A substation flood situation closed-loop treatment method and device based on digital twinning

Abstract

The application discloses a substation waterlogging closed-loop treatment method and device based on digital twinning, the method comprising: determining a waterlogging risk parameter according to environmental data, in-station water level data and drainage working condition data, inputting the waterlogging risk parameter into a pre-constructed waterlogging causal reasoning model for reasoning to obtain a waterlogging risk probability, a waterlogging risk level and a waterlogging causal path, obtaining at least one disaster-affected object in the substation, determining a disaster-affected state corresponding to each disaster-affected object according to the waterlogging risk probability and the waterlogging risk level, inputting the disaster-affected state into a pre-constructed digital twinning model of the substation for display, and generating a waterlogging treatment work order according to the disaster-affected state and the waterlogging causal path, obtaining receipt data of the waterlogging treatment work order, and taking the receipt data as a waterlogging closed-loop treatment result of the substation. The risk tracing capability is improved, the risk reasoning result is accurately mapped to a specific disaster-affected object in the substation, and a complete closed loop from risk information to emergency treatment is formed.

Description

Technical Field

[0001] This application belongs to the field of power facility maintenance, specifically relating to a closed-loop method and device for handling flooding in substations based on digital twins. Background Technology

[0002] As a critical hub in the power system, the safe and stable operation of substations directly affects the reliability of regional power supply. Currently, flood risk assessment of substations mainly relies on empirical judgment methods with fixed thresholds, issuing simple warnings by setting rainfall or water level warning lines, or using traditional statistical regression models and black-box machine learning models for prediction.

[0003] However, fixed threshold methods are ill-suited to the diverse environments and complex operating conditions of different substations, easily leading to missed or false alarms. While black-box models improve prediction accuracy to some extent, they fail to reveal the intrinsic relationships between disaster-causing factors, lack risk tracing capabilities, and result in a disconnect between risk information and on-site emergency response. Furthermore, multi-source heterogeneous data from meteorological, hydrological, and geographical sources, as well as substation water levels and drainage data, exhibit significant differences in sampling frequency, temporal granularity, spatial resolution, and data format, making comprehensive utilization difficult. Simultaneously, existing technologies often limit risk inference results to macro-level risk level outputs, failing to accurately map them to specific disaster-bearing objects within the substation. Summary of the Invention

[0004] The purpose of this application is to provide a closed-loop method and apparatus for handling flooding in substations based on digital twins.

[0005] In a first aspect, embodiments of this application provide a closed-loop substation flood situation handling method based on digital twins, the method comprising: The environmental data, water level data, and drainage condition data of the substation are obtained, and flood risk parameters are determined based on the environmental data, water level data, and drainage condition data. The flood risk parameters are input into a pre-built flood causal reasoning model for reasoning to obtain the flood risk probability, flood risk level, and flood causal path; At least one disaster-bearing object in the substation is identified, and the disaster-bearing status corresponding to the disaster-bearing object is determined according to the flood risk probability and the flood risk level. The disaster-bearing status is input into the pre-built digital twin model of the substation, and the disaster-bearing status corresponding to the at least one disaster-bearing object is displayed through a visual interface. A flood situation handling work order is generated based on the disaster-bearing status and the flood situation causal path. Obtain the receipt data of the flood control work order, and use the receipt data as the closed-loop handling result of the flood situation at the substation.

[0006] Secondly, embodiments of this application provide a digital twin-based substation flood control closed-loop management device, the device comprising: The parameter acquisition module is used to acquire environmental data, water level data and drainage condition data of the substation, and determine flood risk parameters based on the environmental data, water level data and drainage condition data. The risk reasoning module is used to input the flood risk parameters into a pre-built flood causal reasoning model for reasoning, and obtain the flood risk probability, flood risk level and flood causal path; The status acquisition module is used to acquire at least one disaster-bearing object in the substation and determine the disaster-bearing status of the disaster-bearing object according to the flood risk probability and the flood risk level. The work order generation module is used to input the disaster-bearing status into the pre-built digital twin model of the substation, display the disaster-bearing status corresponding to the at least one disaster-bearing object through a visual interface, and generate a flood handling work order based on the disaster-bearing status and the flood causal path. The result acquisition module is used to acquire the receipt data of the flood control work order and use the receipt data as the closed-loop handling result of the flood situation at the substation.

[0007] Thirdly, embodiments of this application provide an electronic device, including a processor, a memory, and a program or instructions stored in the memory and capable of running on the processor, wherein the program or instructions, when executed by the processor, implement the method described above.

[0008] Fourthly, a readable storage medium on which a program or instructions are stored, which, when executed by a processor, implement the method described above.

[0009] The embodiments of this application have the following advantages: In this embodiment, by acquiring environmental data, water level data, and drainage condition data of the substation, flood risk parameters are determined based on these data. This allows for the comprehensive utilization of multi-source heterogeneous data. Furthermore, by inputting the flood risk parameters into a pre-constructed flood causal reasoning model, the application obtains flood risk probability, flood risk level, and flood causal path, revealing the intrinsic correlation between disaster-causing factors and improving risk tracing capabilities. Finally, by acquiring at least one disaster-bearing object within the substation, determining its corresponding disaster-bearing status based on the flood risk probability and level, and inputting this status into a pre-constructed digital twin model of the substation, the application displays the disaster-bearing status of at least one disaster-bearing object through a visual interface. Based on the disaster-bearing status and flood causal path, a flood handling work order is generated, accurately mapping the risk reasoning results to specific disaster-bearing objects within the substation, thus avoiding a disconnect between risk information and on-site emergency response. This application obtains the receipt data of flood control work orders and uses the receipt data as the result of the closed-loop handling of flooding at substations, thus forming a complete closed loop from risk information to emergency response. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0011] Figure 1 It is a flowchart of the steps involved in processing data using related technologies; Figure 2 This is a flowchart illustrating the steps of a closed-loop substation flood control method based on digital twins, as provided in one embodiment of this application. Figure 3 This is a schematic diagram of the structure of a substation flood control system according to an embodiment of this application; Figure 4 This is a logic flowchart of a substation flood control method provided in one embodiment of this application; Figure 5 This is a flowchart illustrating the steps of constructing and updating a causal Bayesian network according to an embodiment of this application; Figure 6 This is a flowchart illustrating the steps of risk level mapping according to an embodiment of this application; Figure 7 This is a flowchart illustrating the steps of receipt data processing according to an embodiment of this application; Figure 8 This is a schematic diagram of the structure of a digital twin-based substation flood control closed-loop device provided in one embodiment of this application. Detailed Implementation

[0012] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the various embodiments of this application will be described in detail below with reference to the accompanying drawings. However, those skilled in the art will understand that many technical details have been presented in the various embodiments of this application to enable readers to better understand this application. However, the technical solutions claimed in this application can be implemented even without these technical details and various changes and updates based on the following embodiments. The division of the various embodiments below is for the convenience of description and should not constitute any limitation on the specific implementation of this application. The various embodiments can be combined with and referenced by each other without contradiction.

[0013] Existing methods for assessing flood risks at substations are insufficient to meet the closed-loop operational needs of the entire process, from routine static assessments and pre-flood warnings to flood control and post-disaster recovery. Traditional fixed threshold or experience-based methods are ill-suited to the diverse environments and complex operating conditions of different substations, and are prone to underreporting or false alarms. While black-box machine learning models improve prediction accuracy to some extent, they fail to reveal the disaster-causing mechanisms resulting from the coupling of multiple factors such as rainfall, external water levels, site topography, and drainage capacity. They lack the ability to trace the causal origins of risks and are unable to output interpretable risk analysis results.

[0014] Meanwhile, heterogeneous data from multiple sources, including meteorological, hydrological, and geographical data, as well as station water levels and drainage data, suffer from inconsistent temporal granularity, missing information, and noise interference. Existing technologies lack data governance and fusion mechanisms based on a unified spatiotemporal benchmark. Furthermore, current solutions often limit risk inference results to macro-level risk level outputs, failing to accurately map them to the status changes and alarm levels of specific disaster-affected objects such as control buildings, underground cable trenches, primary equipment, and drainage facilities. They also cannot form a closed-loop processing mechanism through digital twin linkage display and work order receipt evidence chains. Moreover, they lack a model self-updating mechanism triggered by performance degradation, data drift, or structural stability thresholds, making the model prone to degradation with environmental changes and lacking long-term effectiveness.

[0015] Reference Figure 1 The flowchart illustrates the steps involved in processing data using a related technique.

[0016] The data processing flow of the relevant technology solution is to access multi-source data, perform risk analysis by predicting through correlation models or judging by fixed thresholds, and then output alarms for display and manual handling.

[0017] Multi-source data access refers to collecting data from multiple sources, including meteorological, hydrological, geographical, and station water level data. Correlation model prediction refers to using statistical correlation rather than causal models for risk calculation; fixed threshold discrimination refers to making simple judgments based on pre-set rainfall or water level warning lines; risk analysis refers to assessing the likelihood of substation flooding; alarm display refers to presenting the risk assessment results to maintenance personnel through an interface; and manual handling refers to maintenance personnel manually implementing emergency measures based on experience.

[0018] The solutions of related technologies lack an interpretable causal network construction and continuous update mechanism, making it difficult to trace the source of risks; due to the lack of a closed-loop evidence chain and threshold-triggered update mechanism, an auditable disposal loop cannot be formed, and the model is prone to performance degradation with environmental changes.

[0019] Therefore, this application provides a digital twin-based closed-loop management method and device for substation flooding, which can determine flood risk parameters based on multi-source heterogeneous data, realizing the comprehensive utilization of multi-source heterogeneous data. It can reveal the intrinsic correlation between disaster-causing factors, improve risk tracing capabilities, and accurately map risk inference results to specific disaster-bearing objects within the substation, avoiding the disconnect between risk information and on-site emergency response, forming a complete closed loop from risk information to emergency response.

[0020] Reference Figure 2 The diagram illustrates a flowchart of a closed-loop substation flood control method based on digital twins, according to an embodiment of this application.

[0021] The method may specifically include the following steps: Step 201: Obtain environmental data, water level data, and drainage condition data of the substation, and determine flood risk parameters based on the environmental data, water level data, and drainage condition data.

[0022] Meteorological data can refer to various observational information reflecting dynamic changes in the external atmospheric environment, including minute-level rainfall intensity, hourly cumulative rainfall, and rainfall trends retrieved from meteorological radar echoes, all from meteorological monitoring networks. Minute-level rainfall intensity is used to capture the instantaneous impact of short-duration heavy rainfall, hourly cumulative rainfall reflects medium-term drainage pressure, and meteorological radar echo data is used to identify the movement paths and development trends of convective rainstorms.

[0023] Hydrological data refers to observational information reflecting the current status and trends of water bodies surrounding a substation, including water levels in surrounding rivers, tidal fluctuations, and historical hydrological records. Water levels in surrounding rivers are used to assess the risk of flooding or backflow during external floods, tidal fluctuations affect the outflow capacity of drainage outlets, and historical hydrological records provide a reference for regional flood frequency analysis.

[0024] Geographic information data refers to spatial information describing the natural geographical features of a substation site and its surrounding area, including micro-topographic undulations, slope and aspect, catchment area, and land use types obtained through high-precision digital elevation models. Micro-topographic undulations determine the runoff collection path, slope and aspect affect runoff generation and collection velocity, catchment area reflects the scale of external water inflow, and land use types are used to distinguish between permeable and impermeable surfaces and assign runoff generation coefficients.

[0025] Substation water level data refers to real-time water information collected by monitoring equipment such as water immersion sensors and cable trench water level gauges deployed inside the substation. This includes the water depth at the lowest point within the substation, the water level in the cable trench, and changes in the water level around critical equipment. The water depth at the lowest point within the substation reflects the overall severity of flooding, the water level in the cable trench indicates the risk of water ingress into the underground pipe network system, and the water level around critical equipment is used to determine whether the equipment faces a direct threat of immersion.

