Event prediction model construction method and device for multi-infrastructure system

Abstract

The invention discloses an event prediction model construction method and device for a multi-infrastructure system. The method comprises the following steps: acquiring basic data of each infrastructure in the multi-infrastructure system; performing analog simulation on the basic data of each infrastructure to obtain operation data of each infrastructure; coupling the space-time diagram network constructed by each infrastructure, and establishing a space-time diagram network model; and carrying out migration training on the space-time diagram network model according to the operation data and real-time data obtained by monitoring to obtain an event prediction model. The event prediction model constructed by the scheme of the invention can scientifically and accurately evaluate and predict the response and evolution of the multi-infrastructure system in the event, and helps to deal with urban disaster events.

Description

Technical Field

[0001] The present invention relates to the field of smart city management technology, and in particular to a method and device for constructing an event prediction model for a multi-infrastructure system. Background Art

[0002] Infrastructure is the core pillar for maintaining the normal operation and functionality of a city, and its performance is directly related to the safety and resilience of the city.

[0003] Under the impact of various events, the interconnected coupling relationships among different infrastructure systems (such as stormwater pipe networks, transportation networks, and power supply networks) have a crucial impact on the propagation path and impact range of disasters. This coupling relationship has a significant impact on the overall performance of infrastructure. In particular, when faced with the impact of disasters and other events, the interconnectedness between different systems can lead to cascading effects that exacerbate the impact of disasters and other events. However, existing simulation models often focus only on a single infrastructure system, ignoring the complex spatiotemporal coupling relationships between multiple infrastructure networks. This makes it difficult to scientifically and accurately assess and predict the response and evolution of multiple infrastructure systems during events, hindering the response and management of urban disaster events. Summary of the Invention

[0004] In response to the above-mentioned shortcomings, the present invention provides a method and device for constructing an event prediction model for a multi-infrastructure system. The provided event prediction model can scientifically and accurately evaluate and predict the response and evolution of the multi-infrastructure system in an event, and help to respond to and handle urban disaster events.

[0005] An embodiment of the present invention provides a method for constructing an event prediction model for a multi-infrastructure system, the method comprising: Collect basic data of each infrastructure in a multi-infrastructure system; Simulate the basic data of each infrastructure separately to obtain the operation data of each infrastructure; Couple the space-time graph networks constructed by various infrastructures to establish a space-time graph network model; The spatiotemporal graph network model is subjected to migration training based on the operating data and the real-time data obtained by monitoring to obtain an event prediction model.

[0006] Preferably, collecting basic data of each infrastructure in the multi-infrastructure system includes: Collect infrastructure network data, urban land use data, demographic data, building data, meteorological data, and flood disaster data within the forecast area as basic data for urban facilities; Collect pipeline attributes, inspection well information, urban river information, and discharge outlet location information as basic data for the stormwater pipe network; collect grid structure information, electricity price information, population energy demand pattern information, building energy efficiency data, and distributed energy system information as basic data for the power supply network; Collect road network information, vehicle information, public transportation information, and traffic detection point information as basic data for the transportation network; The collected basic data of urban facilities, basic data of rainwater pipe network, basic data of power supply network and basic data of transportation network are standardized to obtain the basic data of various infrastructures.

[0007] Preferably, the basic data of each infrastructure is simulated separately to obtain the operation data of each infrastructure, including: Determine the input data based on the basic data of the stormwater pipe network, use the preset stormwater runoff model to simulate the stormwater pipe network and urban river network, and obtain the corresponding output data; Determine the input data based on the basic data of the traffic network, use the preset multi-agent traffic simulation model to simulate the traffic network, and obtain the corresponding output data; Determine input data based on the basic data of the power supply network, simulate the power supply network using a preset urban energy analysis model, and obtain corresponding output data; The obtained output data and the corresponding input data are used as the operation data of the rainwater pipe network.

[0008] Furthermore, the input data of the storm runoff model includes rainfall data, topographic data, land use data, pipe network structure data, river data and initial condition data, and the corresponding output data includes pipe flow, pipe water level and surface water depth; The input data of the multi-agent traffic simulation model includes road network data, population travel data, vehicle information, public transportation data and road water accumulation information, and the corresponding output data includes traffic flow and speed distribution; The input data of the urban energy analysis model include building energy consumption data, power supply network data, energy prices, population energy demand pattern information, distributed energy system information and road waterlogging information, and the corresponding output data include power supply capacity and energy consumption.

[0009] Preferably, the spatiotemporal graph networks constructed by various infrastructures are coupled to establish a spatiotemporal graph network model, including: modeling the graph structures according to the spatial locations of the stormwater pipe network, the transportation network, and the power supply network, and establishing a surface elevation network of facility coupling according to the surface elevation data; Constructing a data embedding layer for encoding data in the graph network and the surface elevation network; Constructing a temporal learning layer based on a preset time series model to capture the temporal dependency of the data embedding layer, and constructing a spatial learning layer based on a preset self-attention mechanism to capture the spatial topological association of the data embedding layer; A fully connected layer is constructed to map the spatiotemporal features output by the temporal learning layer and the spatial learning layer into prediction indicators, thereby completing the construction of the spatiotemporal graph network model.

