Power communication network risk assessment method based on risk gene and space-time large model

Abstract

The invention relates to the technical field of power communication network security, in particular to a power communication network risk assessment method based on risk genes and a space-time large model. According to the technical scheme, the power communication network risk assessment method based on the risk gene and the space-time large model comprises the following steps that equipment sensor data, meteorological grid data and service topology data of a power communication transmission network are obtained in real time through a multi-source data acquisition module, and space-time alignment and feature coding are executed; a risk gene map is constructed, the risk gene map comprises a three-layer map structure of an equipment vulnerability gene, an environment threat gene and a business dependence gene, and a cross-space-time risk propagation rule is defined; a multi-modal space-time large model is trained, and a framework combining Transform and a space-time diagram convolutional network is adopted. A biological gene concept is introduced into power communication network risk assessment, gene-level fault traceability is achieved, and risk factors have combinability and interpretability.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power communication network security, and in particular to a power communication network risk assessment method based on a risk gene and a space-time large model. BACKGROUND

[0002] As the core nerve center of the smart grid, the power communication transmission network carries key services such as relay protection, dispatching control, and real-time monitoring. With the advancement of new power system construction, the power grid continues to expand, the complexity of the communication network topology grows exponentially, and it faces serious challenges from multi-dimensional risks such as extreme weather, network attacks, and device aging. Existing power communication network risk assessment methods mainly rely on three types of technology: first, threshold alarm systems, which trigger alarms based on single-index thresholds (such as temperature exceeding a fixed limit), and cannot reflect multi-factor coupled risks; second, static risk assessment models, such as decision trees or Bayesian networks, which rely on historical fault libraries for matching and have long update cycles (usually more than 24 hours), making them difficult to adapt to dynamic changes in the grid topology; third, traditional space-time analysis techniques, which model device status, geographic information, and weather data separately, lacking cross-modal correlation feature mining. These methods have significant limitations: risk prediction granularity is coarse, making it difficult to achieve gene-level fault tracing; space-time data is processed in isolation, failing to effectively predict risk transmission across space and hierarchy; and response strategies rely on human experience, leading to high response delays and misjudgment probabilities. For example, in extreme events such as typhoons or cold waves, existing technologies cannot dynamically capture risk propagation paths, easily leading to cascading power outages.

[0003] The core technical problem to be solved by the present application is how to achieve full-factor perception, accurate quantification, and proactive defense of power communication network risks. Specific problems include: first, how to build a risk model covering the full dimension of "device-environment-service" to support explainable modeling of multi-factor coupled risks; second, how to design a space-time perception deep learning architecture that integrates multi-modal data (such as device time series features, environmental spatial grids, and service topology) to solve the modeling problem of cross-regional risk transmission; third, how to achieve dynamic risk assessment and decision optimization, transforming static "after-the-fact remediation" into intelligent "pre-emptive defense". SUMMARY

[0004] The present application proposes a power communication network risk assessment method based on a risk gene and a space-time large model, solving the problem of high risk misjudgment rates, response delays, and frequent cascading failures caused by the inability of traditional power communication network risk assessment methods to dynamically integrate multi-modal data, accurately model risk transmission paths, and implement proactive defense.

[0005] To solve the above technical problems, the technical solution adopted by the present application is as follows: The power communication network risk assessment method based on a risk gene and a space-time large model comprises the following steps: Real-time acquisition of device sensor data, meteorological grid data and business topology data of power communication transmission network through multi-source data acquisition module, and execution of space-time alignment and feature coding; Constructing a risk gene map including a three-layer map structure of device vulnerability genes, environmental threat genes and business dependence genes, and defining cross-space-time risk propagation rules; Training a multi-modal spatio-temporal large model using a combination of Transformer and spatio-temporal graph convolution network architecture, and fusing device time series features, environmental spatial features and business topology features through cross-modal attention mechanism; Based on the dynamic risk quantification evaluation module, the cascade failure probability is predicted by Monte Carlo simulation, and the resource scheduling strategy is optimized by deep reinforcement learning; Through the three-dimensional visual decision module, the risk heat map and disposal scheme are output, and the risk gene map weight and multi-modal spatio-temporal large model parameters are dynamically updated according to online feedback data.

[0006] Further, the risk assessment method is realized by a system using a hierarchical distributed architecture, including data perception layer, intelligent analysis layer and decision application layer three levels: The data perception layer includes multi-source access gateway, space-time alignment engine and edge computing node, which is used for real-time collection and preprocessing of multi-modal heterogeneous data; The intelligent analysis layer includes risk gene map construction module, multi-modal spatio-temporal large model and dynamic risk assessment engine, which is used for risk gene modeling and multi-modal fusion analysis; The decision application layer includes three-dimensional situation awareness platform, intelligent decision aid and API service gateway, which is used for providing visual interaction and automatic response.

[0007] Further, the step of constructing a risk gene map includes: Extracting entity knowledge from device manuals and fault case library, defining gene attributes and storing graph relationship through graph database; The risk gene map supports dynamic update, and the unrecorded risk mode is identified through anomaly detection algorithm; The gene association rule includes device environment interaction, device business interaction and environment business interaction pairing, and the risk propagation logic is encoded through causal relationship edge and dependence relationship edge.

[0008] Further, the multi-modal spatio-temporal large model consists of five layers: input layer, multi-modal encoding layer, spatio-temporal fusion layer, prediction and optimization layer and output layer: The input layer receives device time series data, environmental spatial grid data and business topology data; The multi-modal encoding layer uses multiple encoders to extract features of each modality and generate unified embedding representation; The spatio-temporal fusion layer adopts an improved cross-modal multi-head attention mechanism and a spatio-temporal graph convolution network to realize multi-modal feature interaction; The prediction and optimization layer predicts the risk distribution based on deep learning and generates a strategy in combination with a global optimization algorithm; The output layer outputs the node-level risk assessment result and generates a global risk heat map.

[0009] Further, the spatio-temporal fusion layer adopts an improved cross-modal multi-head attention mechanism, which includes: Each attention head focuses on specific modal interaction and introduces a learnable gating parameter to suppress irrelevant modal interference; A hierarchical Transformer architecture is adopted, including a basic layer, a cross-modal fusion layer, and a spatio-temporal aggregation layer; Through the collaborative mechanism of Transformer and the spatio-temporal graph convolution network, long-distance dependencies and local topological features are complemented.

[0010] Further, the training of the multi-modal spatio-temporal large model adopts a two-stage training method: In the pre-training stage, the model parameters are initialized using historical failure data, and the multi-modal input data is reconstructed through a masked autoencoder; In the fine-tuning stage, the parameter update direction is constrained through an elastic weight solidification algorithm, and the model parameters are dynamically updated through incremental learning.

[0011] Further, the dynamic risk quantification evaluation module is implemented through a dynamic risk evaluation engine, which includes: A data preprocessing module for multi-modal data cleaning, alignment, and feature encoding; A Monte Carlo simulation module based on a risk gene map and an improved SIR model to perform cascading failure simulation; A reinforcement learning decision module that generates resource scheduling strategies using a proximal policy optimization algorithm; A model update module for incremental training and dynamic updating of the gene map; An API service gateway providing risk assessment result query and strategy execution interfaces.

[0012] Further, the improved SIR model adopted by the Monte Carlo simulation module includes: Multi-state extension, dividing device states into four states: normal, degraded, failure, and recovery; Dynamic propagation rate modeling, adjusting the propagation rate based on real-time environmental data; Spatial heterogeneity factor, calculating the geographical risk weight through the distance between the device location and the threat source; Hierarchical recovery mechanism, dividing the recovery process into three stages: emergency repair, temporary recovery, and permanent repair.

[0013] Further, the method applied to the typhoon disaster prevention scene includes: Match the typhoon wind speed, cable suspension height and service redundancy gene through the risk gene map; Predict the cascading failure path using a multi-modal spatio-temporal large model; Generate the optimal defense strategy and display the risk heat map in real time through a three-dimensional visualization platform.

[0014] Further, the method applied to the extreme cold disaster prevention scene includes: Match the icing thickness, low temperature and cable bearing gene through the risk gene map; Derive the cascading failure probability using Monte Carlo simulation; Output the deicing and routing switching strategy through the reinforcement learning decision module; Support dynamic updating of the gene map, add new extreme freezing related rules and adjust the weight.

