Regional hazardous waste risk pattern prediction method and system based on spatial clustering and time sequence evolution

Abstract

The invention provides a regional hazardous waste risk pattern prediction method and system based on spatial clustering and time sequence evolution, and relates to the technical field of environmental big data. The method comprises the following steps: acquiring multi-source service data to extract node yield load and storage margin attributes, and constructing a risk potential energy vector; calculating topological correlation intensity in combination with a historical transfer record, forming a transfer path impedance coefficient, obtaining a circulation resistance parameter by coupling potential energy decline, mapping the circulation resistance parameter into a circulation probability, and generating a time sequence dynamic circulation tensor; segmenting to obtain a high-risk cluster; and constructing a risk conduction topological graph based on the cluster centroid, identifying a key hub, searching a risk cascade key path, extracting a hotspot drift vector, establishing a dual-drive evolution prediction model, and outputting a regional risk pattern prediction result. The problems that in the prior art, modeling of a risk dynamic conduction mechanism is insufficient, space-time analysis dimensions are mutually separated, and a prediction model lacks physical network constraints are solved.

Description

Technical Field

[0001] This invention relates to the field of environmental big data technology, and in particular to a method and system for predicting regional hazardous waste risk patterns based on spatial clustering and temporal evolution. Background Technology

[0002] With the acceleration of industrialization, the generation of hazardous waste has been increasing year by year, making its environmental risk prevention and control a key focus of regional environmental management. Current hazardous waste supervision mainly relies on the combination of electronic manifest systems and geographic information systems (GIS). By collecting structured data such as enterprise declarations, transfer manifests, and disposal records, it achieves information-based recording and querying of the entire lifecycle of hazardous waste. Existing regulatory platforms can typically display basic location information of enterprises and use statistical methods to summarize and statistically analyze the generation and disposal of hazardous waste within each administrative region, generating static heat maps or statistical reports, thereby achieving point-to-point flow monitoring and post-event traceability.

[0003] To enhance the proactiveness and predictability of regulation, recent technological trends have shifted towards big data integration and intelligent prediction. Researchers are attempting to introduce complex network theory to analyze the structural characteristics of hazardous waste circulation networks and identify key nodes, such as large-scale disposal centers. Simultaneously, they are combining machine learning algorithms, such as random forests, linear regression, or long short-term memory neural networks, to predict future hazardous waste generation based on historical data. Furthermore, with the development of digital twin technology, single-map visualization based on geographic information systems has become an industry standard, aiming to assist regulatory decision-making through multi-scale spatial representation.

[0004] However, existing regional hazardous waste risk prediction methods still have significant limitations in practical applications. First, they lack in-depth modeling of the dynamic transmission mechanism of risk. Existing technologies mostly treat hazardous waste networks as static physical connections, ignoring the risk spillover and cascading effects caused by factors such as disposal capacity saturation and transportation disruptions between nodes. When a core disposal unit fails, the risk rapidly accumulates upstream along the network, and conventional statistical models cannot capture this nonlinear dynamic process based on network topology. Second, the spatiotemporal analysis dimensions are fragmented. Existing methods typically use spatial clustering to identify current hotspots and time-series forecasting to predict total trends, treating these as two independent steps, lacking a unified spatiotemporal coupling analysis framework. This makes it difficult for models to accurately predict the continuous migration paths of high-risk areas in geographic space, and fails to explain how risk hotspots move from one region to another over time. Third, the prediction results lack physical constraints. Purely data-driven algorithms often extrapolate values ​​based solely on historical curves, without fully considering the capacity constraints of real-world physical networks, such as storage capacity limits and cross-regional transfer restrictions. This can lead to predictions that are numerically accurate but logically contradictory, making it difficult to guide actual resource scheduling and risk mitigation deployment. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, the purpose of this invention is to provide a method and system for predicting regional hazardous waste risk patterns based on spatial clustering and temporal evolution. This invention solves the problems in existing technologies, such as insufficient modeling of the dynamic risk transmission mechanism, fragmented spatiotemporal analysis dimensions, and lack of physical network constraints in the prediction model.

[0006] To achieve the above objectives, the present invention provides the following solution:

[0007] A method for predicting regional hazardous waste risk patterns based on spatial clustering and temporal evolution includes:

[0008] Acquire multi-source business data of the target area, and extract attributes of nodes in the target area based on the multi-source business data to obtain a node attribute set, wherein the nodes in the node attribute set have production load attributes and storage capacity attributes.

[0009] Based on the coupling relationship between the production load attribute and the storage margin attribute of each node in the node attribute set, calculate the risk potential vector of each node.

[0010] Based on the historical transfer records extracted from the multi-source business data and the risk potential vector, the transfer path impedance coefficient between each node is calculated;

[0011] The flow resistance parameters between each node are determined based on the risk potential energy vector and the transfer path impedance coefficient, and the geographic coordinates of the nodes are obtained from the multi-source business data to construct regional spatial distance constraints.

[0012] A time-series dynamic flow tensor is generated based on the regional spatial distance constraint and the flow resistance parameter.

[0013] Within a preset three-dimensional mapping domain, a hazardous waste risk evolution spatiotemporal cube is constructed based on the temporal dynamic flow tensor, and the hazardous waste risk evolution spatiotemporal cube is subjected to voxel density segmentation to obtain high-risk clusters.

[0014] Based on the high-risk clusters, a risk transmission topology map is constructed, and the network centrality algorithm is used to analyze the risk transmission topology map to obtain the risk cascading critical path;

[0015] The centroid displacement of the high-risk clusters in the time series is traced sequentially according to the risk cascade critical path to extract the risk hotspot drift vector. The risk hotspot drift vector is then corrected by topological constraints and extrapolated by temporal inertia to obtain a dual-drive evolution prediction model.

[0016] The regional risk evolution is calculated based on the dual-drive evolution prediction model to obtain the prediction results of the regional hazardous waste risk pattern.

[0017] A regional hazardous waste risk pattern prediction system based on spatial clustering and temporal evolution includes:

[0018] The data acquisition and attribute extraction module is used to acquire multi-source business data of the target area, and extract attributes of the nodes in the target area based on the multi-source business data to obtain a node attribute set, wherein the nodes in the node attribute set have production load attributes and storage capacity attributes.

[0019] The risk potential vector calculation module is used to calculate the risk potential vector of each node based on the coupling relationship between the production load attribute and the storage margin attribute of each node in the node attribute set.

[0020] The transfer path impedance calculation module is used to calculate the transfer path impedance coefficient between each node based on the historical transfer records extracted from the multi-source service data and the risk potential vector.

[0021] The circulation resistance and spatial constraint construction module is used to determine the circulation resistance parameters between each node based on the risk potential energy vector and the transfer path impedance coefficient, and to obtain the node geographical coordinates from the multi-source business data to construct regional spatial distance constraints.

[0022] A time-series dynamic flow tensor generation module is used to generate a time-series dynamic flow tensor based on the regional spatial distance constraint and the flow resistance parameter.

[0023] The spatiotemporal cube construction and clustering module is used to construct a hazardous waste risk evolution spatiotemporal cube based on the temporal dynamic flow tensor within a preset three-dimensional mapping domain, and to perform voxel density segmentation on the hazardous waste risk evolution spatiotemporal cube to obtain high-risk clusters.

[0024] The risk topology parsing and path identification module is used to construct a risk transmission topology map based on the high-risk clusters, and to parse the risk transmission topology map using a network centrality algorithm to obtain the risk cascading critical paths;

[0025] The dual-drive evolution prediction model construction module is used to sequentially track the centroid displacement of the high-risk clusters in the time series according to the risk cascade critical path, so as to extract the risk hotspot drift vector, and perform topological constraint correction and temporal inertial extrapolation on the risk hotspot drift vector to obtain the dual-drive evolution prediction model.

[0026] The regional risk pattern prediction module is used to perform evolution calculations on regional risks based on the dual-drive evolution prediction model to obtain the regional hazardous waste risk pattern prediction results.

[0027] The present invention discloses the following technical effects:

[0028] This invention provides a method and system for predicting regional hazardous waste risk patterns based on spatial clustering and temporal evolution. By constructing a risk potential energy vector and transmission topology map coupling production load and storage margin, this invention effectively solves the problem of existing technologies neglecting dynamic risk transmission and cascading effects, accurately simulating the reverse backlog mechanism under network congestion. Simultaneously, by generating a risk evolution spatiotemporal cube using a temporal dynamic flow tensor and combining it with a dual-drive evolution prediction model, it overcomes the shortcomings of traditional methods' fragmented spatiotemporal analysis, achieving quantitative tracking of the continuous drift paths of risk hotspots. Furthermore, by introducing regional spatial distance constraints and flow dynamic equations, it compensates for the lack of physical constraints in purely data-driven models, ensuring that the prediction results conform to real-world physical limitations and significantly improving the robustness of risk warnings and the accuracy of regulatory decisions. Attached Figure Description

[0029] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments 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 drawings can be obtained based on these drawings without creative effort.