[0010] Furthermore, the graph structure is modeled based on the spatial locations of the stormwater pipe network, transportation network, and power supply network, and a facility-coupled surface elevation network is established based on the surface elevation data, including: A graph network of the rainwater pipe network is constructed, using the spatial locations of inspection wells, overflow outlets, water outlets, pipe intersections, and river intersections of the rainwater pipe network as nodes, rainwater pipe and river connection nodes as edges in the graph, and parameters of each facility in the rainwater pipe network as attributes of the nodes and edges. A graph network of the transportation network is constructed using intersections of transportation road networks, transfer points between transportation networks and public transportation, transfer points for public transportation, urban land, and entrances and exits of urban land as nodes; roads, public transportation lines, and lines connecting the center of urban land to the entrances and exits as edges in the graph; and parameters of each facility in the transportation network as attributes of the nodes and edges. A graph network of the power supply network is constructed using power facilities and urban land as nodes, power lines as edges, and parameters of each facility in the power supply network as attributes of the nodes and edges. The surface elevation data is divided into watershed areas, and each watershed area is used as a node in the graph. Connections are established with other facilities in other networks at the spatial location of the watershed area to construct the surface elevation network.

[0011] Preferably, constructing a data embedding layer for encoding the data in the graph network and the surface elevation network comprises: Constructing an attribute feature embedding layer for encoding attribute data in the graph network and the surface elevation network into a vector space; constructing a time period embedding layer for encoding time features in the graph network and the surface elevation network; A fully connected layer is used to embed the time series data segmented according to time in the graph network and the surface elevation network as a spatiotemporal adaptive embedding layer.

[0012] Preferably, the spatiotemporal graph network model is subjected to migration training based on the operation data and the real-time data obtained by monitoring to obtain an event prediction model, including: The running data is used as a simulation data set, and divided into a training set and a validation set according to a preset ratio; Set the initial hyperparameters for transfer training, including learning rate, batch size, network depth, and number of attention heads; Using the spatiotemporal graph network model to learn the spatiotemporal correlation rules of the training set, and detecting whether the prediction error of the spatiotemporal graph network model meets the preset accuracy requirements based on the validation set; When the accuracy requirements are not met, the network structure or hyperparameters of different network layers in the spatiotemporal graph network model are adjusted, retrained, and the prediction error of the spatiotemporal graph network model is retested to see whether it meets the preset accuracy requirements; when the accuracy requirements are met, the network structure and hyperparameters of the current network layer are saved as the initial event prediction model; real-time data of various infrastructures are obtained through sensors deployed in the city as a real-time training set, and the initial event prediction model is iteratively trained for a preset number of times through a reset learning rate to obtain the event prediction model.

[0013] Furthermore, detecting whether the prediction error of the spatiotemporal graph network model meets the preset accuracy requirement based on the validation set includes: Calculating a prediction value using the spatiotemporal graph network model based on the input data in the validation set, wherein the prediction value indicators include predicted flow, predicted water depth, predicted traffic speed, and predicted power load; Calculate the difference between the predicted value of the spatiotemporal graph network model and the corresponding indicator of the output data in the validation set based on a preset indicator measurement model to obtain flow error, water depth error, traffic speed error, and power load error respectively; When the flow error, the water depth error, the traffic speed error, or the power load error does not meet the corresponding set tolerance range, it is determined that the accuracy requirement is not met; When the flow error, the water depth error, the traffic speed error, and the power load error all fall within corresponding set tolerance ranges, it is determined that the accuracy requirement is met.

[0014] An embodiment of the present invention further provides an event prediction model construction device for a multi-infrastructure system, the device comprising: a collection module for collecting basic data of each infrastructure in the multi-infrastructure system; The simulation module is used to simulate the basic data of each infrastructure separately to obtain the operation data of each infrastructure; the coupling module is used to couple the space-time graph network constructed by each infrastructure to establish a space-time graph network model; The training module is used to perform migration training on the spatiotemporal graph network model based on the operating data and the real-time data obtained by monitoring to obtain an event prediction model.

