Drainage basin water environment quality dynamic evaluation method and system based on space-time diagram neural network

Abstract

The invention relates to the technical field of watershed ecological monitoring and intelligent evaluation, in particular to a watershed water environment quality dynamic evaluation method and system based on a space-time diagram neural network, and the method comprises the steps: associating N adjacent monitoring stations for a target monitoring station, and carrying out the monitoring data collection at the target monitoring station; preprocessing the monitoring data of the monitoring station, performing space-time alignment on the preprocessed monitoring data, and transmitting the monitoring data after space-time alignment to a cloud platform; and the cloud platform constructs a drainage basin space-time diagram according to the monitoring data, inputs the drainage basin space-time diagram into a pre-trained space-time diagram neural network model, and predicts to obtain an evaluation result reflecting a short-term emergency situation and a long-term development trend. According to the method, the watershed water habitat quality evaluation can be fused with multi-source spatio-temporal data, dynamic modeling of watershed associated features, real-time evaluation and trend prediction.

Description

Technical Field

[0001] This application relates to the field of watershed ecological monitoring and intelligent assessment technology, and in particular to a method and system for dynamic assessment of watershed water environment quality based on spatiotemporal graph neural networks. Background Technology

[0002] A watershed aquatic habitat refers to a specific space suitable for the survival and reproduction of aquatic organisms, formed by water bodies (oceans, lakes, rivers, wetlands, etc.) and their surrounding environment. Watershed aquatic habitat quality assessment is a core foundation of watershed ecological management. However, current watershed aquatic habitat quality assessment technologies are ill-suited to the characteristics of watersheds, which are characterized by strong spatiotemporal correlation, rapid dynamic changes, and complex influencing factors. Therefore, current watershed aquatic habitat quality assessment technologies suffer from the following problems: 1. Fragmented temporal and spatial relationships, resulting in low assessment accuracy: Traditional watershed aquatic habitat quality assessment methods (e.g., single-site index mean method, static weighted method) ignore the spatial relationships between different monitoring stations within the watershed (e.g., the impact of upstream pollution on downstream areas, hydrological coupling between tributaries and main streams) and temporal dynamics (e.g., seasonal changes, short-term impacts of hydrological events on aquatic habitats), leading to assessment results that fail to reflect the overall aquatic habitat quality of the watershed, with assessment errors exceeding 20%.

[0003] 2. Lagging dynamic response and weak predictive ability: Existing watershed aquatic habitat quality assessment systems are mostly based on historical data for post-event assessment (e.g., monthly / quarterly assessments), lacking the ability to dynamically model real-time data. They cannot capture the short-term impact of events such as sudden pollution and rainstorm runoff on watershed aquatic habitats, and the trend prediction lag time exceeds 7 days, making it difficult to support emergency decision-making.

[0004] 3. Difficulty in multi-dimensional data fusion and low information utilization: Existing watershed aquatic habitat monitoring data covers multiple dimensions such as "water quality physicochemical properties (e.g., DO, COD), biological communities (e.g., plankton, benthic animals), hydrometeorology (e.g., flow velocity, rainfall), and topography (e.g., watershed slope, land use)". However, existing watershed aquatic habitat quality assessment methods often use simple splicing or single-dimensional assessment of monitoring data, failing to fully explore the synergistic relationships between monitoring data, resulting in an information utilization rate of less than 50%. This leads to assessment dimensions that are singular, one-sided, and have low accuracy.

[0005] Therefore, how can we enable watershed aquatic habitat quality assessment to integrate multi-source spatiotemporal data, dynamically model watershed correlation characteristics, and conduct real-time assessment and trend prediction to solve the technical problems of "spatiotemporal fragmentation, delayed response, and insufficient integration" in existing technologies? Summary of the Invention

[0006] This application provides a method and system for dynamic assessment of watershed water environment quality based on spatiotemporal graph neural networks, so that the assessment of watershed aquatic habitat quality can integrate multi-source spatiotemporal data, dynamically model watershed correlation characteristics, and perform real-time assessment and trend prediction.

[0007] To solve the above-mentioned technical problems, this application provides the following technical solution: A dynamic assessment method for watershed water environment quality based on spatiotemporal graph neural network includes the following steps: associating N adjacent monitoring stations with a target monitoring station and collecting monitoring data at the target monitoring station; preprocessing the monitoring data from the monitoring stations, spatiotemporally aligning the preprocessed monitoring data, and transmitting the spatiotemporally aligned monitoring data to a cloud platform; the cloud platform constructs a watershed spatiotemporal graph based on the monitoring data, and inputs the watershed spatiotemporal graph into a pre-trained spatiotemporal graph neural network model to predict assessment results that reflect short-term emergencies and long-term development trends.

