Energy supply risk identification and early warning system based on artificial intelligence

Abstract

The application belongs to the technical field of energy supply risk identification, and more particularly to an energy supply risk identification and early warning system based on artificial intelligence, which takes a module manager as a core scheduling unit, realizes hot plug and dynamic loading of energy-specific risk identification modules by means of a built-in module registration list, a scheduling rule library and a standardized configuration file; secondly, the energy-specific risk identification module is configured with a special identification unit and an AI model optimized by field data training for each energy, which can accurately extract energy-specific risk features and output evaluation values; and then by embedding a cross-energy general risk factor library in the general risk assessment module, analyzing the influence of general factors such as policies, macroeconomics and extreme weather on different energies, integrating the general and exclusive risk assessment results through a weighted fusion algorithm, realizing dynamic identification and comprehensive analysis of cross-energy chain risks, and solving the problems that the existing system cannot capture risk transmission effects and the early warning is delayed.

Description

Technical Field

[0001] This application belongs to the field of energy supply risk identification technology, and more specifically, relates to an artificial intelligence-based energy supply risk identification and early warning system. Background Technology

[0002] Energy is the core support for national economic development and normal social operation. Its stable supply is directly related to national energy security, industrial production efficiency and people's livelihood. Among them, electricity, oil, natural gas and coal are the core energy types in my country's energy system, forming a multi-dimensional and multi-level energy supply network. The various energy sources are interconnected and mutually influential in the production, transmission, storage and consumption links, forming a complex energy ecosystem.

[0003] Against the backdrop of profound adjustments in the global energy landscape, frequent extreme weather events, escalating geopolitical conflicts, and continuous fluctuations in energy demand, the risk factors facing energy supply are becoming increasingly diversified and complex, placing higher demands on the accurate identification and timely early warning of energy supply risks. As a core technical link in the field of energy management, energy supply risk identification and early warning aims to capture potential risks and hidden dangers in various energy supply processes in advance, providing a scientific basis for energy dispatch, emergency response, and strategic decision-making, thereby minimizing the impact of risk events on the stability of energy supply.

[0004] Currently, the industry has developed a series of technical solutions and related systems for identifying and warning of energy supply risks. However, existing solutions and systems generally have limitations in their design, often being independently designed and developed for the supply characteristics of a single energy type, lacking a unified technical framework and collaborative analysis system covering multiple energy types such as electricity, oil, natural gas, and coal. Specifically, the supply risks of different energy types are driven by different core factors, resulting in significant limitations in the analytical dimensions and data sources of existing single-energy early warning systems. For example, electricity supply risks are mainly affected by extreme weather (such as typhoons, rainstorms, and freezing), peak fluctuations in electricity demand, grid equipment failures, and the instability of new energy power generation. Existing electricity risk early warning systems mostly build data collection and analysis models around these factors. Oil supply risks, on the other hand, are more related to geopolitical conflicts, international oil price fluctuations, and transportation risks. The risks associated with these systems are related to factors such as disruptions to shipping routes (e.g., blocked shipping routes, pipeline leaks) and refinery production failures. Corresponding early warning systems also focus on monitoring and assessing these risk factors. More importantly, various energy sources are not isolated but have a close chain reaction effect. The supply risk of a single energy source can easily spread to other energy sectors, triggering systemic risks across energy sources. For example, power outages may lead to the shutdown of energy processing facilities such as refineries and natural gas processing plants, resulting in shortages of oil and natural gas. Conversely, insufficient oil and natural gas supply can also affect the stability of thermal power generation and heat supply, forming a closed-loop risk transmission. Existing single-energy early warning systems lack the ability to integrate cross-energy data and analyze chain risks, making it difficult to effectively capture the risk transmission process across energy sources. This results in delayed early warnings and an inability to cope with complex systemic energy supply risks. Summary of the Invention

[0005] This invention provides an artificial intelligence-based energy supply risk identification and early warning system, which aims to solve the technical problem that current single energy early warning systems cannot effectively capture the risk transmission process across energy sources, resulting in delayed early warnings and difficulty in dealing with complex systemic energy supply risks.

[0006] This invention provides an energy supply risk identification and early warning system based on artificial intelligence. The system adopts a modular structure formed by combining a fixed core framework with pluggable energy-specific modules. Each module achieves bidirectional communication through a preset standardized data interaction interface. The system includes the following modules: Data Acquisition and Fusion Module: Configures multi-source data acquisition interfaces, collects heterogeneous data from multiple sources, cleans, standardizes, and extracts features from the collected data, generates a risk feature dataset in a unified format, and provides a standardized data structure for other modules to use; Module Manager: The core scheduling unit of the system, with a built-in module registration list and scheduling rule library. Based on the energy type to be monitored and the real-time risk scenario context, it dynamically loads the corresponding energy-specific risk identification module from the module library. The registration, adaptation and loading of new energy type modules are completed through standardized configuration files, realizing hot-swappable module expansion. Energy-specific risk identification module: It has a pluggable structure and is equipped with a dedicated risk identification unit for each energy type. Each dedicated risk identification unit has a built-in artificial intelligence model trained and optimized with data from the corresponding energy field. The model extracts the specific risk characteristics of the corresponding energy and outputs the risk assessment value. General Risk Assessment Module: It has a built-in cross-energy general risk factor library, which includes factors related to policy changes, macroeconomic fluctuations, and extreme weather. It analyzes the impact of each general factor on different energy sources to generate general risk assessment results. It interacts with the energy-specific risk identification module and uses a weighted fusion algorithm to integrate the two types of risk assessment results to output comprehensive risk data. Early warning generation module: It presets multi-level early warning judgment benchmarks and generates graded early warning signals based on the comprehensive risk data output by the general risk assessment module and the judgment benchmarks.