[0026] Drainage status data refers to the real-time operating status and engineering parameters of the substation's drainage system and flood control facilities. This data originates from a real-time operation monitoring system and includes the start / stop status of drainage pumps, operating current, drainage flow rate, cumulative load time, and the unobstructed status of drainage outlets. It also covers engineering parameters of passive defense facilities such as the height of retaining walls and the status of floodgates. The operating status of drainage pumps determines the active drainage capacity, the unobstructed status of drainage outlets affects whether external drainage is obstructed, and the height of retaining walls and the status of floodgates constitute a physical defense against external floods.

[0027] It should be noted that the above-mentioned multi-source data had significant differences at the time of initial access: meteorological data were recorded in minutes or hours, the sampling frequency of the sensors in the station varied from seconds to days, the geographic information data was static spatial data, and the hydrological data had different observation frequencies.

[0028] In view of the significant differences in sampling frequency, temporal granularity, spatial resolution and data format among the aforementioned multi-source data, this application implements the following standardized processing procedure: Regarding time-dimensional alignment, rainfall intensity data provided by meteorological monitoring networks are typically recorded in five-minute or hourly units, while the sampling frequency of in-station water immersion sensors and cable trench water level gauges can reach the second level, and drainage pump operation status monitoring data is in the minute level. However, the equipment status data recorded by manual inspections may only be available once a day. This application uses an exponentially weighted moving average to smooth the second-level high-frequency water level fluctuation data, effectively filtering out instantaneous fluctuations in the water surface or electrical noise interference from sensors, extracting the water level change trend signal, and uniformly reducing the sampling frequency to the minute level. For low-frequency discontinuous data such as daily manual inspection water level records and hourly cumulative meteorological rainfall, linear interpolation is comprehensively used to process stable trend segments, while Lagrange interpolation or cubic spline interpolation is used to process nonlinear and drastic changes such as sudden rainfall changes and rapid water level rises. Polynomial fitting of adjacent observation points is used to locally maintain the nonlinear characteristics of the data, ultimately constructing a continuous, gap-free standard time series with a uniform five-minute time interval.

[0029] Regarding data integrity restoration, geographic information data such as high-resolution digital elevation models may experience localized data loss due to cloud cover or sensor strip failures. This application utilizes the spatial autocorrelation of pixels within the surrounding geographic neighborhood of the missing area and employs a distance-weighted Kriging spatial interpolation algorithm to fill in the blank areas, maximizing the restoration of the continuity of surface micro-topographic undulations and slope aspect. For outlier spikes generated by water level sensors within the station due to signal interference and transmission packet loss, an anomaly detection algorithm using the median absolute deviation algorithm is employed. The median deviation from the median is used to measure the degree of dispersion, or the interquartile range (the difference between the upper and lower quartiles) is used to measure the degree of data dispersion and assist in outlier identification. Outliers deviating from the normal range are accurately identified and removed, and then a long short-term memory network regression model trained based on historical normal operation data is used to complete the data for the removed points.

[0030] Regarding the unification of spatial references, geographic information data from different sources may use the WGS-84 coordinate system used by the Global Positioning System, the Beijing 54 coordinate system commonly used in my country's historical surveying and mapping, or the Xi'an 80 coordinate system. This application performs high-precision coordinate projection transformation on all geographic information data, uniformly transforming them to the nationally unified universal transverse Mercator projection coordinate system. At the same time, it performs geocoding processing to ensure that the site topographic data, catchment area boundaries, and river location information extracted from the digital elevation model are accurately matched with the coordinates of the actual physical site of the substation, the main transformer within the station, the control building, cable trenches, and other key facilities at the centimeter level.

[0031] Regarding data format standardization, all binary monitoring data output by the SCADA system, vector files stored in the geographic information system, tabular files of manual inspection records, and JSON data returned by the meteorological service interface are standardized and converted into a columnar storage format for large-scale analysis. Variable naming conventions are unified to ensure that all water level data are in centimeters, rainfall intensity is in millimeters per hour, and drainage flow rate is in cubic meters per hour.

[0032] Through the above process, the original multi-source heterogeneous data is transformed into a high-quality fusion dataset with a unified five-minute time granularity, a unified projected coordinate system, and a unified unit definition and data format. This dataset can provide spatiotemporally accurate aligned and complete and reliable data input for subsequent causal Bayesian network construction and risk inference.

[0033] In this embodiment, flood risk parameters can be determined based on environmental data, substation water level data, and drainage operation data. These flood risk parameters can refer to comprehensive feature indicators extracted from multi-source data fusion, and are used to quantitatively express the current flood risk level faced by the substation.

[0034] Step 202: Input the flood risk parameters into the pre-built flood causal reasoning model for reasoning to obtain the flood risk probability, flood risk level and flood causal path.

[0035] In this embodiment, flood risk parameters can be input variables into a pre-constructed flood causal inference model. The flood causal inference model is a Bayesian network built from historical data, employing a directed acyclic graph structure to express causal dependencies between variables and utilizing conditional probability tables for probabilistic inference. After calculating the input parameters, the flood causal inference model outputs the flood risk probability, flood risk level, and flood causal path.

[0036] Among these, the flood risk probability represents the numerical result output by the model, indicating the likelihood of flooding occurring at the substation. The flood risk level represents the different levels of risk classification based on water depth thresholds, used to distinguish the severity of the risk. The flood causal path represents the risk transmission path revealed by the model, showing how key disaster-causing factors ultimately lead to flooding through intermediate variables.

[0037] Step 203: Obtain at least one disaster-bearing object in the substation, and determine the disaster-bearing status corresponding to each disaster-bearing object based on the probability and level of flood risk.

[0038] In this embodiment, the disaster-bearing objects are physical entities within the substation that may be affected by flooding. Each disaster-bearing object is configured with a corresponding object threshold, such as the threshold height of the control building object, the cable trench sealing height of the underground facility object, the bottom height of the mechanism box of the primary equipment object, and the upper limit of the drainage capacity of the drainage facility object. The object thresholds can be set according to the actual situation.

[0039] Among them, the disaster-bearing objects can refer to the physical entities inside the substation that may be affected by flooding, including control building objects, underground facilities objects, primary equipment objects, drainage facilities objects, etc., and each object has a configurable threshold.

[0040] The object threshold can be a risk assessment limit that can be pre-configured for each disaster-affected object, including the height of the control building threshold, the height of the cable trench blockage, the height of the critical equipment, and the upper limit of the drainage capacity.

[0041] After acquiring the disaster-affected objects, the probability and level of flood risk can be compared item by item with the pre-configured threshold values ​​for each object. Flood risk levels are divided into different grades based on water depth thresholds, each corresponding to a specific water depth range. Based on the comparison results, the disaster-affected status of each object is determined, including normal, watchful, alarm, severe, and inoperable states. When the water depth corresponding to a flood risk level exceeds the object's threshold, the disaster-affected object enters the corresponding alarm or severe state; when the water depth exceeds the object's preset limit, the disaster-affected object is determined to be in an inoperable state. The disaster-affected status is used to support subsequent digital twin-based interactive display and work order processing.

[0042] Step 204: Input the disaster status into the pre-built digital twin model of the substation, display the disaster status corresponding to at least one disaster-bearing object through a visualization interface, and generate a flood handling work order based on the disaster status and the causal path of the flood situation.

[0043] In this embodiment, the digital twin model is a virtual mirror constructed based on the physical entity of the substation. A digital twin refers to a virtual mirror that establishes a three-dimensional geometric model of the physical entity of the substation and overlays physical simulations, equipment behavior and rule-based contingency plans, and links it with real-time monitoring data for the display and simulation of flood situation.

[0044] Digital twin models can employ a four-dimensional fusion architecture integrating geometric, physical, behavioral, and rule-based models. The geometric model, built upon 2D substation drawings and 3D model data, accurately replicates the 3D spatial morphology of disaster-affected objects such as control buildings, cable trenches, primary equipment, and drainage facilities, including building outlines, equipment locations, pipeline routes, and terrain undulations. The physical model, based on digital elevation model terrain data, extrapolates the inundation range and water depth at different water levels, determining the flood convergence path and distribution area within the station. The behavioral model, based on historical experience and equipment parameters, establishes the correspondence between the state of disaster-affected objects and inundation depth, defining multi-level failure thresholds and state transition logic for each disaster-affected object. The rule-based model digitizes emergency plans, forming response rules and work order generation logic.

[0045] In this embodiment, after inputting the disaster-affected status into the digital twin model, the disaster-affected status corresponding to at least one disaster-affected object can be dynamically displayed in three dimensions through a visualization interface. The visualization interface loads a three-dimensional scene of the substation constructed from the geometric model, and calls the physical model to render the inundation range under the current water level according to the disaster-affected status. A semi-transparent blue water body can be used to mark the waterlogged area. Simultaneously, the behavioral model is invoked to link the status of the disaster-affected object with the three-dimensional model, assigning status labels to disaster-affected objects such as control building objects, underground facility objects, primary equipment objects, and drainage facility objects. Normal status is displayed in green, watchful status in yellow, alarm status in orange, severe status in red, and failure status in gray with a flashing warning. The visualization interface synchronously displays a list of affected objects, listing the current status, risk description, and key monitoring data of each disaster-affected object in order of risk level, and highlighting the affected equipment in the three-dimensional scene. Through the above linkage mechanism, abstract disaster-affected status data can be presented as an intuitive three-dimensional diagram, enabling operation and maintenance personnel to clearly grasp the real-time risk status of each area and device in the substation.

[0046] Then, after the disaster-bearing status and flood causal path are input into the rule model, the rule model first analyzes the disaster-bearing status of each disaster-bearing object. For control building objects, underground facility objects, primary equipment objects, and drainage facility objects that are in alarm, severe, or failed states, the rule model extracts the risk level, location, object type, and key monitoring data of these objects. At the same time, the rule model analyzes the flood causal path, identifies the main disaster-causing factors leading to the current risk, such as short-term heavy rainfall contributing the most, drainage pump failure contributing the second most, or external water level backwater being the main cause.

[0047] The rule model pre-builds a digital emergency response plan library. Each emergency response plan includes applicable conditions, target objects, operational steps, required resources, priority, and time requirements. Applicable conditions are defined using rule expressions. For example, when the control building object is in an alarm state and the water depth exceeds the threshold height, and the main disaster-causing factor is continuous rainfall, the water-blocking and sealing plan is matched. When the primary equipment object is in a severe state and the water depth reaches 60 centimeters, and the main disaster-causing factor is drainage failure, the equipment shutdown plan is matched. When multiple disaster-affected objects are simultaneously in a high-risk state and the causal path shows a cascading failure trend, the station-wide evacuation plan is matched.

[0048] The rule model will match the obtained disaster-bearing object status level, risk location, and main disaster-causing factors with the applicable conditions in the emergency plan library one by one, and select one or more candidate plans that meet the current scenario.

[0049] When matching multiple candidate solutions, the rule model can use a hierarchical filtering method to gradually narrow down the candidate range when comprehensively ranking the multiple candidate solutions, and finally determine the optimal solution: The first layer of filtering retains all candidate contingency plans whose risk level reaches the alarm state or above. If there is only one such plan, it is selected directly; if there are multiple plans, the second layer of filtering is initiated.

[0050] The second layer of filtering selects the plan with the lowest resource requirements from the remaining plans. Resource requirements are calculated based on the number of personnel and equipment defined in the plan, with each person counted as one resource unit and each piece of equipment counted as two resource units. The plan with the lowest total number of resource units takes priority. If multiple plans are still tied, the third layer of filtering is applied.

[0051] The third layer of filtering selects the plan with the shortest response time from the remaining plans. The response time is determined based on the time requirement field defined in the plan, with the plan having the shortest completion time taking priority.

[0052] If multiple contingency plans are still listed after the above three-layer filtering, the one with the smallest contingency plan number will be randomly selected as the final contingency plan.