[0008] The above-described method for dynamic assessment of watershed water quality based on spatiotemporal graph neural networks preferably involves associating a target monitoring station with N adjacent monitoring stations, including the following sub-steps: calculating the spatial correlation coefficient between the target monitoring station and monitoring stations within a predetermined range; initially selecting N+M monitoring stations with larger spatial correlation coefficients; calculating the rank correlation coefficient between the target monitoring station and the initially selected monitoring stations based on historical monitoring data of the target monitoring station and historical monitoring data of the initially selected monitoring stations; obtaining the comprehensive correlation coefficient between the target monitoring station and the initially selected monitoring stations based on the spatial and rank correlation coefficients; and selecting N monitoring stations with larger comprehensive correlation coefficients from the N+M monitoring stations to associate with the target monitoring station.

[0009] The above-described method for dynamic assessment of watershed water environment quality based on spatiotemporal graph neural networks preferably includes preprocessing the monitoring data, comprising the following steps: calculating the temporal error value based on the mean and standard deviation of the monitoring data of the target monitoring station at the current moment and the historical monitoring datasets for the same period; if the absolute value of the temporal error value is greater than a first threshold, the monitoring data is preliminarily determined to be abnormal data; combining the monitoring data of the monitoring stations associated with the target monitoring station at the current moment, the spatial error value of the preliminarily abnormal monitoring data is calculated; if the absolute value of the spatial error value is greater than a second threshold, the monitoring data preliminarily determined to be abnormal is verified to be abnormal data, and the abnormal monitoring data is removed; after removing the abnormal monitoring data, the remaining monitoring data is standardized to complete the preprocessing of the monitoring data.

[0010] The above-described method for dynamic assessment of watershed water environment quality based on spatiotemporal graph neural networks preferably aligns monitoring data from different acquisition frequencies to a uniform time granularity based on millisecond-level timestamps to achieve spatiotemporal alignment of the monitoring data.

[0011] The above-described method for dynamic assessment of watershed water environment quality based on spatiotemporal graph neural networks preferably includes a watershed spatiotemporal graph comprising a spatial graph and spatiotemporal sequence features. Constructing the watershed spatiotemporal graph involves the following steps: treating each monitoring station as a node in the spatial graph, and using all monitoring data of a node at the same time as the feature vector of that node; calculating the edge weights between two nodes in the spatial graph based on geographical and physical factors between adjacent nodes; converting the monitoring data of each node into time-series data, and taking the time-series data of each node over the past T hours to form the spatiotemporal sequence features of each node.

[0012] A dynamic assessment system for watershed water environment quality based on a spatiotemporal graph neural network includes: a distributed monitoring subnetwork, edge computing nodes, communication equipment, and a cloud platform. Monitoring stations are deployed according to the watershed hydrological topology, with each target monitoring station associated with N adjacent monitoring stations. The distributed monitoring subnetwork collects monitoring data at the target monitoring stations. The edge computing nodes preprocess the monitoring data and perform spatiotemporal alignment on the preprocessed data. The communication equipment transmits the spatiotemporally aligned monitoring data to the cloud platform. The cloud platform constructs a watershed spatiotemporal graph based on the monitoring data and inputs the watershed spatiotemporal graph into a pre-trained spatiotemporal graph neural network model to predict assessment results reflecting short-term emergencies and long-term development trends.

[0013] In the above-described dynamic assessment system for watershed water environment quality based on spatiotemporal graph neural networks, preferably, edge computing nodes calculate the spatial correlation coefficient between the target monitoring station and monitoring stations within a predetermined range, initially select N+M monitoring stations with larger spatial correlation coefficients, calculate the rank correlation coefficient between the target monitoring station and the initially selected monitoring stations based on historical monitoring data of the target monitoring station and historical monitoring data of the initially selected monitoring stations, obtain the comprehensive correlation coefficient between the target monitoring station and the initially selected monitoring stations based on the spatial correlation coefficient and rank correlation coefficient, and select N monitoring stations with larger comprehensive correlation coefficients from the N+M monitoring stations to associate with the target monitoring station.

[0014] In the above-described dynamic assessment system for watershed water environment quality based on spatiotemporal graph neural networks, preferably, the edge computing node calculates the temporal error value based on the mean and standard deviation of the monitoring data of its target monitoring station at the current moment and the historical monitoring datasets of the same period. If the absolute value of the temporal error value is greater than a first threshold, the edge computing node initially determines that the monitoring data is abnormal. The edge computing node then combines the monitoring data of neighboring monitoring stations associated with the target monitoring station at the current moment to calculate the spatial error value of the initially abnormal monitoring data. If the absolute value of the spatial error value is greater than a second threshold, the edge computing node verifies that the monitoring data initially determined to be abnormal is indeed abnormal, and removes the abnormal monitoring data. After removing the abnormal monitoring data, the edge computing node standardizes the remaining monitoring data to complete the preprocessing of the monitoring data.

[0015] In the above-described dynamic assessment system for watershed water environment quality based on spatiotemporal graph neural networks, preferably, edge computing nodes align monitoring data from different acquisition frequencies to a unified time granularity based on millisecond-level timestamps to complete the spatiotemporal alignment of the monitoring data.