[0007] This invention uses a module manager as the core scheduling unit. Leveraging a built-in module registration list, scheduling rule base, and standardized configuration files, it enables hot-swapping and dynamic loading of energy-specific risk identification modules. This allows for adaptation to new energy types without requiring overall system reconstruction, effectively addressing the shortcomings of poor system adaptability and insufficient scalability. Secondly, the energy-specific risk identification module is equipped with a dedicated identification unit and an AI model trained and optimized with domain data for each energy type. This allows for accurate extraction of energy-specific risk characteristics and output of assessment values, overcoming the limitations of existing single-energy system analysis dimensions. Furthermore, by incorporating a cross-energy general risk factor library into the general risk assessment module, and combining the impact analysis of general factors such as policy, macroeconomics, and extreme weather on different energy types, a weighted fusion algorithm integrates general and specific risk assessments. The results enable dynamic identification and comprehensive assessment of cross-energy chain risks, solving the problems of existing systems being unable to capture risk transmission effects and experiencing early warning delays. Finally, the early warning generation module judges the comprehensive risk data based on multi-level early warning benchmarks, generating graded early warning signals to ensure the real-time and accuracy of early warnings. Therefore, through modular design and cross-energy collaborative analysis mechanisms, this invention not only achieves unified integration and efficient utilization of multi-source data, improving the comprehensiveness, accuracy, and real-time nature of risk identification, but also possesses strong scalability and adaptability, flexibly adapting to different energy types and complex risk scenarios. It effectively addresses cross-energy systemic risks, providing reliable support for energy dispatch, emergency response, and strategic strategies, significantly improving the intelligence and refinement of energy management, and ensuring stable energy supply.

[0008] Preferably, the data acquisition and fusion module has a built-in attention network. The attention network automatically assigns weights based on the historical correlation between each data source and the risk event. Based on the assigned weights, the data after cleaning, format standardization and feature extraction are weighted and fused to generate a unified dimension fused feature vector, thereby obtaining a risk feature dataset in a unified format.

[0009] Preferably, the module manager uses meta-learning technology to construct an intelligent module routing model. The intelligent module routing model is trained based on historical module call records, matching accuracy, processing latency, and risk identification accuracy data under different energy scenarios, and the model parameters are optimized through a scenario learning mechanism. The input of the intelligent module routing model is the concatenation result of the energy type feature vector and the context feature vector, where the energy type feature vector is the energy type to be detected, and the context feature vector is extracted from the metadata output by the data acquisition and fusion module. Based on this, the intelligent module routing model outputs the optimal module combination scheme. The optimal module combination scheme consists of an energy-specific risk identification module and zero or one general risk assessment module, and is generated based on the adaptability score. The adaptability score comprehensively considers the matching degree between the module and the energy type, the historical call success rate, and the processing efficiency.

[0010] Preferably, the energy-specific risk identification module adopts a layered architecture consisting of a basic model layer and a domain fine-tuning layer; The basic model layer is a pre-trained deep neural network. The training data covers historical risk data and domain knowledge of various energy sources. The input is the fusion feature vector output by the data acquisition and fusion module, which is used to extract common risk features of multiple energy sources and output a general risk feature vector. The domain fine-tuning layer serves as a dedicated risk identification unit, configuring exclusive models for different energy characteristics to adapt to the heterogeneous risk requirements of various energy sources.

[0011] Preferably, the energy-specific risk identification module introduces a cross-energy impact factor. The cross-energy impact factor is quantified based on historical risk chain event data, and is calculated by combining the conditional probability of risk events between energy sources with the impact intensity coefficient set by the domain. After normalization, it represents the intensity of the chain effect between different energy risks.

[0012] Preferably, the energy-specific risk identification module adopts a transfer learning mechanism, which uses the parameters obtained by polar training of the basic model layer in historical risk data of multiple energy types as initial parameters, and then uses the domain data of the current energy to be identified to specifically fine-tune the parameters of the domain fine-tuning layer, so as to realize the cross-energy reuse of the basic model layer parameters.

[0013] Preferably, the general risk assessment module has a built-in undirected weighted graph model. The undirected weighted graph model uses each energy type as a node and the correlation strength between energy sources as the edge weight to construct a multi-energy risk propagation network. A risk diffusion model is constructed by simulating the transmission path and impact of general risk factors between nodes. The general risk assessment module incorporates reinforcement learning technology to dynamically adjust the weight values ​​of each general risk factor based on feedback from historical risk handling effects.