[0015] The present invention provides a method and device for constructing an event prediction model for a multi-infrastructure system. The method and device collect basic data from each infrastructure in the multi-infrastructure system; simulate the basic data of each infrastructure separately to obtain the operational data of each infrastructure; couple the spatiotemporal graph networks constructed for each infrastructure to establish a spatiotemporal graph network model; and perform migration training on the spatiotemporal graph network model based on the operational data and real-time data obtained through monitoring to obtain an event prediction model. The event prediction model constructed by the present application scheme can scientifically and accurately evaluate and predict the response and evolution of the multi-infrastructure system during an event, and assist in the response and handling of urban disaster events. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 1 is a flow chart of a method for constructing an event prediction model for a multi-infrastructure system according to an embodiment of the present invention; Figure 2 It is a structural diagram of an event prediction model building device for a multi-infrastructure system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0017] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0018] Urban infrastructure systems are characterized by widespread spatial distribution, complex network topologies, and significant cross-system interactions. For example, in stormwater networks, runoff from rainfall affects water levels and flow, indirectly impacting transportation and power systems through accumulated water. In transportation networks, accumulated water blocks roads, reducing traffic capacity and further impacting logistics and residents' lives. In power systems, water damage to equipment can cause localized or network-wide power outages, exacerbating the failure of other infrastructure.

[0019] How to accurately describe these network effects and reveal the multi-level coupling and co-evolution mechanisms between networks is an urgent problem that needs to be solved in event prediction and risk management.

[0020] See also Figure 1 , is a flow chart of a method for constructing an event prediction model for a multi-infrastructure system according to an embodiment of the present invention, the method comprising steps S1 to S4: Step S1, collecting basic data of each infrastructure in the multi-infrastructure system; Step S2, simulating the basic data of each infrastructure to obtain the operation data of each infrastructure; Step S3, coupling the spatiotemporal graph networks constructed by each infrastructure to establish a spatiotemporal graph network model; Step S4: performing migration training on the spatiotemporal graph network model based on the operating data and the real-time data obtained by monitoring to obtain an event prediction model.

[0021] During the specific implementation of this embodiment, urban disaster management requires a model with broad applicability. However, the complexity of the urban environment and the limitations of monitoring resources pose the following challenges to model development: 1) High difficulty in data acquisition: In real cities, multi-source data (such as rainwater pipe network flow, power consumption, traffic flow, etc.) often have problems such as missing, heterogeneous, and inconsistent. It is difficult to fully describe the system behavior by directly relying on monitoring data; 2) The data scale is huge: City-level data has high temporal and spatial resolution and a complex network structure. Training models directly based on monitoring data is expensive and computationally intensive; 3) Insufficient model migration adaptability: Existing models usually perform well in specific scenarios, but lack cross-regional or dynamic adaptability, making them difficult to be effectively applied to actual disaster management.

[0022] To address the above challenges, it is urgent to develop a transfer learning method based on the combination of simulation data and monitoring data. First, basic data of each infrastructure in the multi-infrastructure system is collected.

[0023] It should be noted that infrastructure may include rainwater pipe networks, power supply networks, transportation networks, etc. By coupling rainwater pipe networks, power supply networks, and transportation networks, it is helpful to resist extreme flood disasters in cities.

[0024] Each infrastructure is modeled based on the data simulation model of each single infrastructure to generate infrastructure operation data.

[0025] After obtaining the operating data of each infrastructure, the space-time graph network constructed by each infrastructure is coupled to establish a space-time graph network model. Based on the method of coupling modeling of multiple infrastructure systems (such as stormwater pipe networks, transportation networks and power supply networks), the spatial distribution characteristics and dynamic interaction processes of multiple systems are incorporated into a unified space-time graph network framework.

[0026] The spatiotemporal graph network model aims to capture the inter-coupling and evolutionary characteristics of urban infrastructure systems in the spatiotemporal dimensions. Because city-level multi-network data is extremely large and exhibits significant spatiotemporal heterogeneity, it is often difficult to train a high-precision model in one go using only real monitoring data. Therefore, the spatiotemporal graph network model is transferred and trained based on the operational data and real-time data obtained from monitoring. This introduces the idea of ​​transfer learning between simulated data and real-time monitored data to generate an event prediction model, which can quickly improve the model's adaptability and robustness to complex scenarios.

[0027] This application dynamically models the interconnected effects of infrastructure systems under disaster impacts, revealing the spatiotemporal coupling between these systems and providing a deeper understanding of the mechanisms by which disasters impact multiple urban systems. The resulting event prediction model can scientifically and accurately assess and predict the response and evolution of multiple infrastructure systems during events. This not only provides a scientific basis for urban infrastructure planning and construction, but also supports disaster emergency management and the optimization of disaster resilience, enhancing the overall resilience of cities and the efficiency of resource allocation.

[0028] In another embodiment of the present invention, when basic data is collected in step S1, the following steps are specifically included: First, basic data on urban facilities are collected, specifically: infrastructure network data within the study area, urban land use data (land use classification, building density, green space distribution), demographic data (population density, age structure, employment situation, travel survey data, daily activity patterns), building data, meteorological data, economic data, and flood disaster data.