[0016] As described above, the watershed water environment quality dynamic assessment system based on spatiotemporal graph neural network preferably includes a watershed spatiotemporal graph comprising a spatial graph and spatiotemporal sequence features. The cloud platform treats each monitoring station as a node in the spatial graph, calculates the edge weights between two nodes in the spatial graph based on the geographical and physical factors between adjacent nodes, converts the monitoring data of each node into time series data, and takes the time series data of each node for the past T hours to form the spatiotemporal sequence features of each node, thereby constructing the watershed spatiotemporal graph.

[0017] Beneficial effects: 1. High assessment accuracy: The spatiotemporal graph neural network model can integrate multi-source data and spatiotemporal correlation, with a relative error of ≤5% in HQI assessment, which is 75% higher than traditional methods (error 20%+), accurately reflecting the overall watershed quality status; 2. Fast dynamic response: Supports real-time evaluation at the 1-hour granularity, enabling real-time evaluation and trend prediction; 3. Strong multi-source fusion capability: It integrates four types of data: water quality, biology, hydrology and topography, increasing the information utilization rate from 50% to 90%, and making the assessment dimensions more comprehensive, avoiding the one-sidedness of single data; 4. High scalability: Supports the addition and removal of monitoring stations (automatic updating of spatiotemporal maps) and the expansion of assessment indicators (such as adding biotoxicity indicators), and can be adapted to watersheds of different sizes (from 100km² to 10,000km²). Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0019] Figure 1 This is a schematic diagram of a watershed water environment quality dynamic assessment system based on a spatiotemporal graph neural network; Figure 2 This is a flowchart of a method for dynamic assessment of watershed water environment quality based on spatiotemporal graph neural networks; Figure 3 This is a flowchart for identifying neighboring monitoring stations associated with a target monitoring station; Figure 4 This is a flowchart for preprocessing monitoring data; Figure 5 This is a flowchart for constructing a watershed spatiotemporal map. Detailed Implementation

[0020] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0021] Example 1 like Figure 1 As shown, this application provides a watershed water environment quality dynamic assessment system 100 based on spatiotemporal graph neural network, including: a distributed monitoring sub-network 110, an edge computing node 120, a communication device 130, and a cloud platform 140; wherein, monitoring stations are deployed according to the watershed hydrological topology (tributary-main stream-estuary), and each monitoring station is associated with N adjacent monitoring stations. At each monitoring station, water quality physicochemical sensors 111, biological monitoring devices 112, hydrological and meteorological devices 113, and topography and land use modules 114 are deployed, thereby forming a distributed monitoring sub-network 110 at each monitoring station. At each monitoring station, an edge computing node 120 (e.g., equipped with an embedded edge computing platform NVIDIA Jetson Xavier NX) and a communication device 130 are also deployed.

[0022] The distributed monitoring sub-network 110 collects monitoring data (water quality physicochemical monitoring data, aquatic organism monitoring data, hydrological monitoring data, and topographic and land use monitoring data) at the monitoring stations. The edge computing node 120 preprocesses the monitoring data collected by the distributed monitoring sub-network 110 at the monitoring stations (e.g., outlier removal, standardization) and performs spatiotemporal alignment on the preprocessed monitoring data. The edge computing node 120 performs partial processing on the monitoring data, thereby reducing the pressure of data transmission to the cloud platform 140 and thus reducing the data processing latency of the cloud platform 140, for example, ≤1 second. The communication device 130 transmits the spatiotemporally aligned monitoring data to the cloud platform 140. The communication device 130 can use "4G / 5G+LoRa (Long Range Radio)" dual-mode communication to ensure the stability of data transmission at remote monitoring stations (e.g., packet loss rate < 0.1%).

[0023] The cloud platform 140 receives the spatiotemporally aligned monitoring data and constructs a watershed spatiotemporal map based on the received monitoring data. The watershed spatiotemporal map is then input into a pre-trained spatiotemporal neural network model to predict assessment results that reflect short-term emergencies and long-term development trends. For example, the cloud platform 140 uses a distributed server cluster (CPU: Intel Xeon Gold 6330, GPU: NVIDIA A100) to support large-scale data storage and computation of the spatiotemporal neural network model.

[0024] Example 2 like Figure 2 As shown, this application provides a method for dynamic assessment of watershed water environment quality based on spatiotemporal graph neural networks, including the following steps: Step S210: Associate N adjacent monitoring stations with the target monitoring station and collect monitoring data at the target monitoring station; Each monitoring station is associated with N neighboring monitoring stations. If the number of neighboring monitoring stations associated with the target monitoring station is too large, interference can easily be introduced when verifying the monitoring data collected by the target monitoring station using the monitoring data collected by the neighboring monitoring stations. If the number of neighboring monitoring stations associated with the target monitoring station is too small, the reliability of verifying the monitoring data collected by the target monitoring station using the monitoring data collected by the neighboring monitoring stations is low. Therefore, in this application, the number of neighboring monitoring stations associated with the target monitoring station is limited. And N is an integer.

