Wind power system information security risk assessment method and device

Abstract

The invention discloses a wind power system information security risk assessment method and device, and the method comprises the steps: building a multi-source data fusion model through collecting the multi-source heterogeneous data of a wind power plant group, obtaining a multi-dimensional data observation tensor, building a wind power system information dynamic behavior baseline model of a self-adaptive space-time diagram neural network based on the multi-dimensional data observation tensor, and carrying out the wind power system information security risk assessment. On the basis of the difference value between the output of the wind power system information dynamic behavior baseline model and actual data, a wind power plant group information collaborative intelligent safety risk assessment model is constructed, and accurate description and intelligent risk assessment of wind power system information behaviors are achieved. Therefore, the information security protection level and the long-term operation stability of the wind power system in a complex and changeable environment are improved, and the technical problems that the security situation of the system is difficult to comprehensively describe and the time-varying property of the operation state of the wind power system and the working condition drift problem are not fully considered in an existing wind power system information security assessment method are solved.

Description

Technical Field

[0001] This invention relates to the field of power system technology, and in particular to a method and apparatus for assessing information security risks in wind power systems. Background Technology

[0002] Wind power systems are characterized by wide distribution, strong equipment heterogeneity, and complex operating environments. In particular, large offshore wind farms often consist of multiple geographically dispersed stations with similar operating modes but local differences, which poses a huge challenge to traditional information security assessment models in terms of data utilization, model adaptability, and collaborative capabilities.

[0003] Existing information security assessment methods for wind power systems have revealed numerous limitations in practical applications. On the one hand, while some machine learning-based methods introduce data analysis capabilities, they typically rely on a single data source—SCADA (Supervisory Control and Data Acquisition) data or network traffic—lacking effective integration of multi-source heterogeneous data such as turbine status, control protocols, environmental information, and external threat intelligence. This results in a narrow assessment perspective and an inability to comprehensively depict the system's security posture. On the other hand, existing information security assessment models for wind power systems generally adopt a "static modeling, long-term use" approach, failing to fully consider the time-varying nature of wind power system operation and the issue of operational condition drift. For example, seasonal wind condition changes, equipment start-up, shutdown, or maintenance operations may cause data fluctuations, and the assessment model cannot distinguish between such normal disturbances and malicious attacks, leading to frequent false alarms. Therefore, constructing a wind power system information security risk assessment method that can deeply integrate multi-source data, dynamically model system behavior, and support collaborative assessment and online updates is a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] This invention provides a method and apparatus for assessing information security risks in wind power systems, which addresses the technical problem that existing methods for assessing information security in wind power systems are unable to fully depict the security status of the system and fail to adequately consider the time-varying nature of the operating status and the drift of the operating conditions.

[0005] In view of this, the first aspect of the present invention provides a method for assessing information security risks in wind power systems, comprising:

[0006] Acquire multi-source heterogeneous data of wind farm clusters and store the multi-source heterogeneous data in a historical database;

[0007] A multi-source heterogeneous data feature extraction and spatiotemporal registration fusion model is used to process the multi-source heterogeneous data to obtain a multi-dimensional data observation tensor;

[0008] Based on the multidimensional data observation tensor, a baseline model of the dynamic behavior of wind power system information based on an adaptive spatiotemporal graph neural network is constructed.

[0009] Based on the difference between the output of the wind power system information dynamic behavior baseline model and the actual data, a wind farm cluster information collaborative intelligent safety risk assessment model is constructed.

[0010] Based on the intelligent security risk assessment model for information collaboration in wind farm clusters, the information security risks of wind power systems are assessed.

[0011] Optionally, the multi-source heterogeneous data includes: wind turbine operating data, station control and communication data, network security monitoring data, and external meteorological and environmental data.

[0012] Optionally, a multi-source heterogeneous data feature extraction and spatiotemporal registration fusion model is used to process the multi-source heterogeneous data to obtain a multi-dimensional data observation tensor, including:

[0013] Preprocessing and spatiotemporal alignment of multi-source heterogeneous data;

[0014] Based on the same time reference and spatial topology, a multidimensional spatiotemporal data cube is constructed. Each data voxel of the multidimensional spatiotemporal data cube corresponds to a multidimensional observation on a device node within a specific time window.

[0015] A multidimensional observation tensor is obtained by processing the multidimensional observations of a multidimensional spatiotemporal data cube using a multi-source heterogeneous data feature extraction and spatiotemporal registration fusion model based on a convolutional autoencoder.

[0016] Optionally, based on the multidimensional data observation tensor, a baseline model of the dynamic behavior of wind power system information based on an adaptive spatiotemporal graph neural network is constructed, including:

[0017] The physical devices and communication interactions of a wind power system are abstracted into a dynamic information network diagram. Where V is the node set, E is the connection relationship, and A is the adjacency matrix representing the strength of the association between nodes;

[0018] On the time branch, a long short-term memory network is used at each node to extract trend, periodic and abrupt features in the time series, capturing the evolutionary pattern of single-node behavior. On the spatial branch, a graph attention network is used to aggregate the node features at each time step, capturing the state influence between spatially adjacent or functionally related nodes, and a gating mechanism is introduced to control the information propagation weights.

[0019] An adaptive spatiotemporal graph neural network is constructed and unsupervised training is performed with the goal of reconstructing future states. By minimizing the error between the predicted values ​​and the actual observed values ​​of the adaptive spatiotemporal graph neural network, a dynamic behavior baseline model is obtained through training.

[0020] Optionally, each safety risk assessment model in the wind farm cluster information collaborative intelligent safety risk assessment model includes an actor network, a commentator network, a target actor network, a target commentator network, a decision preference feature encoder network, and a decision preference feature decoder network. The actor network, commentator network, target actor network, target commentator network, decision preference feature encoder network, and decision preference feature decoder network are all deep neural networks.

[0021] The input to the actor network is the decoded value of the current observation data state and input decision preference features, and the output is the discrete safety risk level and risk assessment confidence of the wind farm.