[0014] Preferably, the general risk assessment module has a built-in risk coupling analysis model. The risk coupling analysis model identifies duplicate features in energy-specific risks and general risks through feature comparison algorithms, removes the identified duplicate features, and uses the analytic hierarchy process to balance the contribution ratio of specific risks and general risks, outputting a comprehensive risk score without redundancy.

[0015] Preferably, the early warning generation module integrates a locally interpretable model, which is a LIME model. By constructing a local linear model, the contribution of each risk factor to the comprehensive risk score is quantified, core influencing factors are screened, and an explanation is generated, wherein the explanation report includes the risk causes, propagation paths, scope of impact, and response suggestions.

[0016] Preferably, the early warning generation module has a built-in adaptive threshold adjustment mechanism. The adaptive threshold adjustment mechanism calculates the early warning accuracy, false alarm rate and economic loss amount of the corresponding risk event based on historical early warning data, and adjusts the judgment threshold of multi-level early warnings using a weighted calculation method.

[0017] The beneficial effects of the invention include: This invention uses a module manager as the core scheduling unit. Leveraging a built-in module registration list, scheduling rule base, and standardized configuration files, it enables hot-swapping and dynamic loading of energy-specific risk identification modules. This allows for adaptation to new energy types without requiring overall system reconstruction, effectively addressing the shortcomings of poor system adaptability and insufficient scalability. Secondly, the energy-specific risk identification module is equipped with a dedicated identification unit and an AI model trained and optimized with domain data for each energy type. This allows for accurate extraction of energy-specific risk characteristics and output of assessment values, overcoming the limitations of existing single-energy system analysis dimensions. Furthermore, by incorporating a cross-energy general risk factor library into the general risk assessment module, and combining the impact analysis of general factors such as policy, macroeconomics, and extreme weather on different energy types, a weighted fusion algorithm integrates general and specific risk assessments. The results enable dynamic identification and comprehensive assessment of cross-energy chain risks, solving the problems of existing systems being unable to capture risk transmission effects and experiencing early warning delays. Finally, the early warning generation module judges the comprehensive risk data based on multi-level early warning benchmarks, generating graded early warning signals to ensure the real-time and accuracy of early warnings. Therefore, through modular design and cross-energy collaborative analysis mechanisms, this invention not only achieves unified integration and efficient utilization of multi-source data, improving the comprehensiveness, accuracy, and real-time nature of risk identification, but also possesses strong scalability and adaptability, flexibly adapting to different energy types and complex risk scenarios. It effectively addresses cross-energy systemic risks, providing reliable support for energy dispatch, emergency response, and strategic strategies, significantly improving the intelligence and refinement of energy management, and ensuring stable energy supply. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 The overall system architecture diagram provided for embodiments of the present invention. Detailed Implementation

[0020] To make the technical problems, technical solutions, and beneficial effects to be solved by this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and are not intended to limit the scope of this application.

[0021] See Figure 1As shown, this embodiment provides an AI-based energy supply risk identification and early warning system. It adopts a modular architecture design, ensuring system stability and module compatibility through a fixed core framework, while adapting to heterogeneous risk requirements across multiple energy sources through pluggable energy-specific modules. Each module achieves bidirectional communication through a pre-defined standardized RESTful API interface, ensuring data consistency and real-time performance. This breaks the rigid limitations of traditional single-energy systems and solves the problems of repetitive module development and cumbersome expansion in multi-energy scenarios. Specifically, the system includes five core modules: a data acquisition and fusion module, a module manager, an energy-specific risk identification module, a general risk assessment module, and an early warning generation module. Each module is independently encapsulated and works collaboratively through a standardized structure.

[0022] The data acquisition and fusion module adopts a distributed acquisition architecture, is configured with multi-source data acquisition interfaces, and supports the acquisition of sensor data, market data, meteorological data, and news text data, as detailed below: Data acquisition phase: An adapter pattern is adopted to design a dedicated adapter for each type of data source, enabling accurate data retrieval and buffering. The acquisition frequency is dynamically adjusted according to the characteristics of the data source. For example, sensor data (electricity load, oil pipeline flow, etc.) is acquired at the second level, market data (energy prices, transaction volume, etc.) is acquired at the minute level, meteorological data (temperature, wind speed, precipitation, etc.) is acquired at the hour level, and news text data (geopolitical events, policy reports, etc.) is acquired in real-time streaming. The acquired data is temporarily stored in a distributed buffer to avoid data loss and supports breakpoint resume.

[0023] Data standardization phase: The collected raw data is cleaned, time-aligned and normalized to eliminate format and unit differences; outliers are filtered by threshold detection and missing values ​​are filled by linear interpolation; all data are mapped to a unified UTC timestamp and data of different frequencies are aligned by a sliding window mechanism (e.g., aggregating second-level sensor data into minute-level averages); numerical data are scaled to the [0, 1] interval using a min-max normalization method.