[0053] After selecting an emergency plan, the rule model instantiates the plan's content into specific flood control work orders. The applicable conditions in the plan are replaced with the actual names and locations of the affected objects, and the operational steps are transformed into specific handling instructions, such as stacking twenty sandbags at the entrance of the control building, starting the No. 2 backup drainage pump, transferring the load of the 10kV No. 3 line to the No. 4 line, and isolating the No. 3 main transformer. Simultaneously, work order priorities are filled in according to risk levels; for example, alarm status corresponds to medium priority, severe status to high priority, and failure status to emergency priority. Completion time limits are filled in according to the time requirements stipulated in the plan: emergency work orders require a response within 15 minutes and completion within 30 minutes, while high-priority work orders require completion within one hour. The work order generation module also generates risk description text based on the main disaster-causing factors in the causal path, explaining that the current risk is mainly caused by continuous heavy rainfall leading to drainage system overload, and recommending priority handling of drainage facilities. The final generated flood control work order includes complete fields such as work order identifier, list of affected objects, risk description, handling recommendations, priority, time requirements, and distribution recipients, and is distributed to the corresponding maintenance personnel for execution via mobile terminals.

[0054] Step 205: Obtain the receipt data of the flood control work order and use the receipt data as the result of the substation's closed-loop flood control process.

[0055] In this embodiment, after a flood control work order is dispatched to the maintenance personnel's terminal, the maintenance personnel complete the on-site handling according to the work order requirements and transmit the handling result data back through the terminal. The receipt data is a set of structured information returned by the maintenance personnel after completing the work order handling, including multiple fields such as work order identifier, status code, time field, evidence identifier, evidence solidification, and verification.

[0056] The work order identifier field records the work order number, site number, and contingency plan template number, used to uniquely identify this handling task. The status code field records the complete lifecycle status of the work order from generation to completion, including status nodes such as generated, dispatched, accepted, executing, completed, rejected, timed out, and suspended. The time field records key time points in the handling process, including dispatch time, acceptance time, start time, and completion time, used for tracking the timeliness of the handling process. The evidence identifier field records the evidence number, evidence type, and collection time collected on-site. Evidence types may include on-site photos, video recordings, equipment operation logs, or sensor snapshots. The evidence solidification field uses cryptographic methods to ensure the tamper-proof and traceability of the evidence. A secure hash algorithm is used to calculate a unique digital digest of the raw data collected on-site, and the raw data is uploaded to storage media, recording its Uniform Resource Identifier as the evidence storage address. The verification field records the conclusion and time of the verification of the returned evidence. The verification method involves retrieving the original data based on the evidence storage address, recalculating the hash value, and comparing it with the returned hash value. If they match, the evidence is deemed tamper-proof and verification is passed. After obtaining all the above-mentioned receipt data, it is stored in the historical knowledge base as the final result of this flood control effort for subsequent post-disaster review, model optimization, and audit traceability.

[0057] In a practical implementation, the receipt data may include the contents of Table 1 below: Table 1 Receipt Data

[0058] In this embodiment, by acquiring environmental data, water level data, and drainage condition data of the substation, flood risk parameters are determined based on these data. This allows for the comprehensive utilization of multi-source heterogeneous data. Furthermore, by inputting the flood risk parameters into a pre-constructed flood causal reasoning model, the application obtains flood risk probability, flood risk level, and flood causal path, revealing the intrinsic correlation between disaster-causing factors and improving risk tracing capabilities. Finally, by acquiring at least one disaster-bearing object within the substation, determining its corresponding disaster-bearing status based on the flood risk probability and level, and inputting this status into a pre-constructed digital twin model of the substation, the application displays the disaster-bearing status of at least one disaster-bearing object through a visual interface. Based on the disaster-bearing status and flood causal path, a flood handling work order is generated, accurately mapping the risk reasoning results to specific disaster-bearing objects within the substation, thus avoiding a disconnect between risk information and on-site emergency response. This application obtains the receipt data of flood control work orders and uses the receipt data as the result of the closed-loop handling of flooding at substations, thus forming a complete closed loop from risk information to emergency response.

[0059] Optionally, environmental data may include meteorological data, hydrological data, and geographic data.

[0060] The steps for determining flood risk parameters based on environmental data, station water level data, and drainage condition data include: S11. Determine the drainage pressure factor based on the rainfall intensity in the meteorological data, the drainage pump operating status and the cumulative operating load time in the drainage condition data; S12. Based on the elevation difference of the substation and the area of ​​the catchment area in the geographical data, and the difference between the real-time water level and the warning water level in the hydrological data, determine the surrounding water accumulation risk factors. S13. Based on the water level data within the station and the reference height of the disaster-bearing object in the drainage condition data, determine the structural exposure factor. S14, at least one of the following factors is used as the flood risk parameter: drainage pressure factor, surrounding water accumulation risk factor, and structural exposure degree factor.

[0061] In this embodiment of the application, the drainage pressure factor can be determined based on the rainfall intensity in the meteorological data, the drainage pump operating status in the drainage condition data, and the cumulative operating load time.

[0062] In practical implementation, rainfall intensity from meteorological data can be used as a quantitative indicator of external water source impact pressure, drainage pump operating status from drainage operating data, including the number of pumps in operation and drainage rate, can be used as a quantitative indicator of current active drainage capacity, and cumulative operating load time can be used as a quantitative indicator of whether the drainage system is in a high-load state for a long period of time.

[0063] The drainage pressure factor is obtained by combining the above variables. The comprehensive calculation process is as follows: First, the rainfall intensity is normalized by dividing the original rainfall intensity value by the region's historical maximum rainfall intensity to obtain the normalized rainfall intensity value, which ranges from zero to one.

[0064] The operating status of the drainage pumps is normalized by dividing the current number of pumps in operation by the total number of pumps installed to obtain the operating ratio, and dividing the current drainage rate by the rated maximum drainage rate to obtain the drainage efficiency. The weighted average of the operating ratio and drainage efficiency is taken as the normalized value of drainage capacity, with a value between zero and one.

[0065] The cumulative operating load time is normalized by dividing the cumulative operating time of the drainage pump in the past six hours by six hours to obtain the normalized load time value, which ranges from zero to one.

[0066] Then, the drainage pressure factor is calculated. The drainage pressure factor equals the normalized rainfall intensity multiplied by the first weight, plus the normalized drainage capacity multiplied by the second weight, plus the normalized load time multiplied by the third weight. The first, second, and third weights are configurable parameters, and can be set to 0.5, 0.3, and 0.2 respectively, with the sum of the weights being one. The drainage pressure factor ranges from zero to one. A higher factor value indicates greater pressure on the substation's drainage system, and that it is approaching or has already exceeded the design drainage capacity limit. When rainfall intensity is high and drainage pumps operate at high loads for extended periods, the normalized rainfall intensity increases, the normalized drainage capacity decreases, and the normalized load time increases, leading to an increase in the drainage pressure factor value, indicating increased pressure on the drainage system.

[0067] In this embodiment of the application, the risk factors of water accumulation in the surrounding area can be determined based on the elevation difference of the substation and the area of ​​the catchment area in the geographical data, and the difference between the real-time water level and the warning water level in the hydrological data.

[0068] In practical implementation, the elevation difference of the substation in the geographic data can be used to reflect the degree of depression of the substation site relative to the surrounding terrain, the area of ​​the catchment area can be used to reflect the scale of external water inflow, and the difference between the real-time water level and the warning water level in the hydrological data can be used to reflect the external flood water level pressure.

[0069] The above factors are combined to calculate the surrounding water accumulation risk factor. The calculation process is as follows: First, the elevation difference of the substation is normalized. The elevation difference is the difference between the average elevation of the substation area and the average elevation of the surrounding area. A positive value indicates that the substation site is higher than the surrounding area, and a negative value indicates that the substation site is lower than the surrounding area. The normalized elevation difference is obtained by dividing the elevation difference by the absolute value of the maximum elevation difference in the area. The value ranges from negative one to positive one. The larger the negative value, the lower the substation site is than the surrounding area and the higher the degree of depression.

[0070] The catchment area is normalized by dividing it by the maximum catchment area, resulting in a normalized value between zero and one. The river level difference is also normalized; it is calculated by subtracting the warning level from the real-time river level. A positive value indicates that the warning level has been exceeded, while a negative value indicates that there is still a safety margin.

[0071] The normalized value of the water level difference is obtained by dividing the river water level difference by the historical highest value exceeding the warning level. The value ranges from negative one to positive one, with a larger positive value indicating a greater exceedance of the warning level.

[0072] Then, the surrounding waterlogging risk factor is calculated. This factor equals the negative normalized value of topographic elevation difference multiplied by the fourth weight, plus the normalized value of catchment area multiplied by the fifth weight, plus the normalized value of water level difference multiplied by the sixth weight. When the normalized value of topographic elevation difference is negative, the negative value becomes positive, reflecting the contribution of low-lying areas. The fourth, fifth, and sixth weights are configurable parameters, which can be set to 0.4, 0.3, and 0.3 respectively, with the sum of the weights being one. The surrounding waterlogging risk factor ranges from zero to one; a higher factor value indicates a greater threat of backflow from external waterlogging or flooding to the substation. When the substation site is low-lying, has a large catchment area, and the river level exceeds the warning level, the normalized value of topographic elevation difference becomes a large negative value, the normalized value of catchment area increases, and the normalized value of water level difference becomes a large positive value, leading to an increase in the surrounding waterlogging risk factor value, indicating an increased risk of backflow from surrounding waterlogging.

[0073] In this embodiment of the application, the structural exposure factor can be determined based on the water level data in the station and the reference height of the disaster-bearing object in the drainage condition data.

[0074] In practical implementation, the current or predicted water depth can be obtained using the water level data within the station, while reference heights of disaster-bearing objects in the drainage condition data can be used, including equipment engineering information such as the threshold height of the control building, the height of the cable trench cover, and the height of the base of key equipment. By comparing the water level data within the station with the reference heights of the disaster-bearing objects, the margin of equipment above the water surface or the degree of submersion exceeding the limit can be calculated, thus obtaining the structural exposure factor.

[0075] The calculation process is as follows: First, the water depth within the station is normalized by dividing the current water depth by the historical maximum water depth of the area to obtain a normalized water depth value, ranging from zero to one. For each disaster-affected object, its exposure level is calculated based on a reference height. The exposure level value equals the normalized water depth value minus the normalized reference height value, which is obtained by dividing the reference height by the maximum reference height, also ranging from zero to one. A positive exposure level value indicates that the water level has exceeded the reference height, and the equipment is submerged; a negative exposure level value indicates that the water level has not yet reached the reference height, and the equipment still has a safety margin.

[0076] The maximum value of the exposure levels of all disaster-affected objects is taken as the structural exposure factor, with a value ranging from negative one to positive one. A higher factor value indicates a higher degree of exposure to risk for critical equipment under the current flooding conditions. A positive value indicates that existing equipment has been flooded, while a smaller absolute negative value indicates a smaller safety margin and is closer to the failure flooding threshold. The structural exposure factor accurately characterizes the vulnerability and sensitivity of core assets under specific flooding scenarios.

[0077] In the embodiments of this application, at least one of the drainage pressure factor, the surrounding water accumulation risk factor, and the structural exposure factor can be used as the flood risk parameter.

[0078] This application addresses the technical challenges of integrating multi-source heterogeneous data and quantifying the coupling relationships between disaster-causing factors in existing technologies by constructing drainage pressure factors, surrounding waterlogging risk factors, and structural exposure factors as flood risk parameters. The drainage pressure factor integrates rainfall intensity and drainage system operational status to quantify the load-bearing capacity of the drainage system; the surrounding waterlogging risk factor integrates topographic elevation, catchment area, and river water level to quantify the threat of external flood backflow; and the structural exposure factor integrates station water level and equipment reference height to quantify the risk of core asset inundation. All three factors achieve a unified dimensional expression of multi-source data through normalization and weighted calculation, possessing clear physical meaning and interpretability. This significantly reduces the dimensionality of the original data, providing high-quality feature input for subsequent causal inference models and achieving an effective transformation from raw monitoring data to interpretable risk features.