[0030] Figure 1 A flowchart of a method for predicting regional hazardous waste risk patterns based on spatial clustering and temporal evolution, provided in an embodiment of the present invention;

[0031] Figure 2 This is a schematic diagram of a regional hazardous waste risk pattern prediction system based on spatial clustering and temporal evolution, provided as an embodiment of the present invention.

[0032] Figure label:

[0033] 1-Data acquisition and attribute extraction module; 2-Risk potential energy vector calculation module; 3-Transfer path impedance calculation module; 4-Flow resistance and spatial constraint construction module; 5-Time-series dynamic flow tensor generation module; 6-Spatiotemporal cube construction and clustering module; 7-Risk topology analysis and path identification module; 8-Dual-drive evolution prediction model construction module. Detailed Implementation

[0034] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0035] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0036] like Figure 1 As shown, this invention provides a method for predicting regional hazardous waste risk patterns based on spatial clustering and temporal evolution, including:

[0037] Step 100: Obtain multi-source business data of the target area, and extract attributes of the nodes in the target area based on the multi-source business data to obtain a node attribute set, wherein the nodes in the node attribute set have production load attributes and storage capacity attributes.

[0038] Step 200: Calculate the risk potential vector of each node based on the coupling relationship between the production load attribute and the storage margin attribute of each node in the node attribute set.

[0039] Step 300: Calculate the transfer path impedance coefficient between each node based on the historical transfer records extracted from the multi-source business data and the risk potential vector;

[0040] Step 400: Determine the flow resistance parameters between each node based on the risk potential energy vector and the transfer path impedance coefficient, and obtain the node geographical coordinates from the multi-source business data to construct regional spatial distance constraints;

[0041] Step 500: Generate a time-series dynamic flow tensor based on the regional spatial distance constraint and the flow resistance parameter;

[0042] Step 600: Within a preset three-dimensional mapping domain, construct a hazardous waste risk evolution spatiotemporal cube based on the temporal dynamic flow tensor, and perform voxel density segmentation on the hazardous waste risk evolution spatiotemporal cube to obtain high-risk clusters;

[0043] Step 700: Based on the high-risk clusters, construct a risk transmission topology map, and use the network centrality algorithm to analyze the risk transmission topology map to obtain the risk cascading critical path;

[0044] Step 800: Track the centroid displacement of the high-risk clusters in the time series according to the risk cascade critical path to extract the risk hotspot drift vector, and perform topological constraint correction and temporal inertial extrapolation on the risk hotspot drift vector to obtain the dual-drive evolution prediction model.

[0045] Step 900: Perform evolution calculations on regional risks based on the dual-drive evolution prediction model to obtain the prediction results of regional hazardous waste risk patterns.

[0046] Furthermore, the specific implementation process of step 100 is as follows:

[0047] In this embodiment, multi-source business data of the target area is first acquired and aggregated around the consistent identifier of the same enterprise in different business systems. Specifically, this embodiment performs field parsing on the multi-source business data, unifying the definitions of key fields such as enterprise name, unified social credit code, manifest number, waste category code, unit of measurement, occurrence time, and flow destination, and performs cross-source deduplication and conflict resolution to eliminate redundant records caused by duplicate reporting, revoked manifests, and supplementary manifests. Subsequently, the parsed data is split into two categories according to business attributes: electronic transfer manifest data and enterprise management plan filing data. A node index is established using the enterprise's unique identifier as the primary key, thereby identifying the target node set such as waste-generating enterprises, storage enterprises, transfer nodes, and disposal and utilization units within the region, providing a consistent data foundation for subsequent attribute extraction and inter-node association calculations.

[0048] In this embodiment, for the extraction of production load attributes, electronic transfer manifest data is used as the core input. Waste generation time-series records for each target node are extracted according to a preset historical period, and the time series are aligned with a uniform time granularity to form a generation vector of the node at different time points. To avoid diluting the current risk assessment with early outdated data, this embodiment implements time decay weighted accumulation on the time-series records, assigning a higher contribution to the generation amount closer to the current time, thereby obtaining a production load attribute that can reflect the recent waste generation pressure. When calculating this attribute, it is based on the manifest generation amount under the same node, the same waste category, or the same risk caliber, allowing for conversion of measurement units and correction of outliers, making the production load attribute comparable within the region and usable for subsequent potential energy modeling.

[0049] In this embodiment, the extraction of the storage balance attribute uses enterprise management plan filing data as the core input. The approved storage limit value of the target node is parsed and combined with the real-time inventory level to form an executable balance calculation. Specifically, this embodiment matches the approved storage limit value and the real-time inventory level according to the same waste category and the same measurement caliber. Inventory data with batch differences or inconsistent reporting frequencies are time-aligned. Then, the difference between the approved storage limit value and the real-time inventory level is calculated to obtain the storage balance attribute. The production load attribute and the storage balance attribute are mapped back to the corresponding target node using the enterprise's unique identifier, forming a node attribute set containing node identifier, attribute value, timestamp, and waste category dimension. This supports subsequent risk potential calculation and time-series evolution modeling.

[0050] Furthermore, the specific implementation process of step 200 is as follows:

[0051] In this embodiment, the risk potential vector is used to couple and characterize the waste generation pressure and storage capacity of a node on the same scale, thereby simultaneously reflecting the combined effect of "recent generation intensity" and "available storage space" on risk in subsequent calculations of transfer impedance and flow resistance. Specifically, this embodiment first summarizes the production load attribute set and storage reserve attribute set of all nodes within the target area based on the node attribute set, and calculates the maximum value of each attribute across all nodes in each attribute set as a normalization benchmark. Then, the production load attribute of each node is normalized to map it to a unified dimensionless interval, eliminating magnitude differences caused by different enterprise sizes and statistical methods. Simultaneously, the storage reserve attribute of each node is also normalized and further converted into a "storage pressure component." This storage pressure component means that the smaller the current storage reserve of a node, the greater the pressure and the higher the risk, thus ensuring that the storage reserve and risk direction are consistent. To ensure stable computation even in extreme cases, such as when the maximum value of the entire domain is zero or when individual node attributes are missing, this embodiment introduces a minimal positive number as a numerical stabilizing term. Its function is to avoid division by zero in normalization operations and suppress numerical divergence. This minimal positive number can be a fixed small quantity that is independent of the attribute dimension and does not participate in manual parameter tuning to change the risk ranking.

[0052] After scaling the two types of components, this embodiment combines the normalized production component and the storage pressure component of a node in a fixed order into a two-dimensional vector, which serves as the risk potential energy vector of that node. The risk potential energy vector describes the comprehensive risk driving force of a node in the dimensions of waste generation and storage, and its function is to provide input containing both intensity and direction information for subsequent risk transmission calculations between nodes. Subsequently, this embodiment calculates the magnitude of this two-dimensional vector as the risk potential energy magnitude. The risk potential energy magnitude is the intensity quantification value of the risk potential energy vector, and its function is to compress the two-dimensional coupling state into a single comparable intensity index, used to represent the overall risk driving force of the node and support the subsequent judgment of the potential energy gradient. Through the above processing, this embodiment ensures that when the production load is high and the storage margin is low, the risk potential energy magnitude increases accordingly; when the production load is low or the storage margin is sufficient, the risk potential energy magnitude decreases accordingly, thus conforming to the physical intuition and business logic of risk sources in regulatory scenarios.

[0053] In this embodiment, the production load attribute is derived from the time-lapse accumulation of waste generation time-series records from electronic transfer manifest data within a preset historical period, and its value is a non-negative real number. The storage balance attribute is derived from the difference between the approved storage limit obtained from the enterprise management plan filing data and the real-time inventory at the current moment, and its value can be a non-negative real number. If a negative value appears, it indicates that the inventory exceeds the approved limit. In this embodiment, it can be regarded as a zero balance and the storage pressure can be maximized to reflect the risk of over-storage. The maximum value of all nodes is obtained by taking the maximum of the corresponding attributes of all nodes at the same moment, under the same waste category or the same risk caliber. Its function is to serve as a normalization benchmark to maintain comparability within the region. The numerical stability term originates from a fixed small quantity pre-defined in this embodiment. Its function is only for numerical stabilization and does not change the relative magnitude relationship of attributes. For example, when the production load of all nodes in the region is zero or the storage margin is zero at a certain moment, the normalization benchmark is zero. This embodiment avoids calculation failure and keeps the normalization result at or close to zero by introducing this fixed small quantity. The two components of the risk potential energy vector come from the normalized production component and the storage pressure component, respectively. Their functions correspond to the waste production drive and the storage constraint drive, respectively. The risk potential energy modulus is obtained by calculating the above two-dimensional components according to the second norm. Its function is to output a unified risk intensity index for subsequent path impedance and flow resistance coupling solution.