[0025] Among them, such as Figure 3 As shown, identifying the monitoring stations associated with a target monitoring station includes the following sub-steps: Step S211: Calculate the spatial correlation coefficient between the target monitoring station and the monitoring stations within its predetermined range, and initially select N+M monitoring stations with larger spatial correlation coefficients; To avoid low correlation due to terrain / regional differences, this application calculates the spatial correlation coefficient between the target monitoring station and each monitoring station within the predetermined range of the target monitoring station, and initially selects N+M monitoring stations with larger spatial correlation coefficients in descending order of spatial correlation coefficient.

[0026] Specifically, the formula for calculating the spatial correlation coefficient is as follows: ; in, For the target monitoring station and its predetermined range Spatial correlation coefficient between monitoring stations; The average location information of the target monitoring stations; For the target monitoring station Location information; For the first Average location information of each monitoring station; For the first The first monitoring station Location information; The quantity of location information; in formula (1) ,and .

[0027] Step S212: Based on the historical monitoring data of the target monitoring station and the historical monitoring data of the initially selected monitoring stations, calculate the rank correlation coefficient between the target monitoring station and the initially selected monitoring stations; Historical monitoring data from the target monitoring stations are categorized and arranged in ascending order within each category, and ranks are assigned to the sorted historical monitoring data for the target monitoring stations. Similarly, historical monitoring data from the initially selected monitoring stations are categorized and arranged in ascending order within each category, and ranks are assigned to the sorted historical monitoring data for the initially selected monitoring stations. The difference between the ranks of the historical monitoring data for the target monitoring stations and the ranks of the historical monitoring data for the initially selected monitoring stations is calculated. The historical monitoring data for the initially selected monitoring stations and the target monitoring stations refer to monitoring data from the same historical point in time.

[0028] Specifically, the formula for calculating the difference in rank is as follows: ; in, For the target monitoring station The rank of historical monitoring data and the initial selection of the first The first monitoring station The difference in rank between the ranks of historical monitoring data; For the target monitoring station The rank of historical monitoring data; For the initial selection of the first The first monitoring station The rank of each historical monitoring data point; in formula (2) .

[0029] Then, based on the difference between the ranks of the historical monitoring data of the target monitoring station and the ranks of the historical monitoring data of the initially selected monitoring stations, the rank correlation coefficient between the target monitoring station and the initially selected monitoring stations is calculated.

[0030] Specifically, the formula for calculating the rank correlation coefficient is as follows: ; in, For the target monitoring sites and the initially selected first The rank correlation coefficient of each monitoring station; The quantity of historical monitoring data; in formula (3) .

[0031] Step S213: Obtain the comprehensive correlation coefficient between the target monitoring station and the initially selected monitoring station based on the spatial correlation coefficient and rank correlation coefficient between the target monitoring station and the initially selected monitoring station; The target monitoring stations are compared with the initially selected... Spatial correlation coefficient between monitoring stations Multiply by the spatial coefficient weight value, and combine the target monitoring station with the initially selected first... Rank correlation coefficient of each monitoring station Multiply by the rank coefficient weight value, and then sum the two products as the target monitoring station and the initially selected first... The overall correlation coefficient between the monitoring stations.

[0032] Specifically, the formula for calculating the comprehensive correlation coefficient is as follows: ; in, For the target monitoring sites and the initially selected first The overall correlation coefficient between the monitoring stations; These are the spatial coefficient weight values; The rank coefficient weight value; in formula (4) .

[0033] Step S214: Select N monitoring stations with the largest comprehensive correlation coefficient from N+M monitoring stations to associate them with the target monitoring station; In the calculation After calculating the comprehensive correlation coefficient, N monitoring stations with larger comprehensive correlation coefficients are selected from the N+M monitoring stations in descending order of comprehensive correlation coefficient. Then, the distributed monitoring sub-network 110 in these N monitoring stations is connected to the edge computing node 120 in the target monitoring station to realize the association between these N monitoring stations and the target monitoring station. The distributed monitoring sub-network 110 in these N monitoring stations will share the collected monitoring data with the edge computing node 120 in the interconnected target monitoring station.

