A distribution network state assessment method and system based on artificial intelligence

Abstract

The present invention provides a distribution network state assessment method and system based on artificial intelligence. The method includes acquiring the operating state data of each distribution network device in real time; extracting features from the operating state data of each distribution network device according to a preset operating state feature extraction model to obtain multi-mode operating features of the device; weightedly fusing the multi-mode operating features of each distribution network device according to a particle swarm optimization algorithm and an adaptive gradient information strategy optimization algorithm to obtain a device state fusion feature; constructing a distribution network state topology map based on the target distribution network topology according to the device state fusion features of each distribution network device, and analyzing the distribution network state topology map according to a preset graph neural network model to obtain a distribution network operating state assessment result. The present invention effectively improves the comprehensiveness and accuracy of distribution network state assessment by adaptively fusion analysis of multi-source data features of distribution network devices combined with topological correlation analysis between devices.

Description

Technical Field

[0001] The present invention relates to the technical field of distribution network operating status assessment, and in particular to a distribution network status assessment method and system based on artificial intelligence. Background Art

[0002] Distribution networks provide essential electrical energy for social production and life, serving as crucial infrastructure for significantly improving people's quality of life and supporting information transmission. Their safe and stable operation directly impacts their ability to maintain and maintain power supply quality, which in turn impacts the normal functioning of social production and life. Therefore, reliable operational status assessment of distribution networks is crucial for maintaining the safe and stable operation of power systems and ensuring the quality, safety, and economic efficiency of distribution network power supply.

[0003] Due to the numerous and mutually coupled influencing factors in distribution network operation, such as distribution network equipment, the environment, and external uncertainties, the distribution network's operating state exhibits nonlinear dynamic evolutionary characteristics. To simplify the analysis problem, existing distribution network operating state assessments typically use monitoring data collected by sensors to individually analyze the operating state of each distribution network device. While this type of assessment can reflect the health of the distribution network to a certain extent, it ignores the correlation between distribution network devices and relies solely on a single type of sensor data to analyze device status. This makes it difficult to ensure the comprehensiveness and accuracy of device status assessments, nor can it provide a comprehensive state analysis conclusion for the distribution network as a whole. Furthermore, due to the lack of an adaptive adjustment mechanism to cope with the ever-changing grid operating environment, it is difficult to adjust the health status assessment model in real time to reflect the latest changes in the distribution network's health status, limiting the application value of the assessment results. Summary of the Invention

[0004] The purpose of the present invention is to provide a distribution network status assessment method based on artificial intelligence. The distribution network equipment status assessment is realized by extracting multi-source data features of the distribution network equipment and performing weighted fusion according to the adaptive gradient information strategy optimization algorithm. The topological correlation analysis between the equipment is combined to obtain the evaluation result of the operation status of the entire distribution network. On the basis of improving the timeliness, comprehensiveness and reliability of the distribution network equipment status assessment, the comprehensiveness and accuracy of the automatic evaluation of the distribution network status are effectively improved, thereby providing reliable protection for the stable and safe operation of the power system.

[0005] In order to achieve the above objectives, it is necessary to provide an artificial intelligence-based distribution network status assessment method and system to address the above technical problems.

[0006] In a first aspect, an embodiment of the present invention provides a distribution network status assessment method based on artificial intelligence, the method comprising the following steps:

[0007] Real-time acquisition of operating status data of each distribution network device within the target distribution network; the operating status data includes device sensor monitoring data, device image data, and device log text data; the device sensor monitoring data includes the device's corresponding telemetry data, telesignaling data, remote control data, and environmental data;

[0008] According to the preset operation status feature extraction model, feature extraction is performed on the operation status data of each distribution network device to obtain the corresponding multi-mode operation features of the device; the multi-mode operation features of the device include device operation data features, device image data features and device log text features; the device operation data features include telemetry features, telesignaling features, remote control features and environmental features;

[0009] According to the adaptive gradient information strategy optimization algorithm, the multi-mode operation characteristics of each distribution network device are weighted and fused to obtain the corresponding device status fusion characteristics;

[0010] According to the equipment status fusion characteristics of each distribution network equipment, a distribution network status topology map is constructed based on the network topology of the target distribution network, and the distribution network status topology map is analyzed according to a preset graph neural network model to obtain the corresponding distribution network operation status evaluation result.

