Method and system for predicting spatial and temporal distribution of pollutants in industrial gathering area

Abstract

The invention relates to the field of industrial pollution prediction, and discloses a method and system for predicting spatial and temporal distribution of pollutants in an industrial accumulation area, and the method comprises the following steps: carrying out the modeling of a pollutant data source in the industrial accumulation area, constructing a pollutant associated data model, and constructing a process-pollutant-environmental medium three-dimensional associated data model; and in combination with Bayesian optimization and Monte Carlo optimization methods, model time sequence prediction and uncertainty quantification are realized, a high-uncertainty pollutant distribution area is output, and finally, the pollutant concentration of the high-uncertainty pollutant distribution area is predicted. According to the method, through innovative multi-modal data fusion, collaborative prediction of heterogeneous graph neural network and physical information embedding and space-time coupled distribution field prediction, a brand new solution is provided for industrial gathering area environment management, and the method has important theoretical significance and practical value.

Description

Technical Field

[0001] This invention relates to the field of industrial pollution prediction, and in particular to a method and system for predicting the spatiotemporal distribution of pollutants in industrial clusters. Background Technology

[0002] With the deepening of my country's industrialization, industrial clusters have become the main carriers supporting the development of manufacturing, but the environmental problems they bring are becoming increasingly prominent. Traditional pollution monitoring methods mainly rely on periodic sampling or event-driven approaches, which suffer from serious lag and widespread monitoring blind spots. The "smoothing effect" of traditional spatial interpolation methods (such as Kriging interpolation) cannot capture the patchy characteristics of pollutants, and machine learning models such as random forests are difficult to effectively handle the complex correlation between equipment, processes, and pollutants in industrial clusters, resulting in delayed identification of new pollutants and insufficient accuracy in pollution diffusion prediction. In addition, existing methods lack the ability to dynamically model the spatiotemporal evolution of pollutants, making it difficult to adapt to the time-varying characteristics of production activities in industrial clusters and unable to provide accurate pollution risk warnings.

[0003] To address the aforementioned technical bottlenecks, this invention proposes an innovative method and system for predicting the spatiotemporal distribution of pollutants in industrial clusters. This system overcomes the limitations of traditional pollution monitoring methods by constructing a complete technical architecture of "data-model-prediction-verification-application," achieving a paradigm shift from "passive monitoring" to "active prediction." The core innovations of this invention are: First, through multimodal data fusion technology, it deeply integrates industrial process data, equipment operating parameters, measured pollutant data, and geospatial information to construct a high-precision pollutant correlation data model; second, it innovatively proposes a three-dimensional heterogeneous graph neural network model, combined with a relation-aware graph attention mechanism (R-GAT), to accurately characterize the dynamic process of pollutant generation, migration, transformation, and decay; third, by embedding physical constraints of chemical reaction kinetics, it achieves a deep integration of physical mechanisms and data-driven approaches, significantly improving the model's interpretability and prediction accuracy; finally, by combining Bayesian optimization and Monte Carlo uncertainty quantification algorithms, it achieves high-precision prediction of the spatiotemporal distribution of pollutants and uncertainty assessment, providing a quantitative basis for environmental risk management.

[0004] Compared with existing technologies, this invention improves the prediction accuracy (R² value) by more than 0.2, while effectively identifying areas with high uncertainty in pollutant distribution, providing scientific decision support for environmental management in industrial clusters. This invention not only overcomes the limitations of traditional pollution monitoring methods but also provides a novel technical approach for pollution prevention and control in industrial clusters, possessing significant theoretical and practical value. Summary of the Invention

[0005] This invention overcomes the shortcomings of the prior art and provides a method and system for predicting the spatiotemporal distribution of pollutants in industrial clusters.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: The first aspect of this invention provides a method for predicting the spatiotemporal distribution of pollutants in industrial clusters, comprising the following steps: Identify the target industrial cluster and acquire its process production characteristics, pollutant physicochemical properties, and environmental media characteristics. Use OCR technology to parse and clean the text data, extracting equipment parameters, process parameters, material flow information, and output characteristics. Construct structured triplet information of process-pollutant-environmental media based on semantic similarity matching algorithm. Import the structured triplet information into the attribute graph model to construct a target pollutant association data model. In the target pollutant association data model, process unit nodes, pollutant nodes, and environmental media nodes are set up; the edges between nodes are defined, including process-equipment edges (carrying equipment parameters), process-pollutant edges (carrying the probability distribution of pollution generation coefficients), and pollutant-environmental media edges (carrying migration coefficients); a relation-aware graph attention mechanism (R-GAT) is introduced to perform weighted aggregation of the relationships between nodes and generate an association embedding vector of process features-pollution physicochemical properties-environmental media. The pollutant generation-migration-transformation-decay pathway is modeled as a chemical reaction kinetic system; a gridded instantaneous pollutant concentration is generated through a fully connected decoder network; the chemical reaction kinetic system and instantaneous pollutant concentration are converted into time-series input data and time-series encoded through a multi-layer encoder; a joint loss function (data fitting loss, physical constraint loss, regularization loss) is designed and combined with gradient descent to train the model, resulting in a target neural network model for pollution time-series prediction. Gaussian regression is performed using Bayesian optimization, and model parameters with residual values ​​closest to preset values ​​are collected in real time; Monte Carlo random sampling is performed to obtain all predicted distribution locations of pollutants within the target aggregation park; the entropy and standard deviation of each predicted distribution location are calculated, and when a preset threshold is met, the corresponding location is marked as a high-uncertainty pollutant distribution area. The distribution area of ​​highly uncertain pollutants is divided into grids to construct sub-regions of highly uncertain pollutant distribution; field pollutant sampling is conducted to construct a measured pollutant fingerprint spectrum; environmental media characteristics, hydrogeological driving factors, and equipment density-weighted process characteristics are extracted as spatial dimension features, and production cycle status is retrieved through historical data networks as temporal dimension features; the spatial and temporal dimension features are imported into the target neural network model for pollution time series prediction, and the probability density function of pollutant concentration is output to construct a pollutant fingerprint prediction spectrum; the model's qualification is judged by cosine similarity analysis with the measured spectrum, and if it is not qualified, the model is updated online.