[0015] The positive effects of the present application are: The biological gene concept is introduced into the power communication network risk assessment, which has the ability to trace the gene level failure, makes the risk factors have combinability and explainability, and supports accurate positioning of the risk source from the device board level.

[0016] Secondly, the risk gene map supports dynamic updating, identifies unrecorded risk patterns through an anomaly detection algorithm, can quickly adapt to the dynamic changes of the power grid topology, realizes minute-level risk assessment updating, and greatly improves the adaptability of the system.

[0017] Through the cross-modal attention mechanism and spatio-temporal fusion layer of the multi-modal spatio-temporal large model, the interactive dependence of devices, environment and business is effectively captured, the risk transmission path simulation is supported, and the discovery time of new risk patterns is significantly shortened.

[0018] Fourthly, the multi-modal data fusion technology overcomes the fragmentation of traditional methods, can reflect the multi-factor coupled risk, predict the risk transmission across space and level, and thus reduce the probability of power outage accidents caused by power grid cascading failures.

[0019] Finally, the traditional "after-the-fact remedy" of manual disposal is transformed into the "pre-emptive defense" of the intelligent system, the resource scheduling strategy is optimized through deep reinforcement learning, the economic loss of the power system caused by extreme events is reduced, and the resilience and reliability of the power grid are improved. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 It is a structure schematic diagram of the risk quantitative assessment system in the embodiment of the present application; Figure 2 It is a risk gene map schematic diagram in the embodiment of the present application; Figure 3 A multi-modal spatio-temporal large model framework for embodiments of the present application. DETAILED DESCRIPTION

[0021] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor also fall within the scope of protection of the present application.

[0022] Embodiment 1 The power communication network risk assessment method based on risk genes and a spatio-temporal large model includes the following steps: Real-time acquisition of device sensor data, meteorological grid data and business topology data of the power communication transmission network through a multi-source data acquisition module, and execution of spatio-temporal alignment and feature coding; Construction of a risk gene map, the risk gene map including a three-layer map structure of device vulnerability genes, environmental threat genes and business dependency genes, and definition of cross-spatio-temporal risk propagation rules; Training of a multi-modal spatio-temporal large model, using an architecture combining Transformer and spatio-temporal graph convolution network, and fusion of device time series features, environmental spatial features and business topology features through a cross-modal attention mechanism; Based on a dynamic risk quantification assessment module, execution of Monte Carlo simulation to predict cascading failure probability, and adoption of deep reinforcement learning to optimize resource scheduling strategies; Output of a risk heat map and a disposal scheme through a three-dimensional visual decision module, and dynamic update of risk gene map weights and multi-modal spatio-temporal large model parameters according to online feedback data.

[0023] The power communication network risk assessment method based on risk genes and a spatio-temporal large model includes the following specific implementation manners. Real-time device sensor data such as optical transmission device error rate and time delay, meteorological grid data such as typhoon wind speed and rainfall, and business topology data including SDH or OTN link traffic matrix are acquired through a multi-source data acquisition module, and a spatio-temporal alignment engine is used to realize nanosecond-level time synchronization and geographical space mapping. A risk gene map is constructed, a graph database is used to store a three-layer structure of device vulnerability genes such as optical cable aging rate, environmental threat genes such as typhoon grade, and business dependence genes such as business SLA grade, and gene pairing rules such as device and environment interaction are defined. A multi-modal spatio-temporal large model is trained, and a structure combining a Transformer and a spatio-temporal graph convolution network ST-GCN is used to fuse time sequence, space and topology features through a cross-modal attention mechanism. A dynamic risk quantification assessment module is used to perform Monte Carlo simulation, an improved SIR model is used to predict the probability of cascading failures, and a deep reinforcement learning algorithm such as PPO is used to optimize resource scheduling strategies. Finally, a three-dimensional visual decision module outputs a risk heat map and a disposal scheme, and dynamically updates gene map weights and model parameters according to online feedback data to realize minute-level risk assessment updates. In a typhoon and cold wave scene, the method reduces the business interruption time by 83%, and the risk identification accuracy is increased to 91%.

[0024] The present application realizes full-factor perception and accurate quantification of power communication network risks by integrating multi-source data and advanced models, can actively identify cross-spatio-temporal risk transmission paths, and thus significantly improves the defense efficiency and safety of the power grid. The method introduces the concept of risk genes, makes the risk assessment interpretable and dynamically adaptable, supports full-chain risk tracing from the device level to the business level, effectively reduces the probability of cascading failures, and enhances the resilience and reliability of the power system.

[0025] The risk assessment method is implemented by a system, which adopts a hierarchical distributed architecture including a data perception layer, an intelligent analysis layer and a decision application layer: The data perception layer includes a multi-source access gateway, a spatio-temporal alignment engine and an edge computing node, which are used to collect and preprocess multi-modal heterogeneous data in real time; The intelligent analysis layer includes a risk gene map construction module, a multi-modal spatio-temporal large model and a dynamic risk assessment engine, which are used to realize risk gene modeling and multi-modal fusion analysis; The decision application layer includes a three-dimensional situation awareness platform, an intelligent decision aid and an API service gateway, which are used to provide visual interaction and automated response.

[0026] The specific implementation of the system using a layered distributed architecture includes a data perception layer including a multi-source access gateway supporting MQTT and Modbus protocols for collecting device data, a time-space alignment engine for realizing time synchronization based on PTP protocol, and an edge computing node deploying a lightweight anomaly detection model for real-time preprocessing of multi-modal heterogeneous data; an intelligent analysis layer including a risk gene map construction module extracting entity knowledge from device manuals and fault case libraries, a multi-modal time-space large model using a five-layer architecture, and a dynamic risk assessment engine combining Monte Carlo simulation and reinforcement learning for realizing risk modeling and fusion analysis; and a decision application layer including a three-dimensional situation awareness platform dynamically rendering a risk heat map, an intelligent decision-making auxiliary associated emergency plan library, and an API service gateway providing RESTful interfaces for integration with power grid dispatching systems. The system interacts through nine modules, supports dynamic updating of gene maps and online strategy optimization, such as automatic switching of routes and dispatching of UPS power supply vehicles in typhoon defense.

[0027] The system uses a layered distributed architecture to realize the collaborative work of data collection, analysis and decision-making through modular design, greatly improving the processing efficiency and system scalability. The layered structure supports real-time data stream processing and multi-source heterogeneous data fusion, ensuring the timeliness and accuracy of risk assessment results, while enhancing the robustness and maintenance convenience of the system, providing one-stop intelligent support for power grid operation and maintenance.

[0028] The steps of constructing the risk gene map include: Extracting entity knowledge from device manuals and fault case libraries, defining gene attributes and storing map relationships through a graph database; The risk gene map supports dynamic updating through anomaly detection algorithms to identify unrecorded risk patterns; Gene association rules include device-environment interaction, device-business interaction and environment-business interaction pairing, and risk propagation logic is encoded through causal relationship edges and dependency relationship edges.

[0029] The specific embodiment of constructing the risk gene map includes extracting entity knowledge such as optical amplifiers and lightning intensity from device manuals and failure case libraries, defining gene attributes such as the quantitative formula of the device gene, and storing the map relationship through a graph database such as Neo4j, for example, the relationship that the base station A points to the risk event optical path interruption with the vulnerability optical fiber aging rate greater than twenty-five percent; the map supports dynamic updating, identifies unrecorded risk patterns through an anomaly detection algorithm, for example, node embedding similarity calculation triggers structural adjustment, and sets an online verification interface for A or B testing, and rolls back to the historical stable version when the difference exceeds five percent; the gene association rule includes device and environment interaction, for example, typhoon wind speed greater than thirty meters per second causes optical cable dancing risk, device and business interaction such as node betweenness centrality greater than zero point six affects the expansion of business interruption range, and environment and business interaction pairing, through causal relationship edges to store trigger threshold and confidence, and dependence relationship edges to store impact coefficient and topological weight, to encode risk propagation logic and realize interpretable modeling.