[0054] Furthermore, the specific implementation process of step 300 is as follows:

[0055] In this embodiment, a topological association matrix between nodes is first constructed based on historical transfer records from multi-source business data. This matrix characterizes the existing flow relationships and business stickiness of hazardous waste among enterprise nodes within a region. Specifically, this embodiment uses electronic transfer manifest data as the primary source. Each transfer record is mapped to a directed association between a starting node and an ending node according to the enterprise's unique identifier. The number of transfers or the cumulative transfer amount within a preset statistical period is used as the association strength to fill the corresponding cells in the topological association matrix. When different waste categories or different disposal destinations exist for the same pair of nodes, this embodiment can construct sub-matrices according to waste categories and then summarize them into a total matrix using a unified standard to ensure that the topological association matrix can truly reflect the stability and frequency of the business link. To convert business stickiness into "resistance" semantics, this embodiment performs a reciprocal transformation on the topological association matrix, making the resistance corresponding to the channel with stronger association smaller. Furthermore, global normalization is performed to ensure that the basic channel resistance values ​​fall within a comparable range, thereby obtaining a set of basic channel resistance values ​​characterizing the business stickiness between nodes.

[0056] In this embodiment, after obtaining the basic channel resistance value, a risk potential energy vector is introduced to characterize the degree of resistance of the transfer path at the risk-driven level. Specifically, for any candidate transfer channel from the output node to the input node, this embodiment reads the risk potential energy vectors of the output node and the input node respectively, and calculates the difference in their risk potential energy magnitudes to reflect the potential energy drop trend along the channel. Simultaneously, the cosine of the direction angle between the two vectors is calculated to reflect the consistency of risk composition. The magnitude difference is used to characterize whether the transfer from the output node to the input node conforms to the natural gradient of "from high-risk driving to low-risk release," and the cosine of the direction angle is used to characterize whether the coupling direction of the risks at the two nodes, composed of the production load component and the storage pressure component, is consistent. Based on these two types of quantification results, this embodiment constructs a reverse repulsive force function of the potential energy gradient, mapping the reverse upward trend and directional inconsistency to a potential energy resistance value. This ensures that when the potential energy of the input node is stronger than that of the output node or the risk composition directions of the two nodes differ significantly, the resistance corresponding to the channel increases, thereby suppressing the contribution of paths that do not conform to the risk gradient in subsequent flow modeling.

[0057] In this embodiment, to obtain the transfer path impedance coefficient between nodes, the basic channel resistance value and the potential energy resistance value are combined to form a comprehensive impedance characterization that simultaneously includes business stickiness and risk gradient resistance. Specifically, this embodiment performs weighted summation on the two types of resistance values ​​after scaling them to obtain the transfer path impedance coefficient. The purpose of weighted summation is to integrate historical business channel constraints and risk potential energy constraints into a single impedance parameter that can be used for subsequent flow dynamics solutions, and to ensure that the impedance coefficient increases with the increase of basic channel resistance and with the increase of potential energy resistance. When there is no historical transfer record between a pair of nodes, this embodiment sets the corresponding topological association strength to zero and makes the basic channel resistance reach a high value or directly marks the channel as unreachable, so as to avoid generating unreliable virtual paths in the absence of business support.

[0058] Specifically, in this embodiment, the inverse repulsion function is used to simultaneously integrate historical business channel constraints and risk potential gradient adversarial constraints when calculating the impedance coefficient of the transfer path between nodes. This means that it quantifies the difficulty of transferring a directed channel from the starting node to the ending node. Its function is to increase the comprehensive impedance of the channel when there are fewer or less stable historical transfers, and when the risk potential energy of the ending node is higher or the risk composition direction is more inconsistent, thereby suppressing paths that do not conform to the natural risk gradient or lack business support from participating in subsequent flow modeling. In specific implementation, this embodiment first extracts historical transfer records within a preset statistical period from the electronic transfer manifest data. Each record is mapped to a directed association between the starting enterprise and the ending enterprise according to the enterprise's unique identifier. The number of transfers or the cumulative manifest volume are used as the topological association strength and written into the topological association strength table. This table then forms a global topological association matrix. Subsequently, this embodiment performs a reciprocal transformation on the topological association strength and performs global normalization to obtain the basic channel resistance value, making the resistance of channels with more frequent transfers smaller and the values ​​comparable. Meanwhile, the risk potential energy vector and risk potential energy magnitude output in this embodiment are respectively taken as the difference between the risk potential energy magnitudes of the starting node and the ending node to characterize the upslope of potential energy along the channel, and the cosine of the angle between the risk potential energy vectors of the two nodes is calculated to characterize the consistency of the risk composition direction. The risk potential energy vector is a two-dimensional vector composed of the normalized production load component and the normalized storage pressure component. Its function is to simultaneously express the node risk intensity and the risk source composition direction. The risk potential energy magnitude is the intensity index of the risk potential energy vector. Its function is to compare the magnitude of the overall risk driving force of the node. In this embodiment, the upslope of potential energy is mapped exponentially to the intensity term of potential energy resistance, and the inconsistency of direction is mapped to a dimensionless penalty term. The two are multiplied to obtain the potential energy resistance value, so that when the potential energy of the ending node is stronger than that of the starting node or when the difference in risk composition between the two nodes increases, the resistance increases rapidly. Finally, the basic channel resistance value and the potential energy resistance value are superimposed to obtain the transfer path impedance coefficient of the directed channel. To ensure numerical stability, this embodiment introduces a very small positive number as a stabilizing term. The source of this term is a fixed small quantity preset in this embodiment. Its function is to avoid division by zero or numerical divergence in normalization and similarity calculation without changing the relative order between channels.

[0059] For example, in this embodiment, using electronic transfer manifest data from a certain statistical period as the source, the historical transfer records from node 1 to node 2 are statistically analyzed to obtain a transfer count of 40, and the transfer count from node 1 to node 3 is statistically analyzed to obtain a transfer count of 5. Furthermore, the maximum transfer count among all node pairs in the entire domain is 50. Therefore, the topological association strength between node 1 and node 2 is set to 40, and the maximum value in the entire domain is set to 50. The topological association strength between node 1 and node 3 is set to 5. Simultaneously, this embodiment obtains the risk potential vector values ​​of node 1 as 0.80 and 0.60, and based on this, the risk potential magnitude is 1. The risk potential vector values ​​of node 2 are 0.30 and 0.20, and the risk potential magnitude is 0.3606. The risk potential vector values ​​of node 3 are 0.90 and 0.90, and the risk potential magnitude is 1.2728. These values ​​are obtained by combining node production load and storage margin after normalization and storage pressure conversion, respectively. Regarding the numerical stability term, this embodiment takes... The fixed small value is 0.000001, which is a constant preset in this embodiment. Accordingly, from node 1 to node 2, since the potential energy modulus at the endpoint is smaller than that at the starting point, the potential energy upslope is 0 and the potential energy resistance term is close to 0. Therefore, the impedance of this channel is mainly determined by the resistance of the basic channel, and the resistance of the basic channel is relatively small due to the large number of historical transfers (40). From node 1 to node 3, since the potential energy modulus at the endpoint (1.2728) is larger than that at the starting point (1.0000), the potential energy upslope is 0.2728 and amplified by exponential mapping. At the same time, the similarity of the risk potential energy vector directions of node 1 and node 3 is approximately 0.9970, calculated by the dot product of the two vectors and the modulus. This makes the direction inconsistency penalty close to 0 but still non-zero. Therefore, the potential energy resistance is non-zero and is superimposed on the resistance of the basic channel, making the transfer path impedance coefficient from node 1 to node 3 significantly larger than that from node 1 to node 2. This reflects the suppression effect of this embodiment on the low-frequency channel and the reverse potential energy channel.

[0060] Furthermore, the specific implementation process of step 400 is as follows:

[0061] In this embodiment, the latitude and longitude geographic coordinates of each node are first parsed from multi-source business data to establish a spatial geometric foundation, and a global distance matrix is ​​constructed accordingly for subsequent spatial constraint calculations. Specifically, this embodiment uses the enterprise management plan filing information and the factory site coordinates registered in the regulatory ledger as the main data source. When multiple coordinate records exist, the latest filed or verified coordinates are used as the node coordinates, and abnormal coordinates are corrected to avoid invalid points such as latitude and longitude crossing boundaries or falling into the sea area. Subsequently, this embodiment calculates the spherical transmission distance between any two nodes using the spherical distance calculation method, and summarizes the distances of all node pairs to form a global distance matrix. The global distance matrix represents the degree of spatial separation between any two nodes within the region. Its function is to provide distance penalties and accessibility determination criteria that conform to geographical reality in subsequent flow modeling, thereby avoiding the generation of cross-domain paths that lack transportation feasibility based solely on business stickiness or potential energy relationships.