[0007] By combining spatiotemporal dimensions, pollutant concentration coupling prediction processing is performed on areas with high uncertainty in pollutant distribution.

[0008] Furthermore, in a preferred embodiment of the present invention, the step of performing multi-source data fusion and three-dimensional correlation model construction and modeling processing on the industrial cluster to obtain the target pollutant correlation data model specifically includes: Identify industrial clusters, designate them as target industrial clusters, acquire multi-source data from these industrial clusters, fuse them, and then create a three-dimensional correlation model. A data processing terminal is introduced to extract feature data from all sources of target pollutant data. The feature data includes text recognition data and electronic data. The text recognition data is parsed and cleaned using OCR methods, and preprocessed pollutant feature data is obtained by combining the electronic data. The pretreatment process production characteristics include equipment parameters (collecting equipment parameters such as reactor type (fixed bed / fluidized bed), heat exchanger model, and separation device specifications to establish an equipment operation status database), corresponding process parameters (extraction temperature (°C), pressure (MPa), reaction time (h), load rate (%), and other operating conditions to construct a process parameter time-series feature vector), and material flow information (quantifying the input / output flow rates of raw materials (such as benzene and toluene), intermediate products (such as nitrobenzene), and final products to generate a material balance matrix). Based on the above conditions, the production characteristics are generated (statistically calculating the product output (kg / m³) per unit time (h / d). 3 (And by-product generation, and establish production-process correlation features). The above equipment parameter features, process parameter features, material flow features and production features are integrated in multiple dimensions. Feature selection algorithms (such as LASSO regression, random forest) are used to screen key feature variables, and a comprehensive feature vector including equipment status, process dynamics, material flow and production efficiency is constructed to provide high-dimensional input features for subsequent pollutant generation prediction.

[0009] Modeling the physicochemical properties of pollution includes physicochemical properties (melting point (°C), boiling point (°C), Henry's constant (Pa·m³ / mol), octanol-water partition coefficient (log Kow)), environmental behavior (soil transport coefficient (Kd, L / kg), groundwater diffusion coefficient (D, m² / s), sediment adsorption capacity (Koc, L / kg)), and toxicity parameters (LC50, mg / L, half-life (t)). 1 / 2 d), bioaccumulation factor (BCF, L / kg) and environmental media parameters (pollutants in soil (permeability coefficient Ks, m / s), groundwater (dispersion αL, m), solid waste (adsorption capacity Q)). max (mmol / g), migration and transformation parameters in sediments (deposition rate Rd, m / y).

[0010] Concentration information modeling involves integrating real-time monitoring data (μg / m³), historical concentration trends (μg / m³ / d), and spatial interpolation results (gridized concentration distribution) to construct a dynamic concentration field. This quantifies the differences in pollutant concentration distribution in soil (μg / kg), groundwater (μg / L), and sediment (μg / g).

[0011] Within the data processing terminal, a multimodal heterogeneous graph neural network model is constructed. Specifically, this includes the construction of a process feature map, using process units (reactors, distillation columns, pump systems, etc.) as nodes and physical relationships between units (such as pipeline connections and thermal integration) as edges. Equipment parameters (reactor volume, material), operating parameters (temperature / pressure gradient, residence time), and material flow parameters (raw material input rate, reflux ratio) are integrated. Multi-order neighborhood features of the process topology are extracted through graph convolution and pooling operations to generate a collaborative embedding vector of equipment-process-material flow, achieving a low-dimensional dense representation of the production process. For pollutant attribute construction, pollutant types are used as nodes, with physicochemical property relationships (such as logKow similarity and functional group co-occurrence probability) and environmental migration coupling relationships (such as the synergistic effect of soil-water partition coefficient Kd and diffusion coefficient D) as edges. A relational graph convolutional network (R-GCN) is used to learn the propagation potential and fate characteristics of pollutants in water-soil multi-media, generating a pollutant-media-toxicity parameter association embedding vector. By establishing a dynamic correlation between process characteristics and pollutant properties through an attention mechanism, the generation-migration-transformation pathway of pollutants in the process is simulated, and dynamic correlation across spectra is achieved.

[0012] The system successively constructs part-of-speech analysis units, syntactic structure similarity units, string similarity units, and rule similarity units. Based on the unit order, it performs comprehensive text similarity analysis on the preprocessed pollutant feature data to generate structured triplet information of process production characteristics, physicochemical properties of pollution, and environmental media within the target industrial cluster. Within the data processing terminal, based on the attribute graph model, all the structured triplet information of process production characteristics, pollution physicochemical properties, and environmental media is imported into the attribute graph model for data model construction. This results in a data model that can query other structured data based on different structured data. At the same time, a semantic similarity matching algorithm is introduced to perform semantic association updates on the constructed data model and generate a target pollutant association data model. The semantic association update enables the constructed data model to automatically associate the corresponding structural data based on the newly input pollutant information. The association rule is to calculate the semantic similarity between the newly input pollutant information and all structural data. If the semantic similarity is greater than the set standard threshold, association processing can be performed.

[0013] Furthermore, in a preferred embodiment of the present invention, the step of constructing a heterogeneous graph neural network model of process production characteristics-pollution physicochemical properties-environmental media by combining the target pollutant association data model specifically includes: Within the target pollutant association data model, process unit nodes and pollutant nodes are set. The nodes are set based on all process features and pollutant attributes in the target pollutant association data model, and the edges between different nodes are defined. Among them, the process-equipment side carries equipment parameters (reactor volume, material), process parameters (temperature / pressure gradient, residence time), material flow parameters (raw material input rate, reflux ratio), the physicochemical property correlation side carries log Kow similarity and functional group co-occurrence probability, and the environmental migration coupling side carries the synergistic effect of soil-water partition coefficient Kd and diffusion coefficient D; A graph attention mechanism is introduced to update the target pollutant association data model after defining the edges between different nodes. This model is used to train the generation-migration-transformation-attenuation routes of different pollutants, resulting in a heterogeneous graph neural network model of process production characteristics, pollution physicochemical properties, and environmental media.