[0030] The construction of the risk gene map enhances the depth and flexibility of risk modeling, supports rapid risk tracing and pattern recognition through visual association rules. The map can dynamically adapt to new risk patterns, improve the accuracy and updateability of risk assessment, and enable operation and maintenance personnel to intuitively understand risk transmission logic, thereby optimizing defense strategies.

[0031] The multi-modal spatio-temporal large model is composed of five layers: an input layer, a multi-modal encoding layer, a spatio-temporal fusion layer, a prediction and optimization layer, and an output layer. The input layer receives device time series data, environmental spatial grid data, and business topology data. The multi-modal encoding layer uses multiple encoders to extract features of each modality and generate a unified embedding representation. The spatio-temporal fusion layer uses an improved cross-modal multi-head attention mechanism and a spatio-temporal graph convolution network to realize multi-modal feature interaction. The prediction and optimization layer predicts risk distribution based on deep learning and generates strategies combined with global optimization algorithms. The output layer outputs node-level risk assessment results and generates a global risk heat map.

[0032] The multi-modal spatio-temporal large model is composed of five layers. The input layer receives device time series data, such as time series tensors of optical power and bit error rate, environmental spatial grid data such as typhoon wind speed grid, and business topology data in the form of adjacency matrix in GraphML, and unifies the reference through a spatio-temporal alignment engine; the multi-modal encoding layer uses device time series encoder to extract features using one-dimensional CNN and LSTM, environmental space encoder uses ConvLSTM and attention mechanism, and business topology encoder uses GraphSAGE to generate node embedding, output unified embedding representation dimension five hundred and twelve; the spatio-temporal fusion layer uses an improved cross-modal multi-head attention mechanism, such as device and environment attention head, and a spatio-temporal graph convolutional network ST-GCN to realize feature interaction; the prediction and optimization layer generates strategies based on deep learning to predict risk distribution, such as node failure probability in the next twenty-four hours, and combines genetic algorithm; the output layer outputs node-level risk assessment through Sigmoid activation and generates a global risk heat map for API interface call. The model accurately predicts the probability of optical cable breakage in a cold wave scenario to eighty-nine point four five percent.

[0033] The five-layer architecture of the multi-modal spatio-temporal large model realizes efficient fusion of multi-source features, and improves the accuracy and generalization ability of risk prediction through deep learning methods. The model supports long time series and spatial topology analysis, can capture complex risk dependency relationships, and thus generate more reliable risk heat maps and decision recommendations, enhancing the situational awareness and early warning capability of the power grid.

[0034] The spatio-temporal fusion layer uses an improved cross-modal multi-head attention mechanism, including: Each attention head focuses on specific modal interaction and introduces learnable gating parameters to suppress irrelevant modal interference; A hierarchical Transformer architecture is used, including a basic layer, a cross-modal fusion layer, and a spatio-temporal aggregation layer; The collaborative mechanism of Transformer and spatio-temporal graph convolutional network realizes the complementarity of long-distance dependence and local topology features.

[0035] The specific implementation of the spatio-temporal fusion layer using the improved cross-modal multi-head attention mechanism includes each attention head focusing on specific modal interaction, such as the impact of a typhoon on a device with a device query and an environment key-value calculation table, and introducing a learnable gating parameter to suppress irrelevant modal interference, such as when the device and business attention weights are greater than zero. nine, priority is given to protecting critical optical cables; using a hierarchical Transformer architecture, including four layers of standard attention in the basic layer to extract modal internal features, two layers of gated attention in the cross-modal fusion layer to force modal interaction, and one layer of spatial and temporal pooling in the spatio-temporal aggregation layer; through the collaborative mechanism of Transformer and spatio-temporal graph convolutional network ST-GCN, Transformer handles long-distance dependencies such as global typhoon wind field data, ST-GCN captures local topological features such as optical cable physical connection strength, and through a feature fusion formula, the complementary effect is achieved, and in the typhoon scenario, the optical cable fracture prediction F1 value is improved to 0.93.

[0036] The improved cross-modal attention mechanism optimizes the interaction efficiency of multi-modal data by focusing on specific modal interaction to reduce irrelevant interference and improve the quality of feature fusion. This mechanism, combined with a hierarchical architecture, can effectively capture long-distance dependencies and local features, enhancing the model's ability to analyze complex risk scenarios and improving the continuity and reliability of risk assessment.

[0037] The training of the multi-modal spatio-temporal large model uses a two-stage training method: In the pre-training stage, historical failure data is used to initialize model parameters, and a masked autoencoder is used to reconstruct multi-modal input data; In the fine-tuning stage, the elastic weight consolidation algorithm is used to constrain the parameter update direction, and incremental learning is used to dynamically update the model parameters.

[0038] The specific implementation of the two-stage training method of the multi-modal spatio-temporal large model is that in the pre-training stage, historical failure data such as multi-year failure records is used to initialize model parameters, and a masked autoencoder is used to reconstruct multi-modal input data, with random masking of 15% of device time series data and 20% of environment grid pixels. The loss function includes mean square error device environment and graph reconstruction loss business topology; in the fine-tuning stage, the elastic weight consolidation EWC algorithm is used to constrain the parameter update direction to prevent catastrophic forgetting, and incremental learning is used to dynamically update the model parameters. New data is received every 15 minutes, and joint learning of risk classification cross-entropy loss and strategy regression mean absolute error loss is performed. Combining spatio-temporal adversarial sample generation such as FGSM attack to simulate abnormal data improves robustness, and after online learning of the model, the risk prediction error is reduced by 15%.

[0039] The two-stage training method improves the adaptability and stability of the model, ensuring its compatibility with historical patterns and emerging risks through pre-training and fine-tuning processes. The elastic weight solidification mechanism prevents the model from forgetting important features, supports continuous learning, and enables the risk assessment system to dynamically evolve, adapt to changes in the power grid environment, and enhance long-term practicality.

[0040] The dynamic risk quantification assessment module is implemented through a dynamic risk assessment engine, which includes: A data preprocessing module for multi-modal data cleaning, alignment, and feature encoding; A Monte Carlo simulation module that performs cascading failure simulation based on the risk gene map and an improved SIR model; A reinforcement learning decision module that generates resource scheduling strategies using the Proximal Policy Optimization (PPO) algorithm; A model update module that implements incremental training and dynamic updating of the gene map; An API service gateway that provides risk assessment result query and strategy execution interfaces.

[0041] The specific implementation of the dynamic risk quantification assessment module through the dynamic risk assessment engine includes the data preprocessing module for multi-modal data cleaning, alignment, and feature encoding, such as normalizing device data to the range of zero to one. The Monte Carlo simulation module performs cascading failure simulation based on the risk gene map and an improved SIR model, such as parallel simulation of ten thousand times to output node failure probability. The reinforcement learning decision module generates resource scheduling strategies using the Proximal Policy Optimization (PPO) algorithm, with action spaces including route switching and dispatching repair teams. The model update module implements incremental training using a ring buffer to store twenty-four hours of data, and dynamically updates the gene map when the similarity is less than 0.7. The API service gateway provides risk assessment result query, such as a RESTful interface returning JSON format work order instructions, achieving second-level response and resource scheduling optimization in typhoon scenarios.

[0042] The dynamic risk assessment engine achieves automation of risk quantification and strategy generation through the integration of simulation and reinforcement learning, greatly improving response speed and decision quality. The engine supports real-time data stream processing and incremental updating, ensuring the timeliness and accuracy of risk assessment results, providing reliable technical support for power grid defense.

[0043] The improved SIR model used by the Monte Carlo simulation module includes: Multi-state extension, dividing device states into four states: normal, degraded, failure, and recovery; Dynamic propagation rate modeling, adjusting the propagation rate based on real-time environmental data; Spatial heterogeneity factor, calculating geographic risk weights by the distance between device locations and threat sources; A hierarchical recovery mechanism divides the recovery process into three stages: emergency repair, temporary recovery, and permanent repair.

[0044] The implementation of the improved SIR model by the Monte Carlo simulation module includes multi-state expansion, dividing device states into normal, degraded, failure, and recovery, such as optical cable entering a degraded state with a probability when the wind speed is greater than thirty meters per second; dynamic propagation rate modeling, adjusting the propagation rate based on real-time environmental data, such as using a wind speed impact coefficient of 0.15; spatial heterogeneity factor, calculating the geographical risk weight by the distance between the device and the threat source, such as a high-risk radius of fifty kilometers; hierarchical recovery mechanism, dividing the recovery process into emergency repair with a duration of two to four hours, temporary recovery for four to eight hours, and permanent repair for twenty-four to forty-eight hours, and simulating the recovery process through state transition equations, accurately deducing the probability of cascading failures up to seventy-one point three percent in a cold wave scenario.