[0062] In this embodiment, after obtaining the global distance matrix, a regional spatial distance constraint is constructed by introducing the hazardous waste transportation limit radius to reflect the boundary of transportation capacity and the distance decay law. Specifically, this embodiment reads the transportation limit radius determined by regulatory rules or business experience as a hard constraint threshold. This limit radius is used to limit the maximum acceptable transportation range of hazardous waste under normal transportation conditions and serves as the basis for judging spatial accessibility. Within the limit radius, this embodiment uses a Gaussian decay method to spatially weight the distance between nodes. The smaller the distance, the greater the weight, and the weight decreases smoothly as the distance increases, thus obtaining a distance weight matrix to express the spatial feasibility of a channel. For node pairs exceeding the limit radius, this embodiment resets the distance weight to zero to directly eliminate inaccessible channels, thus obtaining a regional spatial distance constraint matrix. This matrix represents a constraint table after applying spatial accessibility and distance penalties to all node pairs. Its function is to perform spatial screening and intensity suppression on potential channels in the subsequent circulation probability or circulation tensor generation stage to ensure that the prediction results conform to the transportation radius and geographical proximity.

[0063] In this embodiment, after the spatial constraints are determined, the flow resistance parameters between each node are calculated based on the risk potential energy vector and the transfer path impedance coefficient to form a dynamic quantization input that can be directly used for subsequent flow probability mapping and timing modulation. Specifically, for any directed channel, this embodiment first calculates the gradient projection value of the risk potential energy along the channel direction based on the risk potential energy vectors of the starting and ending nodes and the direction of the spatial connection between the two nodes. This gradient projection value is used to characterize the effective decrease of the risk driving force along the channel direction. The larger the gradient projection value, the more obvious the decrease of risk potential energy from the starting point to the ending point and the more consistent it is with the risk release direction. Subsequently, this embodiment substitutes the gradient projection value and the transfer path impedance coefficient into a preset flow dynamics equation for coupled solution. The flow dynamics equation is used to uniformly map the channel's service impedance and potential energy driving force into a single flow resistance value output, and ensures that the flow resistance increases accordingly when the transfer path impedance coefficient increases and decreases accordingly when the gradient projection value increases, thereby achieving positive and negative correlation at the numerical level. Finally, this embodiment writes the output value as the flow resistance parameter between nodes into the flow resistance matrix, and together with the regional spatial distance constraint matrix, it serves as the key input for the subsequent generation of the time-series dynamic flow tensor, ensuring that the flow modeling simultaneously satisfies the three core constraints of spatial accessibility, historical channel constraints, and risk gradient driving.

[0064] Specifically, in this embodiment, the flow dynamics equation is used to couple the business channel impedance between nodes with the risk potential energy gradient drive into a quantifiable flow resistance parameter. This means that the overall degree of obstruction during the flow of hazardous waste between nodes is numerically characterized. Its function is to provide a monotonically consistent input in the subsequent mapping stage from resistance to probability, ensuring that the more unstable the historical channel or the stronger the risk resistance, the greater the channel resistance, while the more significant the decrease in risk potential energy from the starting point to the ending point, the smaller the channel resistance. In specific implementation, for any directed channel from the starting node to the ending node, this embodiment first calls the risk potential energy modulus of the two nodes obtained in step 200 and calculates the potential energy decrease. That is, the risk potential energy modulus of the starting node is subtracted from the risk potential energy modulus of the ending node and compared with zero to take the larger value, ensuring that the potential energy decrease is non-negative and avoiding the introduction of a negative driving force when the potential energy of the ending node is higher than that of the starting node. The potential energy decrease is the effective potential energy difference along the channel direction that can be used to drive risk release; its function is to suppress reverse potential energy flow and strengthen forward potential energy flow as a driving force term. Subsequently, in this embodiment, the transfer path impedance coefficient obtained in step 300 is used as a resistance term and substituted into the flow dynamics equation along with the aforementioned potential energy decrease for coupled calculation. That is, the flow resistance parameter is output by dividing the impedance coefficient by the sum of the potential energy decrease and the numerical stability term, thereby ensuring at the numerical level that the flow resistance parameter increases when the impedance coefficient increases and decreases when the potential energy decreases. The numerical stability term is derived from a fixed small quantity preset in this embodiment, and its function is only to avoid division by zero when the potential energy decrease is zero and to maintain calculation stability. It is not used to manually adjust the channel sorting.

[0065] For example, in this embodiment, the transfer path impedance coefficient from node 1 to node 2 is 0.80 from the reverse repulsion function calculation result in step 300, the risk potential energy modulus of node 1 is 1.30 from the calculation result after coupling production load and storage margin in step 200, and the risk potential energy modulus of node 2 is 0.50. Then, the potential energy decrease is taken as 1.30 minus 0.50 and compared with 0, taking the larger value of 0.80. The numerical stability term is taken as a fixed small amount of 0.000001 preset in this embodiment. Therefore, the flow resistance parameter from node 1 to node 2 is equal to 0.80 divided by 0.80 plus 0.000001, which is approximately 0.999999. Further, if the transfer path impedance coefficient from node 1 to node 3 is 1.20 and the risk potential energy modulus of node 3 is 1. If the potential energy decrease is 1.30 minus 1.50 and compared with 0, it becomes 0. The numerical stability term remains 0.000001. Thus, the flow resistance parameter from node 1 to node 3 is equal to 1.20 divided by 0.000001, which is approximately 1200000. This reflects that the channel is given a great flow resistance due to the higher risk potential energy at the termination node, and the flow probability calculation in the subsequent calculation approaches the point of no occurrence. The above impedance coefficient is the channel comprehensive impedance obtained by the combined effect of historical transfer records and potential energy resistance. The risk potential energy modulus is the normalized coupling result of node production load and storage margin. The numerical stability term is a fixed constant preset in this embodiment. Its function is to ensure calculation stability and to ensure that the channel can still obtain a comparable resistance output when the potential energy decrease is zero.

[0066] Furthermore, the specific implementation process of step 500 is as follows:

[0067] In this embodiment, the flow resistance parameter matrix obtained in step 400 is first converted into an interpretable basic flow probability matrix so as to express the potential interaction possibility between nodes in the form of probability intensity in the time series tensor. Specifically, this embodiment constructs a mapping relationship from resistance to probability, so that the greater the flow resistance, the smaller the corresponding probability and the decay is monotonically continuous, thereby avoiding the instability caused by hard thresholds. In this embodiment, the flow resistance parameters of each pair of nodes are input into the mapping relationship item by item to perform a negative exponential decay operation to obtain the basic flow probability matrix under the condition of no spatial constraints. The meaning of this matrix is ​​a potential flow intensity table derived only from the channel resistance. Its function is to provide a benchmark interaction intensity of a uniform scale before the subsequent introduction of spatial distance constraints and time modulation, and to output a probability close to zero for channels with extremely large resistance to suppress unreasonable paths.

[0068] In this embodiment, after obtaining the basic flow probability matrix, it is coupled element-wise with the regional spatial distance constraint to form a spatially corrected flow matrix, thereby explicitly injecting the transportation limit radius and distance attenuation law into the channel strength. Specifically, this embodiment performs a matrix Hadamard product operation on the basic flow probability matrix and the regional spatial distance constraint matrix, so that the probability of channels with zero spatial constraints is directly reduced to zero after spatial correction to indicate unreachability, and the probability strength of channels with non-zero spatial constraints is reduced or maintained according to the distance weight ratio. The meaning of the spatially corrected flow matrix is ​​a table of inter-node flow benchmark strength that simultaneously satisfies the channel resistance constraint and the spatial reachability constraint. Its function is to serve as a time benchmark slice for subsequent three-dimensional tensors, ensuring that the tensor does not generate cross-domain flow relationships that exceed the transportation radius or are geographically unreasonable at each time step.

[0069] In this embodiment, the time window length and time step are then set and a three-dimensional tensor structure is constructed to describe the dynamic process of the evolution of the flow intensity between nodes over time. Specifically, this embodiment determines the time window length to cover the total time domain length of prediction and backtracking according to regulatory needs, and determines the time step length to control the time resolution of discretization. Then, the time dimension is superimposed on the two-dimensional space formed by the starting node dimension and the ending node dimension to form a three-dimensional tensor framework. This embodiment uses the spatially corrected flow matrix as a reference slice to copy and map to each time step, thereby maintaining the temporal consistency of channel strength without additional disturbances. Furthermore, the change rate of the node attribute set at different times is introduced to dynamically modulate each time step slice, so that when the recent output load of a node increases or the storage margin decreases, the strength of the inlet and outlet channels associated with it is enhanced or suppressed at the corresponding time step, thereby reflecting the changes of risk-driven changes in the time dimension. Through the above-mentioned reference mapping and change rate modulation, this embodiment generates a temporal dynamic flow tensor. The meaning of this tensor is a three-dimensional data structure that uniformly encodes the potential flow intensity of nodes in the region at each time step. Its function is to provide directly callable temporal input for subsequent spatiotemporal cube projection, density estimation, and high-risk cluster identification.