[0011] In a second aspect, an embodiment of the present invention provides a distribution network status assessment system based on artificial intelligence, the system comprising:

[0012] A data acquisition module is used to obtain the operating status data of each distribution network device in the target distribution network in real time; the operating status data includes device sensor monitoring data, device image data and device log text data; the device sensor monitoring data includes the corresponding telemetry data, telesignaling data, remote control data and environmental data of the device;

[0013] A feature extraction module is used to extract features from the operating status data of each distribution network device according to a preset operating status feature extraction model to obtain corresponding multi-mode operating features of the device; the multi-mode operating features of the device include device operating data features, device image data features, and device log text features; the device operating data features include telemetry features, telesignaling features, remote control features, and environmental features;

[0014] The feature fusion module is used to perform weighted fusion on the multi-mode operation features of each distribution network device based on the adaptive gradient information strategy optimization algorithm to obtain the corresponding device state fusion features;

[0015] The status assessment module is used to construct a distribution network status topology map based on the network topology of the target distribution network according to the device status fusion characteristics of each distribution network device, and analyze the distribution network status topology map according to the preset graph neural network model to obtain the corresponding distribution network operation status assessment result.

[0016] The above-mentioned application provides a distribution network status assessment method and system based on artificial intelligence. The method realizes real-time acquisition of the operating status data of each distribution network device in the target distribution network, including device sensor monitoring data, device image data and device log text data. According to the preset operating status feature extraction model, the operating status data of each distribution network device is respectively extracted to obtain the corresponding device multi-mode operating features including device operating data features, device image data features and device log text features. Then, according to the adaptive gradient information strategy optimization algorithm, the device multi-mode operating features of each distribution network device are weighted fused to obtain the corresponding device status fusion features. According to the device status fusion features of each distribution network device, a distribution network status topology map is constructed based on the network topology of the target distribution network, and the distribution network status topology map is analyzed according to the preset graph neural network model to obtain the corresponding distribution network operating status assessment result. Technical solution. Compared with the existing technology, this artificial intelligence-based distribution network status assessment method is based on the adaptive dynamic fusion analysis of multi-source data characteristics of distribution network equipment combined with the topological correlation analysis between equipment. On the basis of improving the timeliness, comprehensiveness and reliability of distribution network equipment status assessment, it effectively improves the comprehensiveness and accuracy of automatic distribution network status assessment, thereby providing reliable protection for the stable and safe operation of the power system. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 1 is a flow chart of a method for evaluating the state of a distribution network based on artificial intelligence according to an embodiment of the present invention;

[0018] Figure 2 Schematic diagram of the structure of the distribution network status assessment system based on artificial intelligence in an embodiment of the present invention. DETAILED DESCRIPTION

[0019] In order to make the purpose, technical solutions and beneficial effects of this application more clear, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. Obviously, the embodiments described below are part of the embodiments of the present invention and are only used to illustrate the present invention, but are not used to limit the scope of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0020] The artificial intelligence-based distribution network status assessment method provided by the present invention can be understood as addressing the current situation where existing distribution network status assessment methods struggle to produce comprehensive, reliable, and adaptive operating status assessment results that adapt to changing operating environments. This proposed method utilizes feature extraction from multiple sources, including sensor monitoring data (telemetry data, telesignaling data, remote control data, and environmental data), device image data, and device operation log text, to assess distribution network device status through weighted fusion using a particle swarm optimization algorithm and an adaptive gradient information strategy optimization algorithm. This method also combines topological correlation analysis between devices to produce an overall distribution network operating status assessment result. The following examples will provide a detailed explanation of the present invention's artificial intelligence-based distribution network status assessment method.