[0034] The distributed monitoring subnetwork 110 includes: water quality physicochemical sensors 111, biological monitoring equipment 112, hydrological and meteorological equipment 113, and a topography and land use module 114. The water quality physicochemical sensors 111 are used to sample water at monitoring stations according to a predetermined sampling frequency, obtaining water quality physicochemical monitoring data such as DO, COD, ammonia nitrogen, total phosphorus, pH, and turbidity. Optionally, the sampling frequency is 5 minutes / time, and the monitoring accuracy complies with the "Technical Specification for Automatic Monitoring of Surface Water" (HJ 915-2017). The biological monitoring equipment 112 is used to sample aquatic organisms at monitoring stations according to a predetermined sampling frequency, obtaining aquatic organism monitoring data such as chlorophyll, algae density, and microbial content. Optionally, the biological monitoring equipment 112 includes: a fully automatic plankton imager (resolution 0.5 μm / pixel), a benthic organism sampler, etc. Optionally, the sampling frequency is 24 hours / time. Hydrological and meteorological equipment 113 is used to conduct hydrological and meteorological sampling at monitoring stations according to a predetermined sampling frequency, obtaining hydrological monitoring data such as water level, flow velocity, flow rate, water depth, and sediment content, and meteorological monitoring data such as air temperature, humidity, precipitation, wind speed, wind direction, air pressure, sunshine duration, evaporation, and radiation. Optionally, hydrological and meteorological equipment 113 includes: Doppler current meter, water level gauge, and rainfall sensor. Optionally, the sampling frequency is 10 minutes / time. Topography and land use module 114 is used to conduct topography and land use sampling at monitoring stations according to a predetermined sampling frequency, obtaining topography and land use monitoring data such as watershed topography and land use types. Optionally, topography and land use module 114 includes: UAVs, remote sensing satellites, etc., wherein the UAV (equipped with a multispectral camera) acquires watershed topography data quarterly, and the remote sensing satellite (e.g., Sentinel-2) acquires land use types (e.g., cultivated land, forest land, construction land).

[0035] Step S220: Preprocess the monitoring data from the monitoring stations, perform spatiotemporal alignment on the preprocessed monitoring data, and transmit the spatiotemporally aligned monitoring data to the cloud platform; Among them, such as Figure 4 As shown, edge computing node 120 preprocesses the monitoring data, including the following steps: Step S221: The edge computing node calculates the time dimension error value based on the monitoring data of its target monitoring station at the current moment and the mean and standard deviation of the historical monitoring dataset for the same period. At time t, the distributed monitoring subnetwork 110 of the target monitoring station collects monitoring data. The edge computing node 120 of the target monitoring station calls its stored historical monitoring dataset for the same period and calculates the mean of the historical monitoring dataset. and standard deviation Then, the edge computing node 120 of the target monitoring station uses the monitoring data... The average of the monitoring datasets from the same historical period and standard deviation Calculate the time-dimensional error value.

[0036] Specifically, the formula for calculating the time-dimensional error value is as follows: ; in, The monitoring data collected by the distributed monitoring sub-network 110 of the target monitoring station at time t. The time dimension error value.

[0037] Step S222: If the absolute value of the time dimension error is greater than the first threshold, the edge computing node initially determines that the monitoring data is abnormal data; The monitoring data collected at time t is calculated on edge computing node 120. Time dimension error value Then, determine the monitoring data collected at time t. Time dimension error value With the first threshold If the relationship, Then, edge computing node 120 initially determines the monitoring data. If the data is abnormal, proceed to step S223; if Then edge computing node 120 determines the monitoring data. Normal, process terminated.

[0038] Step S223: The edge computing node combines the monitoring data of neighboring monitoring stations associated with the target monitoring station at the current moment to calculate the spatial dimension error value of the preliminary anomaly monitoring data; At time t, the distributed monitoring subnetwork 110 of neighboring monitoring stations associated with the target monitoring station also collects monitoring data. The edge computing node 120 of the target monitoring station calls the monitoring data of these neighboring monitoring stations associated with the target monitoring station. And compare it with the monitoring data of the target monitoring station that was initially determined to be abnormal. By comparison, monitoring data that was initially determined to be abnormal was obtained. The empty dimension error value.

[0039] Specifically, the formula for calculating the spatial dimension error value is as follows: ; in, The monitoring data collected by the distributed monitoring sub-network 110 of the target monitoring station at time t. The spatial dimension error value; For the first Monitoring data collected by each monitoring station at time t; For the first The distance weight between each monitoring station and the target monitoring station; in formula (6) .

[0040] Step S224: If the absolute value of the spatial dimension error value is greater than the second threshold, the edge computing node verification is initially determined to be abnormal monitoring data. The abnormal monitoring data is removed. The monitoring data collected at time t is calculated on edge computing node 120. Spatial dimension error value Then, determine the monitoring data collected at time t. Spatial dimension error value With the second threshold If the relationship, The monitoring data from edge computing node 120 was initially identified as abnormal. If the data is indeed abnormal, remove the abnormal monitoring data; if The monitoring data from edge computing node 120 was initially identified as abnormal. This is normal data.