[0024] Feature extraction and fusion stage: First, for standardized numerical data (such as sensor, market, and meteorological data), principal component analysis is used for dimensionality reduction to extract key feature vectors. For example, peak trend features are extracted from electricity load data, and volatility features are extracted from oil price data. Simultaneously, for news text data, natural language processing techniques (such as the BERT model) are used for sentiment analysis and event recognition to generate text feature vectors. For example, a pre-trained language model maps the text to fixed-dimensional vectors; for instance, the [CLS] tag output of BERT is used as the text feature vector with a dimension of [missing value]. (e.g., 768-dimensional). Sentiment analysis outputs risk polarity (positive / negative), and event identification extracts key entities (e.g., "geopolitical conflict").

[0025] After feature extraction is completed, the text feature vectors and numerical feature vectors are merged to form a preliminary multimodal feature set; The attention network built into the data acquisition and fusion module then automatically calculates the weights of each data source, prioritizing data sources relevant to the current risk context. For example, during hurricane season, meteorological data has a higher weight; during periods of geopolitical tension, news data has a higher weight. The specific fusion process is as follows: Let the number of data sources be N, and the feature vector of each data source be... (in ), dimension Attention weights The calculation expression is as follows: ; In the formula: This represents the query vector, which indicates current context information (such as season, event type, or user-specified energy type), with dimensions of [dimensions missing]. (e.g., 64-dimensional), dynamically generated based on environmental parameters (e.g., calendar data); Represents a learnable matrix; This represents the feature vector of the i-th data source; Then, a weighted fusion calculation is performed based on the calculated attention weights: ; In the formula: This represents the fused feature vector; In this embodiment, the 128-bit feature vectors of the four data sources are multiplied by their corresponding attention weights, and then the main dimensions are weighted and summed to obtain a 128-bit fused feature vector z. After fusion, the vector is normalized to ensure that all dimension values ​​are in the range of 0 to 1. At the same time, metadata (including the weight distribution of each data source, feature extraction accuracy, and text key event labels) is generated and encapsulated with the fused feature vector z in JSON format. This is then pushed to the module manager through a standardized RESTful API interface to provide unified data input for subsequent module scheduling and risk identification.

[0026] In this embodiment, weights are dynamically allocated based on the historical correlation between different data sources and risk events, prioritizing the enhancement of the influence of highly correlated data sources and reducing interference from noisy data. At the same time, multimodal fusion is used to cover both structured and unstructured data, thereby improving the comprehensiveness of risk identification.

[0027] The module manager incorporates a module registration list, a scheduling rule base, and a meta-learning model. Furthermore, it constructs an intelligent module routing model based on meta-learning technology within the module manager. This model learns the adaptation rules of energy type, context, and module combination from historical call data, enabling dynamic module loading without manual configuration—the intelligent module mechanism. Simultaneously, it implements a hot-swappable expansion mechanism through standardized configuration files and reflection mechanisms, solving the problems of rigid expansion and cumbersome energy type adaptation in existing systems. Specifically: Intelligent Module Routing: Assume the registered module combinations in the system are as follows Each combination It consists of one energy-specific risk identification module and one or two general risk assessment modules, for example... Includes a power risk module and a general risk assessment module. Composed of an oil risk module; It consists of a natural gas risk module and a general risk assessment module; m is the total number of combinations; By inputting a feature vector S, which is composed of energy type features and context features, the output module group suitability score is used to select the optimal module combination. The specific expression is as follows: ; In the formula: This indicates the suitability score for the module group; This represents the Sigmoid activation function; Represents the learnable weight matrix of the meta-learning model; Represents the feature vector of energy type; The context feature vector is extracted based on metadata output from the data acquisition and fusion module (such as season, presence of breaking news events, energy demand periods, etc.). All features are normalized to the range of 0 to 1 using min-max normalization. Represents module combination The inherent feature vector includes features such as module type, historical call success rate, average processing latency, and historical risk identification accuracy of energy adaptation. All features are normalized to the range of 0 to 1. This indicates the bias term.

[0028] The optimal module combination is selected based on the module adaptability score, that is, the combination with the highest score is selected as the optimal module combination.

[0029] Furthermore, in this embodiment, to improve the meta-model's adaptability to new energy types and contexts, episodic learning is used for meta-training, as detailed below: The historical module call data, including energy type, context, module combination, risk identification accuracy, processing latency, etc., is divided into multiple independent episodes, each episode containing a support set and a query set; On the support set of each episode, the metamodel learns the local mapping patterns of energy type, context, and module combination, and updates the temporary parameters. On the query set, based on the patterns learned from the support set, the loss between the appropriate matching score of the module group and the actual calling effect is calculated, and the global parameters W and b of the meta-model are optimized through gradient descent. Repeat the above meta-training process until the loss of the meta-model converges on the query set of all scenarios, ensuring that the model has the ability to quickly adapt to new scenarios with a small amount of data; for example, after adding the hydrogen energy module, only a small amount of hydrogen energy-related call data is needed to optimize the scoring strategy.