[0021] In one embodiment, Figure 1 As shown, a distribution network state assessment method based on artificial intelligence is provided, comprising the following steps:

[0022] S11. Acquire operating status data of each distribution network device in the target distribution network in real time; wherein the operating status data can be understood as data related to the operating status of the power equipment in the distribution network collected from multiple aspects. In this embodiment, preferably, the operating status data is set to include device sensor monitoring data, device image data, and device log text data, etc.; wherein the device sensor monitoring data includes telemetry data, telesignaling data, remote control data, and environmental data corresponding to the device; correspondingly, the telemetry data, telesignaling data, and remote control data can be understood as current, voltage, input, output, and control monitoring data of the distribution network operating status; environmental data can be understood as device operation-related data obtained by monitoring various sensors reasonably deployed at key nodes of the power grid according to distribution network requirements, including vibration data and temperature data collected based on vibration sensors, temperature sensors, etc., and is not specifically limited here; device image data can be understood as image data collected by deploying image acquisition equipment, including device appearance image data and infrared image data, etc., which can cover device operating status and health status information; device log text data can be understood as text data storing information records such as device operation, faults, and maintenance, and is not specifically limited here.

[0023] S12. According to the preset operating status feature extraction model, feature extraction is performed on the operating status data of each distribution network device to obtain corresponding multi-mode operating features of the device; wherein, the preset operating status feature extraction model can be understood as a network model that can be used to effectively extract features from multi-source data in the operating status data of the above-mentioned each distribution network device. Preferably, the preset operating status feature extraction model includes a first feature extraction model, a second feature extraction model and a third feature extraction model in parallel; that is, the multi-mode operating features of the device obtained by feature extraction of the operating status data according to the preset operating status feature extraction model include device operating data features, device image data features and device log text features; the device operating data features include telemetry features, telesignaling features, remote control features and environmental features.

[0024] Specifically, the step of extracting features from the operating status data of each distribution network device according to a preset operating status feature extraction model to obtain corresponding multi-mode operating features of the device includes:

[0025] According to the first feature extraction model, feature extraction is performed on the device sensor monitoring data of each distribution network device to obtain corresponding device operation data features; wherein the first feature extraction model can be understood as a network model that can capture the key features of various types of device sensor monitoring data. In this embodiment, preferably, the first feature extraction model is set to include a preset neural network and an operation data feature fusion module connected in sequence, wherein the preset neural network model can select the required network structure according to actual application requirements and the type of collected device sensor monitoring data, such as a Transformer model or a long short-term memory network model; correspondingly, the operation data feature fusion module can be understood as a module that can splice and fuse the features of the extracted various types of device sensor monitoring data, such as a fully connected layer; it should be noted that before using the first feature extraction model to extract operation data features, the device sensor monitoring data needs to be preprocessed including data cleaning, denoising and calibration processing, so as to facilitate reliable feature extraction using the network model, and the preset neural network and operation data feature fusion module in the first feature extraction model can be selected according to actual needs, without specific limitation here. The corresponding feature extraction process can refer to the relevant technical implementation of the corresponding selected model and module, which will not be described in detail here.

[0026] According to the second feature extraction model, feature extraction is performed on the device image data of each distribution network device to obtain corresponding device image data features; the device image data features include operating status indicator light features, alarm indicator light features, device appearance integrity features, corrosion features, sealing features, and connection part tightness features; wherein, the second feature extraction model can, in principle, adopt any network model that can extract image features, but in order to ensure the efficiency and comprehensiveness of device image data feature extraction, this embodiment preferably sets the second feature extraction model to include a convolutional neural network, a topological data analysis module, a sparse coding module, and an image feature fusion module to obtain deep-level features of the device image and enhance the richness and distinctiveness of the feature representation. Specifically, according to the second feature extraction model, feature extraction is performed on the device image data of each distribution network device to obtain corresponding device image data features, including the following steps:

[0027] The device image data is preprocessed to obtain corresponding preprocessed device image data; wherein the device image data can be understood as a device image that can reflect the physical state of a transformer, circuit breaker, DTU (Distribution Terminal Unit), FTU (Feeder Terminal Unit), TTU (Transformer Terminal Unit), fault indicator or other distribution facilities; the corresponding preprocessing may include grayscale, denoising, normalization, image enhancement, edge filling and other processing to avoid problems such as gradient explosion in subsequent model analysis while also improving the accuracy of subsequent feature extraction.