[0029] Collect basic data on the rainwater pipe network, including pipe attributes (diameter, slope, starting and ending elevations, material, and Manning roughness coefficient), inspection wells (location coordinates, elevation, type), urban rivers (cross-section shape, slope, elevation, material, and roughness coefficient), and discharge outlet locations. Collect basic data on the power supply network. For the power supply network, it is necessary to be detailed to the grid structure (the location of substations, transmission lines, distribution lines and the corresponding equipment parameters, electricity price information, population energy demand patterns (electricity load curves of residential, commercial and industrial users), building energy efficiency data (energy consumption level of buildings, power of electrical equipment), and distributed energy systems (installed capacity and distribution of solar energy).

[0030] Collect basic data of the transportation network. For the transportation network, it is necessary to be detailed to the road network (road grade, number of lanes, width), vehicle information (vehicle type, quantity, price, fuel type and price), public transportation (bus routes, stops, schedules, operating hours), and traffic monitoring points (location and data of traffic flow monitoring equipment).

[0031] It should be noted that in this embodiment, the rainwater pipe network, power supply network, and transportation network are used as infrastructure. In other embodiments, the infrastructure can be installed and adjusted according to predicted demand, and basic data of other infrastructure can also be used.

[0032] The acquired multi-source data is then standardized through: 1) Invalid data removal: Missing, duplicate, or abnormal data is checked and removed according to data specifications. 2) Spatial georeferencing: All spatial data is aligned to the same coordinate system (such as CGCS2000) to ensure spatial consistency across different datasets. 3) Attribute data standardization: Attribute field names, units, and formats are standardized across datasets to ensure data consistency and comparability.

[0033] By constructing a spatiotemporal graph network model that integrates geographic information (such as surface elevation and watershed) with network topological characteristics (such as node characteristics and edge weights), and dynamically modeling the functional loss of nodes (such as buildings, power stations, and traffic intersections) and the conduction effect of edges during the propagation of disasters, we can better reveal the multi-level coupling and co-evolution mechanisms between networks.

[0034] In another embodiment provided by the present invention, the step S2 specifically includes the following steps: Generate infrastructure operation data based on each single infrastructure data simulation model.

[0035] When simulating the stormwater pipe network and the urban river network, the input data is determined based on the basic data of the stormwater pipe network. The stormwater runoff model is used to simulate the stormwater pipe network and the urban river network to obtain the corresponding output data. It should be noted that the storm runoff model includes simulation models such as Mike Floor.

[0036] When conducting traffic network simulation, the input data is determined based on the basic data of the traffic network, and the traffic network is simulated using a multi-agent traffic simulation model to establish the relationship between road water depth and traffic free flow.

[0037] It should be noted that the multi-agent traffic simulation model includes simulation models such as MATSim.

[0038] When simulating the power supply network, the input data is determined based on the basic data of the power supply network. The power supply network is simulated using the urban energy analysis model to establish the relationship between the depth of water accumulation on land or roads and the switches of power supply facilities, and the corresponding output data is obtained. It should be noted that the city energy analysis model includes simulation tool models such as City Energy Analyst.

[0039] The obtained output data and the corresponding input data are used as the operating data of the rainwater pipe network, that is, the input data and output data of the storm runoff model, the input data and output data of the multi-agent traffic simulation model, and the input data and output data of the urban energy analysis model are used as the operating data of the rainwater pipe network.

[0040] In another embodiment provided by the present invention, when simulating a rainwater pipe network and an urban river network, the input data of the stormwater runoff model used includes rainfall data, topographic data, land use data, pipe network structure data, river data, and initial condition data, and the corresponding output data includes pipe flow, pipe water level, and surface water depth; Among them, rainfall data includes the spatiotemporal distribution data of designed rainstorm processes or historical rainfall events. Terrain data, that is, the surface elevation model (DEM), is used to determine the surface runoff path. Land use data, including the underlying surface parameters of different land use types, such as permeability and roughness coefficient. Pipeline network structure data, including the spatial position and attribute parameters of pipelines, inspection wells, overflows, etc. River data, including the spatial position, cross-sectional morphology and attribute parameters of the river. Initial conditions include the initial water level of the pipeline network, the initial soil moisture content, etc. The pipeline flow and water level in the output data, that is, the flow and water level changes in the pipeline at each time step. Surface water depth, that is, the change of surface water depth in a specific area over time.

[0041] The input data of the multi-agent traffic simulation model includes road network data, population travel data, vehicle information, public transportation data, and road water accumulation information, and the corresponding output data includes traffic flow and speed distribution.

[0042] Among them, road network data includes road node and link information, the length, capacity, free flow speed, number of lanes, and allowed traffic modes of each link. Population travel data includes each agent's activity plan, including activity type, start and end time, location, and travel mode selection. Vehicle information includes vehicle type, quantity, price, fuel type and price, etc. Public transportation data includes bus routes, stops, schedules, operating hours, etc. Road waterlogging information includes road waterlogging depth data obtained from stormwater pipe network simulation. The traffic flow in the output data includes the traffic flow of each link in the road network at each time step. The speed distribution includes the average vehicle speed of each link in the road network at each time step.