[0070] Specifically, the expression for the resistance-probability mapping function is:

[0071] ;

[0072] in, For node indexing; For node indexing; For nodes under no spatial distance constraints The basic turnover probability; For nodes The flow resistance parameters; It is an exponential function.

[0073] Furthermore, the specific implementation process of step 600 is as follows:

[0074] In this embodiment, a three-dimensional mapping domain for risk evolution representation is first constructed, and the continuous spatiotemporal range is discretized into computable standard units using voxelization, thereby providing a unified spatial container for subsequent density estimation and cluster identification. Specifically, this embodiment uses the longitude and latitude corresponding to the geographic coordinates of nodes as planar coordinate axes and the time series from step 500 as vertical coordinate axes to construct the three-dimensional mapping domain, and determines the spatial and temporal boundaries of the mapping domain based on the target area span and time window length. Subsequently, this embodiment sets fixed length, width, and height dimensions for voxel units, where the spatial dimensions are used to limit the geographic grid scale covered by each voxel, and the temporal dimensions are used to limit the temporal resolution covered by each voxel. Based on this, the three-dimensional mapping domain is meshed to obtain a set of standard voxel units. The standard voxel unit is the smallest discrete computational unit within the three-dimensional mapping domain, and its function is to carry the risk density value formed by the accumulation of flow intensity and support connectivity clustering in subsequent steps.

[0075] In this embodiment, after obtaining the standard voxel unit set, the flow intensity in the temporal dynamic flow tensor is projected as a weight onto the three-dimensional mapping domain, thereby mapping the flow relationship between nodes into a spatiotemporal point cloud or raster weight field that can be spatially statistically analyzed. Specifically, this embodiment analyzes the flow intensity value between the starting node and the ending node in the temporal dynamic flow tensor at each time step, and determines the projection position of the flow event in the plane by combining the geographical coordinates of the two nodes. This embodiment can set the projection position as the midpoint of the line connecting the coordinates of the starting node and the ending node, or take the weighted position on the line according to the flow direction, so as to ensure that the projection point can reflect the spatial occupancy of the channel; at the same time, the time step is used as the time coordinate of the projection point, and the flow intensity is accumulated as a weight into the voxel unit corresponding to the projection point. Through the above projection and accumulation, this embodiment forms an initial voxel weight distribution, so that high-frequency and high-intensity flow forms a weight cluster in the three-dimensional mapping domain, providing the original signal for subsequent smooth interpolation and density modeling.

[0076] In this embodiment, spatiotemporal kernel density estimation is then used to smooth the voxel weight distribution to obtain a continuous risk density field, and high-risk clusters are further identified through voxelized density segmentation. Specifically, this embodiment uses each standard voxel unit as the estimation object, and performs kernel function weighted accumulation of the projection weights in its neighborhood according to spatial and temporal distances to obtain the risk density value of the voxel, thereby constructing a spatiotemporal cube of hazardous waste risk evolution. The spatiotemporal cube means a three-dimensional array that stores risk density values ​​with voxels as indices, and its function is to uniformly express the continuous evolution of regional risk in space and time. After obtaining the risk density value, this embodiment uses an adaptive threshold segmentation method to estimate the global background noise threshold, and removes voxels with risk densities below the threshold to obtain a high-risk retained voxel set, so as to reduce the interference of random fluctuations and sparse noise on the clustering results. Finally, this embodiment performs three-dimensional connected component labeling on the high-risk retained voxel set, and merges spatially adjacent and temporally continuous voxels into independent connected regions according to preset connectivity rules to obtain high-risk clusters, so that risk hotspots can be stably extracted in the form of clusters and can be directly used for subsequent topology map construction and hotspot drift tracking.

[0077] Furthermore, the specific implementation process of step 700 is as follows:

[0078] In this embodiment, firstly, a node set of the risk transmission topology map is constructed based on the high-risk clusters output in step 600, and the spatiotemporal density hotspots are converted from the voxel domain into a discrete topological representation that can be used for network analysis. Specifically, this embodiment extracts the geometric centroid of each high-risk cluster in the three-dimensional mapping domain as a network topology node. The geometric centroid means the average position or density-weighted average of all high-risk voxels in the cluster in the spatial and temporal coordinates. Its function is to represent the location and occurrence time of a risk hotspot cluster with a single point, thereby reducing the network size and maintaining the interpretability of hotspot location. When the same hotspot cluster has a long time span, this embodiment can extract multiple centroid nodes by time slice to maintain the temporal resolution of risk evolution. Subsequently, this embodiment calculates the cumulative flow value between any two network topology nodes based on the time-series dynamic flow tensor. Specifically, within a time window covering the spatiotemporal range of the two clusters, the flow intensity of channels falling into the spatial neighborhood of the starting cluster and pointing to the spatial neighborhood of the ending cluster is summarized, and the intensity is accumulated over time to obtain the cumulative flow value, so that it can reflect the overall scale of risk transmission from upstream hotspots to downstream hotspots.

[0079] In this embodiment, after obtaining the cumulative flux value, it is normalized into a risk transmission probability matrix, and a directed weighted graph is constructed based on this matrix to facilitate the subsequent application of network centrality to identify critical hubs and critical paths. Specifically, for each starting network topology node, the cumulative flux value pointing to all downstream nodes is normalized row by row, so that the sum of probabilities in each row is 1, thereby obtaining the risk transmission probability matrix. This matrix represents the relative probability distribution of transmission from any hotspot node to other hotspot nodes, and its function is to compare transmission preferences and support probability gradient search under different total flux scales. Subsequently, this embodiment uses network topology nodes as vertices, non-zero elements in the risk transmission probability matrix as directed edges, and the corresponding probabilities as edge weights to construct a risk transmission topology graph. To avoid weak connections formed by noisy flux affecting the topology structure, this embodiment can set a minimum retention threshold for the cumulative flux value or probability, retaining only edges exceeding the threshold into the graph, thereby improving network sparsity and the interpretability of critical chains.

[0080] In this embodiment, the risk transmission topology is then analyzed using weighted betweenness centrality to identify key hub nodes, and a maximum probability gradient search is performed starting from these key hub nodes to generate critical paths for risk cascading. Specifically, this embodiment calculates the betweenness centrality of each network topology node on the directed weighted graph. The weighted betweenness centrality refers to the degree to which a node is located on multiple high-probability transmission paths, and its function is to identify hotspot nodes that play a bridging and convergence role in risk transmission. In this embodiment, nodes with betweenness centrality higher than a preset threshold are identified as key hub nodes. The preset threshold can be automatically determined by the quantiles of the centrality distribution or by the mean plus standard deviation to reduce manual parameter tuning. Subsequently, this embodiment takes the key hub node as the starting point and selects downstream nodes from the outgoing edge set corresponding to the risk transmission probability matrix according to the maximum probability gradient strategy. That is, at each step, it prioritizes extending along the direction with the slowest probability decrease or the highest probability, and records the formation of potential risk chains during the extension process. In order to ensure the availability and non-redundancy of the chain, this embodiment performs closed-loop detection on the potential risk chain to avoid returning to the visited node and forming a loop. At the same time, it performs redundancy pruning on the chain segment with parallel branches or low contribution edges, and finally outputs the risk cascade critical path, which can represent the most likely transmission channel for the spread of risk from the main hotspot to the secondary hotspot in the region, and provides a topological constraint skeleton for subsequent hotspot drift vector extraction and evolution prediction.

[0081] Furthermore, the specific implementation process of step 800 is as follows:

[0082] In this embodiment, high-risk clusters are first tracked time-by-time to obtain stable centroid motion trajectories, thus providing a reproducible geometric basis for hotspot drift vector extraction. Specifically, this embodiment divides high-risk clusters into multiple spatial sections along the time dimension. Each spatial section consists of high-risk retained voxels corresponding to that time moment, and density-weighted centroid calculation is performed using the risk density value of the voxels as weights, so that voxels with higher risk density contribute more to the centroid position, thereby suppressing the influence of boundary noise on localization. Subsequently, this embodiment connects the risk-weighted centroids of adjacent time steps in chronological order to obtain the original centroid displacement vector sequence. The original centroid displacement vector represents the actual spatial displacement direction and displacement amplitude of the hotspot between adjacent time steps. Its function is to characterize the direct drift trend of the hotspot without introducing topological constraints and predictors, and to serve as the common input for subsequent topological matching and inertial extrapolation.