[0022] The input to the critic network is a set of states and decisions sampled from all neural network models within the same group, and the output is a state-behavior value function.

[0023] The input to the target actor network is the observed data state at the next time step and the decoded value of the input decision preference features. The output is the discrete safety risk level and risk assessment confidence of the wind farm at the next time step.

[0024] The input to the critic network is the set of data states and decisions for the next time step sampled from all neural network models within the same group, and the output is the target state behavior value function for the next time step.

[0025] The input to the decision preference feature encoder network is a continuous decision history sequence constructed from multiple continuous experience tuples, and the output is the decision preference features of the neural network model.

[0026] The input to the decision preference feature decoder network is the decision preference feature encoding, and the output is the decoded neural network model decision preference features.

[0027] Optionally, before assessing the information security risks of the wind power system based on the wind farm cluster information collaborative intelligent security risk assessment model, the following steps are also included:

[0028] The training method for the intelligent safety risk assessment model based on wind farm cluster information collaboration includes the following steps:

[0029] S1. Initialize the safety risk assessment model for each wind farm by reading the real-time observation data stored in the corresponding data environment. and the maximum number of observation data reads, K;

[0030] S2. The actor network of the safety risk assessment model for each wind farm reads the real-time observation data status and the decoded value of the input decision preference characteristics of the wind farm, and outputs the risk level and risk assessment confidence of each wind farm.

[0031] S3. The data environment of each wind farm station is constructed based on the risk level and risk assessment confidence of each wind farm station output by the safety risk assessment model of each wind farm station. The reward function related to the decision bias of the safety risk assessment model of each wind farm station and the discrete risk assessment level and risk assessment confidence is calculated and fed back to the observation data of the next wind farm station.

[0032] S4. The safety risk assessment model for each wind farm is based on the current observation data of each wind farm, the risk level and risk assessment confidence of each wind farm, the reward function related to the decision bias of the safety risk assessment model of each wind farm and the discrete risk assessment level and risk assessment confidence, and the observation data of the next wind farm. The model constructs an experience tuple and saves it to the experience pool.

[0033] S5. The safety risk assessment model of each wind farm station extracts a preset number of continuous experience tuples from the experience pool to construct the decision history sequence of the safety risk assessment model of each wind farm station. Based on the constructed decision history sequence, the safety risk assessment model of each wind farm station updates the encoder network parameters and decoder network parameters of the output safety risk assessment decision preference characteristics.

[0034] S6. Group the decision preference features of the safety risk assessment models of each wind farm based on the K-means clustering algorithm;

[0035] S7. Increment the number of times the safety risk assessment model of each wind farm station reads the observation data by 1. Determine whether the number of times the safety risk assessment model of each wind farm station reads the observation data has reached the maximum number of observation data reads. If yes, complete the training of the intelligent safety risk assessment model based on wind farm group information collaboration. Otherwise, return to step S2.

[0036] Optionally, the objective function and decision variables of the wind farm cluster information collaborative intelligent safety risk assessment model are:

[0037]

[0038]

[0039]

[0040] Where N represents the number of safety risk assessment models for wind farms, and K represents the maximum number of observation data reads for each safety risk assessment model for wind farms. To determine the true safety level of the wind farm's safety risk assessment model i during its k-th safety risk assessment. Let be the real-time decision value of the safety risk assessment model i for the k-th safety risk assessment of the wind farm, and M be the total number of safety risk assessment levels. This is the set of decision values ​​for the safety risk assessment model of each wind farm in the wind farm cluster.

[0041] A second aspect of the present invention provides a wind power system information security risk assessment device, comprising:

[0042] The data acquisition module is used to acquire multi-source heterogeneous data of wind farm clusters and store the multi-source heterogeneous data in a historical database.

[0043] The fusion module is used to process multi-source heterogeneous data using a multi-source heterogeneous data feature extraction and spatiotemporal registration fusion model to obtain multi-dimensional data observation tensors;

[0044] The first modeling module is used to construct a baseline model of the dynamic behavior of wind power system information based on an adaptive spatiotemporal graph neural network, based on multidimensional data observation tensors.

[0045] The second modeling module is used to construct a wind farm cluster information collaborative intelligent safety risk assessment model based on the difference between the output of the wind power system information dynamic behavior baseline model and the actual data.

[0046] The assessment module is used to assess the information security risks of wind power systems based on the wind farm cluster information collaborative intelligent security risk assessment model.

[0047] Optionally, the fusion module is specifically used for:

[0048] Preprocessing and spatiotemporal alignment of multi-source heterogeneous data;

[0049] Based on the same time reference and spatial topology, a multidimensional spatiotemporal data cube is constructed. Each data voxel of the multidimensional spatiotemporal data cube corresponds to a multidimensional observation on a device node within a specific time window.

[0050] A multidimensional observation tensor is obtained by processing the multidimensional observations of a multidimensional spatiotemporal data cube using a multi-source heterogeneous data feature extraction and spatiotemporal registration fusion model based on a convolutional autoencoder.

[0051] Optionally, the first modeling module is specifically used for:

[0052] The physical devices and communication interactions of a wind power system are abstracted into a dynamic information network diagram. Where V is the node set, E is the connection relationship, and A is the adjacency matrix representing the strength of the association between nodes;

[0053] On the time branch, a long short-term memory network is used at each node to extract trend, periodic and abrupt features in the time series, capturing the evolutionary pattern of single-node behavior. On the spatial branch, a graph attention network is used to aggregate the node features at each time step, capturing the state influence between spatially adjacent or functionally related nodes, and a gating mechanism is introduced to control the information propagation weights.

[0054] An adaptive spatiotemporal graph neural network is constructed and unsupervised training is performed with the goal of reconstructing future states. By minimizing the error between the predicted values ​​and the actual observed values ​​of the adaptive spatiotemporal graph neural network, a dynamic behavior baseline model is obtained through training.