[0043] The input data of the urban energy analysis model includes building energy consumption data, power supply network data, energy prices, population energy demand pattern information, distributed energy system information, and road waterlogging information. The corresponding output data includes power supply capacity and energy consumption.

[0044] Building energy consumption data includes the energy consumption characteristics of each building, including building type, area, power consumption of electrical equipment, and energy efficiency rating. Power supply network data includes the location, capacity, and connectivity of substations, transmission lines, and distribution lines. Energy price information, namely electricity price information for different time periods. Population energy demand pattern information, including electricity load curves for residential, commercial, and industrial users. Distributed energy system information, including the installed capacity and distribution of solar and wind power. Road waterlogging information, namely road waterlogging depth data obtained from stormwater network simulations, is used to assess the impact on power facilities.

[0045] The output data includes power supply capacity, which refers to the power supply capacity of the power supply network facilities. Energy consumption, which refers to the total amount and distribution of energy consumption by various users.

[0046] Based on the integration of the input data and simulation outputs of each model, the resulting spatiotemporal dataset includes five types of data: meteorological data, namely rainfall data, flow and water level data of stormwater pipes and urban river networks, flow and speed data of transportation networks, power supply capacity and user consumption data of the power supply network, and surface water accumulation data.

[0047] In another embodiment provided by the present invention, the step S3 specifically includes the following steps: Modeling of the spatiotemporal graph network is carried out based on various infrastructures, including the design of the graph structure and the design of the spatiotemporal graph network structure.

[0048] When designing the graph structure, the graph structure is modeled based on the spatial locations of the stormwater pipe network, transportation network, and power supply network, and a facility-coupled surface elevation network is established based on surface elevation data. After the coupled infrastructure network graph is constructed, the spatiotemporal graph structure is designed, and a suitable graph neural network structure is designed to learn the dynamic spatiotemporal processes occurring on the graph.

[0049] A data embedding layer is constructed to encode the data in the graph network and the surface elevation network.

[0050] Construct a temporal learning layer for the spatiotemporal graph network. The spatiotemporal graph network learning layer is divided into a temporal learning layer and a spatial learning layer. A temporal learning layer is constructed by using a Transformer or LSTM-based temporal model to capture temporal dependencies. A spatial learning layer is constructed by using a GCN (Graph Convolutional Network) or a Transformer-based self-attention mechanism to capture spatial topological associations.

[0051] A fully connected layer is constructed to map the spatiotemporal features output by the temporal learning layer and the spatial learning layer into prediction indicators as the output of the prediction layer, thereby completing the construction of the spatiotemporal graph network model.

[0052] This application innovatively proposes a method for modeling the coupling of multiple infrastructure systems (such as stormwater networks, transportation networks, and power supply networks) by incorporating the spatial distribution characteristics and dynamic interactions of these systems into a unified spatiotemporal graph network framework. This method integrates the node and edge characteristics of each system, accurately depicting the complex relationships between them, and providing an effective tool for collaborative multi-network analysis.

[0053] In another embodiment provided by the present invention, when modeling a graph structure and establishing a facility-coupled surface elevation network based on surface elevation data, the following steps are specifically included: For the stormwater pipe and urban river network, the spatial locations of the stormwater system's manholes, overflows, outlets, pipe intersections, and river intersections are abstracted as nodes. The stormwater pipe and river connections are used as edges in the graph, and the parameters of each facility are used as node and edge attributes to construct a graph network for the stormwater pipe network.

[0054] For the transportation network, nodes are abstracted as intersections of transportation road networks, transfer points between transportation networks and public transportation, transfer points between public transportation, urban land, and entrances and exits of urban land. The edges of the graph are formed by connecting roads, public transportation routes, and lines from the center of urban land to entrances and exits, and the parameters of each facility are used as attributes of nodes and edges to construct a graph network of the transportation network. For the power supply network, power facilities and urban land are abstracted as a node. The power line is used as an edge in the graph, and the parameters of each facility are used as the attributes of the node and edge to construct the graph network of the power supply network. For the surface elevation network, the surface elevation data is divided into watersheds, and each watershed is used as a point in the diagram. Connections are established with other facilities (points) in the spatial location of the watershed in other networks. This connection is the coupling connection between different infrastructure networks to construct the surface elevation network.

[0055] In another embodiment provided by the present invention, the constructed data embedding layer includes an attribute feature embedding layer, a time period embedding layer (daily, weekly, yearly) embedding layer, and a spatio-temporal adaptive embedding layer.

[0056] The attribute data of the nodes in the graph network and the surface elevation network are encoded into a vector space through an attribute feature embedding layer.