[0014] Furthermore, in a preferred embodiment of the present invention, the physical information constraint is applied to the heterogeneous graph neural network model of process production characteristics-pollution physicochemical properties-environmental media, and Bayesian optimization and Monte Carlo optimization methods are combined to achieve model time series prediction and uncertainty quantification, and output the distribution area of ​​high-uncertainty pollutants, specifically as follows: Physical information is embedded into a heterogeneous graph neural network model of process production characteristics, pollution physicochemical properties, and environmental media. The physical information constraint embedding includes the abstract construction of physical mechanisms and the calculation of instantaneous pollution concentration. The abstract construction of the physical mechanism is to model different pollutants in the generation-migration-transformation-degradation route as a chemical reaction kinetic system; The instantaneous pollution concentration is calculated by introducing a fully connected decoder network into a gridded instantaneous pollution concentration based on the pollutant nodes in the heterogeneous graph neural network model of process production characteristics-pollution physicochemical properties-environmental media. In the heterogeneous graph neural network model of process production characteristics-pollution physicochemical properties-environmental media, the chemical reaction kinetic system and instantaneous pollution concentration are converted into time-series input data. The time-series input data is encoded by a multi-layer encoder. A loss function is designed to calculate the data fitting loss, physical constraint loss and regularization loss of the encoded time-series input data in real time. At the same time, the gradient descent method is combined to run during real-time calculation to prevent overfitting. A standard time for real-time calculation is preset. The calculation stops when the real-time calculation time for data fitting loss, physical constraint loss and regularization loss reaches the corresponding standard time. This yields a heterogeneous graph neural network model of process production characteristics-pollution physicochemical properties-environmental medium that satisfies the time series prediction, and is calibrated as the target neural network model for pollution time series prediction. By combining Bayesian optimization and Monte Carlo methods, the target neural network model for pollution time series prediction is updated to achieve time series prediction and uncertainty quantification, and output the distribution area of ​​pollutants with high uncertainty.

[0015] Furthermore, in a preferred embodiment of the present invention, the step of combining Bayesian optimization and Monte Carlo methods to update the target neural network model for pollution time series prediction, thereby achieving model time series prediction and uncertainty quantification, and identifying the distribution area of ​​high-uncertainty pollutants, specifically involves: In the data processing terminal, the model framework of the target neural network model for pollution time series prediction is rebuilt using the PINN training framework. Within the PINN training framework, Gaussian regression is performed on the target neural network model for pollution time series prediction using Bayesian optimization. The Gaussian regression is used to train the target neural network model for pollution time series prediction. During the regression training process, the model parameters whose residual values ​​are closest to the preset values ​​are collected in real time until the number of regressions equals the preset number of regressions. After Bayesian optimization, the target neural network model for pollution time series prediction is subjected to Monte Carlo processing and random Monte Carlo sampling to obtain all predicted distribution locations of pollutants within the target cluster area. Among these, all predicted distribution locations are uncertain pollutant distribution locations. The entropy and standard deviation of all predicted distribution locations are calculated using a pollution time-series prediction target neural network model. If either the entropy or standard deviation of a predicted distribution location meets a preset value, the corresponding predicted distribution location is marked as a high-uncertainty pollutant distribution area. The high-uncertainty pollutant distribution area refers to the area where the judged concentration of the pollutant is greater than the standard value.

[0016] Furthermore, in a preferred embodiment of the present invention, the step of combining spatiotemporal dimensions to perform pollutant concentration coupling prediction processing on highly uncertain pollutant distribution areas specifically includes: Within the data processing terminal, a high-uncertainty pollutant distribution area is defined, and the high-uncertainty pollutant distribution area is divided into gridded sub-regions of high-uncertainty pollutant distribution. Within the target industrial cluster, on-site pollutant sampling and processing were carried out in the corresponding sub-regions with high uncertainty pollutant distribution. Quantitative processing of pollutant concentration and determination of pollutant type were performed in the sub-regions with high uncertainty pollutant distribution, and a pollutant fingerprint measured spectrum was constructed. Meanwhile, different sub-regions with high uncertainty pollutant distribution are embedded with environmental media characteristics, hydrogeological driving factors, and equipment density-weighted process characteristics to generate spatial dimension characteristics of the sub-regions with high uncertainty pollutant distribution. Among them, environmental media characteristics, hydrogeological driving factors, and equipment density-weighted process characteristics were obtained by searching historical data networks; Simultaneously, by retrieving the production cycle status within the target industrial cluster through historical data networks, the temporal dimension characteristics of the high-uncertainty pollutant distribution sub-region are defined. Within the target neural network model for pollution time series prediction, the spatial and temporal features of the high-uncertainty pollutant distribution sub-regions are imported and combined for calculation, outputting the probability density function of pollutant concentration in the high-uncertainty pollutant distribution sub-regions, which is used to construct the pollutant fingerprint prediction spectrum. The pollutant fingerprint prediction spectrum describes the predicted pollutant concentration and prediction type in the high-uncertainty pollutant distribution sub-region, and performs cosine similarity analysis with the measured pollutant fingerprint spectrum. If the cosine similarity between the predicted pollutant fingerprint spectrum and the measured pollutant fingerprint spectrum is not less than a preset value, the pollutant time series prediction target neural network model is deemed qualified. If it is less than the preset value, the pollutant time series prediction target neural network model is updated online to ensure that the cosine similarity between the output predicted pollutant fingerprint spectrum and the measured pollutant fingerprint spectrum is not less than the preset value.

[0017] One method for online updating of the target neural network model for pollution time series prediction is to introduce parallel techniques of meta-learning and online learning to train the target neural network model for pollution time series prediction in parallel. Using a qualified pollution time-series prediction target neural network model, pollutant fingerprint prediction spectra are constructed for all high-uncertainty pollutant distribution sub-regions within industrial clusters.