[0045] The improved SIR model more realistically simulates the failure propagation process through multi-state expansion and dynamic propagation rate modeling, enhancing the accuracy and practicality of risk deduction. The model considers spatial heterogeneity and hierarchical recovery mechanisms, reflecting the complexity of actual operation and maintenance scenarios, thereby supporting more effective emergency planning and resource scheduling.

[0046] The method applied to the typhoon disaster defense scenario includes: Matching typhoon wind speed, optical cable suspension height, and business redundancy genes through risk gene mapping; Predicting the cascading failure path using a multi-modal spatio-temporal large model; Generating an optimal defense strategy and displaying the risk heat map in real-time through a three-dimensional visualization platform.

[0047] The implementation of the method applied to the typhoon disaster defense scenario includes matching typhoon wind speed, such as environmental gene wind speed greater than thirty meters per second, optical cable suspension height, such as device gene height greater than fifty meters, and business redundancy genes, such as business gene without backup routes, triggering associated rules, such as when the wind speed is greater than thirty-two point six meters per second and the optical cable suspension height is greater than fifty meters, the optical cable dancing risk is high at level three; using a multi-modal spatio-temporal large model to predict the cascading failure path, such as inputting the time series features of the past six hours of optical cable strain value and spatial features of the typhoon center distance, and outputting the risk heat map identifying high-risk node risk values greater than zero point eight five and the propagation chain caused by the typhoon landing leading to optical cable fracture and business interruption; generating an optimal defense strategy, such as enabling microwave links and dispatching de-icing robots through reinforcement learning decision-making, and displaying the risk heat map in real-time through a three-dimensional visualization platform, and the actual measurement improves the failure warning lead time to five point eight hours.

[0048] In the typhoon disaster prevention scenario, the method realizes early risk identification and accurate early warning through gene map matching and model prediction, greatly improving the disaster prevention capability of the power grid. In terms of effect, it can optimize the defense strategy, reduce disaster losses, and enhance business continuity.

[0049] The method applied to the extreme cold disaster prevention scenario includes: Match the icing thickness, low temperature, and optical cable bearing genes through the risk gene map; Use Monte Carlo simulation to deduce the probability of cascading failure; Output the deicing and routing switching strategy through the reinforcement learning decision module; Support dynamic updating of the gene map, add extreme freezing related rules and adjust the weight.

[0050] The specific implementation of the method applied to the extreme cold disaster prevention scenario includes matching the icing thickness, such as the environmental gene thickness being greater than fifteen millimeters, the low temperature, such as the environmental gene temperature being lower than minus ten degrees Celsius, and the optical cable bearing gene, such as the equipment gene bearing being greater than eighty kilograms, activating rules, such as the icing thickness being twenty-two millimeters, the initial value of the fracture probability being fifty-three point five five percent, and after temperature correction, rising to eighty-nine point four five percent; using Monte Carlo simulation to deduce the probability of cascading failure, based on the improved SIR model to simulate the optical cable fracture, the business packet loss rate rises to sixty-five percent, and the key path optical cable A fracture causes the communication interruption of substation B; outputting the deicing and routing switching strategy through the reinforcement learning decision module, such as dispatching the deicing robot and switching to the microwave link, the expected risk reduction is fifty-eight percent; supporting dynamic updating of the gene map, adding extreme freezing related rules and adjusting the weight, the actual emergency response time is shortened from forty-five minutes to eight minutes.

[0051] In the extreme cold prevention scenario, the method supports rapid risk response and strategy adjustment through dynamic gene updating and simulation deduction, effectively preventing cascading failures. It improves the adaptability of the power grid to extreme weather, ensures the stable operation of critical business, and enhances the resilience and sustainability of the overall power grid.

[0052] Embodiment 2 Based on the risk gene and spatiotemporal model, the power communication network risk assessment method in embodiment 1 constructs a three-layer graph structure of device vulnerability genes, environmental threat genes, and business dependency genes, defines cross-spatiotemporal risk propagation rules, combines Transformer and spatiotemporal graph convolution network (ST-GCN) to fuse device time series data, environmental spatial grid, and business topology features, uses Monte Carlo simulation and deep reinforcement learning to realize dynamic risk probability calculation and optimal disposal strategy generation, and displays the risk heat map and emergency plan in real time through a three-dimensional visualization platform.

[0053] The system includes three levels of "data awareness layer / intelligent analysis layer / decision application layer" and nine modules of "multi-source access gateway / space-time alignment engine / edge computing node / risk gene map construction module / multi-modal space-time large model / dynamic risk assessment engine / three-dimensional situation awareness platform / intelligent decision support / API service gateway", supporting multi-source data space-time alignment, dynamic update of gene map and online policy optimization.

[0054] The biological gene concept is introduced into the risk assessment of power communication transmission network for the first time, and a multi-modal space-time large model is used to solve the problem of complex multi-dimensional risk comprehensive evaluation, breaking through the limitations of traditional risk assessment methods, significantly improving the risk assessment accuracy and defense efficiency, and greatly reducing the power outage accidents caused by power grid cascading failures.

[0055] As shown in Figure 1 The power communication network risk assessment method based on risk genes and space-time large models uses a risk quantification evaluation system, adopts a hierarchical distributed architecture, including three levels and nine modules: 1) Data awareness layer: real-time collection and preprocessing of multi-modal heterogeneous data, establishing a unified space-time reference, including multi-source access gateway / space-time alignment engine / edge computing node.

[0056] 2) Intelligent analysis layer: risk gene modeling, multi-modal fusion analysis, and dynamic risk assessment, including risk gene map construction module / multi-modal space-time large model / dynamic risk assessment engine.

[0057] 3) Decision application layer: providing visual interaction and automated response, including three-dimensional situation awareness platform / intelligent decision support / API service gateway.

[0058] The three levels of the system / nine modules and their interaction logic are as follows: Data awareness layer: real-time collection and preprocessing of multi-modal heterogeneous data, establishing a unified space-time reference.

[0059] 1) Multi-source access gateway: supports protocols: MQTT / Modbus (device data), Kafka (log stream), API (weather / geographic information). Data source types: Device status data: optical transmission device error rate, time delay, CPU load Network traffic data: SDH / OTN link traffic matrix, DDoS attack features Environmental data: meteorological warning (typhoon path, rainfall), geological disaster monitoring (deformation sensor) Business data: communication business SLA level, routing dependency table 2) Space-time alignment engine: unified geographic space mapping, based on PTP protocol to achieve nanosecond-level time alignment.

[0060] 3) Edge computing node: Deploy lightweight anomaly detection models to achieve data cleaning and missing value filling.

[0061] Intelligent analysis layer: Realize risk gene modeling, multi-modal fusion analysis and dynamic risk assessment.

[0062] 1) Risk gene graph construction module: Knowledge extraction: Extract entities (such as "optical amplifier" and "lightning intensity") from device manuals and fault case libraries, and define gene attributes: vulnerability genes (device inherent attributes), threat genes (external environment), and propagation genes (topological connections).

[0063] Graph storage: Use graph database storage. Example relationship: (Base station A)-[Vulnerability: optical fiber aging rate > 25%]->(Risk event: optical path interruption).

[0064] The risk gene graph construction module further comprises: Gene dynamic update unit, for identifying unrecorded risk patterns through anomaly detection algorithm, and triggering graph structure adjustment based on node embedding similarity; Online verification interface, supporting A / B test to compare the risk prediction accuracy difference between new and old graph versions, and rolling back to the historical stable version when the difference exceeds 5%.

[0065] 2) Multi-modal spatio-temporal large model Model structure: Input layer: Receive and preprocess multi-modal data, unify spatio-temporal reference. Including device time series data interface, environment space grid interface, business topology interface, introducing spatio-temporal alignment engine, completing data normalization.