[0079] Optionally, the causal inference model for flooding is a causal Bayesian network comprising a topological structure and a conditional probability table. In its implementation, the causal Bayesian network is a probabilistic graphical model consisting of a topological structure and a conditional probability table. The topological structure uses a directed acyclic graph to express the causal dependencies between variables. Each node in the graph represents a risk variable, and each directed edge represents a causal relationship from a parent node to a child node. The conditional probability table is attached to each node and is used to quantitatively describe the probability distribution of that node under different combinations of values ​​of its parent node. Through the combination of the topological structure and the conditional probability table, the causal Bayesian network can achieve probabilistic inference from observed evidence to the target node while maintaining the interpretability of the causal relationship.

[0080] The causal reasoning model for flooding is constructed in the following way: S21. Obtain historical environmental data, historical water level data, and historical drainage condition data of the substation, and determine historical flood risk parameters based on the historical environmental data, historical water level data, and historical drainage condition data. S22, determine at least one target risk node based on the preset target risk event, determine at least one disaster-causing factor node and at least one risk factor node based on historical flood risk parameters, and construct the topology structure using the target risk node, disaster-causing factor node and risk factor node as candidate nodes. S23, form a candidate node pair from any two candidate nodes, calculate the connection gain between the two candidate nodes in the candidate node pair, and connect them using an undirected edge structure. The two candidate nodes in the candidate node pair with the largest connection gain are selected; where the two candidate nodes in the candidate node pair include the first candidate node and the second candidate node. S24, calculate the first causal strength value of the first candidate node pointing to the second candidate node according to the preset causal strength calculation formula, and calculate the second causal strength value of the second candidate node pointing to the first candidate node. S25, compare the first causal strength value with the second causal strength value, determine the direction of the undirected edge structure based on the comparison result, obtain the directed edge structure and update the topology; S26. For each node in the topology, based on historical flood risk parameters and edge connections in the topology, calculate the conditional probability distribution of the node under different values ​​of its corresponding parent node, and obtain the conditional probability table for the node; where the parent node is a node that is pointed to by a directed edge.

[0081] In this embodiment, historical environmental data, historical station water level data, and historical drainage condition data are multi-source datasets collected from the substation's historical operation records, covering various scenarios under different rainfall intensities, external water levels, and drainage conditions. These historical data are time-aligned, missing data filled, anomaly-handled, and format-unified according to the data governance process in step 201, resulting in a historical fusion dataset with a unified spatiotemporal benchmark. Based on this historical fusion dataset, the drainage pressure factor, surrounding waterlogging risk factor, and structural exposure factor for historical periods are determined using the factor calculation methods in steps S11-S14, serving as historical flood risk parameters for subsequent model training. These historical flood risk parameters are used for the structure and parameter learning of the causal Bayesian network, enabling the model to mine causal dependencies between variables from historical data.

[0082] Historical environmental data can be a collection of past meteorological, hydrological, and geographical data collected from the substation's historical operation records. Historical station water level data can be a collection of past station water depth monitoring data collected from the substation's historical operation records. Historical drainage operating condition data can be a collection of past drainage system operating status and engineering parameter data collected from the substation's historical operation records. Historical flood risk parameters can be drainage pressure factors, surrounding water accumulation risk factors, and structural exposure factors calculated based on historical multi-source data, used for training causal Bayesian network models.

[0083] In this embodiment, at least one target risk node can be determined based on a preset target risk event, and at least one disaster-causing factor node and at least one risk factor node can be determined based on historical flood risk parameters. The target risk node, disaster-causing factor node, and risk factor node are used as candidate nodes to construct a topology structure.

[0084] In practical implementation, the target risk event is the core event that needs to be predicted in the model, such as flooding in a substation or water ingress into critical equipment. Target risk nodes are determined based on the preset target risk event and are used as output nodes of the network, representing the final prediction result. Historical flood risk parameters include drainage pressure factors, surrounding water accumulation risk factors, structural exposure factors, and various raw monitoring data. Variables that may be associated with the target risk event are selected from the historical flood risk parameters and identified as disaster-causing factor nodes and risk factor nodes, respectively. Disaster-causing factor nodes represent the original driving factors that may lead to flooding, such as rainfall intensity, river level, and topographic elevation difference. Risk factor nodes represent intermediate state variables in the flood development process, such as drainage pressure factors, surrounding water accumulation risk factors, and water depth within the substation. The target risk nodes, disaster-causing factor nodes, and risk factor nodes are collectively used as candidate nodes to form the basic node set for subsequent network topology construction, thus building the initial topology.

[0085] In this embodiment of the application, any two candidate nodes are combined into a candidate node pair, the connection gain between the two candidate nodes in the candidate node pair is calculated, and the two candidate nodes in the candidate node pair with the largest connection gain are connected using an undirected edge structure; wherein, the two candidate nodes in the candidate node pair include a first candidate node and a second candidate node.

[0086] In practical implementation, a greedy strategy can be used to gradually add node connections. Two different nodes are randomly selected from the candidate node set to form a candidate node pair, and this process is repeated for all possible combinations to form multiple candidate node pairs. For each candidate node pair, the connection gain between the two candidate nodes is calculated. The connection gain measures the improvement in the model's interpretability after connecting these two nodes, and can be calculated using metrics such as log-likelihood increment, Bayesian information criterion, or minimum description length. The log-likelihood increment is a scoring metric used in structure learning to select candidate connections, measuring the improvement in the model's interpretability after adding an edge. Among all candidate node pairs, the one with the largest connection gain is selected, and an undirected edge is added to connect the two nodes in this pair. At this point, the edge only shows the connection relationship, and the direction is not yet determined. Through this iterative approach, the node pair that maximizes the improvement in the model's interpretability is added to the network structure each time, gradually building a complete graph structure skeleton.

[0087] In this context, the two candidate nodes in a candidate node pair include a first candidate node and a second candidate node. The first and second candidate nodes are only used to identify two different nodes in the node pair. They do not have a fixed order or primary / secondary relationship. When determining the direction, it is necessary to calculate the causal strength values ​​in the two directions separately.

[0088] In this embodiment, the true causal relationship direction can be determined by calculating the causal strength values ​​in two directions. The preset causal strength calculation formula uses point-state causality as the calculation basis, and point-state causality is defined based on conditional entropy in information theory.

[0089] The first causal strength value, pointing from the first candidate node to the second candidate node, represents the information gain brought about by introducing the information of the first candidate node in determining the state of the second candidate node under the condition of a known target risk event. Specifically, the conditional entropy of the second candidate node under the condition of a known target risk event is first calculated, then the conditional entropy of the second candidate node under the conditions of the known first candidate node and the target risk event is calculated, and the two are subtracted to obtain the first causal strength value.

[0090] Similarly, the second causal strength value of the second candidate node pointing to the first candidate node represents the amount of information gain brought about by introducing the information of the second candidate node to determine the state of the first candidate node under the condition of known target risk event.

[0091] In the specific implementation, the preset formula for calculating causal strength is as follows: PC(xi→xj|y)=H(xj|y)-H(xj|xi,y) Where PC(xi→xj|y) represents the causal strength value of variable xj pointed to by variable xi, given the target risk event y. xi represents the pointing variable in the first or second candidate node, and xj represents the pointed variable in the first or second candidate node. y represents the target risk event corresponding to the target risk node.

[0092] H(xj|y) represents the conditional entropy of variable xj given the target risk event y. It quantifies the uncertainty of variable xj when only the outcome of the target risk event is known. A larger conditional entropy value indicates greater uncertainty and makes the state of variable xj more difficult to predict.

[0093] H(xj|xi,y) represents the conditional entropy of variable xj given the variables xi and the target risk event y. It is used to quantify the uncertainty of variable xj when both the value of variable xi and the outcome of the target risk event are known.

[0094] The difference between H(xj|y) and H(xj|xi,y) represents the reduction in uncertainty regarding the state of the determined variable xj after introducing information from variable xi. In other words, it's the information gain that variable xi provides for the explanatory variable xj. The larger this difference, the stronger the causal explanatory power of variable xi for variable xj, meaning a greater likelihood of a causal relationship from xi to xj.

[0095] In practical implementation, the calculation of point-state causality depends on the estimation of conditional entropy. Different conditional entropy estimation methods are required for different types of variables.

[0096] For discrete variables, frequency statistics can be used to estimate the probability distribution of each value state, and then the entropy value can be calculated based on the probability distribution.

[0097] For continuous variables, one of the following three options can be selected based on the sample size and computational resources: The first approach is adaptive binning discretization. This approach first determines the number of bins based on the sample size. This can be achieved using the Sturgess rule (the number of bins equals one plus the logarithm of the sample size base 2) or the Freedman-Diaconis rule. Then, the continuous variable is divided into several discrete intervals, with each sample point mapped to a corresponding interval number, transforming the continuous variable into a discrete variable. Finally, the frequency of each interval is calculated using methods for discrete variables, and the entropy value is calculated. This approach is suitable for scenarios with a sample size of less than one thousand or where computational resources are limited.

[0098] The steps to calculate H(xj|y) are as follows: Discretize variables xj and y separately into bins to obtain discretized xj_disc and y_disc. Statistically analyze the joint frequency distribution of xj_disc and y_disc, calculate the conditional probability of xj_disc taking each value given the values ​​of y_disc, and calculate the conditional entropy based on these conditional probabilities to obtain H(xj|y).

[0099] The steps to calculate H(xj|xi,y) are as follows: The variables xj, xi, and y are discretized into bins, resulting in discretized xj_disc, xi_disc, and y_disc. The joint frequency distribution of the three variables is statistically analyzed, and the conditional probability of xj_disc taking each value given the values ​​of xi_disc and y_disc is calculated. Based on these conditional probabilities, the conditional entropy is calculated to obtain H(xj|xi,y).

[0100] The second approach is kernel density estimation. This approach first selects a kernel function, such as a Gaussian kernel or an Epanechnikov kernel. Then, it determines the kernel window width, which can be achieved using the Scott rule (window width equal to the standard deviation multiplied by the negative fifth root of the sample size) or adaptively using the Silverman rule. Next, the kernel density estimation method is used to estimate the probability density function of the continuous variable, and the differential entropy is calculated based on the estimated probability density function. This approach is suitable for scenarios with a sample size of one thousand or more that require high-precision causal orientation determination.

[0101] The steps to calculate H(xj|y) are as follows: The joint probability density function of variables xj and y is estimated using the two-dimensional kernel density estimation method. The marginal probability density function of y is obtained by integrating over the xj direction. The conditional probability density function of xj is obtained by dividing the joint density by the marginal density. The differential entropy is calculated based on the conditional probability density function to obtain H(xj|y).

[0102] The steps to calculate H(xj|xi,y) are as follows: The joint probability density function of variables xj, xi, and y is estimated using the three-dimensional kernel density estimation method. The joint marginal density function of xi and y is obtained by integrating over the xj direction. The conditional probability density function of xj given xi and y is obtained by dividing the three-dimensional joint density by the joint marginal density of xi and y. The differential entropy is calculated based on the conditional probability density function to obtain H(xj|xi,y).

[0103] The third approach is a hybrid approach. For high-dimensional scenarios, the condition variables can be reduced in dimensionality or binned first to decrease their dimension. Then, the kernel density estimation method can be used to calculate the differential entropy of the target variable to balance computational complexity and estimation accuracy.

[0104] The steps to calculate H(xj|y) are as follows: Discretize the variable y by binning to obtain y_disc. For each interval of y_disc, select the corresponding sample subset, and use the one-dimensional kernel density estimation method to estimate the conditional probability density function of the variable xj in the subset, calculating the conditional differential entropy within each interval. Sum the conditional differential entropies of all intervals according to the sample proportion of each interval to obtain H(xj|y).