[0028] The image data of the pre-processing device is input into the convolutional neural network for feature extraction to obtain the corresponding first image feature; wherein the convolutional neural network can be understood as a CNN model, assuming that the model parameters are recorded as , then the first image feature (feature map) extracted by the convolutional neural network can be expressed as:

[0029]

[0030] in, Represents pre-processed device image data; is the feature map extracted by the CNN model, and are the height and width of the feature map, is the feature dimension;

[0031] The pre-processed device image data is input into the topology data analysis module for topology extraction, obtaining corresponding second image features. The topology data analysis module can be understood as a functional module capable of performing topology data analysis (TDA), and the topology data analysis module extracts topological features based on a persistent homology algorithm. The corresponding second image features include global features, local features, and topological structures. The specific process for obtaining the second image features is as follows:

[0032] 1) At a given threshold Next, according to the pre-processing device image data , construct the corresponding simplicial complex and obtain the corresponding Vietoris-Rips complex , expressed as:

[0033]

[0034] 2) According to Compute the persistent homology of the Vietoris-Rips complex and use it to extract topological features:

[0035]

[0036] Where,

[0037]

[0038]

[0039] Among them, ker is kernel, which refers to the set of all elements mapped to zero. For a linear map or an operator For example, ker is all the Elements im is image, which refers to all the images that can be processed by the previous operator. The set of elements to be mapped, for a linear map or an operator For example, im is all satisfied Elements A collection of yes Dimension Chain Group, represents the boundary operator, Dimensions representing homology;

[0040] 3) Based on the persistent homology, a persistent barcode is calculated, and the persistent barcode is used to represent the result of the persistent homology, and the persistent homology result is used as the final second image feature:

[0041]

[0042] in, Indicates the The birth and death times of each topological feature.

[0043] The first image feature and the second image feature are input into the image feature fusion module for splicing and fusion to obtain the device image data feature. The specific process of obtaining the device image data feature is as follows:

[0044] First, the first image feature and the second image feature are spliced ​​together to obtain the spliced ​​feature representation:

[0045]

[0046] in, Represents splicing features; represents the feature concatenation function; and represent the first image feature and the second image feature respectively;

[0047] Then the spliced ​​features obtained above are input into the fully connected layer for fusion to obtain the final feature representation, that is, the device image data features Expressed as:

[0048]

[0049] in, Represents device image data characteristics; Represents the activation function. Common activation functions include ReLU and Sigmoid, which are used to introduce nonlinearity. and They represent the weight matrix of the fully connected layer and the bias term of the fully connected layer respectively.

[0050] This embodiment uses a convolutional neural network and data topology analysis to extract image features separately and fuse them to obtain device image data features. The method can not only use CNN to extract local features (hierarchical features) from the image and well capture the spatial structure information and texture features of the image, but also use data topology analysis TDA to extract global topological structural features in the image, such as connectivity and holes. It quantifies topological features through persistent homology (PHS) to improve robustness to noise and has strong feature representation capabilities. That is, the combination of the above two algorithms can better extract local and global features of the image, while also improving the robustness and effectiveness of image feature representation.