[0066] Multi-modal encoding layer: Extract features of each modality to generate unified embedding representation. Including device time series encoder, environment space encoder, business topology encoder, introducing cross-modal projection, realizing spatio-temporal position encoding.

[0067] Spatio-temporal fusion layer: Multi-modal feature interaction and joint modeling. Including cross-modal gated attention mechanism and spatio-temporal graph convolution network (ST-GCN), introducing sparse attention, realizing residual connection to prevent gradient vanishing.

[0068] Prediction and optimization layer: Risk quantification calculation and strategy generation. Including deep learning prediction and global optimization, based on trained model output long-term risk trend and key vulnerable points.

[0069] Output layer: Result adaptation and visualization interface. Including node-level risk assessment, global risk heat map, strategy execution interface, realizing three-dimensional geographic information rendering and dynamic threshold adjustment.

[0070] Training strategy: Pre-training phase: Train the base model using historical failure data.

[0071] Online learning: Dynamically update model parameters through EWC.

[0072] 3) Dynamic risk assessment engine Composed of five sub-modules: Data preprocessing module: multi-modal data cleaning, alignment, feature encoding Monte Carlo simulation module: cascading failure simulation based on risk genes and improved SIR model Reinforcement learning decision module: resource scheduling strategy generation and optimization Model update module: incremental training and dynamic gene map update API service gateway: provides risk assessment result query and strategy execution interface Three processing stages: Stage 1: Basic risk value calculation Stage 2: Monte Carlo risk propagation simulation, using an improved SIR model to simulate risk diffusion Stage 3: Deep reinforcement learning strategy optimization Decision application layer: provides visual interaction and automated response.

[0073] 1) Three-dimensional situational awareness platform: realize geographic information superposition. Dynamic rendering elements: risk heat map (red / yellow / green three-color early warning).

[0074] 2) Intelligent decision-making assistance: Emergency plan library: associate risk level and disposal strategy (such as P1 level risk triggering automatic routing switching) Resource scheduling optimization: generate optimal repair path based on genetic algorithm (reduce average response time) 3) API service gateway: provide RESTful interface to support integration with power grid dispatching system (EMS) and operation and maintenance work order system.

[0075] The risk quantification evaluation method proposed by the present application comprises the following steps: Step S1: Obtain device sensor data, meteorological grid data and business topology data through a multi-source data acquisition module, perform spatio-temporal alignment and feature encoding; Step S2: Construct a risk gene map, define the vulnerability quantification rules of device gene nodes, the threat exposure calculation model of environmental gene nodes and the criticality correction factor of business gene nodes; Step S3: Train a multi-modal spatio-temporal model, use a cross-modal attention mechanism to fuse device time series features, environmental spatial features and business topology features, and capture risk transmission rules through a spatio-temporal graph convolution network; Step S4: Based on the dynamic risk quantification evaluation module, perform Monte Carlo simulation to predict the probability of cascading failures, and optimize the resource scheduling strategy using deep reinforcement learning; Step S5: Output the risk heat map and treatment scheme through the three-dimensional visual decision module, and update the risk gene atlas weight and large model parameters according to the online feedback data.

[0076] The training method of the multi-modal spatio-temporal large model in step S3 includes: 1) Pre-training phase: initialize model parameters using historical failure data, and reconstruct multi-modal input data using a mask autoencoder; 2) Online learning phase: constrain the parameter update direction through the elastic weight consolidation (EWC) algorithm to retain the memory ability for historical risk patterns.

[0077] The risk gene atlas is a three-dimensional atlas composed of three types of risk genes: device genes, environment genes, and business genes, including environment-device interaction, device-device interaction, device-business interaction, and business-business interaction. The risk propagation logic is encoded through different combination rules of three-dimensional genes, realizing global coupling analysis of device-environment-business.

[0078] The risk gene list includes the following examples: 1) Device genes: hardware aging genes, performance degradation genes, etc. 2) Environment genes: meteorological disaster genes, geological risk genes, human threat genes, etc. 3) Business genes: business criticality genes, topology dependence genes, redundancy genes, etc.

[0079] Risk gene atlas: presents the correlation relationship of device genes, environment genes, and business genes.

[0080] The relevant definitions are as follows: Risk gene / event: Primary gene: inherent attribute of device / environment / business, such as "average wind speed ≥ 20 m / s".

[0081] Secondary gene: triggered by primary gene and transformed into secondary gene, such as primary gene "average wind speed ≥ 20 m / s" triggering primary event "optical cable dancing", and transformed into secondary gene "optical cable mechanical stress > safety limit", causing subsequent secondary event "optical cable fracture".

[0082] Primary event: triggered directly by primary gene.

[0083] Secondary event: triggered by secondary gene transformed from primary event.

[0084] Risk gene pairing: device-environment interaction pairing, device-business interaction pairing, environment-business interaction pairing, causal relationship pairing, dependency relationship pairing, AND logic relationship pairing, OR logic relationship pairing, etc., such as "typhoon wind speed 12 (environmental gene) → optical cable dancing risk (device gene)" relationship pairing.

[0085] Risk gene relationship chain: composed of multiple paired risk genes, including causal relationship chain, dependency relationship chain, etc., such as "high wind speed (primary gene) → optical cable dancing (primary event / secondary gene) → optical cable fracture (secondary event / secondary gene) → business interruption (secondary event)" composed of risk gene causal relationship chain.

[0086] The relevant operation mechanism is as follows: Risk gene activation: for example, typhoon wind speed gene (environment) + optical cable suspension height gene (device) → trigger "optical cable three-level dancing" risk event.

[0087] Risk relationship transmission: for example, high wind speed (primary gene) → optical cable dancing (primary event / secondary gene) → optical cable fracture (secondary event / secondary gene) → business interruption (secondary event).

[0088] Risk gene storage: store risk rules through risk gene map, predict risk propagation through multi-modal spatio-temporal large model, and generate disposal strategy (such as route switching).

[0089] Risk map update: new risk mode (such as no backup route) triggers gene map expansion, and optimizes defense strategy through reinforcement learning. The risk gene map is shown in Figure 2 .

[0090] 1) Topology type: directed attribute graph 2) Node type: Environment gene node (rhombus): represents external environmental threats, examples: typhoon level, lightning density, geological disaster index.

[0091] Device gene node (circle): represents the inherent risk attributes of the device, examples: optical cable aging rate, power module MTBF, heat dissipation efficiency.

[0092] Business gene node (hexagon): represents business dependency relationship, examples: dispatch telephone business priority, protection control channel redundancy.

[0093] 3) Edge type: Causal relationship edge (solid arrow): represents the risk triggering relationship, examples: high wind speed (environmental gene) → causes → optical cable dancing (device gene / risk event).

[0094] Dependency edge (dashed arrow): represents the business impact path, example: node betweenness centrality (business gene) → impact → protection misoperation risk (business gene).

[0095] 4) Technology feature annotation Dynamic weight identification: edge weight represents the correlation strength through numerical annotation and color depth, example: the edge weight between optical cable aging rate > 40% and optical cable fracture risk is 0.73, dark red Gene combination rule: define the "gene mutation" condition: when 3 or more related genes exceed the threshold at the same time, trigger high-risk events, example: [communication interruption risk] + [differential protection service packet loss] + [business interruption impact expansion] = protection misoperation high risk.

[0096] Dynamic update: real-time receive device state data, update gene correlation weight through online learning, update cycle: incremental graph optimization is performed every 15 minutes.

[0097] 5) Comparison of causal relationship / dependency relationship edges and differentiated application in risk assessment The comparison of causal relationship edges and dependency relationship edges is as follows: Dimension Causal Relation Edge Dependency Relation Edge Definition Represents a one-way action relationship where a certain risk gene (cause) directly triggers or exacerbates another risk event (effect) Represents the mutual influence between two risk genes in terms of business or function, reflecting the indirect link of risk transmission Directionality One-way arrow (from cause to effect), clearly indicating the direction of risk transmission Bidirectional or undirected, indicating mutual dependence or co-occurrence Storage attributes - Trigger threshold (e.g. temperature > 45°C) - Confidence (0~1) - Time decay factor (λ = 0.1 / day) - Impact coefficient (e.g. business interruption impact expands 3.2 times) - Redundancy identifier (True / False) - Topological weight Update mechanism Adjust confidence dynamically according to real-time monitoring data (e.g. certain causal relationship triggers 3 times in a row, confidence +0.15) Automatically update with changes in business topology (e.g. after adding a backup route, the dependency relationship impact coefficient decreases) Application of causal relationship edges Risk tracing: locate the root cause of the risk event (such as the root cause of optical cable fracture is typhoon + aging).