[0030] For the registration and management of new modules, during system initialization in this embodiment, the adapted energy-specific risk identification module and general risk assessment module are automatically registered to the module registration list. The module registration list records information such as module ID, name, adapted energy, API address, and input / output dimensions. When adding a new energy module (such as hydrogen energy), there is no need to modify the core code; only a standardized JSON configuration file needs to be submitted. The module manager scans the configuration file directory every 30 seconds. After detecting a new file, it parses the content through reflection and verifies the availability of the API interface (by sending a test fusion feature vector). Once the verification is successful, the module instance is automatically loaded and the module registration list is updated, enabling hot-swappable expansion.

[0031] Specifically, in this embodiment, the modular management process is as follows: Receive the fusion feature vector z (including timestamp and feature value), metadata (such as data energy weight distribution and event tags) output by the data acquisition and fusion module, and the target energy type E (such as oil and hydrogen energy) specified by the user. Encode the user-specified energy type E into an energy type feature vector. Contextual features are extracted from metadata, such as timestamps corresponding to seasons and event tags corresponding to the existence of geopolitical events. These features are then normalized to form a contextual feature vector. splicing and The input feature vector is obtained. ; The meta-model calculates the fit score for all candidate module combinations based on the aforementioned fit score calculation formula, and selects the combination with the highest score. Then the module manager sends the highest-rated combination through a standardized interface. All modules in the process send data requests, distributing the fused feature vector z to each module.

[0032] The energy-specific risk identification module has a pluggable structure and adopts a layered architecture of a basic model layer and a domain fine-tuning layer to balance the need for reuse of general features of multiple energy sources and accurate extraction of specific risk features. It also introduces cross-energy influencing factors to quantify the chain effect between energy sources and combines transfer learning technology to solve the problem of insufficient historical data of new energy sources. This solves the technical bottleneck that a single AI model cannot adapt to the heterogeneous risks of multiple energy sources and achieves knowledge transfer and efficient training.

[0033] The energy-specific risk identification module includes multiple modules, each corresponding to a type of energy and having a built-in dedicated risk identification unit. The basic model layer is constructed using a deep neural network. The training data covers historical risk data of multiple energy sources such as electricity, oil, and natural gas (e.g., supply fluctuations, fault records, price changes, etc.). By extracting common risk features of multiple energy sources (e.g., supply and demand imbalances, correlations of sudden events, etc.), that is, by inputting the fusion feature vector output by the data acquisition and fusion module, a general risk feature vector is output.

[0034] The domain fine-tuning layer is configured with a dedicated model for different energy characteristics to achieve the extraction of specific risk features. For example, the domain fine-tuning layer for power risks uses a graph neural network to adapt to the characteristics of the power grid topology. It inputs general feature vectors and data on the connection relationship between power grid nodes to focus on identifying risks such as peak load overload and line fault propagation. The domain fine-tuning layer for oil risks uses a long short-term memory network to adapt to the time-series characteristics of the oil supply chain. It inputs oil pipeline flow and time-series data of geopolitical events to identify risks such as supply chain disruptions and abnormal price fluctuations.

[0035] During model training, the parameters obtained from training the basic model layer on energy sources with sufficient data (electricity, oil) are used as initial parameters. For new energy sources such as hydrogen, the weight matrix and bias terms of the domain fine-tuning layer are fine-tuned using a small amount of domain data (pilot supply data, safety incident records). This eliminates the need to train the model from scratch, significantly reducing reliance on historical data and enabling knowledge transfer.

[0036] In this embodiment, a cross-energy impact factor is introduced to quantify the inter-energy chain effect. Modules share information through this cross-energy impact factor, and by combining general features and domain features, an energy-specific risk assessment value is output. The specific expression is as follows: ; In the formula: This represents the cross-energy influence factor of energy i on energy j; This indicates the total number of energy types supported by the system; This represents the probability that, after a risk event occurs in energy i in historical data, a risk event will subsequently occur in energy j. This can be expressed as... Calculated; This represents the influence intensity coefficient of energy i on energy j, which is set based on domain knowledge. For example, the influence intensity of electricity on oil refining is 0.8, and the influence intensity of oil on natural gas transportation is 0.3. This indicates the relationship between energy j and all other energy sources. The maximum value is used for normalization.

[0037] The final comprehensive risk identification score output by the energy-specific risk identification module is calculated based on the basic model features, domain-specific features, and cross-energy impact factors, as shown in the following expression: ; In the formula: This represents the comprehensive risk assessment score for target energy j. , , The weights are denoted by 1, and the system is optimized using historical risk data. This represents the mean of the general risk feature vector; This represents the mean of a domain-specific risk feature vector. This represents the comprehensive risk assessment score for other energy sources i, output by the corresponding energy-specific risk assessment module; N represents the total number of energy types supported by the system.

[0038] In this embodiment, the layered AI architecture design enables modules to reuse features while capturing specific risks; and the transfer learning technology reduces the training data requirements of the new energy module; effectively solving the pain point of insufficient historical risk data for new energy; and by introducing cross-energy influencing factors to quantify the chain effect between energy sources, the one-sidedness of single-energy risk assessment is avoided, which is in line with the trend of multi-energy integrated management and provides related data support for subsequent comprehensive assessment.