[0083] In this embodiment, after obtaining the original centroid displacement vector sequence, it is matched with the risk cascade critical path obtained in step 700 to extract the risk hotspot drift vector, and further topological constraints are introduced to achieve spatial correction. Specifically, this embodiment maps the original centroid displacement vector of each time step to the local out-degree neighborhood corresponding to the current hotspot node in the risk transmission topology map, determines the consistency between the displacement direction and the direction of the critical path, thereby determining the topological link segment that best fits the hotspot movement at that time step and defining it as the risk hotspot drift vector. The meaning of the risk hotspot drift vector is the dominant movement direction of the hotspot after being constrained by the critical path. Its function is to keep the drift estimation consistent with the risk transmission skeleton and avoid the hotspot from jumping without conforming to the transmission law under noise disturbance. Subsequently, this embodiment obtains the set of out-degree adjacent edges of the drift vector on the topological graph, calculates the directional similarity between the drift vector and the directional vector of each adjacent edge, and selects the adjacent edge with the highest similarity as the guiding constraint edge. The meaning of the guiding constraint edge is the topological out-degree edge that is most consistent with the current hotspot movement direction. Its function is to provide an interpretable network direction reference for the drift vector. In this embodiment, the drift vector is projected onto the direction of the guiding constraint edge to obtain the spatial topology correction component, thereby filtering out the lateral components that are inconsistent with the topology and enhancing the effective displacement components along the key transmission channels.

[0084] In this embodiment, after obtaining the spatial topology correction component, a dual-drive evolution prediction model is further generated using temporal inertial extrapolation and adaptive fusion to achieve robust prediction of future hotspot displacements. Specifically, this embodiment constructs a time-series sliding window, extracting risk hotspot drift vector sequences from several historical moments as input samples, feeding them into a pre-trained long short-term memory network to extract temporal dependency features and outputting the displacement prediction component for the next moment. The meaning of the temporal inertial extrapolation component is the next drift prediction obtained based on historical motion inertia and periodic patterns. Its function is to compensate for the excessive constraints that may be caused by relying solely on topological projection and to capture the inertial changes of hotspots under regulatory cycles and business rhythms. Subsequently, this embodiment employs an adaptive Kalman filter to assimilate the spatial topology correction component and the temporal inertial extrapolation component and estimate their states. The adaptive Kalman filter dynamically adjusts the fusion strength and outputs the optimal next-time displacement estimate when the prediction and observation are inconsistent. This ensures that the prediction is closer to the topology correction component when the reliability of the topology constraints is higher, and closer to the inertial extrapolation component when the temporal model is more sensitive to short-term changes. Finally, this embodiment uses the fused next-time displacement as the output to form a dual-drive evolution prediction model, which is then used for subsequent evolution calculations of regional risk patterns.

[0085] Specifically, in this embodiment, the dual-drive evolution prediction model is used to adaptively fuse the temporal inertia extrapolation results with the topological constraint correction results, thereby outputting the hotspot drift vector of the next time step to drive the forward evolution of the risk hotspot location. This means that the historical motion inertia and the risk transmission topological skeleton are used to jointly estimate the hotspot drift at the same time. Its function is to maintain prediction stability when the hotspot fluctuates sharply in the short term and to maintain the consistency of prediction direction when the topological constraints are reliable.

[0086] In specific implementation, this embodiment first calculates the original displacement vector based on the risk-weighted centroid sequence obtained in step 800 when the time step is the current moment. That is, the displacement direction and displacement magnitude of the adjacent moment are obtained by subtracting the centroid coordinate vector of the previous moment from the centroid coordinate vector of the current moment. The original displacement is then matched with the critical path of the risk transmission topology map to determine the guiding constraint edge. The unit direction vector corresponding to the guiding constraint edge is then taken as the topology direction reference, thereby obtaining the displacement projection amount corresponding to the topology constraint term. This projection amount is used to characterize the actual movement degree of the hotspot in the topology skeleton direction. At the same time, this embodiment inputs the hotspot drift vector sequence of several historical moments into the pre-trained long short-term memory network and outputs the temporal inertia extrapolation component of the next moment. This component is used to characterize the next drift inferred from historical inertia and rhythmic patterns. Subsequently, this embodiment constructs a fusion gain based on the idea of ​​adaptive Kalman filtering. This fusion gain is jointly determined by the posterior state covariance and the process noise covariance and is compared with the observation noise covariance. This makes the model more dependent on the topological constraint term when the uncertainty is large or the observation noise is small, and more dependent on the temporal inertia extrapolation component when the observation noise is large or the model is more stable.

[0087] In the fusion computation, this embodiment uses the temporal inertial extrapolation component as the reference prediction, and uses the difference between the topological direction projection and the reference prediction as the correction amount. The correction amount is scaled by the fusion gain and then superimposed on the reference prediction to obtain the final output of the hotspot drift vector at the next moment.

[0088] In the above process, the hotspot drift vector is a vectorized expression of the displacement direction and magnitude of the hotspot between adjacent time steps, and its function is to serve as the direct input for hotspot position updates; the temporal inertia extrapolation component is the next displacement prediction given by the Long Short-Term Memory network after extracting temporal features from the historical drift sequence, and its function is to characterize the motion inertia and periodicity of the hotspot; the guiding constraint edge is the topological outgoing edge that is most consistent with the current drift direction, and its function is to provide directional constraints on the network structure; the posterior state covariance matrix, the process noise covariance matrix, and the observation noise covariance matrix are used to characterize the uncertainty of the fusion estimation at the previous time step, the uncertainty of state evolution, and the uncertainty of topological projection observation, respectively. Their sources can be adaptively estimated from the statistical results of historical prediction residuals and topological matching errors, thereby avoiding reliance on manual parameter tuning and ensuring the transferability of the model in different regions and time periods.

[0089] For example, in this embodiment, the current time step is set to 10. Step 800 obtains the risk-weighted centroid coordinate vector of time step 10, which is derived from the high-risk cluster voxel density weighted calculation result of 2.0 and 5.0. The risk-weighted centroid coordinate vector of time step 9 is 1.2 and 4.6, thus obtaining the adjacent displacement vectors as 0.8 and 0.4. The guiding constraint edge corresponding to the hot spot node in the risk transmission topology is determined by the principle of maximum cosine similarity. Its unit direction vector is derived from the normalized result of the centroid coordinate difference between the two ends of the edge, which is 0.8944 and 0.4472. Therefore, the projection of the displacement vector in the topological direction is 0.8 multiplied by 0.8944 plus 0.4 multiplied by 0.4472 to obtain 0.8944, and the displacement components of the corresponding topological constraint term are 0.8944 multiplied by 0.8944 and 0.8944 multiplied by 0.4472 to obtain 0.8000 and 0.4000.

[0090] At the same time step, the input sequence of the Long Short-Term Memory (LSTM) network comes from the historical hotspot drift vector records from time steps 6 to 10. Its output next-time temporal inertia extrapolation components are 0.6 and 0.3. The covariance is derived from the historical fusion residual statistics. In this embodiment, the posterior state covariance at time step 10 is taken as 0.20, the process noise covariance as 0.05, and the observation noise covariance as 0.10. Therefore, the fusion gain is obtained by dividing 0.20 + 0.05 by 0.20 + 0.05 + 0.10. The value reaches 0.7143. The correction amount is obtained by subtracting the inertial extrapolation components 0.6 and 0.3 from the topological constraint terms 0.8000 and 0.4000, resulting in 0.2000 and 0.1000. After fusion gain scaling, these values ​​become 0.1429 and 0.0714. The final output hotspot drift vector for the next time step is obtained by adding 0.6 and 0.3 to 0.1429 and 0.0714, resulting in 0.7429 and 0.3714. This achieves the adaptive fusion prediction of the dual-drive evolution prediction model described in this embodiment.

[0091] Furthermore, the specific implementation process of step 900 is as follows:

[0092] In this embodiment, the hotspot drift vector for the next time step is first output based on the dual-drive evolution prediction model obtained in step 800. The centroid position of high-risk clusters is then updated forward accordingly to achieve continuous evolution calculation of risk hotspots in both spatial and temporal dimensions. Specifically, this embodiment uses each high-risk cluster identified in step 600 as the evolution object, reads its risk-weighted centroid coordinates at the current time step as the initial state, and uses the drift vector for the next time step output by the dual-drive evolution prediction model as the state increment. The centroid coordinates are then updated by displacement to obtain the predicted centroid position for the next time step. When multiple clusters exist on the same risk cascade critical path, this embodiment updates the predicted positions of hotspots from upstream to downstream sequentially according to the topological order of the critical path to maintain consistency between hotspot evolution and the risk transmission framework, avoiding unreasonable jumps in downstream hotspots before upstream hotspots.

[0093] In this embodiment, after updating the predicted centroid position, the predicted centroid position is back-projected onto the three-dimensional mapping domain constructed in step 600, and the risk density field is updated synchronously, thereby forming a visualized and quantifiable regional risk pattern prediction result. Specifically, this embodiment uses the predicted centroid position as the center and reconstructs or translates the voxel occupancy range of the corresponding cluster on its corresponding spatial grid and time layer. This embodiment can adopt a rigid translation method that keeps the voxel shape of the cluster unchanged to improve computational stability, or it can modulate the voxel density amplitude according to the node attribute change rate and channel strength change in step 500 to reflect the enhancement or decay of risk intensity. Subsequently, this embodiment superimposes the predicted voxel sets of all clusters in the next time step to form a predicted risk density cube, and projects the predicted risk density cube into a two-dimensional risk pattern layer according to a preset output resolution or aggregates it into district and county-level risk statistics results according to administrative divisions, thereby simultaneously meeting the output requirements of spatial display and management decision-making.