[0018] A second aspect of this invention also provides a system for predicting the spatiotemporal distribution of pollutants in industrial clusters. The distribution prediction system integrates a high-performance computing architecture and a data storage module, including a non-volatile memory consisting of a DDR4 RDIMM memory module with ECC verification and an NVMe solid-state storage array using 3D NAND flash memory, and a multi-core processor based on the Zen4 microarchitecture. The memory contains a distribution prediction method program with a prediction engine. When the program is executed in parallel through a superscalar pipeline execution unit within the processor, the following steps are achieved: The sources of pollutant data in industrial clusters are identified and modeled to obtain a data model of the correlation of target pollutants. By combining the target pollutant association data model, a heterogeneous graph neural network model of process production characteristics, pollutant physicochemical properties, and environmental media is constructed. Physical information constraints are applied to the heterogeneous graph neural network model of process production characteristics-pollution physicochemical properties-environmental media. By combining Bayesian optimization and Monte Carlo optimization methods, the model can achieve time series prediction and uncertainty quantification, and output the distribution area of ​​pollutants with high uncertainty. By combining spatiotemporal dimensions, pollutant concentration coupling prediction processing is performed on areas with high uncertainty in pollutant distribution.

[0019] This invention addresses the technical deficiencies in the background technology and offers the following beneficial effects: It models pollutant data sources in industrial clusters, constructs a pollutant correlation data model, and uses this model to build a three-dimensional correlation data model of process-pollutant-environmental media; combining Bayesian optimization and Monte Carlo optimization methods, it achieves time-series prediction and uncertainty quantification of the model, outputs the distribution areas of high-uncertainty pollutants, and finally predicts the pollutant concentrations in these high-uncertainty pollutant distribution areas. This invention provides a novel solution for environmental management in industrial clusters through innovative multimodal data fusion, collaborative prediction using heterogeneous graph neural networks and physical information embedding, and spatiotemporally coupled distribution field prediction, possessing significant theoretical and practical value. Attached Figure Description

[0020] 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 of the present invention. For those skilled in the art, other embodiments can be obtained from these drawings without creative effort.

[0021] Figure 1 A flowchart of a method for predicting the spatiotemporal distribution of pollutants in industrial clusters is shown. Figure 2A flowchart illustrating the method for outputting the distribution area of ​​pollutants with high uncertainty is shown. Figure 3 A program view of a system for predicting the spatiotemporal distribution of pollutants in an industrial cluster is shown. Detailed Implementation

[0022] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0024] Figure 1 A flowchart illustrating a method for predicting the spatiotemporal distribution of pollutants in industrial clusters is shown, including the following steps: The sources of pollutant data in industrial clusters are identified and modeled to obtain a data model of the correlation of target pollutants. By combining the target pollutant association data model, a heterogeneous graph neural network model of process production characteristics, pollutant physicochemical properties, and environmental media is constructed. Physical information constraints are applied to the heterogeneous graph neural network model of process production characteristics-pollution physicochemical properties-environmental media. By combining Bayesian optimization and Monte Carlo optimization methods, the model can achieve time series prediction and uncertainty quantification, and output the distribution area of ​​pollutants with high uncertainty. By combining spatiotemporal dimensions, pollutant concentration coupling prediction processing is performed on areas with high uncertainty in pollutant distribution.

[0025] Furthermore, in a preferred embodiment of the present invention, the step of determining the source of pollutant data and modeling the industrial cluster to obtain a target pollutant association data model specifically involves: Identify industrial clusters, designate them as target industrial clusters, and acquire all data sources of pollutants within these industrial clusters, designating them as target pollutant data sources. A data processing terminal is introduced to extract feature data from all sources of target pollutant data. The feature data includes text recognition data and electronic data. The text recognition data is parsed and cleaned using OCR methods, and preprocessed pollutant feature data is obtained by combining the electronic data. Among them, the equipment parameters of the pretreatment process production characteristics (collecting equipment parameters such as reactor type (fixed bed / fluidized bed), heat exchanger model, separation device specifications, etc., and establishing an equipment operation status database), the corresponding process parameters (extraction temperature (°C), pressure (MPa), reaction time (h), load rate (%), etc., constructing a process parameter time-series feature vector) and material flow information (quantifying the input / output flow of raw materials (such as benzene, toluene), intermediate products (such as nitrobenzene) and final products, generating a material balance matrix), based on the above conditions, generate production characteristics (statistically calculating the product output (kg / m³) per unit time (h / d). 3 (And by-product generation, and establish production-process correlation features). The above equipment parameter features, process parameter features, material flow features and production features are integrated in multiple dimensions. Feature selection algorithms (such as LASSO regression, random forest) are used to screen key feature variables, and a comprehensive feature vector including equipment status, process dynamics, material flow and production efficiency is constructed to provide high-dimensional input features for subsequent pollutant generation prediction.

[0026] Modeling the physicochemical properties of pollution includes physicochemical properties (melting point (°C), boiling point (°C), Henry's constant (Pa·m³ / mol), octanol-water partition coefficient (log Kow)), environmental behavior (soil migration coefficient (Kd, L / kg), groundwater diffusion coefficient (D, m² / s), sediment adsorption capacity (Koc, L / kg)), toxicity parameters (LC50, mg / L, half-life (t1 / 2, d), bioaccumulation factor (BCF, L / kg)), and environmental media parameters (pollution in soil (permeability coefficient Ks, m / s), groundwater (dispersion αL, m), and solid waste (adsorption capacity Q)). max (mmol / g), migration and transformation parameters in sediments (deposition rate Rd, m / y).

[0027] Concentration information modeling involves integrating real-time monitoring data (μg / m³), historical concentration trends (μg / m³ / d), and spatial interpolation results (gridized concentration distribution) to construct a dynamic concentration field. This quantifies the differences in pollutant concentration distribution in soil (μg / kg), groundwater (μg / L), and sediment (μg / g).

[0028] Within the data processing terminal, part-of-speech analysis units, syntactic structure similarity units, string similarity units, and rule similarity units are constructed sequentially. Based on the unit order, the preprocessed pollutant feature data is subjected to comprehensive text similarity analysis to generate structured triplet information of process production characteristics, pollution physicochemical properties, and environmental media within the target industrial cluster. Within the data processing terminal, based on the attribute graph model, all the structured triplet information of process production characteristics, pollution physicochemical properties, and environmental media is imported into the attribute graph model for data model construction. This results in a data model that can query other structured data based on different structured data. At the same time, a semantic similarity matching algorithm is introduced to perform semantic association updates on the constructed data model and generate a target pollutant association data model. The semantic association update enables the constructed data model to automatically associate the corresponding structural data based on the newly input pollutant information. The association rule is to calculate the semantic similarity between the newly input pollutant information and all structural data. If the semantic similarity is greater than the set standard threshold, association processing can be performed.