[0051] According to the third feature extraction model, feature extraction is performed on the device log text data of each distribution network device to obtain corresponding device log text features; the device log text features include system log features, fault diagnosis log features, maintenance record features and operation record features; wherein, in principle, the third feature extraction model can adopt any network model that can extract text features, but in order to ensure the effective extraction of semantic information and capture contextual semantic relationships, this embodiment preferably sets the second feature extraction model to include a popular learning module, a hypergraph neural network and a text feature fusion module to better handle nonlinear relationships and high-order feature interactions, thereby ensuring the comprehensiveness, accuracy and robustness of text feature extraction. Specifically, according to the third feature extraction model, feature extraction is performed on the device log text data of each distribution network device to obtain corresponding device log text features, including the following steps:

[0052] The device log text data is converted into word vectors to obtain corresponding text word vector embeddings, and position codes are added to the text word vector embeddings to obtain reconstructed text word vector embedding data; wherein, the device log text data is a data set of device operation logs of target power grid equipment collected in the power system. Since the device operation log data is text data, it is necessary to first perform text preprocessing including word segmentation, removal of useless words, etc., and then perform word vector conversion to obtain corresponding text word vector embeddings; it should be noted that the acquisition of text word vector embeddings and the addition of position codes can be implemented by referring to relevant existing technologies and will not be described in detail here.

[0053] The reconstructed text word vector embedding data is input into the popular learning module for dimensionality reduction processing to obtain the corresponding first text feature; wherein, the popular learning module can be understood as a functional module that uses popular learning methods such as Locally Linear Embedding (LLE) or t-SNE (t-distributed stochastic neighbor embedding) to reduce the dimensionality of high-dimensional text embedding. Assume Based on the high-dimensional embedding matrix obtained from the above-mentioned reconstructed text word vector embedding data, the following describes the process of obtaining the first text feature using the local linear embedding (LLE) method as an example:

[0054] When using the Local Linear Embedding (LLE) method, the reconstruction weight matrix is ​​first calculated by finding the nearest neighbors of each data point:

[0055]

[0056] Then, we obtain the reduced-dimensional embedding (first text feature) by solving the following optimization problem:

[0057]

[0058] in, is the weight matrix; and Respectively represent the i-th and j-th samples of the original data point matrix in the high-dimensional space; express right Contribution coefficient of reconstruction; is the embedding after dimensionality reduction; and Respectively and The corresponding low-dimensional space samples. To solve this optimization problem, we first need to initialize , can be initialized by a specific random method, and then iterated by gradient descent or other numerical optimization techniques. until it finally converges.

[0059] According to the reconstructed text word vector embedding data, a log text hypergraph is constructed based on the context information of the device log text data, and the log text hypergraph is input into the hypergraph neural network for feature extraction to obtain the corresponding second text feature; wherein, the log text hypergraph can be understood as each word is embedded into a graph node, and a hyperedge is constructed based on the contextual semantic association of the text word, and each hyperedge is connected to multiple related words to obtain a graph. The specific implementation of constructing the log text hypergraph based on the context information of the device log text data can refer to the construction method of the related text hypergraph, which will not be described in detail here; the corresponding second text feature acquisition process is:

[0060] First, based on the log text hypergraph, we obtain the corresponding hypergraph adjacency matrix, hyperedge degree matrix, and vertex degree matrix. Then, according to the hypergraph convolution formula of the following hypergraph convolutional neural network, we extract high-order relationship features based on the hypergraph adjacency matrix, hyperedge degree matrix, and vertex degree matrix:

[0061]

[0062] in, and Respectively represent the nodes in Layer and Characteristics of the layer; 、 and denote the hyperedge degree matrix, vertex degree matrix and hypergraph adjacency matrix respectively; represents the weight matrix; represents the activation function;

[0063] Finally, the outputs of each hypergraph convolutional layer are aggregated to generate a high-order feature representation of the text, that is, the second text feature is obtained:

[0064]

[0065] in, represents the node weight vector; N represents the number of nodes; Represents the second text feature.

[0066] The first text feature and the second text feature are input into the text feature fusion module for splicing and fusion to obtain the device log text feature; wherein the text feature fusion module can be understood as a functional module that first splices the first text feature and the second text feature and then performs full-connection fusion processing. The specific process of obtaining the device log text feature is not described in detail here.