[0094] In this embodiment, to ensure the interpretability and usability of the prediction results, consistency verification and result solidification are performed on the predicted risk pattern during the output stage. Specifically, this embodiment checks whether the predicted centroid location falls within the reachable range allowed by the regional spatial distance constraints, and prunes or reverts prediction results that cross boundaries or fall into inaccessible areas to the nearest reachable boundary or to the previous stable position to ensure that the prediction conforms to the transportation radius and geographical boundaries. At the same time, this embodiment records the critical path attribution, upstream source nodes, and main transmission direction corresponding to each predicted hotspot based on the risk transmission topology map, thereby attaching risk link explanation information when outputting the risk pattern, enabling regulators to trace the potential transmission source and diffusion direction of the risk hotspot. Finally, this embodiment outputs the risk pattern sequence obtained from the next time step or multiple rolling predictions as the regional hazardous waste risk pattern prediction result, and can generate risk hotspot migration trajectories, risk intensity change curves, and a list of key attention areas according to the time series to support subsequent early warning issuance and regulatory scheduling.

[0095] like Figure 2 As shown, this embodiment also provides a regional hazardous waste risk pattern prediction system based on spatial clustering and temporal evolution, including:

[0096] The data acquisition and attribute extraction module 1 is used to acquire multi-source business data of the target area, and extract attributes of the nodes in the target area based on the multi-source business data to obtain a node attribute set, wherein the nodes in the node attribute set have production load attributes and storage capacity attributes.

[0097] Risk potential vector calculation module 2 is used to calculate the risk potential vector of each node based on the coupling relationship between the production load attribute and the storage margin attribute of each node in the node attribute set.

[0098] The transfer path impedance calculation module 3 is used to calculate the transfer path impedance coefficient between each node based on the historical transfer records extracted from the multi-source business data and the risk potential vector.

[0099] The circulation resistance and spatial constraint construction module 4 is used to determine the circulation resistance parameters between each node based on the risk potential energy vector and the transfer path impedance coefficient, and to obtain the node geographical coordinates from the multi-source business data to construct regional spatial distance constraints.

[0100] The temporal dynamic flow tensor generation module 5 is used to generate a temporal dynamic flow tensor based on the regional spatial distance constraint and the flow resistance parameter.

[0101] The spatiotemporal cube construction and clustering module 6 is used to construct a hazardous waste risk evolution spatiotemporal cube based on the temporal dynamic flow tensor within a preset three-dimensional mapping domain, and to perform voxel density segmentation on the hazardous waste risk evolution spatiotemporal cube to obtain high-risk clusters.

[0102] The risk topology parsing and path identification module 7 is used to construct a risk transmission topology map based on the high-risk clusters, and to parse the risk transmission topology map using a network centrality algorithm to obtain the risk cascading critical path;

[0103] The dual-drive evolution prediction model construction module 8 is used to sequentially track the centroid displacement of the high-risk clusters in the time series according to the risk cascade critical path, so as to extract the risk hotspot drift vector, and perform topological constraint correction and temporal inertial extrapolation on the risk hotspot drift vector to obtain the dual-drive evolution prediction model.

[0104] The regional risk pattern prediction module is used to perform evolution calculations on regional risks based on the dual-drive evolution prediction model to obtain the regional hazardous waste risk pattern prediction results.

[0105] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0106] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for predicting regional hazardous waste risk patterns based on spatial clustering and temporal evolution, characterized in that, include: Acquire multi-source business data of the target area, and extract attributes of nodes in the target area based on the multi-source business data to obtain a node attribute set, wherein the nodes in the node attribute set have production load attributes and storage capacity attributes. Based on the coupling relationship between the production load attribute and the storage margin attribute of each node in the node attribute set, calculate the risk potential vector of each node. Based on the historical transfer records extracted from the multi-source business data and the risk potential vector, the transfer path impedance coefficient between each node is calculated; The flow resistance parameters between each node are determined based on the risk potential energy vector and the transfer path impedance coefficient, and the geographic coordinates of the nodes are obtained from the multi-source business data to construct regional spatial distance constraints. A time-series dynamic flow tensor is generated based on the regional spatial distance constraint and the flow resistance parameter. Within a preset three-dimensional mapping domain, a hazardous waste risk evolution spatiotemporal cube is constructed based on the temporal dynamic flow tensor, and the hazardous waste risk evolution spatiotemporal cube is subjected to voxel density segmentation to obtain high-risk clusters. Based on the high-risk clusters, a risk transmission topology map is constructed, and the network centrality algorithm is used to analyze the risk transmission topology map to obtain the risk cascading critical path; The centroid displacement of the high-risk clusters in the time series is traced sequentially according to the risk cascade critical path to extract the risk hotspot drift vector. The risk hotspot drift vector is then corrected by topological constraints and extrapolated by temporal inertia to obtain a dual-drive evolution prediction model. The regional risk evolution is calculated based on the dual-drive evolution prediction model to obtain the prediction results of the regional hazardous waste risk pattern.

2. The method for predicting regional hazardous waste risk patterns based on spatial clustering and temporal evolution according to claim 1, characterized in that, The process involves acquiring multi-source service data for the target area and extracting attributes from nodes within the target area based on the multi-source service data to obtain a node attribute set, including: The multi-source business data is parsed and deduplicated to separate electronic transfer form data and enterprise management plan filing data, and the target node is identified from the multi-source business data using the enterprise's unique identifier. Extract the waste generation time-series records of the target node within a preset historical period from the electronic transfer manifest data, and perform time decay weighted accumulation on the waste generation time-series records to obtain the production load attribute; The approved storage limit of the target node is parsed from the enterprise management plan filing data, and the real-time inventory of the target node at the current moment is extracted. The difference between the approved storage limit and the real-time inventory is calculated to obtain the storage balance attribute. The production load attribute and the storage reserve attribute are mapped to the corresponding target nodes to obtain the node attribute set.

3. The method for predicting regional hazardous waste risk patterns based on spatial clustering and temporal evolution according to claim 1, characterized in that, The expression for calculating the risk potential vector is: ； in, For node indexing; This is the node index used to retrieve the maximum value of all nodes; For nodes The production load attribute value; For nodes The storage balance attribute value; For nodes The risk potential vector; For nodes The risk potential energy modulus; This is for calculating the maximum value. To prevent extremely small positive numbers with a denominator of zero; This is a transpose operation; This refers to the 2-norm operation.

4. The method for predicting regional hazardous waste risk patterns based on spatial clustering and temporal evolution according to claim 1, characterized in that, The step of calculating the transfer path impedance coefficient between each node based on the historical transfer records extracted from the multi-source business data and the risk potential vector includes: The historical transfer records are analyzed to obtain the topological correlation matrix, and the topological correlation matrix is ​​subjected to reciprocal operation and normalization to obtain the basic channel resistance value characterizing the business stickiness between nodes. Obtain the risk potential energy vector of the output node as the starting point of the transfer and the risk potential energy vector of the input node as the ending point of the transfer, and calculate the corresponding magnitude difference and direction angle cosine value. Construct the inverse repulsive force function of the potential energy gradient based on the difference in modulus and the cosine of the angle between the directions, so as to calculate the potential energy resistance value. The transfer path impedance coefficient between each node is obtained by weighted summing of the basic channel resistance value and the potential energy resistance value. The expression for the reverse repulsion function is: ； in, For node indexing; For node indexing; This is the node index used to retrieve the maximum value of all nodes; This is the node index used to retrieve the maximum value of all nodes; For historical transfer records by nodes To the node The strength of the topological association formed by the number of transfers or the cumulative number of joint orders; For nodes The transfer path impedance coefficient; For nodes The risk potential vector; For nodes The risk potential vector; For nodes The risk potential energy modulus; For nodes The risk potential energy modulus; max is the maximum value operation; To prevent extremely small positive numbers with a denominator of zero; It is an exponential function; This is a transpose operation; This refers to the 2-norm operation.