[0105] The steps to calculate H(xj|xi,y) are as follows: Dimensionality reduction is performed on variables xi and y, for example, by using principal component analysis to obtain the first principal component, resulting in the dimension-reduced variable z. Variable z is then discretized by binning, yielding z_disc. For each interval of z_disc, a corresponding subset of samples is selected. The conditional probability density function of variable xj in this subset is estimated using a one-dimensional kernel density estimation method, and the conditional differential entropy within each interval is calculated. The conditional differential entropies of all intervals are then weighted and summed according to the sample proportions of each interval to obtain H(xj|xi,y).

[0106] The above three solutions are all optional implementation methods and do not constitute any limitation on this application.

[0107] S25, compare the first causal strength value with the second causal strength value, determine the direction of the undirected edge structure based on the comparison result, obtain the directed edge structure and update the topology.

[0108] In the specific implementation, the causal strength values ​​in the two directions quantify the strength of causal influence in different directions and are used for subsequent direction determination. When the first causal strength value is greater than the second causal strength value, it indicates that the first candidate node has a stronger causal explanatory power for the second candidate node, and the causal relationship from the first candidate node to the second candidate node should be established. When the second causal strength value is greater than the first causal strength value, it indicates that the second candidate node has a stronger causal explanatory power for the first candidate node, and the causal relationship from the second candidate node to the first candidate node should be established.

[0109] In this embodiment of the application, for each node in the topology, the conditional probability distribution of the node under different values ​​of its corresponding parent node is calculated based on historical flood risk parameters and edge connection relationships of the topology, thus obtaining the conditional probability table corresponding to the node; wherein, the parent node is a node that is pointed to by a directed edge.

[0110] In the specific implementation, the topology has already determined the directed edge connections between nodes through a causal discovery algorithm. Each node may have zero or one or more parent nodes pointing to it. A parent node is the upstream node that points to the current node through a directed edge. For each node in the topology, historical data sequences corresponding to that node and all its parent nodes are extracted from historical flood risk parameters. The conditional probability of the node taking each possible state is calculated given that all its parent nodes have different value combinations. The conditional probability is calculated by counting the frequency of each state of the node in the samples that satisfy a specific combination of parent node values, dividing the frequency by the total number of samples under that combination of parent node values, and obtaining the estimated conditional probability value. The conditional probabilities of the node under all combinations of parent node values ​​are summarized in tabular form, which is the conditional probability table for that node. The conditional probability table is the quantitative basis for probabilistic inference in causal Bayesian networks. When the observational evidence of certain nodes is known, probability information can be transmitted along the directed edge direction through the conditional probability table to calculate the posterior probability of the target node.

[0111] In this context, a parent node is a node to which a directed edge points. Specifically, in a directed acyclic graph (DAG), if there exists a directed edge from node A to node B, then node A is the parent node of node B, and node B is the child node of node A. The value of the parent node directly affects the probability distribution of the child nodes, and this influence is quantified using a conditional probability table.

[0112] This application addresses the technical problem that existing black-box models cannot reveal the intrinsic relationships between disaster-causing factors and lack risk tracing capabilities by constructing a causal Bayesian network as a causal inference model for flood conditions. Through the combination of topological structure and conditional probability tables, the causal Bayesian network can not only realize probabilistic inference from observed evidence to target nodes, outputting flood risk probability and risk level, but also trace back the causal path with the highest contribution, clearly showing the complete chain of risk transmission from disaster-causing factors through intermediate variables to target nodes, thus achieving interpretable flood risk analysis.

[0113] Optionally, the method further includes the following steps: S31. If a loop structure is formed in the topology after adding a directed edge structure, then delete the directed edge structure with the smallest connection gain or the smallest causal strength value from the several directed edge structures corresponding to the loop structure. S32, delete directed edge structures in the topology whose causal strength value is lower than a preset threshold or whose causal strength value is negative.

[0114] In the embodiments of this application, if the newly added edge structure causes a loop in the topology during the process of gradually adding directed edge structures, that is, if there is a path that starts from a certain node, passes through several directed edges, and can return to the node, then the loop structure needs to be processed to restore the characteristics of a directed acyclic graph.

[0115] From all the directed edges constituting the loop, the edge with the smallest connection gain or the edge with the smallest causal strength value can be selected and deleted. Connection gain is the scoring criterion used when the edge was initially added to the network, reflecting its contribution to the model's explanatory power. Causal strength value is the quantified result of causal strength calculated when the edge's direction was determined, reflecting the strength of the causal relationship represented by the edge. Deleting the edge with the smallest contribution allows for the elimination of loops while preserving, as far as possible, the causal connections that contribute most to the model's explanatory power.

[0116] In this embodiment, after the initial formation of the topology, all directed edges can be screened, deleting those with excessively low or negative causal strength values. A negative causal strength value indicates that, under known target risk events, introducing variable information to the pointed-to variable actually increases the uncertainty of the pointed-to variable's state. This situation lacks positive causal explanatory power in a physical sense and should be eliminated. A preset threshold can be used to determine whether the causal strength value meets the retention standard. Edges with causal strength values ​​below this threshold indicate that their causal explanatory power for the pointed-to variable is very weak, possibly caused by data noise rather than a true causal relationship, and should also be eliminated. Pruning refers to a processing mechanism that removes weak causal edges or redundant edges during network updates to improve network stability and interpretability. Through pruning, weak causal edges and false causal relationships in the network can be removed, improving the network's simplicity, stability, and interpretability.

[0117] The preset threshold ε can be set in one of the following ways: The first method is to use a fixed threshold. The preset threshold ε can be between 0.01 and 0.1, for example, it can be 0.05. This method is suitable for situations where the business scenario is relatively stable and the data distribution characteristics do not change much.

[0118] The second approach is an adaptive threshold based on sample size. The preset threshold ε is equal to a constant c divided by the square root of the current sample size. The constant c typically ranges from 0.5 to 2.0, and the current sample size is the number of samples currently used for modeling. This approach is suitable for scenarios where the data volume changes dynamically. When the sample size is small, the threshold automatically increases, and pruning is more stringent to avoid overfitting. When the sample size increases, the threshold automatically decreases, preserving more potential causal edges to fully extract information from the data.

[0119] The third method is quantile thresholding. The preset threshold ε is the p-th percentile of the causal strength values ​​of all candidate edge structures, such as the 10th or 20th percentile. This method first sorts all candidate edge causal strength values ​​from smallest to largest, and uses the value at the p-th percentile as the threshold. Edges with causal strength values ​​lower than this threshold are deleted. This method is suitable for scenarios requiring control over network density, and the number of edges ultimately retained can be precisely controlled by adjusting the percentile p.

[0120] The specific value of the threshold ε can be configured according to business needs and model stability requirements, using the method described above.

[0121] This application addresses the technical challenges of loop structures disrupting the directed acyclic graph (DAG) characteristics and weak causal edges and spurious causal relationships affecting network interpretability during the construction of causal Bayesian networks through loop destruction and pruning techniques. Loop destruction removes edges with the minimum connection gain or causal strength when loops form, ensuring the network always meets the DAG requirements. Pruning uses fixed thresholds, adaptive sample size thresholds, or quantile thresholds to remove edges with excessively low or negative causal strength, eliminating noise interference and spurious causal connections. This significantly improves the network's simplicity, stability, and interpretability, resulting in clearer and more reliable causal paths.

[0122] Optionally, the steps of inputting flood risk parameters into a pre-built flood causal reasoning model to obtain the flood risk probability, flood risk level, and flood causal path include: S41, assign the flood risk parameters to the corresponding disaster-causing factor nodes; S42, based on the flood risk parameters and the conditional probability table corresponding to the disaster-causing factor nodes, the probability is propagated along the directed edge structure of the topology; S43, In the probability propagation process, the expected value corresponding to the risk factor node is calculated as the water accumulation depth, and the posterior probability corresponding to the target risk node is calculated as the flood risk probability. S44, determine the flood risk level based on the depth of water accumulation; S45. Calculate the contribution of each disaster-causing factor node to the posterior probability corresponding to the target risk node. Based on the contribution, trace back from the target risk node along the directed edge structure to obtain the causal path of the flood situation.

[0123] In this embodiment of the application, flood risk parameters can be assigned to the corresponding disaster-causing factor nodes.

[0124] In practical implementation, flood risk parameters include drainage pressure factors, surrounding waterlogging risk factors, and structural exposure factors. These factors have already established correspondences with specific disaster-causing factor nodes during the model building phase. The flood risk parameter values ​​calculated at the current moment are assigned to the corresponding disaster-causing factor nodes as known observational evidence input into the model. After the disaster-causing factor nodes receive their assigned values, their states are fixed to the current observation values, eliminating uncertainty and serving as the starting point for probability propagation.

[0125] In this embodiment, probability propagation can be performed along the directed edge structure of the topology based on the flood risk parameters and the conditional probability table corresponding to the disaster-causing factor nodes.

[0126] In the implementation, each non-input node is accompanied by a conditional probability table, recording the conditional probability of the node taking each state under different combinations of its parent node's values. Starting from the already assigned hazard factor node, the posterior probability distribution of each downstream node is calculated sequentially along the direction of the directed edges. For each node, its own probability distribution is calculated based on the current probability distributions of all its parent nodes and its conditional probability table. This process propagates downwards layer by layer until the target risk node is reached.

[0127] In this embodiment of the application, during the probability propagation process, the expected value corresponding to the risk factor node can be calculated as the water depth, and the posterior probability corresponding to the target risk node can be calculated as the flood risk probability.

[0128] In practical implementation, risk factor nodes, such as the water depth node within a station, have probability distributions that represent the likelihood of occurrence across different value intervals. The expected value of this node is obtained by multiplying the water depth value represented by each interval by the probability corresponding to that interval and summing the results. The target risk node represents the final risk event, and its posterior probability reflects the likelihood of this event occurring under the current observational evidence. This posterior probability is then output as the flood risk probability.

[0129] In this embodiment of the application, the flood risk level can be determined based on the depth of water accumulation.

[0130] In practice, the water depth can be divided into different levels based on a preset water depth threshold.

[0131] For example, d represents the depth of the water, in centimeters (cm); if the input is in meters (m), it will be converted to centimeters (cm) first.

[0132] RiskLevel(d)=L4 (d<27cm); L3 (27cm≤d<60cm); L2 (60cm≤d<100cm); L1 (d≥100cm).

[0133] In other words, a water depth of less than 27 centimeters is classified as Level 4 risk; a water depth between 27 and 60 centimeters is classified as Level 3 risk; a water depth between 60 and 100 centimeters is classified as Level 2 risk; and a water depth of more than 100 centimeters is classified as Level 1 risk. The calculated water depth is compared with these thresholds to determine the current flood risk level.

[0134] In the embodiments of this application, the contribution of each disaster-causing factor node to the posterior probability corresponding to the target risk node can be calculated separately. Based on the contribution, the causal path of the flood situation can be obtained by tracing back from the target risk node along the directed edge structure.

[0135] In the specific implementation, for each disaster-causing factor node, the change in the posterior probability of the target risk node after removing the observed evidence of that node is calculated. This change represents the contribution of that disaster-causing factor node to the final risk outcome. The greater the contribution, the stronger the dominant role of the disaster-causing factor in the current risk. Starting from the target risk node, the process traces backward along the path with the largest contribution, that is, each time a parent node pointing to the current node with the largest contribution is selected to continue tracing upstream until the disaster-causing factor node is reached. This path is the causal path of the flood situation, clearly demonstrating the complete chain of risk transmission from key disaster-causing factors through intermediate variables to the final outcome.