[0098] Propagation prediction: predict the cascading failure path based on the causal chain (such as typhoon → tower tilt → optical cable fracture).

[0099] Dynamic alarm: trigger early warning when the causal chain confidence exceeds the threshold (such as confidence > 0.8, then upgrade to P1 event).

[0100] Application of dependency relationship edges Business impact analysis: calculate the interruption probability of key business (such as 3 optical cables dependent on dispatch telephone business are all interrupted).

[0101] Redundancy design verification: check whether the dependency path meets the N-1 redundancy requirement (such as a single path dependent on a certain business is marked as high risk).

[0102] Resource scheduling optimization: preferentially protect high-dependence coefficient business nodes (such as P0 level business dependent power supply module needs to be preferentially backed up).

[0103] Example of risk gene list 1. Device vulnerability gene: Device vulnerability genes include: hardware aging gene, performance degradation gene, etc. Primary genes, also include physical change amplitude gene, etc. Secondary genes / events, specific examples are as follows: 1) (Primary gene) Hardware aging gene: cable aging rate, power module life, chip cumulative working time.

[0104] 2) (Primary gene) Performance degradation gene: optical power decay, time delay jitter, bit error rate.

[0105] 3) (Secondary gene / event) Physical change amplitude gene: line dance 2, Environmental threat gene 1) (Primary gene) Meteorological disaster gene: typhoon wind speed, lightning density, ice thickness.

[0106] 2) (Primary gene) Geological risk gene: surface deformation rate, seismic intensity.

[0107] 3) (Primary gene) Human threat gene: network attack frequency, construction damage risk 3, Business dependency gene 1) (Primary gene) Business criticality gene Gene name Quantitative index Classification threshold (example) Data source SLA level Business assurance level P0 (protection control), P1 (dispatch phone), P2 (video monitoring) Business management system Business traffic peak Bandwidth utilization rate (%) Low (<60), Medium (60-90), High (>90) Traffic probe 2) (Primary gene) Topology dependency gene Gene name Quantitative index Classification threshold (example) Data source Betweenness Centrality Betweenness Centrality Low (<0.3), Medium (0.3-0.6), High (>0.6) Network topology analysis Node degree Number of connection links Low (<3), Medium (3-5), High (>5) Network topology database 3) (Primary gene) Redundancy gene Gene name Quantitative index Classification threshold (example) Data source Backup route availability Number of redundant paths Low (no redundancy), high (N+1 redundancy) SDN controller state query Power redundancy configuration UPS / battery backup time (h) Low (<2), Medium (2-4), High (>4) Power monitoring system The multi-modal spatio-temporal large model is composed of five layers of input layer, multi-modal encoding layer, spatio-temporal fusion layer, prediction and optimization layer, and output layer, adopts the architecture of Transformer+spatio-temporal graph convolutional neural network (ST-GCN), fuses time series feature extraction and spatial topology analysis, and proposes a spatio-temporal attention mechanism to solve the modeling problem of cross-regional risk transmission of power communication networks. The framework of the multi-modal spatio-temporal large model (as shown in Figure 3 ): input layer → multi-modal encoding layer → spatio-temporal fusion layer → prediction and optimization layer → output layer.

[0108] The data input and output and core operations among the five levels of the multi-modal spatio-temporal large model are shown in the following table: Hierarchy Input dimension Output dimension Core operation Input layer Raw multi-modal data Aligned tensor / matrix Unified and normalized in time and space Multi-modal encoding layer B×T×D / B×H×W×C / Business GraphML topology matrix B×Seq×512 CNN / LSTM / GraphSAGE Spacetime fusion layer B×Seq×512 B×N×512 Cross-modal attention + ST-GCN Prediction and optimization layer B×N×512 B×N×1 (probability) Deep learning prediction + global optimization algorithm Output layer B×N×1 API / visualization instructions Sigmoid / deconvolution / JSON packaging Embodiment 3 On the basis of embodiment 1: the technical scheme of the present application is completely explained in detail, richly and coherently through the specific implementation mode. The implementation mode covers the whole process details of system architecture, data collection, model construction, evaluation engine to application scene, ensuring the integrity and implementability of the technical scheme.

[0109] 1. System overall architecture and data collection implementation mode The system adopts a hierarchical distributed architecture, including a data perception layer, an intelligent analysis layer, and a decision application layer. The data perception layer collects multi-modal heterogeneous data of the power communication transmission network in real time through a multi-source access gateway, including: Device sensor data: optical transmission device error rate (e.g., BER value below 1e-9 is normal), latency (standard deviation less than 5ms is low risk), CPU load (percentage), memory usage, device temperature, etc. Data is collected through MQTT / Modbus protocol, with a sampling frequency of 1 second / second.

[0110] Weather grid data: typhoon wind speed (maximum sustained wind speed classification, e.g., less than 17.2m / s is low risk), rainfall (mm / h), lightning density (times / square kilometer), surface deformation rate (mm / year), etc. Data comes from the meteorological bureau API and lightning positioning system, with a spatial resolution of 1km x 1km grid and an update frequency of 5 minutes / second.

[0111] Business topology data: SDH / OTN link flow matrix, business SLA level (e.g., P0 level for protection control business), routing dependency table, node betweenness centrality (e.g., greater than 0.6 is high risk), etc. Data format is GraphML, containing adjacency matrix and edge attributes (e.g., bandwidth, real-time traffic, latency).

[0112] The spatio-temporal alignment engine realizes nanosecond-level time synchronization based on the PTP protocol, maps device coordinates (longitude, latitude) to three-dimensional Cartesian coordinates, and unifies the spatial reference through GeoHash encoding (12 bits). The edge computing node deploys a lightweight anomaly detection model (e.g., based on LSTM), performs data cleaning (outlier removal), missing value filling (sliding window mean interpolation), and normalization processing (Min-Max standardization to the range [0, 1]), ensuring accurate alignment of multi-modal data in the spatio-temporal dimension. The data perception layer outputs the aligned tensor, such as device time series data as a B x T x D dimensional tensor (B is the batch size, T is the time step, and D is the feature dimension), and environmental spatial data as a B x H x W x C dimensional tensor (H and W are the spatial grid height and width, and C is the channel number).

[0113] 2. Risk gene map construction implementation The risk gene map construction module extracts entity knowledge from device manuals, fault case libraries, and historical event libraries, establishing a three-layer graph structure: Equipment vulnerability genes include hardware aging genes (e.g., optical cable aging rate = current attenuation / design value, thresholds are low <20%, medium 20-40%, high >40%), performance degradation genes (e.g., optical power attenuation normal range -3~-15dBm), and physical change amplitude genes (e.g., line galloping classification: Level 1 galloping vibration amplitude ≤10cm, frequency ≤1 time / second). Gene attributes are encoded through metadata, for example, gene ID is "DEV_GENE_001", quantification formula is "optical cable aging rate = current attenuation / design value", and risk threshold is defined as ">40% triggers warning".

[0114] Environmental threat genes include meteorological disaster genes (such as typhoon wind speed classification: low <17.2m / s, medium 17.2-32.6m / s, high >32.6m / s), geological risk genes (such as the ground surface deformation rate threshold <5mm / year is low risk), and human threat genes (such as the network attack frequency threshold <10 times / day is low risk).

[0115] Business-dependent genes: including business-critical genes (such as SLA levels P0, P1, P2), topology-dependent genes (such as node betweenness centrality threshold >0.6 indicating high risk), and redundancy genes (such as backup route availability indicators True / False).