[0055] As can be seen from the above technical solutions, the wind power system information security risk assessment method provided by the present invention has the following advantages:

[0056] The present invention provides a method for information security risk assessment of wind power systems. This method collects multi-source heterogeneous data from wind farm clusters, constructs a multi-source data fusion model to obtain a multi-dimensional data observation tensor, and builds an adaptive spatiotemporal graph neural network-based baseline model of the dynamic behavior of wind power system information based on the multi-dimensional data observation tensor. Based on the difference between the output of the baseline model and the actual data, a collaborative intelligent security risk assessment model for wind farm cluster information is constructed. This method achieves accurate characterization and intelligent risk assessment of wind power system information behavior, thereby improving the information security protection level and long-term operational stability of wind power systems in complex and variable environments. It solves the technical problems of existing wind power system information security assessment methods, which struggle to comprehensively characterize the system's security posture and fail to fully consider the time-varying nature of wind power system operating states and the drift of operating conditions. Attached Figure Description

[0057] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, 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 the present invention. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0058] Figure 1 This is a flowchart illustrating a wind power system information security risk assessment method provided in an embodiment of the present invention;

[0059] Figure 2 This is a training framework diagram of the intelligent safety risk assessment model based on wind farm cluster information collaboration provided in this embodiment of the invention;

[0060] Figure 3 This is a self-organizing framework diagram of the neural network model for wind farm safety risk assessment based on K-means algorithm grouping provided in this embodiment of the invention;

[0061] Figure 4 This is a schematic diagram of the structure of a wind power system information security risk assessment device provided in an embodiment of the present invention. Detailed Implementation

[0062] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0063] For easier understanding, please refer to Figure 1 This invention provides an embodiment of a wind power system information security risk assessment method, comprising:

[0064] Step 101: Obtain multi-source heterogeneous data of the wind farm cluster and store the multi-source heterogeneous data in the historical database.

[0065] It should be noted that a wind farm cluster consists of multiple wind farm stations, each equipped with the following facilities and systems: 1) Wind turbines, each equipped with sensors to monitor operating parameters such as wind speed, wind direction, rotational speed, and temperature; 2) A station control center, responsible for collecting and processing data from each wind turbine and performing real-time monitoring and management through a SCADA system. It supports multiple communication protocols (such as IEC-104 and Modbus) for data transmission; 3) Network security monitoring equipment, including industrial firewalls, intrusion detection systems (IDS), and traffic mirroring devices, used to monitor the security status of network communications; 4) External environment monitoring equipment, such as satellite cloud image receivers, anemometers, and automatic identification systems for ships, providing external meteorological and environmental information. Furthermore, the safety risk levels used for wind farm cluster safety risk assessment in this invention consist of five levels: Level 1 (extremely high risk), Level 2 (high risk), Level 3 (medium risk), Level 4 (low risk), and Level 5 (acceptable risk or safe). Specifically, Level 1 (extremely high risk) indicates that the system has serious information security vulnerabilities or threats, with large-scale data tampering or attacks; Level 2 (high risk) indicates that the security issues in the system lead to significant functional impairments or data loss. Although they may not cause immediate catastrophic consequences, if not addressed promptly, they will quickly escalate into more serious problems, affecting the normal operation of the system and service provision; Level 3 (medium risk) indicates that the system has a certain level of security threats that affect some functions or information services of the system, but will not cause major damage to the overall operation; Level 4 (low risk) indicates that the system has only minor security threats that are unlikely to have a significant negative impact on the stability or performance of the system. However, these risks still need to be monitored to ensure that they do not accumulate into larger problems; Level 5 (acceptable risk or safe) indicates that the information network is within the normal range or has slight fluctuations.

[0066] In this embodiment of the invention, multi-source heterogeneous data of the wind farm cluster is first collected. This multi-source heterogeneous data includes wind turbine operating data, site control and communication data, network security monitoring data, and external meteorological and environmental data. Wind turbine operating data directly reflects the real-time operating status of the turbines, specifically covering SCADA real-time operating data (such as active / reactive power of the turbines for assessing power generation efficiency, wind speed and direction for assessing the potential impact of the external environment on the operation of the wind farm cluster units, turbine speed for assessing the operating status of the mechanical transmission system, and operating temperatures of core components such as gearboxes and generators for assessing the risk of overheating in core equipment). Site control and communication data includes remote monitoring data based on the IEC-104 protocol (used to assess the stability of command transmission between the wind farm cluster and the dispatch center), status feedback, and remote start / stop control functions, as well as device-to-device communication data based on the Modbus protocol (used to assess the communication stability of various core components within the wind farm). Network security monitoring data for wind farms includes (e.g., industrial firewall operation logs for real-time detection of potential risks such as external network interception and abnormal access; traffic mirroring data for extracting communication characteristics and detecting malicious code transmission in real-time using traffic analysis tools). External meteorological and environmental data, such as satellite cloud imagery for predicting regional meteorological changes in wind farm clusters; wind speed and direction meter data for real-time assessment of local wind energy resource distribution; and ship AIS and navigation data for assessing ship collision risks and potential risks in wind farm construction and operation.

[0067] In one embodiment, at the multi-source heterogeneous data acquisition execution level, based on the operating characteristics and management needs of the wind farm cluster, a reasonable time window parameter is set, i.e., a fixed time interval for acquiring multi-source heterogeneous data. After synchronously acquiring multi-source heterogeneous data within the set time window, the data is stored in real time to the historical database. In this embodiment of the invention, taking into account the multi-timescale characteristics of multi-source heterogeneous data acquisition within various wind farm clusters (such as the typical SCADA real-time data acquisition interval of 1-10 seconds and meteorological forecast data of 15-60 minutes), a unified time window for acquiring multi-source heterogeneous data is set to 3 minutes.

[0068] Step 102: Use a multi-source heterogeneous data feature extraction and spatiotemporal registration fusion model to process the multi-source heterogeneous data and obtain the multidimensional data observation tensor.