[0029] It should be noted that the target pollutant data sources include, but are not limited to, environmental impact assessment reports, site investigation reports, and process flow diagrams within industrial clusters. Information is extracted from these data sources, including equipment type lists (reactors, distillation towers, etc.), process parameters (temperature, pressure), and material flow information (raw materials / intermediates / products). If paper reports exist, they are converted into processable text data using OCR (Optical Character Recognition). The text is then cleaned, removing irrelevant characters, headers, and footers, and segmented into paragraphs and sentences to obtain preprocessed pollutant feature data. A multi-pattern matching mechanism is designed to address the unique structure of industrial environmental impact assessment reports, including part-of-speech tagging, syntactic similarity, string similarity, and rule similarity units. Weighted similarity scoring is used to achieve accurate parsing of complex industrial text. Logical relationships exist between different units, creating interconnected structures that generate structured tripartite information on equipment, processes, and pollutants within the target industrial cluster. For example: Equipment: Distillation tower; Process parameter: Temperature, value: 150℃; Process: Catalytic cracking; Pollutant: Benzene series compounds. All structured triplet information on equipment, process, and contaminants is aggregated to construct a knowledge database. An attribute graph model is used to store entity relationships, such as equipment A → process B → contaminant C. Finally, a semantic association update method is introduced to realize the interrelationships between data, facilitating data retrieval.

[0030] Furthermore, in a preferred embodiment of the present invention, the step of constructing a heterogeneous graph neural network model of process production characteristics-pollution physicochemical properties-environmental media by combining the target pollutant association data model specifically includes: Within the target pollutant association data model, equipment nodes, process nodes, and pollutant nodes are set up. The nodes are set up based on the structured triplet information of all equipment-process-pollutant in the target pollutant association data model, and the edges between different nodes are defined. The edges between nodes include equipment-process edges, process-contaminant edges, and contaminant-related edges. The edge weights carried by the equipment-process edges are operating parameters, representing the temperature, pressure, and load rate when the equipment is operating. The edge weight carried by the process-pollutant edge is the probability distribution of the pollution generation coefficient, which represents the amount of pollutants produced during process operation. The edge weights carried by the pollutant association edges are reaction kinetic parameters, representing the transformation association parameters between different pollutants; A graph attention mechanism is introduced to update the target pollutant association data model after defining the edges between different nodes. This model is used to train the generation-migration-transformation-attenuation routes of different pollutants, resulting in a heterogeneous graph neural network model of process production characteristics, pollution physicochemical properties, and environmental media.

[0031] It should be noted that constructing a heterogeneous graph neural network model of process production characteristics, pollution physicochemical properties, and environmental media to obtain the generation-migration-transformation-degradation routes of different pollutants first requires defining nodes and edges. Edge types include equipment-process edges, process-pollutant edges, and pollutant-related edges. Based on the edge weights of different edges, combined with a relation-aware graph attention mechanism, edge types are distinguished during neighbor aggregation. The contribution of different process routes to pollutant generation is learned, i.e., the generation-migration-transformation-degradation routes of different pollutants are generated. This model is used to construct the heterogeneous graph neural network model of process production characteristics, pollution physicochemical properties, and environmental media, laying the foundation for the subsequent construction of a predictive model for the spatiotemporal distribution of pollutants.

[0032] Furthermore, in a preferred embodiment of the present invention, the step of combining spatiotemporal dimensions to perform pollutant concentration coupling prediction processing on highly uncertain pollutant distribution areas specifically includes: Within the data processing terminal, a high-uncertainty pollutant distribution area is defined, and the high-uncertainty pollutant distribution area is divided into gridded sub-regions of high-uncertainty pollutant distribution. Within the target industrial cluster, on-site pollutant sampling and processing were carried out in the corresponding sub-regions with high uncertainty pollutant distribution. Quantitative processing of pollutant concentration and determination of pollutant type were performed in the sub-regions with high uncertainty pollutant distribution, and a pollutant fingerprint measured spectrum was constructed. Meanwhile, different sub-regions with high uncertainty pollutant distribution are embedded with environmental media characteristics, hydrogeological driving factors, and equipment density-weighted process characteristics to generate spatial dimension characteristics of the sub-regions with high uncertainty pollutant distribution. Among them, environmental media characteristics, hydrogeological driving factors, and equipment density-weighted process characteristics were obtained by searching historical data networks; Simultaneously, by retrieving the production cycle status within the target industrial cluster through historical data networks, the temporal dimension characteristics of the high-uncertainty pollutant distribution sub-region are defined. Within the target neural network model for pollution time series prediction, the spatial and temporal features of the high-uncertainty pollutant distribution sub-regions are imported and combined for calculation, outputting the probability density function of pollutant concentration in the high-uncertainty pollutant distribution sub-regions, which is used to construct the pollutant fingerprint prediction spectrum. The pollutant fingerprint prediction spectrum describes the predicted pollutant concentration and prediction type in the high-uncertainty pollutant distribution sub-region, and performs cosine similarity analysis with the measured pollutant fingerprint spectrum. If the cosine similarity between the predicted pollutant fingerprint spectrum and the measured pollutant fingerprint spectrum is not less than a preset value, the pollutant time series prediction target neural network model is deemed qualified. If it is less than the preset value, the pollutant time series prediction target neural network model is updated online to ensure that the cosine similarity between the output predicted pollutant fingerprint spectrum and the measured pollutant fingerprint spectrum is not less than the preset value.

[0033] One method for online updating of the target neural network model for pollution time series prediction is to introduce parallel techniques of meta-learning and online learning to train the target neural network model for pollution time series prediction in parallel. Using a qualified pollution time-series prediction target neural network model, pollutant fingerprint prediction spectra are constructed for all high-uncertainty pollutant distribution sub-regions within industrial clusters.