[0116] The graph is stored using a graph database (such as Neo4j), and node types include environmental gene nodes (diamonds), device gene nodes (circles), and business gene nodes (hexagons). Edge relationships include causal edges (solid arrows, storing trigger thresholds, confidence levels, and time decay factors) and dependency edges (dashed arrows, storing influence coefficients, redundancy indicators, and topological weights). Gene association rules are defined using the Cypher query language, for example: MATCH (e1:EnvGene {id:"ENV_GENE_001"}) WHERE e1.value>30 MATCH (d:DeviceGene {id:"DEV_GENE_001"}) WHERE d.value>40 CREATE (e1)-[:CAUSES {confidence: 0.85}]->(risk:RiskEvent {name: "Optical Cable Dancing Risk"}) The gene dynamic updating unit identifies unrecorded risk patterns through an anomaly detection algorithm (such as similarity calculation based on node embedding, similarity <0.7 to create a new node) and sets up an online verification interface for A / B testing. When the difference in risk prediction accuracy between the new and old graph versions exceeds 5%, it is rolled back to the historical stable version. The graph supports gene pairing combination rules, including device-environment interaction (such as "typhoon wind speed >30m / s AND optical cable suspension height >50m → optical cable dancing risk"), device-service interaction (such as "node betweenness centrality >0.6 AND no backup route → business interruption impact expands 3.2 times"), and environment-service interaction pairing.

[0117] 3. Multimodal spatiotemporal large model construction and training implementation The multimodal spatiotemporal large model adopts a five-layer architecture, which is implemented as follows: Input layer: receives multimodal data, including device time series data (BxTxD dimensional tensor, T=12 time steps corresponding to past 60 minutes of data, D=5 features such as optical power, bit error rate), environmental spatial data (BxHxWxC dimensional tensor, H=50 latitude grids, W=60 longitude grids, C=4 channels such as wind speed, rainfall), and business topology data (adjacency matrix and flow matrix in GraphML format). The input layer unifies the reference through a spatiotemporal alignment engine, for example, the time axis is aligned to the UTC time zone using UNIX timestamps, and the spatial axis is mapped through GeoHash encoding.

[0118] Multimodal encoding layer: dedicated encoders are used to extract features: device time series encoder uses 1D-CNN (convolution kernel size 3, step 1) and LSTM (hidden layer 512 dimensions) to extract local temporal features, output dimension BxTx512; environmental spatial encoder uses ConvLSTM (convolution kernel 5x5) and spatial attention mechanism to extract spatiotemporal features, output flattened to Bx(HxW)x512; business topology encoder uses GraphSAGE (aggregation function is mean pooling) and node betweenness calculation to generate node embedding, output dimension BxNx512 (N is the number of nodes). The encoding layer aligns the embedding dimension to d_model=512 through cross-modal projection (linear layer) and injects spatiotemporal position encoding: time position encoding uses the sine function PE(t,2i)=sin(t / 10000^(2i / d_model)), and spatial position encoding maps GeoHash encoding to a vector through an Embedding layer.

[0119] Space-time fusion layer: improved cross-modal multi-head attention mechanism (8 heads), each head focuses on specific modal interaction, e.g. head 1 calculates the attention weight of device Q and environment K / V, the gating parameter is a learnable variable (initial value 0.5). Hierarchical Transformer architecture includes basic layer (4 layers of standard multi-head attention + LayerNorm + FFN), cross-modal fusion layer (2 layers of gated attention) and space-time aggregation layer (1 layer of spatial and temporal pooling). Transformer and ST-GCN collaboration: Transformer handles long-range dependencies (such as typhoon wind field data), ST-GCN captures local topological features through spatial convolution (kernel 3x3) and temporal convolution (kernel 3), and outputs fused features BxNx512.

[0120] Prediction and optimization layer: based on deep learning to predict risk distribution, using fully connected layer and Sigmoid activation to output node-level failure probability (BxNx1); combined with global optimization algorithm (such as genetic algorithm to optimize device replacement plan, linear programming to allocate bandwidth), the objective function is to minimize operation and maintenance cost and maximize risk reduction.

[0121] Output layer: reconstruct the spatial grid to generate a global risk heat map (BxHxWx1) through deconvolution, and output the JSON format result through RESTful API, such as {"risk_level":"P2", "affected_nodes":["base station A"],"confidence":0.88}.

[0122] Model training uses a two-stage method: pre-training stage uses historical failure data (n samples), reconstructs the input through a masked autoencoder (mask rate 15% device data, 20% environment data), and the loss function is MSE reconstruction loss; fine-tuning stage uses elastic weight consolidation (EWC) algorithm to constrain important parameter updates, and the loss function is 0.6 cross-entropy loss (risk classification) + 0.3 MAE loss (strategy regression) + 0.1 * graph reconstruction loss. Online learning mechanism triggers incremental update every 15 minutes, using a ring buffer to store the latest 24 hours of data.

[0123] 4. Dynamic risk assessment engine implementation The dynamic risk assessment engine uses a microservice architecture, including the following sub-modules: Data preprocessing module: performs multi-modal data cleaning (based on 3σ rule to remove outliers), alignment (dynamic time warping algorithm) and feature encoding (one-hot encoding of categorical variables).

[0124] Monte Carlo simulation module: Perform cascading failure simulation based on risk gene map and improved SIR model. The improved SIR model is extended to multiple states (normal S, degraded D, failure I, and recovery R), and the state transition probability is P(S→D)=1-e^(-λ_dΔt), where λ_d=α·W(t)+β·R(t) (W(t) is the real-time wind speed, and R(t) is the rainfall intensity). The dynamic transmission rate is modeled as β(t)=β_0·[1+γ·w(t) / W_threshold] (γ=0.15, W_threshold=25m / s). The spatial heterogeneity factor is ρ(x,y)=1 / (1+λ·(D(x,y)-D_0)) (D_0=50km, λ=0.1). The hierarchical recovery mechanism includes emergency repair (probability 0.3, time consumption 2-4h), temporary recovery (probability 0.5, time consumption 4-8h), and permanent repair (probability 0.2, time consumption 24-48h). The simulation module performs 10,000 Monte Carlo simulations in parallel, and outputs the node failure probability distribution and key propagation paths (such as "Base Station A→Optical Cable X→Substation B").

[0125] Reinforcement learning decision module: Adopt Proximal Policy Optimization (PPO) algorithm, state space includes device risk value matrix, resource deployment state (repair truck location, backup route bandwidth), action space includes discrete action (route switching, device restart) and continuous action (power adjustment amplitude -20%~+20%). Reward function is Reward=w1·∑ΔR_reduced-w2·Cost_action-w3·T_downtime (weights w1=0.6, w2=0.3, w3=0.1).

[0126] Model update module: Implement incremental training, preserve historical pattern memory through elastic weight solidification; gene map dynamic update based on node embedding similarity, add new gene node when similarity <0.7.

[0127] API service gateway: Provide RESTful interface, such as POST / risk / assessment query risk assessment, return JSON work order instruction.

[0128] 5. Typical application scenario implementation Application scenario one: Typhoon disaster prevention Data input: Real-time typhoon wind speed 52m / s, rainfall 50mm / h, device optical cable suspension height 55m, business redundancy=0 (no backup route).

[0129] Genetic profile activation: matching environmental gene ENV GENE 001 (wind speed > 30 m / s), device gene DEV GENE 001 (suspension height > 50 m), business gene BIZ GENE 002 (no backup route), trigger rule chain: IF wind speed > 32.6 m / s AND suspension height > 50 m THEN cable dance risk = high risk level 3 -> cable break risk > 79% -> service interruption risk > 83%.

[0130] Model prediction: multi-modal spatio-temporal large model inputs time series features (past 6 hours cable strain values), spatial features (typhoon center distance), outputs risk heat map identifies high-risk nodes (risk value > 0.85), and cascading failure path is "typhoon landing -> cable dance -> cable break -> substation B communication interruption -> protection misoperation".

[0131] Decision and execution: reinforcement learning outputs optimal strategy (switch backup route + deploy UPS power supply vehicle), expected risk reduction is 58%, and cost is 150,000 yuan. Switching routes through API calls to SDN controllers (latency < 200 ms), and real-time risk heat map is displayed on a three-dimensional visualization platform (red zone is typhoon eye wall area).

[0132] Application scenario two: extreme cold disaster prevention Data input: temperature -15.3℃, ice thickness 22mm, cable bearing 85kg (design limit 100kg).

