution decomposition graph is as follows: The contribution tensor construction unit associates the dynamic pollution field characterization, pollution source inventory data, and transport path labels simulated by the reverse trajectory model to construct a high-order contribution tensor, whose dimensions correspond to the spatial grid, time step, pollution source type, and transport path category, respectively. The tensor decomposition unit uses a nonnegative tensor decomposition algorithm to reduce the dimensionality and decompose the high-order contribution tensor, resulting in a set of low-rank spatiotemporal pattern factors, source contribution factors, and path contribution factors. The contribution quantification and map generation unit performs reverse mapping and proportional allocation of the various factors obtained from the decomposition with the dynamic air quality comprehensive index and spatiotemporal cumulative flux, quantitatively calculates the percentage contribution of various pollution sources and transport paths to the final integral result, and integrates them in the form of heat map, streamline map and stacked chart to generate a multidimensional contribution decomposition map.

Images