[0067] This embodiment uses manifold learning and hypergraph neural networks for text feature extraction. It not only reduces the dimensionality of high-dimensional data through manifold learning while retaining the intrinsic structure and feature relationships of the data to improve the quality of feature extraction, but also captures higher-order and complex relationships in text data based on the hypergraph neural network by constructing a hypergraph, providing richer text feature representation, thereby ensuring improved accuracy and richness of text feature extraction while enhancing the robustness and generalization ability of the text feature extraction model.

[0068] S13. According to the adaptive gradient information strategy optimization algorithm, weighted fusion is performed on the multi-mode operation characteristics of each distribution network device to obtain the corresponding device state fusion characteristics; wherein weighted fusion can be understood as a process of dynamically optimizing the fusion weights of the multi-mode operation characteristics of the device by combining the particle swarm optimization algorithm and the adaptive gradient information strategy optimization algorithm, and performing weighted summation on the multi-mode operation characteristics of the device based on the obtained optimal fusion weight vector; specifically, the step of weighted fusion of the multi-mode operation characteristics of each distribution network device according to the adaptive gradient information strategy optimization algorithm to obtain the corresponding device state fusion characteristics includes:

[0069] The multi-mode operation characteristics of the device are normalized to obtain corresponding multi-mode operation characteristics to be fused; wherein the normalization can be achieved by the following expression:

[0070]

[0071] Where, Indicates the mth feature in the multi-mode operation feature of the device. and Represents the mean and standard deviation of the mth feature; Represents the mth feature among the standardized multi-mode operation features to be fused.

[0072] Establish a device state classification task based on the multi-mode operation features to be fused, and take maximizing the device state classification prediction accuracy as the optimization goal. The device state classification task is iteratively solved by the particle swarm optimization algorithm to obtain the corresponding initial fusion weight vector. The specific process of obtaining the initial fusion weight vector can be understood as first initializing the particle swarm. Each particle represents a feature fusion weight vector configuration and has a corresponding position (i.e., weight). and speed Two attributes, the weight of each particle is expressed as:

[0073]

[0074] in, Represents the fusion weight vector corresponding to the i-th particle; 、 and Respectively represent the fusion weight values ​​of the i particles corresponding to the equipment operation data features, equipment image data features and equipment log text features;

[0075] Based on the purpose of merging the multi-mode operation features of the distribution network equipment for the health status assessment of the distribution network equipment, a device status classification task based on the fusion features is established, and the fitness function is used to evaluate the quality of each particle. This is used to measure the performance of the fusion features on the validation set, and then optimize the fusion weight of the features. During the entire optimization process, the classification accuracy is used as the fitness function to update the speed and position of each particle, and the iteration is repeated to gradually optimize the weight. After several iterations, the required initial fusion weight vector is obtained. , 、 and They respectively represent the initial fusion weight values ​​of the device operation data features, device image data features and device log text features obtained by the particle swarm algorithm.

[0076] It should be noted that, in this embodiment, initializing multiple particles provides a wider search space for optimal solution search, which can accelerate the possibility and speed of finding the global optimal solution, while also avoiding local optimal solutions and ensuring global optimization capability.