[0069] It should be noted that the collected multi-source heterogeneous data undergoes preprocessing, including data cleaning, missing value imputation, unit normalization, and temporal resampling. After preprocessing, a multi-dimensional spatiotemporal data cube is constructed based on a unified time reference and spatial topology. Each data voxel in the multi-dimensional spatiotemporal data cube corresponds to a multi-dimensional observation on a specific device node within a specific time window. Finally, a multi-source heterogeneous data feature extraction and spatiotemporal registration fusion model based on a convolutional autoencoder is used to process the multi-dimensional observations of the multi-dimensional spatiotemporal data cube, obtaining a multi-dimensional data observation tensor. The multi-source heterogeneous data feature extraction and spatiotemporal registration fusion model based on a convolutional autoencoder consists of an encoder and a decoder, employing a three-dimensional convolutional neural network structure to simultaneously capture the local correlations and hierarchical features of the data in the temporal, spatial, and feature dimensions, ultimately outputting a unified multi-dimensional data observation tensor.

[0070] Step 103: Based on the multidimensional data observation tensor, construct a baseline model of the dynamic behavior of wind power system information based on an adaptive spatiotemporal graph neural network.

[0071] It should be noted that the physical equipment and communication interaction relationships of the wind power system are abstracted into a dynamic information network diagram. Where V is the node set (including key information nodes such as wind turbines, SCADA servers, remote terminal units, substation controllers, network firewalls, and monitoring centers), E is the connection relationship (logical or physical connection relationships such as control command transmission and data reporting), and A is the adjacency matrix representing the strength of the association between nodes. Two learnable embedding matrices are defined: and , An embedding matrix for a learnable spatial representation. Let be the set of real numbers. The learnable temporal representation embedding matrix is ​​given by O, where O represents the total number of nodes in the wind power system and d represents the dimension of the embedding vector. The adaptive adjacency matrix is ​​calculated as follows:

[0072]

[0073] in, The matrix is ​​an adaptive adjacency matrix, and Softmax and ReLU are activation functions, respectively.

[0074] The final graph neural network connection structure is derived from the prior physical topology. With adaptive adjacency matrix It is formed by fusion, and is represented as:

[0075]

[0076] in, This is the connection structure of a graph neural network.

[0077] A dual-branch encoder is used to extract spatiotemporal features. In the temporal branch, a long short-term memory network is used at each node to extract trend, periodic and abrupt features in the time series, capturing the evolutionary patterns of single-node behavior. In the spatial branch, a graph attention network is used to aggregate the node features at each time t, capturing the state influence between spatially adjacent or functionally related nodes, and a gating mechanism is introduced to control the information propagation weights.

[0078] Finally, a three-layer stacked adaptive spatiotemporal graph neural network is constructed, with each layer sequentially performing "temporal encoding → spatial propagation → gating update" to achieve deep fusion of spatiotemporal features and output a high-order hidden state. ,in, For time steps, The hidden state dimension is used. Unsupervised training is performed with the goal of reconstructing future states. By minimizing the error between the predicted values ​​and the actual observed values ​​of the adaptive spatiotemporal graph neural network, a dynamic behavior baseline model is obtained. By taking the unified multidimensional data observation tensor output by the multi-source heterogeneous data feature extraction and spatiotemporal registration fusion model as input, the output can be a prediction of the dynamic behavior of information networks, enabling the model to learn the spatiotemporal behavior characteristics and system evolution trends under normal operating conditions.

[0079] Step 104: Based on the difference between the output of the wind power system information dynamic behavior baseline model and the actual data, construct a wind farm cluster information collaborative intelligent safety risk assessment model.

[0080] It should be noted that a collaborative intelligent safety risk assessment model for wind farm clusters is constructed by calculating the difference between the output of the dynamic behavior baseline model and the real-time observation data. Each wind farm is equipped with a safety risk assessment model based on deep reinforcement learning.

[0081] The optimization problem of the wind farm cluster information-coordinated intelligent safety risk assessment model includes the objective function and decision variables. The objective function and decision variables are:

[0082]

[0083]

[0084]

[0085] Where N is the number of safety risk assessment models for wind farms (equal to the number of wind farms), and K is the maximum number of observation data reads for each wind farm's safety risk assessment model. To determine the true safety level of the wind farm's safety risk assessment model i during its k-th safety risk assessment. Let be the real-time decision value of the safety risk assessment model i for the k-th safety risk assessment of the wind farm, and M be the total number of safety risk assessment levels. This represents the set of decision values ​​for the safety risk assessment model of each wind farm in a wind farm cluster. In this invention, five levels are set as the safety risk assessment standards for each wind farm; therefore, M=5. 1 to 5 indicate that the safety risk level of each wind farm station output by the neural network model has risen from acceptable risk or safe to extremely high risk level.

[0086] The optimization problem of collaborative intelligent safety risk assessment for wind farms is transformed into a Markov decision process related to safety risk assessment. This includes designing the state, output, and reward function of the deep reinforcement learning-based safety risk assessment model configured for each wind farm. The state design of each deep reinforcement learning-based safety risk assessment model is as follows:

[0087]

[0088] in, This represents the state of the wind farm's safety risk assessment model i during its k-th safety risk assessment. This refers to the actual observation data during the k-th safety risk assessment of wind farm station i. This is a baseline model for dynamic behavior.

[0089] The output of each deep reinforcement learning-based security risk assessment model is designed as follows:

[0090]

[0091] in, The confidence level of the risk assessment when the k-th safety risk assessment is conducted for wind farm site i. This is the output of the safety risk assessment model i for the kth time.

[0092] The reward function for each deep reinforcement learning-based security risk assessment model is designed as follows:

[0093]

[0094] in, Let i be the reward function for the k-th safety risk assessment of the wind farm safety risk assessment model i.

[0095] like Figure 2 As shown, this diagram illustrates the data interaction environment for constructing safety risk assessment models for each wind farm and the safety risk assessment models for each wind farm. Each safety risk assessment model includes an actor network. Critics Network Target actor network Target Commentator Network Decision preference feature encoder network and decision preference feature decoder network Action Network Critics Network Target actor network Target Commentator Network Decision preference feature encoder network and decision preference feature decoder network Both are deep neural networks, consisting of one input layer, multiple hidden layers, and one output layer.