[0136] This application addresses the technical problems of existing black-box models, such as the inability to quantify the contribution of disaster-causing factors and the difficulty in tracing the source of risk, by assigning flood risk parameters to disaster-causing factor nodes and performing probability propagation along directed edges. The scheme calculates the expected value of risk factors as the water depth during probability propagation, determines the risk level based on water depth thresholds, and simultaneously calculates the target posterior probability as the risk probability. By calculating the contribution of each disaster-causing factor to the posterior probability and tracing back along directed edges, an interpretable causal path can be output from key disaster-causing factors through intermediate variables to the final result, achieving a unification of risk quantification and source tracing analysis.

[0137] Optionally, the step of obtaining at least one disaster-bearing object in the substation and determining the disaster-bearing status of each object based on the probability and level of flood risk includes: S51, obtain at least one disaster-bearing object in the substation, and the preset state threshold corresponding to each disaster-bearing object; wherein, the preset state threshold includes water depth threshold or probability threshold corresponding to different risk levels. S52 compares the probability of flooding risk and the level of flooding risk with the preset state thresholds of each disaster-bearing object; S53. Based on the comparison results, determine the disaster status corresponding to each disaster-bearing object; wherein, the disaster status includes at least one of the following: normal status, attention status, alarm status, critical status, and failure status.

[0138] In this embodiment, at least one disaster-bearing object in the substation and a preset state threshold corresponding to each disaster-bearing object can be obtained; wherein, the preset state threshold includes a water depth threshold or a probability threshold corresponding to different risk levels. A disaster-bearing object refers to any object in the disaster-bearing object set, which can be a regional object, an equipment object, or a functional object. Examples include control building objects, underground cable trench objects, primary equipment objects, and drainage facility objects. Both the object set and the threshold can be configured.

[0139] In this embodiment, the disaster-affected object is a physical entity within the substation that may be affected by flooding, including control building objects, underground facility objects, primary equipment objects, and drainage facility objects. A predefined object list is read, and each disaster-affected object is associated with a preset state threshold. The preset state thresholds are judgment boundaries corresponding to different risk levels, including two types: water depth thresholds and probability thresholds. Water depth thresholds are engineering parameters related to the object's own structure, such as the threshold height of the control building object, the cable trench sealing height of the underground facility object, the bottom height of the mechanism box of the primary equipment object, and the upper limit of the drainage capacity of the drainage facility object. Probability thresholds are pre-set risk probability limits, for example, using an 80% risk probability as the alarm trigger line.

[0140] The control building may include core areas of the secondary systems such as control rooms, relay protection rooms, DC rooms, and communication remote control rooms.

[0141] Underground facilities can include underground spatial structures such as cable trenches, tunnels, openings, and underground entrances.

[0142] Primary equipment can include high-voltage equipment such as mechanism boxes, main transformers, and structural foundations.

[0143] Drainage facilities can include drainage pumps, drainage outlets, flood control facilities such as retaining walls.

[0144] The probability and level of flood risk are compared with the preset state thresholds for each disaster-affected object. The probability of flood risk is a numerical result expressing the likelihood of flooding at the substation. The flood risk level is a macro-level classification based on water depth, corresponding to a specific water depth range. These two macro-risk results are compared item by item with the preset thresholds for each disaster-affected object to determine whether the current risk level has reached or exceeded the risk assessment limit for that object. The preset thresholds may differ for different disaster-affected objects due to their structural characteristics and varying importance.

[0145] Based on the comparison results, the disaster status corresponding to each disaster-bearing object is determined; the disaster status includes at least one of the following: normal status, attention status, alarm status, critical status, and failure status.

[0146] In this embodiment, the specific status of each disaster-affected object can be determined based on the comparison results, and the disaster-affected status is divided into five levels. When the water depth threshold is higher than the upper limit of the current risk level and the risk probability is lower than the probability threshold, the disaster-affected object is in a normal state. When the water depth threshold is within the current risk level range but close to the upper limit, or the risk probability is close to the probability threshold, the disaster-affected object is in a state of concern. When the water depth threshold is within the current risk level range but lower than the upper limit, or the risk probability exceeds the probability threshold, the disaster-affected object is in an alarm state. When the water depth threshold is lower than the lower limit of the current risk level, the disaster-affected object is in a severe state. When the water depth threshold is lower than the lower limit of the current risk level and the water depth exceeds the object's preset limit, the disaster-affected object is determined to be in a failed state. The disaster-affected status is used to support subsequent digital twin linkage display and work order processing.

[0147] In practical implementation, disaster-bearing objects include, but are not limited to, the four typical types of objects shown in Table 2. Each type of disaster-bearing object has a corresponding main data source, typical risk mechanism, and configurable items. The thresholds and fields in Table 2 are for illustrative purposes only; in actual applications, they can be configured and expanded according to the specific circumstances of the substation. Table 2 below provides a comparison example of disaster-bearing objects and configurable items.

[0148] Table 2 Examples of disaster-affected objects and configurable items

[0149] As shown in Table 2, the four types of disaster-bearing objects have different risk mechanisms and configurable items: The control building includes core secondary system areas such as the control room, relay protection room, DC room, and communication remote control room. A typical risk mechanism is that water ingress leads to DC power supply system failure, which in turn triggers a cascading consequence of relay protection device power failure, communication interruption, and loss of control over the entire station's monitoring. Configurable items for the control building include an object list, threshold height, and data source mapping. The threshold height is used to determine whether the water level has reached the danger limit for entering the control building, and the data source mapping is used to associate real-time water level monitoring data with the corresponding control building objects.

[0150] Underground facilities include underground spatial structures such as cable trenches, tunnels, openings, and underground entrances. A typical risk mechanism is the formation of a conduit effect after water ingress, rapidly diverting accumulated water from the station area into the underground pipe network. This water can then penetrate the secondary system through weak points such as cable penetrations, causing hidden disasters. Configurable items for underground facilities include an object list and a cable trench sealing height field. The cable trench sealing height field records the height of the sealing facilities at each cable trench entrance, serving as a basis for determining whether groundwater may enter the secondary system.

[0151] Primary equipment includes high-voltage equipment such as control boxes, main transformers, and structural foundations. The typical risk increases significantly around three water depth thresholds: 27 cm, 60 cm, and 100 cm. 27 cm is the critical height at which water can enter the bottom of the control box; 60 cm is the risk threshold for flooding the main transformer cooling system; and 100 cm is the limit height at which the stability of the structural foundation is threatened. Configurable items for primary equipment include an item list and a critical height field. The critical height field records the ground clearance of critical components of each primary piece of equipment, serving as a basis for determining whether the equipment faces a direct threat of immersion.

[0152] Drainage facilities include flood control structures such as drainage pumps, drainage outlets, and retaining walls. A typical risk mechanism is that backwater from external river levels obstructs the outflow from drainage outlets, or a power outage causes the drainage pumps to malfunction, triggering a chain reaction of continuously rising water levels within the facility. Configurable items for drainage facilities include an object list and backwater status criteria. These criteria define the conditions under which a backwater condition is identified when the external water level is higher than the drainage outlet; for example, a backwater alarm is triggered when the river level exceeds the bottom height of the drainage outlet and continues to rise.

[0153] This application addresses the technical challenge of translating macro-level risk outcomes into concrete physical entities, thus hindering precise flood control efforts, by establishing a mapping mechanism between disaster-affected objects and preset state thresholds. Water depth and probability thresholds are configured for disaster-affected objects such as control buildings, underground facilities, primary equipment, and drainage facilities. The probability and level of flood risk are compared item by item with these thresholds to determine the state of each object, including normal, watchful, alarm, severe, and ineffective. This achieves a refined mapping from macro-level risk to specific object states, enabling flood control command to accurately locate affected facilities and differentiate risk severity, providing a decision-making basis for subsequent digital twin visualization and differentiated emergency response.

[0154] Optionally, the method further includes the following steps: S61. Based on the receipt data and the environmental data, station water level data and drainage condition data corresponding to the flood control work order, an online evaluation sample is generated. S62, Calculate the performance metrics of the causal Bayesian network on online evaluation samples; S63, obtain the value range and frequency of each flood risk parameter in the online assessment sample, as the data distribution characteristics of the online assessment sample; S64, obtain the value range and frequency of each flood risk parameter in the historical sample data corresponding to the online assessment sample, as the data distribution characteristics of the training sample; S65, calculate the difference between the data distribution characteristics of the online evaluation samples and the data distribution characteristics of the training samples; S66: Obtain the current topology of the causal Bayesian network and the historical topology before the update, and count the number and proportion of edge changes between the two as the change in topology. S67. If the performance index is less than or equal to the first preset threshold, or the difference value is greater than or equal to the second preset threshold, or the change in topology is greater than or equal to the third preset threshold, then the causal Bayesian network is updated based on the online evaluation samples.

[0155] In this embodiment, the receipt data is structured information returned by maintenance personnel after completing work order processing, including work order identifier, status code, time field, evidence identifier, evidence solidification, and verification fields. The receipt data can be correlated and matched with the environmental data, station water level data, and drainage condition data that triggered the work order at that time, and reorganized according to the same format and method as historical training data to form a new online evaluation sample. This sample represents real data under the current actual operating scenario and is used for subsequent model performance evaluation and update decisions.

[0156] In this embodiment, input data from online evaluation samples can be substituted into the currently deployed causal Bayesian network to run forward inference and obtain prediction results. The prediction results are then compared with the actual results occurring in the samples to calculate performance metrics such as precision, recall, or F1 score. The F1 score is the harmonic mean of precision and recall, which comprehensively evaluates the model's classification performance under imbalanced sample conditions. Performance metrics reflect the model's performance on new data.

[0157] In this embodiment, for all samples in the online assessment sample set, the minimum and maximum values ​​of each flood risk parameter, such as drainage pressure factor, surrounding waterlogging risk factor, and structural exposure factor, can be statistically analyzed as the value range. The proportion of occurrences of different value ranges for each parameter can be statistically analyzed as the frequency of occurrence. The value range and the frequency of occurrence together constitute the data distribution characteristics of the online assessment samples, which are used for comparison with the distribution of historical training data.

[0158] It can also obtain the historical training sample set used when building the current model, and statistically analyze the value range and frequency of each flood risk parameter in the set to form the data distribution characteristics of the training samples. The data distribution characteristics of the training samples represent the data environment characteristics during model training.

[0159] In this embodiment, methods such as the overall stability index, KL divergence, or quantile difference can be used to calculate the degree of difference between the distribution characteristics of the online evaluation samples and the distribution characteristics of the training samples. The overall stability index is calculated by comparing the frequency of online data with the frequency of training data. The KL divergence measures the amount of information loss between the two probability distributions, and the quantile difference compares the degree of shift between the two distributions at key quantiles. The larger the difference value, the more significant the change in the distribution of the online data relative to the training data.

[0160] Then, the current topology of the current version of the causal Bayesian network can be obtained, including the connections of all nodes and directed edges. The network topology of the previous version or a baseline version is obtained as the historical topology. The differences between the two versions are compared, and the number of newly added edges, deleted edges, and edges whose directions have changed are counted. The total number of these changed edge structures is divided by the total number of edge structures in the baseline version to obtain the change ratio. The number of changes and the change ratio reflect the drastic change in network structure as data is updated.

[0161] In this embodiment of the application, if the performance index is less than or equal to a first preset threshold, or the difference value is greater than or equal to a second preset threshold, or the change in topology is greater than or equal to a third preset threshold, the causal Bayesian network is updated based on the online evaluation samples.

[0162] In the specific implementation, the first preset threshold is set for the performance index. Taking the first preset threshold as 0.75 as an example, an update is triggered when the F1 score is lower than 0.75.

[0163] The second preset threshold is set for the differences in data distribution characteristics. Taking the second preset threshold as 0.25 as an example, an update is triggered when the difference value is greater than 0.25.

[0164] The third preset threshold is for topology changes. For example, if the third preset threshold is 10%, an update will be triggered when the topology change ratio is greater than 10%.

[0165] If any of these conditions are met, it indicates that the current model is no longer suitable for the new data environment or has undergone significant changes, requiring the model update process to be initiated. The update can be performed by incrementally learning the original model based on newly added online evaluation samples, or by retraining the model using accumulated new data. After the update, the new version is recorded, and the baseline metrics and thresholds are updated.