[0034] It should be noted that, firstly, on-site sampling is conducted to obtain standard pollutant control data, i.e., pollutant fingerprint measured spectra. Then, spatial features are extracted by dividing the cluster areas. Environmental media characteristics, hydrogeological driving factors, and equipment density-weighted process characteristics are spatial features. Environmental media characteristics are soil permeability and groundwater depth; hydrogeological driving factors are soil permeability and groundwater depth; equipment density-weighted process characteristics are ∑ equipment capacity × process pollution coefficient. All features are embedded into the model, and time-dimensional features are also embedded into the model in parallel. The time-dimensional features are based on enterprise scheduling and equipment start-up and shutdown time series to predict the pulse release of pollutants under abnormal operating conditions (such as nighttime illegal discharge, maintenance flushing), and combined calculations are performed to output the probability density function of pollutant concentration in sub-regions with high uncertainty pollutant distribution. This is used to construct pollutant fingerprint predicted spectra and to achieve comparison between spectra. The spectra predict the pollutant component spectra of each region for matching and verification with measured concentrations. If the cosine similarity is not less than a preset value, the current model is considered qualified and can predict pollutant distribution and concentration changes with high accuracy; otherwise, online learning updates are required. Meta-learning and online learning are employed to continuously optimize the model. The meta-learning strategy treats the introduction of new processes as a few-sample task, using a support set (a small amount of data on the new process) and a query set (historical data) to quickly fine-tune the model. N-way K-shot task decomposition ensures the model can quickly adapt to new processes with limited data. Specific parameter settings include an inner loop learning rate α=0.001, an outer loop learning rate β=0.01, and a batch size of 32. Online learning uses Progressive Neural Network (PNN) technology, retaining a pre-trained model pool and continuously expanding the network according to task changes, achieving natural accumulation of experience and knowledge reuse. Federated learning technology is also combined to protect enterprise privacy, achieve distributed data updates, and avoid catastrophic forgetting. The combination of these two learning strategies allows for parallel training of the target neural network model for pollution time-series prediction.

[0035] Figure 2 A flowchart illustrating the method for outputting the distribution area of ​​highly uncertain pollutants is shown, including the following steps: S202: Physical information constraints are applied to the heterogeneous graph neural network model of process production characteristics-pollution physicochemical properties-environmental media, and Bayesian optimization and Monte Carlo optimization methods are combined to realize the model's time series prediction and uncertainty quantification, and output the distribution area of ​​pollutants with high uncertainty. S204: Combining Bayesian optimization and Monte Carlo methods, the target neural network model for pollution time series prediction is updated to achieve time series prediction and uncertainty quantification, and the distribution area of ​​pollutants with high uncertainty.

[0036] Furthermore, in a preferred embodiment of the present invention, the physical information constraint is applied to the heterogeneous graph neural network model of process production characteristics-pollution physicochemical properties-environmental media, and Bayesian optimization and Monte Carlo optimization methods are combined to achieve model time series prediction and uncertainty quantification, and output the distribution area of ​​high-uncertainty pollutants, specifically as follows: Physical information is embedded into a heterogeneous graph neural network model of process production characteristics, pollution physicochemical properties, and environmental media. The physical information constraint embedding includes the abstract construction of physical mechanisms and the calculation of instantaneous pollution concentration. The abstract construction of the physical mechanism is to model different pollutants in the generation-migration-transformation-degradation route as a chemical reaction kinetic system; The instantaneous pollution concentration is calculated by introducing a fully connected decoder network into a gridded instantaneous pollution concentration based on the pollutant nodes in the heterogeneous graph neural network model of process production characteristics-pollution physicochemical properties-environmental media. In the heterogeneous graph neural network model of process production characteristics-pollution physicochemical properties-environmental media, the chemical reaction kinetic system and instantaneous pollution concentration are converted into time-series input data. The time-series input data is encoded by a multi-layer encoder. A loss function is designed to calculate the data fitting loss, physical constraint loss and regularization loss of the encoded time-series input data in real time. At the same time, the gradient descent method is combined to run during real-time calculation to prevent overfitting. A standard time for real-time calculation is preset. The calculation stops when the real-time calculation time for data fitting loss, physical constraint loss and regularization loss reaches the corresponding standard time. This yields a heterogeneous graph neural network model of process production characteristics-pollution physicochemical properties-environmental medium that satisfies the time series prediction, and is calibrated as the target neural network model for pollution time series prediction. By combining Bayesian optimization and Monte Carlo methods, the target neural network model for pollution time series prediction is updated to achieve time series prediction and uncertainty quantification, and output the distribution area of ​​pollutants with high uncertainty.

[0037] It should be noted that the pollutant formation process first needs to be modeled as a chemical reaction kinetic system, with differential equation constraints embedded. These differential equation constraints are: Simultaneously, it is necessary to calculate the instantaneous pollution concentration, which can be used for hydrogeological driving factors (such as groundwater flow velocity) and to calculate the source term after advection-diffusion correction, thus maintaining prediction accuracy after model construction. Combining differential equation constraints, these two can be embedded into a heterogeneous graph neural network model of process production characteristics, pollution physicochemical properties, and environmental media to construct a time-series architecture. First, the neural network architecture is constructed, and the time-series input data described is encoded and processed. Data fitting loss, physical constraint loss, and regularization loss are calculated in real time. Simultaneously, gradient descent is used during real-time calculation to prevent overfitting. The resulting model is suitable for predicting concentration changes at different time points in industrial clusters, thus obtaining a model for time-series prediction of pollutant concentrations.

[0038] Furthermore, in a preferred embodiment of the present invention, the step of combining Bayesian optimization and Monte Carlo methods to update the target neural network model for pollution time series prediction, thereby achieving model time series prediction and uncertainty quantification, and identifying the distribution area of ​​high-uncertainty pollutants, specifically involves: In the data processing terminal, the model framework of the target neural network model for pollution time series prediction is rebuilt using the PINN training framework. Within the PINN training framework, Gaussian regression is performed on the target neural network model for pollution time series prediction using Bayesian optimization. The Gaussian regression is used to train the target neural network model for pollution time series prediction. During the regression training process, the model parameters whose residual values ​​are closest to the preset values ​​are collected in real time until the number of regressions equals the preset number of regressions. After Bayesian optimization, the target neural network model for pollution time series prediction is subjected to Monte Carlo processing and random Monte Carlo sampling to obtain all predicted distribution locations of pollutants within the target cluster area. Among these, all predicted distribution locations are uncertain pollutant distribution locations. The entropy and standard deviation of all predicted distribution locations are calculated using a pollution time-series prediction target neural network model. If either the entropy or standard deviation of a predicted distribution location meets a preset value, the corresponding predicted distribution location is marked as a high-uncertainty pollutant distribution area. The high-uncertainty pollutant distribution area refers to the area where the judged concentration of the pollutant is greater than the standard value.