[0057] By constructing a time period embedding layer, the time features of the graph network and the surface elevation network, such as daily period, weekly period, and annual period, are encoded.

[0058] By constructing a spatiotemporal adaptive embedding layer: a fully connected layer (FC(·)) is used to embed the simulated time series data segmented according to the time period.

[0059] This study dynamically models the interconnected effects of disasters on infrastructure systems, revealing the spatiotemporal coupling between these systems and providing a deeper understanding of the mechanisms by which disasters impact multiple urban systems. This capability not only provides a scientific basis for urban infrastructure planning and construction, but also supports disaster emergency management and the optimization of disaster resilience, thereby enhancing overall urban resilience and resource allocation efficiency.

[0060] In another embodiment provided by the present invention, the step S4 specifically includes the following steps: The spatiotemporal graph network model aims to capture the inter-coupling and evolutionary characteristics of urban infrastructure systems across time and space. Because city-level multi-network data is extremely large and exhibits significant spatiotemporal heterogeneity, it is often difficult to train a high-precision model using only real monitoring data. Therefore, the introduction of transfer learning (transfer learning) from simulated data to monitored data can rapidly improve the model's adaptability and robustness to complex scenarios.

[0061] Data preparation includes simulation datasets: generated from simulation results of stormwater pipe networks, traffic networks, and power supply networks. Monitoring datasets include real-time data acquired by various sensors and real-time monitoring systems (flow meters, traffic detectors, electricity meters, etc.) at corresponding nodes.

[0062] Using simulated data, we can initially learn the spatiotemporal correlation patterns of urban stormwater pipe networks, transportation networks, and power supply networks; learn basic spatiotemporal characteristics (such as coupling relationships between nodes and time series patterns) to lay a solid foundation for subsequent migration training.

[0063] The training process includes: Data partitioning: Divide the simulated data into a training set and a validation set according to a preset ratio based on time, and set aside a portion as a test set.

[0064] It should be noted that the preset ratio can be set or adjusted according to actual conditions, and can be set to 7:2:1.

[0065] Set the initial hyperparameters for transfer training, including learning rate, batch size, network depth, number of attention heads (if using the Transformer-based spatiotemporal learning module), etc.

[0066] Using the spatiotemporal graph network model to learn the spatiotemporal correlation rules of the training set, and detecting whether the prediction error of the spatiotemporal graph network model meets the preset accuracy requirements based on the validation set; When the accuracy requirements are not met, the network structure or hyperparameters of different network layers in the spatiotemporal graph network model are adjusted, retrained, and the prediction error of the spatiotemporal graph network model is retested to see whether it meets the preset accuracy requirements; when the accuracy requirements are met, the network structure and hyperparameters of the current network layer are saved as the initial event prediction model, that is, an initial event prediction model with good prediction ability in the simulation data domain is obtained.

[0067] Real-time data of various infrastructures are obtained by sensors deployed in the city as a real-time training set. The initial event prediction model is iteratively trained for a preset number of times by resetting the learning rate to obtain the event prediction model.

[0068] The model is fine-tuned in stages. First, the first few layers of the spatiotemporal graph network (such as the embedding layer and some spatiotemporal learning layers) are "frozen," and only the higher layers (such as the last few layers or the prediction layer) are trained to avoid completely resetting the model parameters. Then, the model structure and parameters pre-trained on simulated data are retained, and the monitoring data is used as a new training set. Multiple rounds of iterations are performed with a small learning rate, allowing the model to retain existing knowledge while being modified for real-world scenarios.

[0069] It should be noted that this case can also be based on the verification and testing of the time prediction model. If the accuracy of the verification set continues to improve, it means that the model is gradually adapting to the real scenario; if over-training causes good results on monitoring data but fails on simulation data or another part of monitoring data, it means that the scale of monitoring data needs to be increased.

[0070] This invention significantly reduces reliance on large-scale monitoring data by combining simulated and monitored data and employing transfer learning techniques. By initially training the model on simulated data and then fine-tuning and optimizing it with limited monitored data, the model not only achieves high prediction accuracy but also significantly reduces the economic and time costs of data collection and processing.

[0071] In another embodiment provided by the present invention, when performing accuracy verification, the following process is specifically included: The spatiotemporal graph network model is used to calculate a prediction value based on the input data in the validation set, and the indicators of the prediction value include predicted flow, predicted water depth, predicted traffic speed and predicted power load.

[0072] Calculating the difference between the predicted value of the spatiotemporal graph network model and the corresponding indicator of the output data in the validation set based on a preset indicator measurement model; The indicator measurement model includes one of the following: in, is the predicted value, y i is the true value, and N is the number of samples.

[0073] The root mean square error (RMSE) is sensitive to outliers and is mainly used to measure the overall level of prediction error.

[0074] The mean absolute error (MAE) is used to average the absolute errors, reflecting the average deviation, and is often used to measure the absolute degree of prediction deviation.