[0041] Step S225: After removing abnormal monitoring data, the edge computing node standardizes the remaining monitoring data to complete the preprocessing of the monitoring data. After removing abnormal monitoring data from the edge computing node 120, missing data can be filled in for the remaining normal monitoring data, for example, using spatiotemporal interpolation. After filling in the missing data, the edge computing node 120 also standardizes the monitoring data to eliminate the dimensions of the monitoring data, thus facilitating subsequent processing. For example, water quality physicochemical monitoring data (e.g., DO, COD) are normalized to the [0,1] interval using min-max normalization; aquatic biological monitoring data (e.g., Shannon-Wiener index) are standardized using Z-score (mean 0, standard deviation 1) to eliminate magnitude differences; topography and land use monitoring data (e.g., cultivated land, forest land, etc.) are encoded into one-hot vectors using type encoding, and watershed slope, altitude, etc. are converted into dimensionless indices.

[0042] In addition, the edge computing node 120 can align monitoring data with different collection frequencies (e.g., 5 minutes / time, 10 minutes / time, 24 hours / time) to a unified time granularity (1 hour / time) based on millisecond-level timestamps, so as to complete the spatiotemporal alignment of the monitoring data and ensure the spatiotemporal consistency of the monitoring data.

[0043] Step S230: The cloud platform constructs a watershed spatiotemporal map based on the monitoring data, and inputs the watershed spatiotemporal map into a pre-trained spatiotemporal map neural network model to predict the assessment results that reflect short-term emergencies and long-term development trends. The cloud platform 140 receives the spatiotemporally aligned monitoring data and constructs a watershed spatiotemporal map based on the received monitoring data. The watershed spatiotemporal map is then input into a pre-trained spatiotemporal map neural network model to predict assessment results that reflect short-term emergencies and long-term development trends. The watershed spatiotemporal map includes spatial map and spatiotemporal sequence features. Through the watershed spatiotemporal map, the spatial correlation between monitoring stations and the temporal dependency between monitoring data can be quantified.

[0044] Among them, such as Figure 5 As shown, the cloud platform 140 constructs a watershed spatiotemporal map, including the following steps: Step S231: Treat each monitoring station as a node in the spatial map, and treat all monitoring data of a node at the same time as the feature vector of that node;

[0045] Each monitoring station in the watershed is taken as a node in the spatial map, and all the monitoring data of a node at the same time (water quality physicochemical (6-dimensional), biological community (3-dimensional), hydrological and meteorological (4-dimensional), topography and land use (5-dimensional)) are taken as the feature vector of that node (18-dimensional feature vector).

[0046] Step S232: Calculate the edge weights between two nodes in the spatial graph based on the geographical and physical factors between two adjacent nodes; The edge weights between two nodes in the spatial map are calculated based on geographical factors (reciprocal of distance difference, elevation difference, terrain shading, etc.) and physical factors (climate similarity, underlying surface similarity, prevailing wind direction, water flow velocity, etc.) between adjacent nodes to reflect the strength of the association between the two nodes. The monitoring data is received by the cloud platform 140, and the geographical and physical factors are historical data pre-stored on the cloud platform 140.

[0047] Specifically, the formula for calculating edge weights is as follows: ; in, For nodes and nodes Edge weights between them; For nodes and nodes Interval One geographical factor; For the first The weight values ​​of each geographical factor , The quantity of geographical factors; For nodes and nodes Interval One physical factor; For the first The weight values ​​of each physical factor , The quantity of physical factors; This represents the total weight value of geographical factors. This represents the total weight value of physical factors, and .

[0048] Step S233: Convert the monitoring data of each node into time series data, take the time series data of each node for the past T hours, and form the spatiotemporal sequence characteristics of each node; The monitoring data of each node is converted into time series data. The time series data of each node for the past 72 hours (1 hour / time, a total of 72 time steps) are taken to form the spatiotemporal sequence features of each node. The dimension of the spatiotemporal sequence features of each node is 18 (features) × 72 (time steps).

[0049] Building upon the above, the spatiotemporal graph neural network model adopts a fusion architecture of "spatial graph convolution (GCN) + temporal convolutional network (TCN)". Specifically, the spatial graph convolution (GCN) uses a two-layer graph convolutional network to aggregate spatial features of the spatial graph at each time step, thereby uncovering upstream-downstream and tributary-mainstream correlation features. The temporal convolutional network (TCN) uses a three-layer temporal convolutional network, combining the aggregated spatial features to perform convolution operations on the spatiotemporal sequence features (72 time steps). Through dilated convolution, it expands the temporal receptive field, capturing short-term emergencies (e.g., a sudden drop in DO within 1 hour) and long-term trends (e.g., COD changes within 72 hours). These short-term emergencies and long-term trends are mapped to the "aquatic habitat quality index (HQI)", where HQI ∈ [0, 100], and the "aquatic habitat quality index (HQI)" is used as the evaluation result.

[0050] In addition, the "Aquatic Habitat Quality Index (HQI)" is divided into Q (e.g., 5) level ranges. After obtaining the "Aquatic Habitat Quality Index (HQI)," the "Aquatic Habitat Quality Index (HQI)" is mapped to the corresponding level range, and the corresponding level range is used as the evaluation result.