[0166] This application addresses the technical challenges of performance degradation and the difficulty in timely detection and updating of models due to prolonged runtime and environmental changes by constructing performance indicator monitoring, data distribution drift detection, and topology change detection mechanisms. It automatically triggers model updates when performance indicators fall below a first threshold, data distribution differences exceed a second threshold, or the proportion of network structure changes exceeds a third threshold. This ensures that the causal Bayesian network can continuously adapt to new data environments, maintain stable prediction accuracy and interpretability, and achieves dynamic optimization throughout the model's lifecycle and long-term effectiveness.

[0167] Reference Figure 3 The diagram illustrates a structural schematic of a substation flood control system according to an embodiment of this application. In this embodiment, a digital twin-based closed-loop substation flood control method can be applied to the substation flood control system.

[0168] The substation flood control system includes a data access and management module, a causal network construction module, a causal network update module, a risk reasoning and classification module, a key object configuration module, a digital twin linkage display module, a work order and disposal suggestion module, a receipt evidence collection module, and a trigger threshold monitoring module.

[0169] The data access governance module is used to access multi-source data and perform time alignment, missing data completion, anomaly handling, field unification, and data version management.

[0170] The causal network construction module is used to select candidate connections based on log-likelihood increments and form a directed acyclic graph based on point-state causality orientation, and to learn or initialize conditional probability tables.

[0171] The causal network update module is used to update the structure based on new data, prune weak causal edges, and perform loop detection and loop breaking processing when incremental updates cause causal edges to form loops in order to maintain a directed acyclic graph and update the conditional probability table.

[0172] The risk reasoning and classification module is used to perform probabilistic reasoning based on observational evidence, and outputs the risk probability and risk level. The causal network structure and conditional probability table used for reasoning are provided or updated by the causal network construction module and the causal network update module.

[0173] The key object configuration module is used to configure the key object set, the mapping relationship between objects and data sources, and object thresholds. It maps the risk reasoning results to object status or alarm information and provides them for use by the digital twin linkage display module.

[0174] The digital twin linkage display module is used to link the geometric model, physical model, behavioral model and rule model, render the flood range, display the affected objects and alarm information, and map the risk reasoning results to the object status based on the object list, threshold and mapping rules of the key object configuration module before displaying them.

[0175] The work order and disposal suggestion module is used to generate disposal suggestions and work orders based on the rule base or contingency plan, and to track the execution status of work orders.

[0176] The receipt evidence collection module is used to collect work order receipt fields and evidence identifiers. The work order receipt fields include a status code and a time field, and the evidence identifiers include evidence number, evidence hash, or evidence storage address.

[0177] The trigger threshold monitoring module is used to monitor performance degradation thresholds and trigger model updates and version switching. Performance metrics include F1 score, which is the harmonic mean of model precision and recall. Optional extensions include monitoring drift metrics or structural stability thresholds.

[0178] Reference Figure 4 The diagram shows a logical flowchart of a substation flood control method according to an embodiment of this application.

[0179] The methods for handling flooding in substations include the following steps: It integrates multi-source data, including meteorological data, hydrological data, geographical data, station water level data, and drainage operation data.

[0180] Perform time alignment and resampling on multi-source data.

[0181] Perform missing data completion and anomaly detection processing, and record the data version.

[0182] This forms a fusion set of variables and feature information.

[0183] Construct or update a causal Bayesian network, select candidate edges by log-likelihood increment, determine direction by point-state causality, and maintain the directed acyclic graph structure by pruning.

[0184] Based on observational evidence, probabilistic reasoning is performed and risk information and key causal paths are output.

[0185] Risk is classified based on a preset water depth threshold, which is configurable.

[0186] The results of risk reasoning are mapped to the status of key objects and displayed in conjunction with a digital twin model.

[0187] Generate handling suggestions or work orders and collect receipt information, including status code, completion time, and evidence identifier.

[0188] Determine if the update threshold has been triggered. If it has, proceed to the receipt data processing flow to execute the update, retraining, canary release, and baseline threshold update processes after the threshold is triggered. Once completed, return to the multi-source data access step. If it has not been triggered, return directly to the multi-source data access step to continue running and form a closed loop.

[0189] Reference Figure 5 The diagram illustrates a flowchart of the steps involved in constructing and updating a causal Bayesian network according to an embodiment of this application.

[0190] The process of building and updating a causal Bayesian network includes the following steps: Construct a set of candidate variables and target nodes.

[0191] The candidate connection gain can be calculated using scoring metrics such as log-likelihood increment or Bayesian information criterion.

[0192] Select the candidate connection with the highest gain and add it to the network structure.

[0193] The causal strength and direction are calculated based on the point-state causality formula.

[0194] Determine whether a ring structure has formed.

[0195] If a cycle is formed, remove the lowest-scoring edge from the set of candidate edges that lead to the cycle, i.e., the one with the smallest log-likelihood increment or the smallest point-state causality, to break the cycle-acyclic graph.

[0196] Pruning is performed to remove causal edges whose point-state causality values ​​are less than or equal to a preset threshold or whose point-state causality values ​​are negative. The specific value of the preset threshold is configured according to business requirements and model stability requirements.

[0197] Learn or update the conditional probability table and output the critical path.

[0198] Determine if there is new data. If there is new data, return to the step of calculating candidate connection gain to continue updating; if there is no new data, output the current network structure and conditional probability table and return to the main process to continue executing the subsequent risk reasoning and classification steps, or end this update subprocess.

[0199] Reference Figure 6 The diagram illustrates a flowchart of risk level mapping steps provided in an embodiment of this application.

[0200] The risk level and key object mapping process includes the following steps: The risk inference output receives probabilistic inference results and key causal paths from the risk inference and classification modules. The probabilistic inference results include the posterior probability distribution of each risk node, while the key causal paths reveal the dominant path from the causative factor through intermediate variables to the target risk node.

[0201] Risk level mapping converts the risk inference output into four risk levels (Level 1 to Level 4) based on preset water depth thresholds. Level 1 risk corresponds to a water depth of over 100 cm, Level 2 risk to 60-100 cm, Level 3 risk to 27-60 cm, and Level 4 risk to below 27 cm. The water depth thresholds can be configured based on the substation's actual flood control standards and historical hydrological data.

[0202] The critical object configuration sets include control building objects, underground facility objects, primary equipment objects, and drainage facility objects. Control building objects include control rooms, relay protection rooms, DC rooms, and communication remote control rooms. Underground facility objects include cable trenches, tunnels, openings, and underground entrances. Primary equipment objects include mechanism boxes, main transformers, and structural foundations. Drainage facility objects include drainage pumps, drainage outlets, and retaining walls. Threshold parameters for each object can be configured according to actual engineering parameters.

[0203] The object state machine outputs results based on the risk level and object thresholds. Object states include normal, watchful, alarm, critical, and failed states. When the risk level reaches Level 1 and the measured parameters of a critical object exceed preset limits, the object is classified as failed. The preset limits for control building objects are the threshold height field value, for underground facility objects the cable trench sealing height field value, for primary equipment objects the critical equipment height field value, and for drainage facility objects the upper limit of drainage capacity field value. When the actual water level exceeds these preset limits, the corresponding object is classified as failed.

[0204] The digital twin linkage display links the digital twin virtual model and renders the flooding range in a 3D scene based on the output of the object state machine. Different colors are used to indicate the status level of each disaster-affected object, and the list of affected objects, alarm information and handling suggestions are displayed.

[0205] Reference Figure 7 The diagram illustrates a flowchart of the steps involved in processing receipt data according to an embodiment of this application.

[0206] Generate handling suggestions or work orders. Based on the emergency plan matched by the rule model, the plan content is instantiated into specific flood handling work orders. The work order includes fields such as work order identifier, list of affected objects, risk description, handling suggestions, priority, time requirements, and dispatch recipients.

[0207] Dispatch work orders and record the dispatch time. The generated work orders are dispatched to the corresponding maintenance personnel via mobile terminals, and the dispatch time is recorded in the system as the starting point for tracking the timeliness of handling.

[0208] The system collects receipt fields and evidence identifiers. To ensure the tamper-proof and auditable nature of the data, when the maintenance terminal submits a work order receipt, a secure hash algorithm is used to calculate the original data after on-site handling, generating a unique digital digest as the digital fingerprint of the handling record. At the same time, the original data is uploaded to the storage medium and a Uniform Resource Identifier (URI) is obtained. The final evidence identifier is a tuple containing the digital digest and the URI.

[0209] Verification receipts and evidence are entered into the database. Original evidence data is retrieved from the storage medium using a Uniform Resource Identifier (URI), and a digital digest is recalculated using the same hash algorithm. This recalculated digest is then compared with the digital digest recorded in the work order receipt. If they match, the evidence is deemed tamper-proof, verified, and stored in the historical knowledge base. If they do not match, an anomaly alarm is triggered, and the evidence is rejected from entering the database, thus constructing a fully credible evidence chain from risk warning to on-site handling and model feedback.

[0210] Create online evaluation samples. The receipt data is correlated and matched with the environmental data, station water level data, and drainage condition data that triggered the work order at that time. The data is then reorganized in the same format as the historical training data to form new online evaluation samples, which are used for subsequent model performance evaluation and update decisions.

[0211] The system determines whether a threshold is triggered. The threshold can be at least one of the following: performance metric threshold, data drift metric threshold, or structural stability metric threshold. For example, the performance metric threshold uses the F1 score; an update is triggered when the F1 score is below 0.75. The data drift metric threshold uses the overall stability index (PSI) as an example; an update is triggered when the PSI is greater than 0.25. The structural stability metric threshold uses the topology change ratio as an example; an update is triggered when the change ratio is greater than 10%.

[0212] If the threshold is triggered, perform model updates, retraining, or canary releases and record the version, updating baseline metrics and thresholds. If the threshold is not triggered, continue running.

[0213] Reference Figure 8 The diagram shows a schematic of a digital twin-based substation flood control closed-loop management device according to an embodiment of this application. The device includes: The parameter acquisition module 801 is used to acquire environmental data, water level data and drainage condition data of the substation, and determine flood risk parameters based on the environmental data, water level data and drainage condition data. The risk reasoning module 802 is used to input the flood risk parameters into a pre-built flood causal reasoning model for reasoning, and obtain the flood risk probability, flood risk level and flood causal path; The status acquisition module 803 is used to acquire at least one disaster-bearing object in the substation and determine the disaster-bearing status of the disaster-bearing object according to the flood risk probability and the flood risk level. The work order generation module 804 is used to input the disaster-bearing status into the pre-built digital twin model of the substation, display the disaster-bearing status corresponding to the at least one disaster-bearing object through a visual interface, and generate a flood handling work order based on the disaster-bearing status and the flood causal path. The result acquisition module 805 is used to acquire the receipt data of the flood control work order and use the receipt data as the flood control closed-loop handling result of the substation.

[0214] Optionally, the environmental data includes meteorological data, hydrological data, and geographic data; the parameter acquisition module 801 includes: The first parameter acquisition submodule is used to determine the drainage pressure factor based on the rainfall intensity in the meteorological data, the drainage pump operating status and the cumulative operating load time in the drainage condition data; The second parameter acquisition submodule is used to determine the surrounding water accumulation risk factor based on the substation topographic elevation difference and catchment area in the geographic data, and the water level difference between the real-time water level and the warning water level in the hydrological data. The third parameter acquisition submodule is used to determine the structural exposure factor based on the water level data in the station and the reference height of the disaster-bearing object in the drainage condition data. The fourth parameter acquisition submodule is used to use at least one of the drainage pressure factor, the surrounding water accumulation risk factor, and the structural exposure factor as the flood risk parameter.