[0133] Genetic profile activation: matching environmental genes ENV GENE 003 (ice thickness > 15mm), ENV GENE 004 (temperature <-10℃), and device gene DEV GENE 003 (bearing > 80kg), triggering rules: when ice thickness is 22mm, the initial breakage probability is 53.55%, after temperature correction (probability +3% per 1℃ decrease) and wind speed correction (+20%), it rises to 89.45%.

[0134] Monte Carlo reasoning: improved SIR model simulates that after cable breakage, service packet loss rate rises to 65%, and cascading failure probability is 71.3%, and the path is "cable A breakage -> substation B communication interruption -> line C overload".

[0135] Decision and execution: reinforcement learning outputs deicing and route switching strategy (dispatches deicing robots + enables microwave link), risk reduction is 58%. Genetic profile dynamically adds rule RULE ICE 002 (triggers PO level risk when ice > 20mm and bearing > 90kg).

[0136] 6. Visualization decision and output implementation The decision application layer dynamically superimposes geographic information (tower GPS coordinates, terrain elevation) through a three-dimensional situation awareness platform (based on WebGL rendering), and the risk heat map adopts red (risk value > 0.8), yellow (0.5-0.8) and green (<0.5) three-color early warning. The intelligent decision assistance module is associated with an emergency plan library (such as P1 level risk triggering automatic routing switching), and generates an optimal repair path (minimizes response time) based on a genetic algorithm. The API service gateway is integrated with a power grid dispatching system (EMS) through a gRPC protocol, and outputs a work order instruction (JSON format) Through the specific embodiments described above, the present application realizes full-chain automation from data collection, risk modeling, dynamic evaluation to decision output, and significantly improves the risk prevention and control capability of the power communication network.

[0137] The above-described embodiments are more detailed and specific, express the preferred embodiments of the present application, and are only used to illustrate the technical ideas and characteristics of the present application. The purpose is to enable those skilled in the art to understand the content of the present application and implement it, but it is not limited to the present application, and the patent range of the present application cannot be limited to the present embodiment. That is, any equivalent changes or modifications made within the spirit disclosed by the present application, without departing from the structure of the present application, the internal improvement of the system and the change between the subsystems, and the transformation, etc., are still within the patent range of the present application.

Claims

1. A risk assessment method for power communication networks based on risk genes and a spatiotemporal large model, characterized in that, Includes the following steps: The multi-source data acquisition module acquires equipment sensor data, meteorological grid data, and service topology data of the power communication transmission network in real time, and performs spatiotemporal alignment and feature encoding. Construct a risk gene map, which includes a three-layer map structure of equipment vulnerability genes, environmental threat genes, and business dependence genes, and define cross-temporal and spatial risk propagation rules; The training of a multimodal spatiotemporal large model adopts an architecture that combines Transformer and spatiotemporal graph convolutional network, and fuses device temporal features, environmental spatial features and business topology features through a cross-modal attention mechanism; Based on the dynamic risk quantification assessment module, Monte Carlo simulation is performed to predict the probability of cascading failures, and deep reinforcement learning is used to optimize resource scheduling strategies. The three-dimensional visualization decision-making module outputs risk heat maps and response plans, and dynamically updates the risk gene map weights and multimodal spatiotemporal large model parameters based on online feedback data.

2. The risk assessment method for power communication networks based on risk genes and a spatiotemporal large model according to claim 1, characterized in that, The risk assessment method is implemented through a system that adopts a layered distributed architecture, comprising three main layers: a data perception layer, an intelligent analysis layer, and a decision application layer. The data perception layer includes a multi-source access gateway, a spatiotemporal alignment engine, and edge computing nodes, which are used to collect and preprocess multimodal heterogeneous data in real time. The intelligent analysis layer includes a risk gene map construction module, a multimodal spatiotemporal large model, and a dynamic risk assessment engine, which are used to realize risk gene modeling and multimodal fusion analysis. The decision application layer includes a 3D situational awareness platform, intelligent decision support, and an API service gateway, which provide visual interaction and automated response.

3. The risk assessment method for power communication networks based on risk genes and a spatiotemporal large model according to claim 1, characterized in that, The steps for constructing the risk gene map include: Extract entity knowledge from equipment manuals and fault case databases, define gene attributes, and store graph relationships through a graph database; The risk gene map supports dynamic updates and identifies unrecorded risk patterns through anomaly detection algorithms. Gene association rules include pairings of device-environment interaction, device-business interaction, and environment-business interaction, and encode risk propagation logic through causal and dependency edges.

4. The risk assessment method for power communication networks based on risk genes and a spatiotemporal large model according to claim 1, characterized in that, The multimodal spatiotemporal large model consists of five layers: an input layer, a multimodal coding layer, a spatiotemporal fusion layer, a prediction and optimization layer, and an output layer. The input layer receives device timing data, environmental spatial grid data, and service topology data; The multimodal coding layer uses multiple encoders to extract features of each modality and generate a unified embedding representation; The spatiotemporal fusion layer employs an improved cross-modal multi-head attention mechanism and a spatiotemporal graph convolutional network to achieve multimodal feature interaction; The prediction and optimization layer uses deep learning to predict risk distribution and combines it with a global optimization algorithm to generate strategies. The output layer outputs node-level risk assessment results and generates a global risk heatmap.

5. The risk assessment method for power communication networks based on risk genes and a spatiotemporal large model according to claim 4, characterized in that, The spatiotemporal fusion layer employs an improved cross-modal multi-head attention mechanism, including: Each attention head focuses on a specific modal interaction and introduces learnable gating parameters to suppress irrelevant modal interference; It adopts a hierarchical Transformer architecture, including a base layer, a cross-modal fusion layer, and a spatiotemporal aggregation layer; The complementary mechanism of long-range dependencies and local topological features is achieved through the collaboration of Transformer and spatiotemporal graph convolutional network.

6. The risk assessment method for power communication networks based on risk genes and a spatiotemporal large model according to claim 4, characterized in that, The training of the multimodal spatiotemporal large model adopts a two-stage training method: During the pre-training phase, historical fault data is used to initialize model parameters, and multimodal input data is reconstructed through a masked autoencoder. During the fine-tuning phase, the parameter update direction is constrained by the elastic weight solidification algorithm, and the model parameters are dynamically updated using incremental learning.

7. The risk assessment method for power communication networks based on risk genes and a spatiotemporal large model according to claim 1, characterized in that, The dynamic risk quantification assessment module is implemented through a dynamic risk assessment engine, which includes: The data preprocessing module is used for multimodal data cleaning, alignment, and feature encoding; The Monte Carlo simulation module performs cascading failure simulations based on the risk gene map and an improved SIR model; The reinforcement learning decision-making module uses a proximal policy optimization algorithm to generate resource scheduling strategies. The model update module enables incremental training and dynamic updates of the gene map; The API service gateway provides interfaces for querying risk assessment results and executing policies.

8. The risk assessment method for power communication networks based on risk genes and a spatiotemporal large model according to claim 7, characterized in that, The improved SIR model used in the Monte Carlo simulation module includes: Multi-state extension divides the device state into four states: normal, degraded, fault, and recovery. Dynamic propagation rate modeling, adjusting the propagation rate based on real-time environmental data; Spatial heterogeneity factor, which calculates geographical risk weights based on the distance between the device location and the threat source; The tiered recovery mechanism divides the recovery process into three stages: emergency repair, temporary recovery, and permanent repair.

9. The risk assessment method for power communication networks based on risk genes and a spatiotemporal large model according to claim 1, characterized in that, When the method is applied to typhoon disaster prevention scenarios, it includes: By matching genes related to typhoon wind speed, optical cable suspension height, and business redundancy through risk gene mapping; Predicting cascading failure paths using a multimodal spatiotemporal large model; Generate the optimal defense strategy and display the risk heat map in real time through a 3D visualization platform.

10. The risk assessment method for power communication networks based on risk genes and spatiotemporal large models according to claim 1, characterized in that, When the method is applied to extreme cold wave disaster prevention scenarios, it includes: Match genes related to icing thickness, low temperature, and optical cable load-bearing capacity using risk gene mapping; Monte Carlo simulation was used to estimate the probability of cascading failures; The reinforcement learning decision module outputs de-icing and route switching strategies. It supports dynamic updates of the gene map, adds rules related to extreme freezing and adjusts the weights.

Images