[0051]

[0052] The following is a specific embodiment. (a) Implementation Scenarios Taking a certain watershed (100 km long, 2000 km²) as an example, the watershed contains 15 monitoring stations (5 main stream monitoring stations and 10 tributary monitoring stations). Potential pollution sources are two upstream chemical industrial parks and midstream agricultural non-point source pollution. Historically, there have been problems with excessive COD and declining biodiversity. Real-time assessment and early warning of aquatic habitat quality are required.

[0053] (II) Implementation Steps 1. System deployment and monitoring data collection: Fifteen monitoring stations were deployed, each integrating water quality physicochemical sensors, biological monitoring equipment, hydrological and meteorological equipment, and topographic and land use modules, thus forming a distributed monitoring sub-network at each station. Each monitoring station also deployed edge computing nodes and communication equipment. The edge computing nodes performed data preprocessing and spatiotemporal alignment, while the communication equipment transmitted data to the cloud platform via 4G / 5G. The collected monitoring data covered 18-dimensional feature vectors, including DO, COD, ammonia nitrogen, Shannon-Wiener index, flow velocity, and rainfall, with a time granularity of 1 hour per instance.

[0054] 2. Construction of watershed spatiotemporal map: Using 15 monitoring stations as nodes, all monitoring data of each node at the same time are used as the 18-dimensional feature vector of that node; the edge weight between two adjacent nodes is calculated; and time series data of nearly 72 hours are taken to form a 18×72-dimensional spatiotemporal sequence feature.

[0055] 3. Evaluation of the spatiotemporal graph neural network model: The spatiotemporal neural network model was trained using historical monitoring data from the basin over the past year (approximately 8760 time steps). During the training process: Spatial Graph Convolutional Network (GCN) aggregates the spatial characteristics of monitoring sites, while Temporal Convolutional Network (TCN) captures temporal dynamics. The model loss function (MSE) converged to 0.005, and the HQI prediction error was ≤3 (relative error ≤5%). Real-time assessment: At 12:00 on [Date], the system output HQI=58 (medium level), of which S6 (downstream of chemical industrial park) HQI=38 (poor level), the main influencing factor being COD=45mg / L (exceeding the standard by 1.5 times).

[0056] The beneficial effects of this application are as follows: 1. High assessment accuracy: The spatiotemporal graph neural network model can integrate multi-source data and spatiotemporal correlation, with a relative error of ≤5% in HQI assessment, which is 75% higher than traditional methods (error 20%+), accurately reflecting the overall watershed quality status; 2. Fast dynamic response: Supports real-time evaluation at the 1-hour granularity, enabling real-time evaluation and trend prediction; 3. Strong multi-source fusion capability: It integrates four types of data: water quality, biology, hydrology and topography, increasing the information utilization rate from 50% to 90%, and making the assessment dimensions more comprehensive, avoiding the one-sidedness of single data; 4. High scalability: Supports the addition and removal of monitoring stations (automatic updating of spatiotemporal maps) and the expansion of assessment indicators (such as adding biotoxicity indicators), and can be adapted to watersheds of different sizes (from 100km² to 10,000km²).

[0057] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0058] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for dynamic assessment of watershed water environment quality based on spatiotemporal graph neural networks, characterized in that, Includes the following steps: Associate the target monitoring station with N adjacent monitoring stations, and collect monitoring data at the target monitoring station; The monitoring data from the monitoring stations is preprocessed, the preprocessed monitoring data is spatiotemporally aligned, and the spatiotemporally aligned monitoring data is transmitted to the cloud platform. The cloud platform constructs a spatiotemporal map of the watershed based on monitoring data, and inputs the spatiotemporal map into a pre-trained spatiotemporal neural network model to predict assessment results that reflect short-term emergencies and long-term development trends.

2. The method for dynamic assessment of watershed water environment quality based on spatiotemporal graph neural networks according to claim 1, characterized in that, Associating a target monitoring station with N adjacent monitoring stations includes the following sub-steps: Calculate the spatial correlation coefficient between the target monitoring station and the monitoring stations within its predetermined range, and initially select N+M monitoring stations with larger spatial correlation coefficients; Based on the historical monitoring data of the target monitoring station and the historical monitoring data of the initially selected monitoring stations, the rank correlation coefficient between the target monitoring station and the initially selected monitoring stations is calculated; The comprehensive correlation coefficient between the target monitoring station and the initially selected monitoring station is obtained based on the spatial correlation coefficient and the rank correlation coefficient between the target monitoring station and the initially selected monitoring station; Select N monitoring stations with the highest overall correlation coefficient from N+M monitoring stations to associate them with the target monitoring station.