[0215] Optionally, the flood causal inference model is a causal Bayesian network including a topological structure and a conditional probability table; the flood causal inference model is constructed in the following manner: The historical environmental data, historical water level data, and historical drainage condition data of the substation are obtained, and historical flood risk parameters are determined based on the historical environmental data, historical water level data, and historical drainage condition data. At least one target risk node is determined based on the preset target risk event, at least one disaster-causing factor node and at least one risk factor node are determined based on the historical flood risk parameters, and the target risk node, the disaster-causing factor node and the risk factor node are used as candidate nodes to construct the topology. Any two candidate nodes are paired to form a candidate node pair. The connection gain between the two candidate nodes in the candidate node pair is calculated, and the two candidate nodes in the candidate node pair with the largest connection gain are connected using an undirected edge structure. The two candidate nodes in the candidate node pair include a first candidate node and a second candidate node. The first causal strength value from the first candidate node to the second candidate node is calculated according to the preset causal strength calculation formula, and the second causal strength value from the second candidate node to the first candidate node is calculated. By comparing the first causal strength value with the second causal strength value, the direction of the undirected edge structure is determined based on the comparison result, the directed edge structure is obtained, and the topology is updated. For each node in the topology, based on the historical flood risk parameters and the edge connection relationships of the topology, the conditional probability distribution of the node under different values ​​of its corresponding parent node is calculated, and the conditional probability table corresponding to the node is obtained; wherein, the parent node is a node that points to the node through a directed edge.

[0216] Optionally, the device further includes: The first pruning module is used to delete the directed edge structure with the smallest connection gain or the smallest causal strength value from the several directed edge structures corresponding to the ring structure if a ring structure is formed in the topology after adding the directed edge structure. The second pruning module is used to delete directed edge structures in the topology whose causal strength value is lower than a preset threshold or whose causal strength value is negative.

[0217] Optionally, the risk reasoning module 802 includes: The first risk acquisition submodule is used to assign the flood risk parameters to the corresponding disaster-causing factor nodes; The second risk acquisition submodule is used to perform probability propagation along the directed edge structure of the topology based on the flood risk parameters and the conditional probability table corresponding to the disaster-causing factor nodes. The third risk acquisition submodule is used to calculate the expected value corresponding to the risk factor node as the water accumulation depth during the probability propagation process, and to calculate the posterior probability corresponding to the target risk node as the flood risk probability. The fourth risk acquisition submodule is used to determine the flood risk level based on the water depth; The fifth risk acquisition submodule is used to calculate the contribution of each disaster-causing factor node to the posterior probability corresponding to the target risk node, and to trace back from the target risk node along the directed edge structure based on the contribution to obtain the flood causal path.

[0218] Optionally, the status acquisition module 803 includes: The first state acquisition submodule is used to acquire at least one disaster-bearing object in the substation, and a preset state threshold corresponding to each disaster-bearing object; wherein, the preset state threshold includes a water depth threshold or a probability threshold corresponding to different risk levels. The second state acquisition submodule is used to compare the flood risk probability and the flood risk level with the preset state thresholds of each of the disaster-bearing objects; The third state acquisition submodule is used to determine the disaster-bearing state corresponding to each of the disaster-bearing objects based on the comparison results; wherein, the disaster-bearing state includes at least one of the following: normal state, attention state, alarm state, critical state, and failure state.

[0219] Optionally, the device further includes: The first update module is used to generate an online evaluation sample based on the environmental data, the station water level data, and the drainage condition data corresponding to the receipt data and the flood control work order; The second update module is used to calculate the performance index of the causal Bayesian network on the online evaluation sample; The third update module is used to obtain the value range and frequency of occurrence of each flood risk parameter in the online assessment sample, as the data distribution characteristics of the online assessment sample; The fourth update module is used to obtain the value range and frequency of occurrence of each flood risk parameter in the historical sample data corresponding to the online assessment sample, as the data distribution characteristics of the training sample; The fifth update module is used to calculate the difference between the data distribution characteristics of the online evaluation samples and the data distribution characteristics of the training samples; The sixth update module is used to obtain the current topology of the causal Bayesian network and the historical topology before the update, and to count the number and proportion of edge changes between the two as the amount of topology change. The seventh update module is used to update the causal Bayesian network based on the online evaluation samples if the performance index is less than or equal to a first preset threshold, or the difference value is greater than or equal to a second preset threshold, or the change in topology is greater than or equal to a third preset threshold.

[0220] As the apparatus embodiment is basically similar to the method embodiment, it is described in a relatively simple manner. For relevant details, please refer to the description of the method embodiment.

[0221] An embodiment of this application also provides an electronic device, which may include a processor, a memory, and a computer program stored in the memory and capable of running on the processor. When the computer program is executed by the processor, it implements the method described above.

[0222] An embodiment of this application also provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, it implements the method described above.

[0223] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use and processing of the relevant data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0224] The above provides a detailed description of a digital twin-based closed-loop substation flood control method and device. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A closed-loop management method for flooding in substations based on digital twins, characterized in that, The method includes: The environmental data, water level data, and drainage condition data of the substation are obtained, and flood risk parameters are determined based on the environmental data, water level data, and drainage condition data. The flood risk parameters are input into a pre-built flood causal reasoning model for reasoning to obtain the flood risk probability, flood risk level, and flood causal path; At least one disaster-bearing object in the substation is identified, and the disaster-bearing status corresponding to the disaster-bearing object is determined according to the flood risk probability and the flood risk level. The disaster-bearing status is input into the pre-built digital twin model of the substation, and the disaster-bearing status corresponding to the at least one disaster-bearing object is displayed through a visual interface. A flood situation handling work order is generated based on the disaster-bearing status and the flood situation causal path. Obtain the receipt data of the flood control work order, and use the receipt data as the closed-loop handling result of the flood situation at the substation.

2. The method according to claim 1, characterized in that, The environmental data includes meteorological data, hydrological data, and geographic data; the steps for determining flood risk parameters based on the environmental data, the station water level data, and the drainage condition data include: The drainage pressure factor is determined based on the rainfall intensity in the meteorological data, the drainage pump operating status and cumulative operating load time in the drainage condition data; Based on the substation topographic elevation difference and catchment area in the geographical data, and the water level difference between the real-time water level and the warning water level in the hydrological data, the surrounding water accumulation risk factors are determined. The structural exposure factor is determined based on the water level data within the station and the reference height of the disaster-bearing object in the drainage condition data. At least one of the drainage pressure factor, the surrounding water accumulation risk factor, and the structural exposure factor is used as the flood risk parameter.

3. The method according to claim 1, characterized in that, The flood causal inference model is a causal Bayesian network including a topological structure and a conditional probability table; the flood causal inference model is constructed in the following manner: The historical environmental data, historical water level data, and historical drainage condition data of the substation are obtained, and historical flood risk parameters are determined based on the historical environmental data, historical water level data, and historical drainage condition data. At least one target risk node is determined based on the preset target risk event, at least one disaster-causing factor node and at least one risk factor node are determined based on the historical flood risk parameters, and the target risk node, the disaster-causing factor node and the risk factor node are used as candidate nodes to construct the topology. Any two candidate nodes are paired to form a candidate node pair. The connection gain between the two candidate nodes in the candidate node pair is calculated, and the two candidate nodes in the candidate node pair with the largest connection gain are connected using an undirected edge structure. The two candidate nodes in the candidate node pair include a first candidate node and a second candidate node. The first causal strength value from the first candidate node to the second candidate node is calculated according to the preset causal strength calculation formula, and the second causal strength value from the second candidate node to the first candidate node is calculated. By comparing the first causal strength value with the second causal strength value, the direction of the undirected edge structure is determined based on the comparison result, the directed edge structure is obtained, and the topology is updated. For each node in the topology, based on the historical flood risk parameters and the edge connection relationships of the topology, the conditional probability distribution of the node under different values ​​of its corresponding parent node is calculated, and the conditional probability table corresponding to the node is obtained; wherein, the parent node is a node that points to the node through a directed edge.

4. The method according to claim 3, characterized in that, The method further includes: If adding the directed edge structure results in a loop structure in the topology, then the directed edge structure with the smallest connection gain or the smallest causal strength value will be deleted from the several directed edge structures corresponding to the loop structure. Delete directed edge structures in the topology whose causal strength value is lower than a preset threshold or whose causal strength value is negative.

5. The method according to claim 3, characterized in that, The steps of inputting the flood risk parameters into a pre-constructed flood causal reasoning model to obtain the flood risk probability, flood risk level, and flood causal path include: The flood risk parameters are assigned to the corresponding disaster-causing factor nodes; Based on the flood risk parameters and the conditional probability table corresponding to the disaster-causing factor nodes, probability propagation is performed along the directed edge structure of the topology. In the probability propagation process, the expected value corresponding to the risk factor node is calculated as the water accumulation depth, and the posterior probability corresponding to the target risk node is calculated as the flood risk probability. The flood risk level is determined based on the water depth. The contribution of each disaster-causing factor node to the posterior probability corresponding to the target risk node is calculated respectively. Based on the contribution, the causal path of the flood situation is obtained by tracing back from the target risk node along the directed edge structure.

6. The method according to claim 1, characterized in that, The steps of acquiring at least one disaster-bearing object in the substation and determining the disaster-bearing status of each disaster-bearing object based on the flood risk probability and the flood risk level include: Obtain at least one disaster-bearing object in the substation, and a preset state threshold corresponding to each disaster-bearing object; wherein, the preset state threshold includes a water depth threshold or a probability threshold corresponding to different risk levels; The probability of flooding risk and the level of flooding risk are compared with the preset state thresholds of each of the disaster-bearing objects; Based on the comparison results, the disaster-bearing status corresponding to each of the disaster-bearing objects is determined; wherein, the disaster-bearing status includes at least one of the following: normal status, attention status, alarm status, critical status, and failure status.

7. The method according to claim 3, characterized in that, The method further includes: Based on the receipt data and the environmental data, station water level data, and drainage condition data corresponding to the flood control work order, an online evaluation sample is generated. Calculate the performance metrics of the causal Bayesian network on the online evaluation samples; The range and frequency of each flood risk parameter in the online assessment sample are obtained as the data distribution characteristics of the online assessment sample; The range and frequency of each flood risk parameter in the historical sample data corresponding to the online assessment sample are obtained as the data distribution characteristics of the training sample; Calculate the difference between the data distribution characteristics of the online evaluation samples and the data distribution characteristics of the training samples; Obtain the current topology of the causal Bayesian network and the historical topology before the update, and count the number and proportion of edge changes between the two as the topology change amount. If the performance index is less than or equal to the first preset threshold, or the difference value is greater than or equal to the second preset threshold, or the change in topology is greater than or equal to the third preset threshold, then the causal Bayesian network is updated based on the online evaluation sample.

8. A substation flood situation closed-loop management device based on digital twin, characterized in that, The device includes: The parameter acquisition module is used to acquire environmental data, water level data and drainage condition data of the substation, and determine flood risk parameters based on the environmental data, water level data and drainage condition data. The risk reasoning module is used to input the flood risk parameters into a pre-built flood causal reasoning model for reasoning, and obtain the flood risk probability, flood risk level and flood causal path; The status acquisition module is used to acquire at least one disaster-bearing object in the substation and determine the disaster-bearing status of the disaster-bearing object according to the flood risk probability and the flood risk level. The work order generation module is used to input the disaster-bearing status into the pre-built digital twin model of the substation, display the disaster-bearing status corresponding to the at least one disaster-bearing object through a visual interface, and generate a flood handling work order based on the disaster-bearing status and the flood causal path. The result acquisition module is used to acquire the receipt data of the flood control work order and use the receipt data as the closed-loop handling result of the flood situation at the substation.

9. An electronic device, characterized in that, It includes a processor, a memory, and a program or instructions stored on the memory and executable on the processor, wherein the program or instructions, when executed by the processor, implement the method as described in any one of claims 1-7.

10. A readable storage medium, characterized in that, The readable storage medium stores a program or instructions that, when executed by a processor, implement the method as described in any one of claims 1-7.

Images