[0096] Action Network The input is the current observation data status. and input decision preference features Decoded value The output is the discrete safety risk level and risk assessment confidence level of the wind farm. .

[0097] Critics Network The input is the set of states and decisions sampled from all neural network models within the same group. The output is the state behavior value function.

[0098] Target Actor Network The input is the observed data status at the next time step and the decoded value of the input decision preference features, and the output is the discrete safety risk level and risk assessment confidence of the wind farm at the next time step.

[0099] Commentator Network The input is the set of data states and decisions for the next time step obtained by sampling all neural network models within the same group, and the output is the target state behavior value function for the next time step.

[0100] Decision Preference Feature Encoder Network The input is a series of consecutive empirical tuples. Constructing a continuous decision-making history sequence The output is the decision preference features of the neural network model. .in, This represents the state of the wind farm's safety risk assessment model i during its (k+1)th safety risk assessment.

[0101] Decision Preference Feature Decoder Network The input is the decision preference feature encoding. The output is the decoded neural network model decision preference features. .

[0102] like Figure 2 As shown, the training steps for the intelligent safety risk assessment model based on wind farm cluster information collaboration are as follows:

[0103] S1. Initialize each wind farm station Security risk assessment model parameters (including actor network parameters) , Commentator network parameters Target actor network parameters Target commentator network parameters Decision preference feature encoder network parameters and decision preference feature decoder network parameters ), experience pools of each neural network model and the decision preference characteristics of each neural network model Configure the number of times each neural network model reads the real-time observation data stored in its corresponding data environment. and maximum number of observation data reads The safety risk assessment models of all wind farms were randomly grouped.

[0104] S2. The actor network of the safety risk assessment model for each wind farm reads the real-time observation data status of the wind farm. and input decision preference features Decoded value Output the risk level and risk assessment confidence level of each wind farm. .

[0105] S3. The data environment of each wind farm is constructed based on the risk level of each wind farm output by the safety risk assessment model of each wind farm. and risk assessment confidence level The reward function related to the safety risk assessment model of each wind farm station and the discrete risk assessment level and risk assessment confidence level decision bias was calculated. Feedback on the next observation data from the wind farm station. .

[0106] S4. The safety risk assessment model for each wind farm is based on the current observation data of each wind farm. Risk level and risk assessment confidence level of wind farms The reward function related to the safety risk assessment model of each wind farm and the decision bias of discrete risk assessment level and risk assessment confidence level. And the observation data of the next wind farm station. Constructing experience tuples And save it to the experience pool. .

[0107] S5. A fixed update frequency is used as the update condition for each neural network model. That is, after each neural network model interacts with the data environment a certain number of times, it is considered to meet the update condition. If the update condition is met, the safety risk assessment model of each wind farm station extracts a preset number B consecutive empirical tuples from the experience pool. Decision history sequence for constructing safety risk assessment models for each wind farm. Based on the constructed decision history sequence The encoder network parameters that update the safety risk assessment model output for each wind farm site, representing the safety risk assessment decision preference characteristics. and decoder network parameters The specific update method uses the gradient update method, and the update formula is as follows:

[0108]

[0109]

[0110]

[0111]

[0112]

[0113] in, and For weight parameters, To update the gradients for the parameters of the decision preference feature encoder network, To update the gradients for the parameters of the decision preference feature decoder network, The joint loss function of the commentator network is used to update the security risk assessment model within the same group. This is the loss function used to update the encoder and decoder networks. As the expected factor, This represents the set of empirical tuples sampled from all neural network models within the same current group. These represent the sets of current observed data state, safety risk assessment decision, risk assessment bias reward function, future data observation state, and current decision preference features for the neural network model within the same group. The state-behavior value function output by the commentator network for the safety risk assessment model i of a wind farm site. for The target value, The set of states for the next time step The corresponding set of actions, As a discount factor, For weight parameters, Let i represent the state value function output by the target commentator network of the safety risk assessment model i for wind farms. Let i be the policy function output by the target actor network of the safety risk assessment model i for wind farms. The output probability distributions of the encoder and decoder networks, To construct a decision history sequence The reward set in.

[0114] Next, based on the decision history sequence, the encoder network outputs the current security risk assessment decision preference features of each neural network model. The decoder reads the security risk assessment decision preference characteristics. Decode and save to the experience pool. Proceed to step S6 to perform self-organizing updates of the neural network model for safety risk assessment of each wind farm.

[0115] S6. Group the decision preference features of the safety risk assessment models for each wind farm station based on the K-means clustering algorithm. For example... Figure 3 As shown, the decision preference features of the safety risk assessment models for each wind farm are grouped based on the K-means clustering algorithm to achieve a self-organizing mechanism for the safety risk assessment models of each wind farm. First, the number of groups for each wind farm's safety risk assessment model is determined. Since this embodiment considers 5 wind farms, the number of self-organizing groups for the neural network models can be set to 2. Second, the K-means algorithm cluster centers are initialized. Then, the distance between each neural network model's decision preference features and the set cluster centers is calculated, and each model is assigned to the group containing the nearest cluster center. Further, the average value of the decision preference features of all neural network models within each group is calculated as the new cluster center. Finally, the above steps are repeated until the cluster center values ​​no longer change significantly, completing the grouping of the neural network models for the safety risk assessment of each wind farm. Subsequent parameter updates and sharing during the training process of each neural network model will be based on this grouping.

[0116] S7. Increment the number of times the safety risk assessment model of each wind farm station reads the observation data by 1. Determine whether the number of times the safety risk assessment model of each wind farm station reads the observation data has reached the maximum number of observation data reads. If yes, complete the training of the intelligent safety risk assessment model based on wind farm group information collaboration. Otherwise, return to step S2.

[0117] In this embodiment of the invention, the number of interactions with the data environment required to satisfy the update conditions of the actor network and the commentator network in the neural network model is the same as the number of interactions required to satisfy the update conditions of the encoder and decoder networks. If the update conditions are met, the neural network models in each group respectively draw a small batch of experience tuples from the experience pool model. Based on the joint loss function in the multi-agent-actor-attention mechanism-critic framework, the parameters of the actor network and critic network of each neural network model in the same group are updated. The specific update formula is as follows:

[0118]

[0119]

[0120] in, This represents the average value of the network behavior of other actors in the same group, excluding the wind farm safety risk assessment model i. The constant term coefficients used when updating the target network.