[0039] It should be noted that after obtaining the model for time-series pollutant concentration prediction, uncertainty quantification is required. Bayesian optimization and Monte Carlo methods are introduced to identify high-uncertainty regions, i.e., areas within industrial clusters with relatively high pollutant concentrations. Bayesian optimization is used for efficient search, ensuring the model meets search accuracy requirements. The Monte Carlo deep ensemble method is employed to quantify cognitive uncertainty in the model's predictions, achieving high-precision spatiotemporal predictions. Furthermore, reliable and interpretable uncertainty measures are provided for the prediction results, offering a quantitative basis for environmental risk management. Subsequently, the resulting target neural network model for pollution time-series prediction can predict regions where pollutant concentrations exceed standard values ​​and can continue to perform spatiotemporal distribution prediction processing within these regions, including concentration and location prediction.

[0040] like Figure 3 As shown, a second aspect of the present invention also provides a spatiotemporal distribution prediction system for pollutants in industrial clusters. The distribution prediction system integrates a high-performance computing architecture and a data storage module, including a non-volatile memory consisting of a DDR4 RDIMM memory module with ECC verification and an NVMe solid-state storage array using 3D NAND flash memory, and a multi-core processor based on the Zen4 microarchitecture. The memory contains a distribution prediction method program with a prediction engine. When the program is executed in parallel through a superscalar pipeline execution unit within the processor, the following steps are achieved: The sources of pollutant data in industrial clusters are identified and modeled to obtain a data model of the correlation of target pollutants. By combining the target pollutant association data model, a heterogeneous graph neural network model of process production characteristics, pollutant physicochemical properties, and environmental media is constructed. Physical information constraints are applied to the heterogeneous graph neural network model of process production characteristics-pollution physicochemical properties-environmental media. By combining Bayesian optimization and Monte Carlo optimization methods, the model can achieve time series prediction and uncertainty quantification, and output the distribution area of ​​pollutants with high uncertainty. By combining spatiotemporal dimensions, pollutant concentration coupling prediction processing is performed on areas with high uncertainty in pollutant distribution.

[0041] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for predicting the spatio-temporal distribution of pollutants in an industrial cluster, characterized in that, The method comprises the following steps: determining the data sources of pollutants in the industrial cluster and modeling to obtain a target pollutant correlation data model; combining the target pollutant correlation data model to construct a process production feature-pollution physical and chemical property-environment medium heterogeneous graph neural network model; physically constraining the process production feature-pollution physical and chemical property-environment medium heterogeneous graph neural network model, and combining Bayesian optimization and Monte Carlo optimization to realize time series prediction and uncertainty quantification of the model, and output a high-uncertainty pollutant distribution area; combining the space-time dimension to perform pollutant concentration coupling prediction processing on the high-uncertainty pollutant distribution area.

2. The method for predicting the spatio-temporal distribution of pollutants in an industrial cluster according to claim 1, characterized in that, The method of determining the data sources of pollutants in the industrial cluster and modeling to obtain a target pollutant correlation data model comprises: determining an industrial cluster, calibrating it as a target industrial cluster, and obtaining all data sources of pollutants in the industrial cluster, calibrating them as target pollutant data sources; introducing a data processing terminal to extract feature data from all target pollutant data sources, including text recognition data and electronic data, and combining an OCR method to perform text analysis and cleaning on the text recognition data, and combining electronic data to obtain preprocessed pollutant feature data, wherein the preprocessed pollutant feature data includes process production features, pollution physical and chemical properties, and environmental medium features; in the data processing terminal, a part-of-speech analysis unit, a syntactic structure similarity unit, a string similarity unit, and a rule similarity unit are constructed in sequence, and according to the unit sequence, the preprocessed pollutant feature data is subjected to comprehensive text similarity analysis to generate structured triple information of process production features-pollution physical and chemical properties-environmental media in the target industrial cluster; in the data processing terminal, based on an attribute graph model, all structured triple information of process production features-pollution physical and chemical properties-environmental media is imported into the attribute graph model for data model construction to obtain a data model that can query other structure data according to different structure data, and a semantic similarity matching algorithm is introduced to update the semantic correlation of the constructed data model to generate a target pollutant correlation data model; wherein the semantic correlation update enables the constructed data model to automatically associate corresponding structure data according to newly input pollutant information, and the association rule is to calculate the semantic similarity between the newly input pollutant information and all structure data, and if the semantic similarity is greater than a set standard threshold, the association processing can be performed.

3. The method for predicting the spatio-temporal distribution of pollutants in an industrial cluster according to claim 1, characterized in that, The method of combining the target pollutant correlation data model to construct a process production feature-pollution physical and chemical property-environment medium heterogeneous graph neural network model comprises: in the target pollutant correlation data model, setting device nodes, process nodes, and pollutant nodes, wherein the nodes are set according to all structured triple information of process production features-pollution physical and chemical properties-environmental media in the target pollutant correlation data model, and the edges between different nodes are defined; The edges between the nodes include process-equipment edges, process-pollutant edges, and pollutant-environment medium edges, the edge weight carried by the process-equipment edge is an equipment parameter representing a volume, material, energy consumption, or the like when the equipment runs a process; The edge weight carried by the process-pollutant edge is a pollutant production coefficient probability distribution representing an amount of pollutants produced when the process runs; The edge weight carried by the pollutant-environment medium edge is a migration coefficient representing a migration ability of the pollutants in different environment media; The graph attention mechanism is introduced to update a target pollutant correlation data model after defining edges between different nodes, to train a generation-migration-transformation-decay route of different pollutants in the model, and to obtain a process production feature-pollutant physical and chemical property-environment medium heterogeneous graph neural network model.