[0039] The general risk assessment module constructs a multi-energy risk propagation model through an undirected weighted graph model to simulate the propagation path of general risks (natural disasters, economic fluctuations, policy changes) among multiple energy sources, capture the impact of cross-energy general risks, dynamically adjust the weights of general risk factors using reinforcement learning to adapt to the differences in risk priority under different scenarios, eliminate risk overlap through a risk coupling analysis model, balance the contribution of specific risks and general risks, solve the problem of general risks being ignored or repeatedly calculated in traditional systems, and improve the accuracy and comprehensiveness of comprehensive risk assessment. Specifically, the general risk assessment module incorporates a cross-energy general risk factor library, an undirected weighted graph model, a reinforcement learning agent, and a risk coupling analysis model. The general risk factor library contains three core factors: policy changes, macroeconomic fluctuations, and extreme weather (corresponding to natural disasters, economic fluctuations, and policy changes). Each factor corresponds to specific characteristic indicators. For example, the policy factor includes the type of policy release and the real-time time; the extreme weather factor includes the type of disaster and the intensity level. The factor data is extracted from the metadata output by the data acquisition and fusion module, and after standardization processing, it is stored in the general risk factor library.

[0040] The undirected weighted graph model includes energy nodes (electricity, oil, etc.) and general risk nodes (policy, economy, weather). The edge weights represent the strength of risk association between nodes (obtained through training with historical data). Based on this, a multi-energy risk propagation network is constructed. The direct impact strength, indirect impact strength, and total impact strength of risks on energy are calculated sequentially based on the undirected weighted graph model to accurately quantify the propagation effect of general risks. The expression for the direct impact strength is as follows: ; In the formula: Indicates general risk energy The direct impact on strength; Indicates a general risk node For energy nodes The edge weights between them reflect the inherent strength of their relationship (e.g., the edge weight between natural disasters and electricity is 0.7, and the edge weight between natural disasters and oil is 0.6). Indicates a general risk node The risk level score is obtained by standardizing the general risk features extracted from the data acquisition and fusion module.

[0041] The formula for calculating the intensity of the indirect influence is as follows: ; In the formula: Indicates general risk energy The intensity of the indirect influence; Indicates energy node For energy nodes The edge weights between energy sources reflect the strength of their interconnectedness; for example, the edge weight between electricity and oil is 0.8. Indicates energy node The exclusive risk score is output by the energy-specific risk identification module.

[0042] The total impact intensity is a weighted sum of direct and indirect impacts, expressed as follows: ; In the formula: Indicates risk to energy The overall intensity of the impact; This indicates the direct influence coefficient weights, which are obtained through historical training.

[0043] The reinforcement learning agent dynamically adjusts the weights of various general risk factors based on feedback from historical risk handling performance. The agent iteratively optimizes the state space (scene features and general risk states), action space (weight adjustment coefficients), and reward function (balancing accuracy and deduplication rate) to select the optimal weight adjustment method, avoiding the rigidity problem of fixed weights. The expression for the reward function is as follows: ; In the formula: Indicates the reward value; This represents the accuracy rate of early warnings after weight adjustment, calculated by the degree of matching between the predicted risk level and the actual risk level. The risk overlap rate is represented by the proportion of overlap between the general risk assessment results and the energy-specific risk assessment results. and This represents the weighting coefficients, which sum to 1. After iterative optimization based on the reinforcement learning agent, the final weight is calculated using the following formula: ; In the formula: Indicates general risk The final weight; This represents the optimal adjustment coefficient for the reinforcement learning agent's selection. Indicates the learning rate; Indicates general risk Weights before adjustment.

[0044] The risk coupling analysis model is used to integrate the assessment results of specific risks and general risks, avoid overlapping or omissions in early warnings, and output energy. The comprehensive risk score is expressed as follows: ; In the formula: Indicates energy Comprehensive risk score; , , This represents the coupling weight coefficients, which sum to 1. Indicates energy A unique risk score is output by the energy-specific risk identification module; Indicates general risk The final weight; Indicates general risk energy The overall intensity of the impact; Indicates energy The risk overlap rate indicates the proportion of overlap between the specific risks and general risks of this energy source.

[0045] In this embodiment, the multi-energy risk propagation model constructed by the undirected weighted graph model accurately simulates the diffusion path of risks in the energy network, making the general risk impact assessment more in line with actual scenarios; the reinforcement learning dynamic weight adjustment improves the adaptability of the entire system in different scenarios and solves the rigidity problem of fixed weights; and the risk coupling analysis model eliminates risk overlap, avoids warning redundancy, and balances the contributions of the two types of risks, providing comprehensive and non-redundant risk data for subsequent warning generation.