[0075] The average accuracy (MAP) can reflect the proportion of the prediction error on the true value scale by relativizing the error.

[0076] By monitoring the performance of the model in real time through the validation set during the training process (such as flow prediction error, water depth prediction error, traffic speed prediction error, power load prediction error, etc.), if the performance does not meet expectations, the network structure or hyperparameters need to be adjusted; this can improve the accuracy of the model.

[0077] When the flow error, the water depth error, the traffic speed error, or the power load error does not meet the corresponding set tolerance range, it is determined that the accuracy requirement is not met; When the flow error, the water depth error, the traffic speed error, and the power load error all fall within corresponding set tolerance ranges, it is determined that the accuracy requirement is met.

[0078] It should be noted that the indicator measurement model specified in this embodiment is only a preferred embodiment. In other embodiments, the indicator error can be calculated using other measurement models.

[0079] After migration training, the model can better adapt to the dynamic changes of the real world. At this time, the model can be deployed to the city operation management platform or cloud, accessing real-time or regularly updated monitoring data. The embodiment of the present invention also provides an event prediction model construction device for a multi-infrastructure system, see Figure 2 , is a schematic diagram of the structure of a device for constructing an event prediction model for a multi-infrastructure system according to an embodiment of the present invention, the device comprising: A collection module, used to collect basic data of each infrastructure in a multi-infrastructure system; The simulation module is used to simulate the basic data of each infrastructure separately to obtain the operation data of each infrastructure; the coupling module is used to couple the space-time graph network constructed by each infrastructure to establish a space-time graph network model; The training module is used to perform migration training on the spatiotemporal graph network model based on the operating data and the real-time data obtained by monitoring to obtain an event prediction model.

[0080] It should be noted that the event prediction model construction device for the multi-infrastructure system provided in the embodiment of the present invention can execute the event prediction model construction method for the multi-infrastructure system described in any of the above embodiments. The specific functions of the event prediction model construction device for the multi-infrastructure system will not be described here.

[0081] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for constructing an event prediction model for a multi-infrastructure system, characterized in that: The method comprises: Collect basic data of each infrastructure in a multi-infrastructure system; Simulate the basic data of each infrastructure separately to obtain the operation data of each infrastructure; Couple the space-time graph networks constructed by various infrastructures to establish a space-time graph network model; The spatiotemporal graph network model is subjected to migration training based on the operating data and the real-time data obtained by monitoring to obtain an event prediction model.

2. The method for constructing an event prediction model for a multi-infrastructure system according to claim 1, wherein: Collect basic data of each infrastructure in a multi-infrastructure system, including: Collect infrastructure network data, urban land use data, demographic data, building data, meteorological data, and flood disaster data within the forecast area as basic data for urban facilities; Collect pipeline attributes, inspection well information, urban river information, and discharge outlet location information as basic data for the stormwater pipe network; collect grid structure information, electricity price information, population energy demand pattern information, building energy efficiency data, and distributed energy system information as basic data for the power supply network; Collect road network information, vehicle information, public transportation information, and traffic detection point information as basic data for the transportation network; The collected basic data of urban facilities, basic data of rainwater pipe network, basic data of power supply network and basic data of transportation network are standardized to obtain the basic data of various infrastructures.

3. The method for constructing an event prediction model for a multi-infrastructure system according to claim 1, wherein: Simulate the basic data of each infrastructure separately to obtain the operation data of each infrastructure, including: Determine the input data based on the basic data of the stormwater pipe network, use the preset stormwater runoff model to simulate the stormwater pipe network and urban river network, and obtain the corresponding output data; Determine the input data based on the basic data of the traffic network, use the preset multi-agent traffic simulation model to simulate the traffic network, and obtain the corresponding output data; Determine input data based on the basic data of the power supply network, simulate the power supply network using a preset urban energy analysis model, and obtain corresponding output data; The obtained output data and the corresponding input data are used as the operation data of the rainwater pipe network.

4. The method for constructing an event prediction model for a multi-infrastructure system according to claim 3, wherein: The input data of the stormwater runoff model includes rainfall data, terrain data, land use data, pipe network structure data, river data and initial condition data, and the corresponding output data includes pipeline flow, pipeline water level and surface water depth; the input data of the multi-agent traffic simulation model includes road network data, population travel data, vehicle information, public transportation data and road water accumulation information, and the corresponding output data includes traffic flow and speed distribution; The input data of the urban energy analysis model include building energy consumption data, power supply network data, energy prices, population energy demand pattern information, distributed energy system information and road waterlogging information, and the corresponding output data include power supply capacity and energy consumption.