[0121] By adopting a closed-loop training and incremental learning approach, when the system topology or operating conditions change, the incremental learning mechanism is triggered, and the affected sub-stations are locally retrained to ensure the continuous effectiveness of the dynamic behavior baseline and risk assessment model deployment, thereby enabling the online real-time deployment of the assessment model.

[0122] Step 105: Based on the wind farm cluster information collaborative intelligent security risk assessment model, assess the information security risks of the wind power system.

[0123] It should be noted that the information security risk assessment results of the wind power system are obtained by using the trained and grouped wind farm cluster information collaborative intelligent security risk assessment model.

[0124] This invention provides a maintenance plan optimization method for power system simulation operation. It collects multi-source heterogeneous data from wind farm clusters, constructs a multi-source data fusion model to obtain a multi-dimensional data observation tensor, and builds an adaptive spatiotemporal graph neural network-based baseline model of wind power system information dynamic behavior based on the multi-dimensional data observation tensor. Based on the difference between the output of the wind power system information dynamic behavior baseline model and the actual data, a wind farm cluster information collaborative intelligent safety risk assessment model is constructed. This achieves accurate characterization and intelligent risk assessment of wind power system information behavior, thereby improving the information security protection level and long-term operational stability of wind power systems in complex and variable environments. It solves the technical problems of existing wind power system information security assessment methods, which are unable to comprehensively characterize the system's security status and fail to fully consider the time-varying nature and operating condition drift of wind power system operation.

[0125] This invention overcomes the limitations of traditional methods that rely solely on a single data source or historical experience by integrating multi-source data from the wind turbine itself, site control, network security monitoring, and the external environment. This not only broadens the assessment perspective but also more accurately reflects the overall security status of the system. The information dynamic behavior baseline model constructed using an adaptive spatiotemporal graph neural network effectively captures the co-evolutionary patterns between information flow, control flow, and physical state of the wind power system under normal operating conditions, providing a solid foundation for subsequent anomaly detection and risk assessment. The offshore wind farm cluster information collaborative intelligent security risk assessment model, designed based on deep reinforcement learning and parameter sharing mechanisms, not only considers the common knowledge among wind farms to improve training efficiency but also takes into account the unique differences between each farm, achieving efficient and personalized security risk assessment. Furthermore, the wind power system information security risk assessment method provided by this invention has good scalability and is applicable to the training and deployment of security risk assessment models for a larger number of wind farm clusters.

[0126] For easier understanding, please refer to Figure 4 This invention provides an embodiment of a wind power system information security risk assessment device, comprising:

[0127] The data acquisition module is used to acquire multi-source heterogeneous data of wind farm clusters and store the multi-source heterogeneous data in a historical database.

[0128] The fusion module is used to process multi-source heterogeneous data using a multi-source heterogeneous data feature extraction and spatiotemporal registration fusion model to obtain multi-dimensional data observation tensors;

[0129] The first modeling module is used to construct a baseline model of the dynamic behavior of wind power system information based on an adaptive spatiotemporal graph neural network, based on multidimensional data observation tensors.

[0130] The second modeling module is used to construct a wind farm cluster information collaborative intelligent safety risk assessment model based on the difference between the output of the wind power system information dynamic behavior baseline model and the actual data.

[0131] The assessment module is used to assess the information security risks of wind power systems based on the wind farm cluster information collaborative intelligent security risk assessment model.

[0132] In one embodiment, the fusion module is specifically used for:

[0133] Preprocessing and spatiotemporal alignment of multi-source heterogeneous data;

[0134] Based on the same time reference and spatial topology, a multidimensional spatiotemporal data cube is constructed. Each data voxel of the multidimensional spatiotemporal data cube corresponds to a multidimensional observation on a device node within a specific time window.

[0135] A multidimensional observation tensor is obtained by processing the multidimensional observations of a multidimensional spatiotemporal data cube using a multi-source heterogeneous data feature extraction and spatiotemporal registration fusion model based on a convolutional autoencoder.

[0136] In one embodiment, the first modeling module is specifically used for:

[0137] The physical devices and communication interactions of a wind power system are abstracted into a dynamic information network diagram. Where V is the node set, E is the connection relationship, and A is the adjacency matrix representing the strength of the association between nodes;

[0138] On the time branch, a long short-term memory network is used at each node to extract trend, periodic and abrupt features in the time series, capturing the evolutionary pattern of single-node behavior. On the spatial branch, a graph attention network is used to aggregate the node features at each time step, capturing the state influence between spatially adjacent or functionally related nodes, and a gating mechanism is introduced to control the information propagation weights.

[0139] An adaptive spatiotemporal graph neural network is constructed and unsupervised training is performed with the goal of reconstructing future states. By minimizing the error between the predicted values ​​and the actual observed values ​​of the adaptive spatiotemporal graph neural network, a dynamic behavior baseline model is obtained through training.

[0140] In one embodiment, the multi-source heterogeneous data includes: wind turbine operating data, site control and communication data, network security monitoring data, and external meteorological and environmental data.

[0141] In one embodiment, each safety risk assessment model in the wind farm cluster information collaborative intelligent safety risk assessment model includes an actor network, a commentator network, a target actor network, a target commentator network, a decision preference feature encoder network, and a decision preference feature decoder network. The actor network, commentator network, target actor network, target commentator network, decision preference feature encoder network, and decision preference feature decoder network are all deep neural networks.

[0142] The input to the actor network is the decoded value of the current observation data state and input decision preference features, and the output is the discrete safety risk level and risk assessment confidence of the wind farm.

[0143] The input to the critic network is a set of states and decisions sampled from all neural network models within the same group, and the output is a state-behavior value function.

[0144] The input to the target actor network is the observed data state at the next time step and the decoded value of the input decision preference features. The output is the discrete safety risk level and risk assessment confidence of the wind farm at the next time step.