[0046] The early warning generation module integrates a LIME model, an adaptive threshold adjustment unit, a risk classification unit, and a multi-channel push unit. The LIME model quantifies the contribution of each feature to the risk level by constructing a linear approximation model in the local feature space corresponding to the comprehensive risk score, as detailed below: With the current comprehensive feature vector Generate a local perturbation sample set centered on [the target location]. ( The number of samples (e.g., 1000) is used to control the perturbation range within ±10% of the original feature value to ensure locality. Weights are assigned based on the similarity (calculated using Euclidean distance) between the perturbed sample and the original feature vector. Higher similarity results in higher weights, as shown in the following formula: ; In the formula: Indicates the first The weights of each perturbed sample; Indicates the first The feature vector of a perturbed sample; Indicates energy The comprehensive risk characteristic vector; This represents the operator for calculating Euclidean distance; Indicates bandwidth parameter; With perturbation sample set The input is the risk level (based on the original comprehensive risk score) and the label is the risk level, combined with the sample weights. Training a linear regression model: ; In the formula: Indicates the predicted risk level of the perturbation sample; The dimension representing the comprehensive risk feature vector; Indicates the first The larger the absolute value of the coefficient of each feature, the stronger the impact of that feature on risk. Indicates the first The first perturbation sample One eigenvalue; Represents the bias term of a linear model; Extracting coefficients of the linear model The absolute values ​​are sorted in descending order, and the top 5 core features are selected as key influencing factors. The contribution percentage of each factor is then calculated. ; In the formula: No. The contribution percentage of each feature ; In this embodiment, sample weights are assigned based on Euclidean distance, a local linear model is trained, the contribution of each feature to the core risk score is quantified, the top 5 core influencing factors are screened, and an explanatory report containing the causes of risk, propagation path, and response suggestions is generated, providing an intuitive explanation of the source of risk to assist decision-making.

[0047] The adaptive threshold adjustment unit has three preset warning thresholds, each corresponding to a low-risk upper limit threshold. Medium-risk upper limit threshold High-risk lower limit threshold (Due to comprehensive risk score) Therefore, it satisfies The initial value is set based on historical household statistics, for example: ， ， ; In the adaptive threshold adjustment unit of this embodiment, the threshold is adjusted based on the historical false alarm rate, accuracy rate, and average loss coefficient of risk events, as expressed below: ; In the formula: This indicates the adjusted target threshold; This indicates the initial threshold before adjustment or the threshold from the previous round. This represents the threshold adjustment coefficient; Indicates the historical accuracy rate of early warnings; This indicates the historical false alarm rate for early warnings; The average loss coefficient for risk events is derived from standardized indicators such as economic losses and supply disruption duration caused by historical risk events. The greater the loss, the closer the loss is to 1.

[0048] Furthermore, the adjusted threshold must satisfy the hierarchical relationship: If the adjustment violates the constraints, then the boundary value is taken.

[0049] The risk grading unit generates a standardized graded early warning report based on an adaptively adjusted threshold and the analysis results of an interpretable AI engine, for example: Low risk: This indicates that energy supply is basically stable, with only minor potential risks, and no urgent intervention is required. Medium risk: This indicates that there is a significant risk of fluctuations in energy supply, requiring close monitoring and preparation of countermeasures; High risk: This indicates a serious risk of energy supply disruption, requiring immediate activation of an emergency response. Extremely high risk: This indicates that an energy supply emergency has occurred, requiring the implementation of the highest level of emergency response plan.

[0050] In this embodiment, the early warning report is encapsulated in JSON format, including energy type, early warning timestamp, comprehensive risk level, early warning level, the top 5 core features output based on the LIME model (such as geopolitical events, contributing 35%; cross-energy impact of power supply disruptions, contributing 25%), and also includes risk disclosure, namely the analysis conclusions of the local linear model and response suggestions.

[0051] The multi-channel push unit provides dashboard push, app push, mobile notification, and API interface push based on the generated early warning report.

[0052] In this embodiment, the LIME model integrates interpretable AI technology, which significantly improves the interpretability of early warnings, helps users understand the causes of risks more efficiently, solves the black box problem of AI early warnings, and enhances the trust in decision-making; and reduces the false alarm rate and false negative rate by adjusting the threshold through an adaptive threshold adjustment mechanism.

[0053] The above are merely preferred embodiments of this application and are not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. An energy supply risk identification and early warning system based on artificial intelligence, characterized in that, The system adopts a modular structure formed by combining a fixed core framework with pluggable energy-specific modules. Each module achieves bidirectional communication through a pre-set standardized data interaction interface. The system includes the following modules: Data Acquisition and Fusion Module: Configures multi-source data acquisition interfaces, collects heterogeneous data from multiple sources, cleans, standardizes, and extracts features from the collected data, generates a risk feature dataset in a unified format, and provides a standardized data structure for other modules to use; Module Manager: The core scheduling unit of the system, with a built-in module registration list and scheduling rule library. Based on the energy type to be monitored and the real-time risk scenario context, it dynamically loads the corresponding energy-specific risk identification module from the module library. The registration, adaptation and loading of new energy type modules are completed through standardized configuration files, realizing hot-swappable module expansion. Energy-specific risk identification module: It has a pluggable structure and is equipped with a dedicated risk identification unit for each energy type. Each dedicated risk identification unit has a built-in artificial intelligence model trained and optimized with data from the corresponding energy field. The model extracts the specific risk characteristics of the corresponding energy and outputs the risk assessment value. General Risk Assessment Module: It has a built-in cross-energy general risk factor library, which includes factors related to policy changes, macroeconomic fluctuations, and extreme weather. It analyzes the impact of each general factor on different energy sources to generate general risk assessment results. It interacts with the energy-specific risk identification module and uses a weighted fusion algorithm to integrate the two types of risk assessment results to output comprehensive risk data. Early warning generation module: It presets multi-level early warning judgment benchmarks and generates graded early warning signals based on the comprehensive risk data output by the general risk assessment module and the judgment benchmarks.