4. The method for predicting the spatio-temporal distribution of pollutants in an industrial cluster according to claim 1, characterized in that, The process production feature-pollutant physical and chemical property-environment medium heterogeneous graph neural network model is physically constrained, and a Bayesian optimization and a Monte Carlo optimization method are combined to realize model time series prediction and uncertainty quantification, and output a high-uncertainty pollutant distribution area, specifically as follows: The process production feature-pollutant physical and chemical property-environment medium heterogeneous graph neural network model is physically embedded, wherein the physical information constraint embedding includes abstract construction of physical mechanisms and calculation of instantaneous pollutant concentration; The abstract construction of the physical mechanism is to model different pollutants in the generation-migration-transformation-decay route as a chemical reaction kinetics system; The instantaneous pollutant concentration calculation is to generate a grid-based instantaneous pollutant concentration based on the pollutant nodes in the process production feature-pollutant physical and chemical property-environment medium heterogeneous graph neural network model by introducing a fully connected decoder network; In the process production feature-pollutant physical and chemical property-environment medium heterogeneous graph neural network model, the chemical reaction kinetics system and the instantaneous pollutant concentration are converted into time series input data, and the time series input data is encoded by a multi-layer encoder, a loss function is designed to calculate data fitting loss, physical constraint loss, and regularization loss in real time, and a gradient descent method is used to prevent overfitting during real-time calculation; A standard time of real-time calculation is preset, and when the time of data fitting loss, physical constraint loss, and regularization loss real-time calculation reaches the corresponding standard time, the calculation is stopped, and a process production feature-pollutant physical and chemical property-environment medium heterogeneous graph neural network model satisfying time series prediction is obtained, which is calibrated as a pollution time series prediction target neural network model; The Bayesian optimization and the Monte Carlo method are combined to update the pollution time series prediction target neural network model, realize model time series prediction and uncertainty quantification, and output a high-uncertainty pollutant distribution area.

5. The method for predicting the spatio-temporal distribution of pollutants in an industrial cluster according to claim 4, characterized in that, The Bayesian optimization and the Monte Carlo method are combined to update the pollution time series prediction target neural network model, realize model time series prediction and uncertainty quantification, and output a high-uncertainty pollutant distribution area. In the data processing terminal, the PINN training framework is used to perform secondary construction of the pollution time series prediction target neural network model. In the PINN training framework, the Gaussian regression is performed on the pollution time series prediction target neural network model through Bayesian optimization. The Gaussian regression is a regression training of the pollution time series prediction target neural network model, and the model parameters closest to the preset value are collected in real time during the regression training process until the regression times equal the preset times, and then the regression training is stopped. After the Bayesian optimization, the pollution time series prediction target neural network model is subjected to Monte Carlo processing, random Monte Carlo sampling is performed, and all the predicted distribution positions of the pollutants in the target industrial park are obtained, wherein all the predicted distribution positions are uncertain pollutant distribution positions. The entropy value and the standard deviation of all the predicted distribution positions are calculated by the pollution time series prediction target neural network model, and if one of the entropy value and the standard deviation of the predicted distribution position satisfies the preset value, the corresponding predicted distribution position is marked as a high-uncertainty pollutant distribution area. The high-uncertainty pollutant distribution area represents an area where the judgment concentration of the pollutant is greater than the standard value.

6. The method for predicting the spatio-temporal distribution of pollutants in an industrial cluster according to claim 1, characterized in that, The high-uncertainty pollutant distribution area is subjected to pollutant concentration coupling prediction processing in combination with the space-time dimension, specifically as follows: In the data processing terminal, the high-uncertainty pollutant distribution area is set, and the high-uncertainty pollutant distribution area is divided into different high-uncertainty pollutant distribution sub-areas. In the target industrial park, the corresponding high-uncertainty pollutant distribution sub-area is subjected to on-site pollutant sampling processing, and the pollutant concentration quantification processing and the pollutant type determination are performed in the high-uncertainty pollutant distribution sub-area to construct a pollutant fingerprint measured spectrum. At the same time, the different high-uncertainty pollutant distribution sub-areas are subjected to embedding processing of environmental medium characteristics, hydrogeological driving factors, and equipment density weighted process characteristics to generate spatial dimension characteristics of the high-uncertainty pollutant distribution sub-area. The environmental medium characteristics, the hydrogeological driving factors, and the equipment density weighted process characteristics are obtained by searching the historical data network. At the same time, the production cycle state in the target industrial park is searched through the historical data network and is defined as the time dimension characteristics of the high-uncertainty pollutant distribution sub-area. In the pollution time series prediction target neural network model, the spatial dimension characteristics and the time dimension characteristics of the high-uncertainty pollutant distribution sub-area are imported and combined for calculation, and the probability density function of the pollutant concentration of the high-uncertainty pollutant distribution sub-area is output, which is used to construct a pollutant fingerprint prediction spectrum. The pollutant fingerprint prediction spectrum describes the predicted value and the predicted type of the pollutant concentration in the high-uncertainty pollutant distribution sub-area, and is subjected to cosine similarity analysis with the pollutant fingerprint measured spectrum. If the cosine similarity between the pollutant fingerprint prediction spectrum and the pollutant fingerprint measured spectrum is not less than the preset value, it is judged that the pollution time sequence prediction target neural network model is qualified, and if it is less than the preset value, the pollution time sequence prediction target neural network model is updated online to ensure that the cosine similarity between the output pollutant fingerprint prediction spectrum and the pollutant fingerprint measured spectrum is not less than the preset value. The method for updating the pollution time sequence prediction target neural network model online is to introduce meta-learning and online learning parallel technology to parallel train the pollution time sequence prediction target neural network model. Through the qualified pollution time sequence prediction target neural network model, pollutant fingerprint prediction spectrum construction is performed on all high-uncertainty pollutant distribution sub-regions in the industrial cluster area.

7. An industrial cluster pollutant spatiotemporal distribution prediction system characterized in that, The distribution prediction integrated high-performance computing architecture and data storage module includes a non-volatile memory composed of an ECC-verified DDR4 RDIMM memory module and an NVMe solid-state storage array using a 3D NAND flash memory, and a multi-core processor based on a Zen4 microarchitecture; the distribution prediction method program with a prediction engine is solidified and deployed in the memory, and when the program is executed in parallel by the superscalar pipeline execution unit in the processor, the distribution prediction steps of any one of claims 1-6 are realized.

Images