[0145] The input to the critic network is the set of data states and decisions for the next time step sampled from all neural network models within the same group, and the output is the target state behavior value function for the next time step.

[0146] The input to the decision preference feature encoder network is a continuous decision history sequence constructed from multiple continuous experience tuples, and the output is the decision preference features of the neural network model.

[0147] The input to the decision preference feature decoder network is the decision preference feature encoding, and the output is the decoded neural network model decision preference features.

[0148] In one embodiment, a training module is also included. The training module is used to train the intelligent safety risk assessment model based on wind farm cluster information collaboration. The training method includes the following steps:

[0149] S1. Initialize the safety risk assessment model for each wind farm by reading the real-time observation data stored in the corresponding data environment. and the maximum number of observation data reads, K;

[0150] S2. The actor network of the safety risk assessment model for each wind farm reads the real-time observation data status and the decoded value of the input decision preference characteristics of the wind farm, and outputs the risk level and risk assessment confidence of each wind farm.

[0151] S3. The data environment of each wind farm station is constructed based on the risk level and risk assessment confidence of each wind farm station output by the safety risk assessment model of each wind farm station. The reward function related to the decision bias of the safety risk assessment model of each wind farm station and the discrete risk assessment level and risk assessment confidence is calculated and fed back to the observation data of the next wind farm station.

[0152] S4. The safety risk assessment model for each wind farm is based on the current observation data of each wind farm, the risk level and risk assessment confidence of each wind farm, the reward function related to the decision bias of the safety risk assessment model of each wind farm and the discrete risk assessment level and risk assessment confidence, and the observation data of the next wind farm. The model constructs an experience tuple and saves it to the experience pool.

[0153] S5. The safety risk assessment model of each wind farm station extracts a preset number of continuous experience tuples from the experience pool to construct the decision history sequence of the safety risk assessment model of each wind farm station. Based on the constructed decision history sequence, the safety risk assessment model of each wind farm station updates the encoder network parameters and decoder network parameters of the output safety risk assessment decision preference characteristics.

[0154] S6. Group the decision preference features of the safety risk assessment models of each wind farm based on the K-means clustering algorithm;

[0155] S7. Increment the number of times the safety risk assessment model of each wind farm station reads the observation data by 1. Determine whether the number of times the safety risk assessment model of each wind farm station reads the observation data has reached the maximum number of observation data reads. If yes, complete the training of the intelligent safety risk assessment model based on wind farm group information collaboration. Otherwise, return to step S2.

[0156] In one embodiment, the objective function and decision variables of the wind farm cluster information collaborative intelligent safety risk assessment model are:

[0157]

[0158]

[0159]

[0160] Where N represents the number of safety risk assessment models for wind farms, and K represents the maximum number of observation data reads for each safety risk assessment model for wind farms. To determine the true safety level of the wind farm's safety risk assessment model i during its k-th safety risk assessment. Let be the real-time decision value of the safety risk assessment model i for the k-th safety risk assessment of the wind farm, and M be the total number of safety risk assessment levels. This is the set of decision values ​​for the safety risk assessment model of each wind farm in the wind farm cluster.

[0161] The wind power system information security risk assessment device provided in this invention is used to execute the wind power system information security risk assessment method provided in this invention. Its principle and the technical effects achieved are the same as those of the wind power system information security risk assessment method provided in this invention, and will not be repeated here.

[0162] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0163] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for assessing information security risks of a wind power system, characterized in that, The method comprises the following steps: acquiring multi-source heterogeneous data of a wind farm group and storing the multi-source heterogeneous data into a historical database; processing the multi-source heterogeneous data by using a multi-source heterogeneous data feature extraction and space-time registration fusion model to obtain a multi-dimensional data observation tensor; constructing a wind power system information dynamic behavior baseline model based on the adaptive space-time graph neural network based on the multi-dimensional data observation tensor; constructing a wind farm group information collaborative intelligent security risk assessment model based on the difference between the output of the wind power system information dynamic behavior baseline model and actual data; evaluating the wind power system information security risk based on the wind farm group information collaborative intelligent security risk assessment model.

2. The wind power system information security risk assessment method of claim 1, wherein, The multi-source heterogeneous data comprises wind turbine body operation data, station control communication data, network security monitoring data and external meteorological and environmental data.

3. The method of claim 1, wherein, The multi-source heterogeneous data is processed by using the multi-source heterogeneous data feature extraction and space-time registration fusion model to obtain the multi-dimensional data observation tensor, which comprises the following steps: preprocessing and space-time alignment of the multi-source heterogeneous data; constructing a multi-dimensional space-time data cube based on the same time reference and space topology structure, each data voxel of the multi-dimensional space-time data cube corresponding to a multi-dimensional observation on a device node within a specific time window; processing the multi-dimensional observation of the multi-dimensional space-time data cube by using the multi-source heterogeneous data feature extraction and space-time registration fusion model based on the convolutional autoencoder to obtain the multi-dimensional data observation tensor.

4. The method of claim 1, wherein, The wind power system information dynamic behavior baseline model based on the adaptive space-time graph neural network is constructed based on the multi-dimensional data observation tensor, which comprises the following steps: A dynamic information network graph is abstracted from physical devices and communication interaction of a wind power system Wherein, V is a node set, E is a connection relationship, and A is an adjacency matrix representing the connection strength between nodes. on the time branch, using the long short-term memory network to extract the trend, period and mutation features in the time series on each node to capture the single-node behavior evolution law, on the space branch, using the graph attention network to aggregate the node features at each time to capture the state influence between adjacent or functionally related nodes in space, and introducing a gating mechanism to control the information propagation weight; constructing the adaptive space-time graph neural network to perform unsupervised training with the goal of reconstructing the future state, training the dynamic behavior baseline model by minimizing the error between the prediction value of the adaptive space-time graph neural network and the real observation value.