2. The energy supply risk identification and early warning system based on artificial intelligence according to claim 1, characterized in that, The data acquisition and fusion module has a built-in attention network. The attention network automatically assigns weights based on the historical correlation between each data source and risk event. Based on the assigned weights, the cleaned, format-standardized, and feature-extracted data are weighted and fused to generate a unified dimension fused feature vector, thereby obtaining a risk feature dataset in a unified format.

3. The energy supply risk identification and early warning system based on artificial intelligence according to claim 1, characterized in that, The module manager uses meta-learning technology to construct an intelligent module routing model. The intelligent module routing model is trained based on historical module call records, matching accuracy, processing latency, and risk identification accuracy data under different energy scenarios. The model parameters are optimized through a scenario learning mechanism. The input of the intelligent module routing model is the concatenation result of the energy type feature vector and the context feature vector, where the energy type feature vector is the energy type to be detected, and the context feature vector is extracted from the metadata output by the data acquisition and fusion module. Based on this, the intelligent module routing model outputs the optimal module combination scheme. The optimal module combination scheme consists of an energy-specific risk identification module and zero or one general risk assessment module, and is generated based on the adaptability score. The adaptability score comprehensively considers the matching degree between the module and the energy type, the historical call success rate, and the processing efficiency.

4. The energy supply risk identification and early warning system based on artificial intelligence according to claim 1, characterized in that, The energy-specific risk identification module adopts a layered architecture consisting of a basic model layer and a domain fine-tuning layer. The basic model layer is a pre-trained deep neural network. The training data covers historical risk data and domain knowledge of various energy sources. The input is the fusion feature vector output by the data acquisition and fusion module, which is used to extract common risk features of multiple energy sources and output a general risk feature vector. The domain fine-tuning layer serves as a dedicated risk identification unit, configuring exclusive models for different energy characteristics to adapt to the heterogeneous risk requirements of various energy sources.

5. The energy supply risk identification and early warning system based on artificial intelligence according to claim 4, characterized in that, The energy-specific risk identification module introduces a cross-energy impact factor. This cross-energy impact factor is quantified based on historical risk chain event data, and is calculated by combining the conditional probability of risk events between energy sources with the impact intensity coefficient set by the domain. After normalization, it represents the intensity of the chain effect between different energy risks.

6. The energy supply risk identification and early warning system based on artificial intelligence according to claim 4, characterized in that, The energy-specific risk identification module adopts a transfer learning mechanism, which uses the parameters obtained by polar training of the basic model layer in historical risk data of multiple energy types as initial parameters, and then uses the domain data of the current energy to be identified to fine-tune the parameters of the domain fine-tuning layer, so as to realize the cross-energy reuse of the basic model layer parameters.

7. The energy supply risk identification and early warning system based on artificial intelligence according to claim 1, characterized in that, The general risk assessment module has a built-in undirected weighted graph model. The undirected weighted graph model uses each energy type as a node and the correlation strength between energy sources as the edge weight to construct a multi-energy risk propagation network. A risk diffusion model is constructed by simulating the transmission path and impact of general risk factors between nodes. The general risk assessment module incorporates reinforcement learning technology to dynamically adjust the weight values ​​of each general risk factor based on feedback from historical risk handling effects.

8. The energy supply risk identification and early warning system based on artificial intelligence according to claim 1, characterized in that, The general risk assessment module has a built-in risk coupling analysis model. The risk coupling analysis model identifies duplicate features in energy-specific risks and general risks through feature comparison algorithms, removes the identified duplicate features, and uses the analytic hierarchy process to balance the contribution ratio of specific risks and general risks, outputting a comprehensive risk score without redundancy.

9. The energy supply risk identification and early warning system based on artificial intelligence according to claim 1, characterized in that, The early warning generation module integrates a locally interpretable model, which is a LIME model. By constructing a local linear model, it quantifies the contribution of each risk factor to the comprehensive risk score, screens core influencing factors, and generates an explanation, including the risk cause, propagation path, scope of impact, and response recommendations.

10. The energy supply risk identification and early warning system based on artificial intelligence according to claim 1, characterized in that, The early warning generation module has a built-in adaptive threshold adjustment mechanism. The adaptive threshold adjustment mechanism calculates the early warning accuracy, false alarm rate and economic loss amount of the corresponding risk event based on historical early warning data, and adjusts the judgment threshold of multi-level early warnings using a weighted calculation method.

Images