3. The method for dynamic assessment of watershed water environment quality based on spatiotemporal graph neural networks according to claim 1 or 2, characterized in that, Preprocessing the monitoring data includes the following steps: The time dimension error value is calculated based on the mean and standard deviation of the target monitoring station's monitoring data at the current moment and the historical monitoring dataset for the same period. If the absolute value of the time-dimensional error is greater than the first threshold, the monitoring data is preliminarily determined to be abnormal data; By combining the monitoring data of the monitoring stations associated with the target monitoring station at the current moment, the spatial dimension error value of the preliminary anomaly monitoring data is calculated; If the absolute value of the spatial dimension error value is greater than the second threshold, then the monitoring data that was initially judged to be abnormal is indeed abnormal, and the abnormal monitoring data is removed. After removing abnormal monitoring data, the remaining monitoring data is standardized to complete the preprocessing of the monitoring data.

4. The method for dynamic assessment of watershed water environment quality based on spatiotemporal graph neural networks according to claim 3, characterized in that, Based on millisecond-level timestamps, monitoring data from different collection frequencies are aligned to a unified time granularity to achieve spatiotemporal alignment of the monitoring data.

5. The method for dynamic assessment of watershed water environment quality based on spatiotemporal graph neural networks according to claim 1 or 2, characterized in that, A watershed spatiotemporal map includes a spatial map and spatiotemporal sequence features. Constructing a watershed spatiotemporal map involves the following steps: Each monitoring station is treated as a node in the spatial map, and all monitoring data of a node at the same time are used as the feature vector of that node. The edge weights between two nodes in the spatial graph are calculated based on geographical and physical factors between adjacent nodes. The monitoring data of each node is converted into time series data, and the time series data of each node for the past T hours are taken to form the spatiotemporal sequence characteristics of each node.

6. A dynamic assessment system for watershed water environment quality based on spatiotemporal graph neural networks, characterized in that, include: Distributed monitoring subnetwork, edge computing nodes, communication equipment, and cloud platform; Among them, monitoring stations are set up according to the watershed hydrological topology, and the target monitoring station is associated with N adjacent monitoring stations. The distributed monitoring sub-network collects monitoring data at the target monitoring station. Edge computing nodes preprocess the monitoring data, perform spatiotemporal alignment on the preprocessed monitoring data, and communication devices transmit the spatiotemporally aligned monitoring data to the cloud platform. The cloud platform constructs a spatiotemporal map of the watershed based on monitoring data, and inputs the spatiotemporal map into a pre-trained spatiotemporal neural network model to predict assessment results that reflect short-term emergencies and long-term development trends.

7. The watershed water environment quality dynamic assessment system based on spatiotemporal graph neural network according to claim 6, characterized in that, Edge computing nodes calculate the spatial correlation coefficient between the target monitoring station and monitoring stations within its predetermined range. They initially select N+M monitoring stations with larger spatial correlation coefficients. Based on the historical monitoring data of the target monitoring station and the historical monitoring data of the initially selected monitoring stations, they calculate the rank correlation coefficient between the target monitoring station and the initially selected monitoring stations. Based on the spatial and rank correlation coefficients between the target monitoring station and the initially selected monitoring stations, they obtain the comprehensive correlation coefficient between the target monitoring station and the initially selected monitoring stations. From the N+M monitoring stations, they select N monitoring stations with larger comprehensive correlation coefficients to associate with the target monitoring station.

8. The watershed water environment quality dynamic assessment system based on spatiotemporal graph neural network according to claim 6 or 7, characterized in that, Edge computing nodes calculate time-dimensional error values ​​based on the mean and standard deviation of the monitoring data of their target monitoring station at the current moment and the historical monitoring datasets for the same period. If the absolute value of the time-dimensional error value is greater than a first threshold, the edge computing node initially determines that the monitoring data is abnormal. The edge computing node then combines the monitoring data of neighboring monitoring stations associated with the target monitoring station at the current moment to calculate the spatial dimension error value of the initially abnormal monitoring data. If the absolute value of the spatial dimension error value is greater than a second threshold, the edge computing node verifies that the monitoring data initially determined to be abnormal is indeed abnormal and removes the abnormal monitoring data. After removing the abnormal monitoring data, the edge computing node standardizes the remaining monitoring data to complete the preprocessing of the monitoring data.

9. The watershed water environment quality dynamic assessment system based on spatiotemporal graph neural network according to claim 8, characterized in that, Edge computing nodes align monitoring data from different collection frequencies to a unified time granularity based on millisecond-level timestamps, thereby achieving spatiotemporal alignment of the monitoring data.

10. The watershed water environment quality dynamic assessment system based on spatiotemporal graph neural network according to claim 6 or 7, characterized in that, The spatiotemporal map of a watershed includes: a spatial map and spatiotemporal sequence features; The cloud platform treats each monitoring station as a node in the spatial map. It calculates the edge weights between two nodes in the spatial map based on the geographical and physical factors between adjacent nodes, and converts the monitoring data of each node into time series data. It then takes the time series data of each node for the past T hours to form the spatiotemporal sequence characteristics of each node in order to construct the spatiotemporal map of the watershed.

Images