5. The method of claim 1, wherein, Each security risk assessment model in the wind farm group information collaborative intelligent security risk assessment model comprises an actor network, a critic network, a target actor network, a target critic network, a decision preference feature encoder network and a decision preference feature decoder network, and the actor network, the critic network, the target actor network, the target critic network, the decision preference feature encoder network and the decision preference feature decoder network are all deep neural networks; the input of the actor network is the current observation data state and the decoded value of the input decision preference feature, and the output is the wind power station discrete security risk level and risk assessment confidence; the input of the critic network is the state and decision set sampled by all neural network models in the same group, and the output is the state behavior value function; the input of the target actor network is the observation data state at the next time and the decoded value of the input decision preference feature, and the output is the wind power station discrete security risk level and risk assessment confidence at the next time; the input of the target critic network is the observation data state at the next time and the decoded value of the input decision preference feature, and the output is the wind power station discrete security risk level and risk assessment confidence at the next time. The input of the critic network is the data state and decision set sampled by all neural network models in the same group at the next moment, and the output is the target state behavior value function at the next moment; The input of the decision preference feature encoder network is a continuous decision history sequence constructed by multiple continuous experience tuples, and the output is a decision preference feature of the neural network model; The input of the decision preference feature decoder network is the decision preference feature code, and the output is the decoded decision preference feature of the neural network model.

6. The wind power system information security risk assessment method of claim 5, wherein, Before the wind farm group information collaborative intelligent security risk assessment model is used to evaluate the information security risk of the wind power system, the following steps are further included: The wind farm group information collaborative intelligent security risk assessment model is trained, and the training method includes the following steps: S1, initializing the safety risk assessment model of each wind power station reads the number of times of real-time observation data saved in the corresponding data environment and the maximum observation data reading number K; S2, the actor network of the security risk assessment model of each wind power station reads the real-time observation data state of the wind power station and the decoded value of the input decision preference feature, and outputs the risk level and risk assessment confidence of each wind power station; S3, the data environment of each wind power station is constructed based on the risk level and risk assessment confidence of each wind power station output by the security risk assessment model of each wind power station, and the reward function related to the decision deviation of the security risk assessment model of each wind power station and the discrete risk assessment level and risk assessment confidence is calculated to feed back the observation data of the next wind power station; S4, the security risk assessment model of each wind power station is based on the current observation data of each wind power station, the risk level and risk assessment confidence of the wind power station, the reward function related to the decision deviation of the security risk assessment model of each wind power station and the discrete risk assessment level and risk assessment confidence, and the observation data of the next wind power station, to construct experience tuples and save them to the experience pool; S5, the security risk assessment model of each wind power station extracts a preset number of continuous experience tuples from the experience pool, constructs a decision history sequence of the security risk assessment model of each wind power station, and updates the encoder network parameters and decoder network parameters of the security risk assessment model of each wind power station based on the constructed decision history sequence; S6, the decision preference features of the security risk assessment model of each wind power station are grouped based on the K-means clustering algorithm; S7, the number of times that the current security risk assessment model of each wind power station reads the observation data is increased by 1, and it is judged whether the number of times that the current security risk assessment model of each wind power station reads the observation data reaches the maximum observation data reading times, if yes, the training of the wind farm group information collaborative intelligent security risk assessment model is completed, otherwise, return to step S2.

7. The method of claim 1, wherein, The objective function and decision variable of the wind farm group information collaborative intelligent security risk assessment model are: Wherein, N is the number of safety risk assessment models of the wind farm station, K is the maximum observation data reading number of the safety risk assessment model of each wind farm station, is the real safety level assessment when the safety risk assessment model i of the wind farm station is performed for the kth time, is the real-time decision value when the safety risk assessment model i of the wind farm station is performed for the kth time, and M is the total number of safety risk assessment levels, is the decision value set of the safety risk assessment model of each wind farm station of the wind farm station group.

8. A wind power system information security risk assessment apparatus, characterized by, It includes: A data acquisition module is configured to acquire multi-source heterogeneous data of a wind farm group and store the multi-source heterogeneous data into a historical database; A fusion module is configured to process the multi-source heterogeneous data by using a multi-source heterogeneous data feature extraction and spatio-temporal registration fusion model to obtain a multi-dimensional data observation tensor; A first modeling module is configured to construct a wind power system information dynamic behavior baseline model based on an adaptive spatio-temporal graph neural network based on the multi-dimensional data observation tensor; The second modeling module is configured to construct a wind farm group information collaborative intelligent security risk assessment model based on a difference between an output of a wind power system information dynamic behavior baseline model and actual data. The evaluation module is configured to evaluate a wind power system information security risk based on the wind farm group information collaborative intelligent security risk assessment model.

9. The wind power system information security risk assessment apparatus according to claim 8, wherein, The fusion module is specifically configured to: perform preprocessing and spatio-temporal alignment on the multi-source heterogeneous data; construct a multi-dimensional spatio-temporal data cube based on the same time reference and spatial topology structure, each data voxel of the multi-dimensional spatio-temporal data cube corresponding to multi-dimensional observation quantities on a device node within a specific time window; perform processing on the multi-dimensional observation quantities of the multi-dimensional spatio-temporal data cube by using a multi-source heterogeneous data feature extraction and spatio-temporal registration fusion model based on a convolutional autoencoder to obtain a multi-dimensional data observation tensor. 10.The wind power system information security risk assessment apparatus of claim 8, wherein, The first modeling module is specifically configured to: A dynamic information network graph is abstracted from physical devices and communication interaction of a wind power system Wherein, V is a node set, E is a connection relationship, and A is an adjacency matrix representing the connection strength between nodes. extract trend, period and mutation features in a time sequence on each node on a time branch by using a long short-term memory network, capture a single-node behavior evolution law, aggregate node features at each time on a spatial branch by using a graph attention network, capture state influences between adjacent or functionally related nodes in space, and introduce a gating mechanism to control information propagation weights; construct an adaptive spatio-temporal graph neural network to perform unsupervised training with a future state as a target, train the dynamic behavior baseline model by minimizing an error between a prediction value of the adaptive spatio-temporal graph neural network and a real observation value.

Images