5. The method for constructing an event prediction model for a multi-infrastructure system according to claim 1, wherein: The spatiotemporal graph networks constructed by various infrastructures are coupled to establish a spatiotemporal graph network model, including: Model the graph structure of the stormwater pipe network, transportation network, and power supply network based on their spatial locations, and establish a facility-coupled surface elevation network based on surface elevation data. Constructing a data embedding layer for encoding data in the graph network and the surface elevation network; Constructing a temporal learning layer based on a preset time series model to capture the temporal dependency of the data embedding layer, and constructing a spatial learning layer based on a preset self-attention mechanism to capture the spatial topological association of the data embedding layer; A fully connected layer is constructed to map the spatiotemporal features output by the temporal learning layer and the spatial learning layer into prediction indicators, thereby completing the construction of the spatiotemporal graph network model.

6. The method for constructing an event prediction model for a multi-infrastructure system according to claim 5, wherein: The graph structures of the stormwater pipe network, transportation network, and power supply network are modeled based on their spatial locations. A facility-coupled surface elevation network is established based on surface elevation data, including: A graph network of the rainwater pipe network is constructed, using the spatial locations of inspection wells, overflow outlets, water outlets, pipe intersections, and river intersections of the rainwater pipe network as nodes, rainwater pipe and river connection nodes as edges in the graph, and parameters of each facility in the rainwater pipe network as attributes of the nodes and edges. A graph network of the transportation network is constructed using intersections of transportation road networks, transfer points between transportation networks and public transportation, transfer points for public transportation, urban land, and entrances and exits of urban land as nodes; roads, public transportation lines, and lines connecting the center of urban land to the entrances and exits as edges in the graph; and parameters of each facility in the transportation network as attributes of the nodes and edges. A graph network of the power supply network is constructed using power facilities and urban land as nodes, power lines as edges, and parameters of each facility in the power supply network as attributes of the nodes and edges. The surface elevation data is divided into watershed areas, and each watershed area is used as a node in the graph. Connections are established with other facilities in other networks at the spatial location of the watershed area to construct the surface elevation network.

7. The method for constructing an event prediction model for a multi-infrastructure system according to claim 5, wherein: Constructing a data embedding layer for encoding the data in the graph network and the surface elevation network, including: Constructing an attribute feature embedding layer for encoding attribute data in the graph network and the surface elevation network into a vector space; constructing a time period embedding layer for encoding time features in the graph network and the surface elevation network; A fully connected layer is used to embed the time series data segmented according to time in the graph network and the surface elevation network as a spatiotemporal adaptive embedding layer.

8. The method for constructing an event prediction model for a multi-infrastructure system according to claim 1, wherein: The spatiotemporal graph network model is subjected to migration training based on the operation data and the real-time data obtained by monitoring to obtain an event prediction model, including: The running data is used as a simulation data set, and divided into a training set and a validation set according to a preset ratio; Set the initial hyperparameters for transfer training, including learning rate, batch size, network depth, and number of attention heads; Using the spatiotemporal graph network model to learn the spatiotemporal correlation rules of the training set, and detecting whether the prediction error of the spatiotemporal graph network model meets the preset accuracy requirements based on the validation set; When the accuracy requirements are not met, the network structure or hyperparameters of different network layers in the spatiotemporal graph network model are adjusted, retrained, and the prediction error of the spatiotemporal graph network model is retested to see whether it meets the preset accuracy requirements; when the accuracy requirements are met, the network structure and hyperparameters of the current network layer are saved as the initial event prediction model; real-time data of various infrastructures are obtained through sensors deployed in the city as a real-time training set, and the initial event prediction model is iteratively trained for a preset number of times through a reset learning rate to obtain the event prediction model.

9. The method for constructing an event prediction model for a multi-infrastructure system according to claim 8, wherein: Detecting whether the prediction error of the spatiotemporal graph network model meets the preset accuracy requirement based on the validation set includes: Calculating a prediction value using the spatiotemporal graph network model based on the input data in the validation set, wherein the prediction value indicators include predicted flow, predicted water depth, predicted traffic speed, and predicted power load; Calculate the difference between the predicted value of the spatiotemporal graph network model and the corresponding indicator of the output data in the validation set based on a preset indicator measurement model to obtain flow error, water depth error, traffic speed error, and power load error respectively; When the flow error, the water depth error, the traffic speed error, or the power load error does not meet the corresponding set tolerance range, it is determined that the accuracy requirement is not met; When the flow error, the water depth error, the traffic speed error, and the power load error all fall within corresponding set tolerance ranges, it is determined that the accuracy requirement is met.

10. An event prediction model construction device for a multi-infrastructure system, characterized in that: The device comprises: A collection module, used to collect basic data of each infrastructure in a multi-infrastructure system; The simulation module is used to simulate the basic data of each infrastructure and obtain the operation data of each infrastructure; The coupling module is used to couple the spatiotemporal graph networks constructed by various infrastructures and establish a spatiotemporal graph network model; The training module is used to perform migration training on the spatiotemporal graph network model based on the operating data and the real-time data obtained by monitoring to obtain an event prediction model.

Images