5. The method for predicting regional hazardous waste risk patterns based on spatial clustering and temporal evolution according to claim 1, characterized in that, The step of determining the flow resistance parameters between nodes based on the risk potential energy vector and the transfer path impedance coefficient, and obtaining the node geographical coordinates from the multi-source business data to construct regional spatial distance constraints, includes: The latitude and longitude geographic coordinates of each node are parsed from the multi-source service data, and the spherical transmission distance of the latitude and longitude geographic coordinates between any two nodes is calculated to obtain the global distance matrix. Obtain the preset hazardous waste transportation limit radius, use the Gaussian decay function to perform spatial weighting calculation on the global distance matrix to obtain the distance weights, and reset the distance weights exceeding the limit radius to zero to obtain the regional spatial distance constraints; Calculate the gradient projection values ​​of the risk potential energy vectors of the starting node and the ending node in the direction of the transfer path, and substitute the gradient projection values ​​and the transfer path impedance coefficient into the preset flow dynamics equation for coupling solution. The flow resistance parameters between the nodes are obtained by outputting quantized values ​​through the flow dynamics equation; wherein, the flow resistance parameters are positively correlated with the transfer path impedance coefficient and negatively correlated with the gradient projection value. The calculation expression for the fluid dynamics equation is as follows: ； in, For node indexing; For node indexing; For nodes The decrease in potential energy; For nodes The risk potential energy modulus; For nodes The risk potential energy modulus; For nodes The transfer path impedance coefficient; For nodes The flow resistance parameter; max is the maximum value calculation; To prevent extremely small positive numbers with a denominator of zero.

6. The method for predicting regional hazardous waste risk patterns based on spatial clustering and temporal evolution according to claim 1, characterized in that, The step of generating a time-series dynamic flow tensor based on the regional spatial distance constraint and the flow resistance parameter includes: Construct a resistance-probability mapping function, input the flow resistance parameter into the resistance-probability mapping function to perform negative exponential decay operation, and obtain the basic flow probability matrix that characterizes the potential interaction possibility between nodes in an unconstrained state; Perform a Hadamard product operation on the basic flow probability matrix and the regional spatial distance constraint to obtain the spatially corrected flow matrix; Set the time window length and time step length, and construct a three-dimensional tensor structure containing the start node dimension, the end node dimension, and the time dimension based on the time window length and time step length; The spatially modified flow matrix is ​​used as a reference slice and mapped to each time step of the three-dimensional tensor structure. The rate of change of the node attribute set at different times is dynamically modulated to the slice values ​​of each time step to obtain the temporal dynamic flow tensor. The expression for the resistance-probability mapping function is: ； in, For node indexing; For node indexing; For nodes under no spatial distance constraints The basic turnover probability; For nodes The flow resistance parameters; It is an exponential function.

7. The method for predicting regional hazardous waste risk patterns based on spatial clustering and temporal evolution according to claim 1, characterized in that, Within a preset three-dimensional mapping domain, a hazardous waste risk evolution spatiotemporal cube is constructed based on the temporal dynamic flow tensor, and the hazardous waste risk evolution spatiotemporal cube is subjected to voxel-based density segmentation to obtain high-risk clusters, including: A three-dimensional mapping domain is constructed with geographical latitude and longitude as the planar coordinate axis and time series as the vertical coordinate axis, and the three-dimensional mapping domain is divided into standard voxel units with fixed length, width and height dimensions; The flow intensity value in the time-series dynamic flow tensor is analyzed, and the flow intensity value is projected as a weight onto the spatial coordinate point corresponding to the three-dimensional mapping domain; A spatiotemporal kernel density estimation algorithm is used to perform smooth interpolation calculations on all the standard voxel units to obtain the risk density value of each standard voxel unit, so as to construct the spatiotemporal cube of hazardous waste risk evolution; An adaptive threshold segmentation algorithm is used to calculate the global background noise threshold of the hazardous waste risk evolution spatiotemporal cube, and standard voxel units with risk density values ​​lower than the global background noise threshold are removed to obtain a high-risk retention voxel set; Three-dimensional connected component labeling is performed on the high-risk reserved voxel set to merge standard voxel units that are spatially adjacent and have continuous risk density values ​​into an independent connected region, thereby obtaining the high-risk cluster.

8. The method for predicting regional hazardous waste risk patterns based on spatial clustering and temporal evolution according to claim 1, characterized in that, Based on the high-risk clusters, a risk transmission topology graph is constructed, and the network centrality algorithm is used to analyze the risk transmission topology graph to obtain the risk cascading critical paths, including: The geometric centroids of each of the high-risk clusters are extracted as network topology nodes, and the cumulative flow throughput between each of the network topology nodes is calculated based on the temporal dynamic flow tensor. The cumulative value of the flow throughput is normalized to obtain the risk transmission probability matrix; Construct the risk transmission topology graph, wherein the vertices of the risk transmission topology graph are network topology nodes, and the directed edges of the risk transmission topology graph are non-zero elements in the risk transmission probability matrix; The weighted betweenness centrality algorithm is used to calculate the betweenness centrality of each network topology node in the risk transmission topology graph, and nodes with betweenness centrality higher than a preset threshold are identified. This yields key hub nodes. Starting from the key hub node, a maximum probability gradient search is performed in the risk transmission topology graph, and downstream nodes are connected sequentially along the direction where the risk transmission probability matrix value decreases the slowest, thus obtaining a potential risk chain; The potential risk chain is subjected to closed-loop detection and redundancy pruning to obtain the risk cascade critical path.

9. The method for predicting regional hazardous waste risk patterns based on spatial clustering and temporal evolution according to claim 1, characterized in that, The process involves sequentially tracing the centroid displacement of the high-risk clusters along the risk cascade critical path over time to extract risk hotspot drift vectors. These vectors are then subjected to topological constraint correction and temporal inertia extrapolation to obtain a dual-drive evolutionary prediction model, comprising: The spatial cross-section of the high-risk cluster at each time step is extracted, and the risk-weighted centroid of each spatial cross-section is calculated using a density-weighted algorithm. The risk-weighted centroids of adjacent time steps are then connected to obtain the original centroid displacement vector. The original centroid displacement vector is matched with the risk cascade critical path in the risk transmission topology map to obtain the risk hotspot drift vector; Obtain the out-degree adjacent edges of the risk hotspot drift vector on the risk transmission topology graph, and calculate the cosine similarity between the risk hotspot drift vector and each of the out-degree adjacent edges to determine the guiding constraint edges; Projecting the risk hotspot drift vector onto the guiding constraint edge yields the spatial topology correction component; A time-series sliding window is constructed, and the risk hotspot drift vector sequence of historical moments is extracted and input into a pre-trained long short-term memory network for temporal feature extraction to obtain the temporal inertial extrapolation component; An adaptive Kalman filter algorithm is used to assimilate the spatial topology correction component and the temporal inertial extrapolation component and estimate their states to generate the dual-drive evolution prediction model. The expression for the dual-drive evolution prediction model is: ； in, For time step index; For time steps The hotspot drift vector; For time steps The temporal inertial extrapolation component; For time steps Risk-weighted centroid coordinate vector of high-risk clusters; For time steps Risk-weighted centroid coordinate vector of high-risk clusters; For time steps The unit direction vector of the guiding constraint edge; For time steps The posterior state covariance matrix; The process noise covariance matrix; To observe the noise covariance matrix; This is a transpose operation; Invert a matrix.

10. A regional hazardous waste risk pattern prediction system based on spatial clustering and temporal evolution, characterized in that, include: The data acquisition and attribute extraction module is used to acquire multi-source business data of the target area, and extract attributes of the nodes in the target area based on the multi-source business data to obtain a node attribute set, wherein the nodes in the node attribute set have production load attributes and storage capacity attributes. The risk potential vector calculation module is used to calculate the risk potential vector of each node based on the coupling relationship between the production load attribute and the storage margin attribute of each node in the node attribute set. The transfer path impedance calculation module is used to calculate the transfer path impedance coefficient between each node based on the historical transfer records extracted from the multi-source service data and the risk potential vector. The circulation resistance and spatial constraint construction module is used to determine the circulation resistance parameters between each node based on the risk potential energy vector and the transfer path impedance coefficient, and to obtain the node geographical coordinates from the multi-source business data to construct regional spatial distance constraints. A time-series dynamic flow tensor generation module is used to generate a time-series dynamic flow tensor based on the regional spatial distance constraint and the flow resistance parameter. The spatiotemporal cube construction and clustering module is used to construct a hazardous waste risk evolution spatiotemporal cube based on the temporal dynamic flow tensor within a preset three-dimensional mapping domain, and to perform voxel density segmentation on the hazardous waste risk evolution spatiotemporal cube to obtain high-risk clusters. The risk topology parsing and path identification module is used to construct a risk transmission topology map based on the high-risk clusters, and to parse the risk transmission topology map using a network centrality algorithm to obtain the risk cascading critical paths; The dual-drive evolution prediction model construction module is used to sequentially track the centroid displacement of the high-risk clusters in the time series according to the risk cascade critical path, so as to extract the risk hotspot drift vector, and perform topological constraint correction and temporal inertial extrapolation on the risk hotspot drift vector to obtain the dual-drive evolution prediction model. The regional risk pattern prediction module is used to perform evolution calculations on regional risks based on the dual-drive evolution prediction model to obtain the regional hazardous waste risk pattern prediction results.

Images