[0077] The global initial fusion weight vector obtained by global search through the particle swarm algorithm can provide a reliable optimization basis for the subsequent dynamic adjustment and local optimization of the initial fusion weight vector based on the Adaptive Gradient-Informed Policy Optimization (AGIPO) algorithm. The AGIPO algorithm in this embodiment can be understood as taking the optimal weight vector obtained by the above-mentioned particle swarm algorithm as the starting point, and after generating a batch of state-action-reward experience data using the preset policy network and the multi-mode operation characteristics to be fused, it provides a more accurate and stable short-term policy improvement direction based on the introduction of the policy update step size combined with the gradient information of the policy network advantage function to guide and fine-tune the target policy. Strategy, so that the target strategy can better approximate the optimal strategy of the improved deep reinforcement learning optimization algorithm. Among them, the policy network is a model for approximating the policy function, which accepts the state as input and outputs the probability distribution of each possible action in the state. That is, in deep reinforcement learning, the policy network is usually a function that maps from the state space to the action probability distribution space, through the parameters To define and update, the parameter θ is an adjustable parameter of the policy network, which determines how the policy network maps the input state to the output action probability distribution. By adjusting θ, the policy network can avoid falling into local optimality, learn better strategies, make better decisions in the environment, lead to higher cumulative rewards, and show more stable and robust behavior. This is also the core of the AGIPO algorithm learning and training. The optimal strategy refers to the strategy that can maximize the expected cumulative return in a given Markov decision process MDP. Here, MDP is a four-tuple ,in is the state space, is the action space, is the state transition probability, is the reward function; the corresponding expected cumulative return is the expected value of the total reward obtained by taking actions according to a certain strategy starting from the initial state, which can be expressed as:

[0078]

[0079] The optimal strategy The following conditions must be met:

[0080]

[0081] In order to facilitate the description of the following embodiments, the current policy network is defined as follows: , past policy networks and Policy Network , among which, and Describe the state and corresponding action strategy respectively:

[0082]

[0083]

[0084]

[0085] Where, The policy network parameters are In this case, the state is s and the action strategy is selected The probability value of The policy network parameters are In this case, the state is s and the action strategy is selected The probability value of Yes The Jacobian of the adjusted transformation ensures that the probability of the new policy distribution remains valid. express The action mapping function of the policy network, represents a random variable; Representation Policy Network The second-order gradient of the advantage function of . The following will describe in detail the process of dynamic optimization of the feature fusion weight vector based on the AGIPO algorithm.

[0086] The state space and action space are defined respectively according to the device state fusion feature and the feature fusion weight adjustment vector to construct a Markov decision process model, and the multi-modal operation feature of the device is weightedly fused according to the initial fusion weight vector to obtain the corresponding initial fusion feature; wherein, the state space in the Markov decision process model can be understood as the data space obtained by taking the feature obtained by fusing the multi-modal operation feature to be fused based on the fusion weight as the state; the corresponding action space can be understood as the data space obtained by constructing the action strategy based on the weight adjustment amount of each modal feature; that is, the state space and action space can be understood as the state input and action output of the policy network in the reinforcement learning algorithm, respectively, and how to adjust the fusion weight is determined according to the state, and the new weight is used to generate the state at the next moment; at the same time, the state is also used to calculate the value function and the advantage function to guide the policy update.

[0087] It should be noted that the reward function of the Markov decision process model in this embodiment can use the F1 score (F1Score), that is, the reward value is given based on the performance of the adjusted fusion features on the current task. Before introducing the application process of the AGIPO algorithm, the value function, advantage function, and advantage gradient used by the AGIPO algorithm are defined and explained below:

[0088] 1) Value Function :

[0089]

[0090] In the formula, the value function It is a statement starting from state s, according to the strategy The expected cumulative reward function that can be obtained by the action; is a discount factor that discounts future rewards. Its value range is [0,1] and is usually set to a value close to 1. is the immediate reward obtained at time t+k, that is, the execution strategy The reward value obtained through the reward function.

[0091] 2) TD residual term , advantage function and advantage gradient , advantage function is used to measure action Superiority over current strategies:

[0092]

[0093] Where,

[0094]

[0095] Correspondingly, the advantage gradient Expressed as:

[0096]

[0097] in, is the TD residual for the action The derivative of is used to measure the execution of the action at the current moment impact on future value; represents the initial advantage gradient value; Represents the TD residual at time t+n, which is used to measure the difference between the current policy and the value function prediction; It is a hyperparameter used to control the trade-off between bias and variance. The introduction of λ can smooth the advantage estimate to obtain a more stable estimation result, that is, λ controls the weight of the time step k; Is a decay factor that makes the contribution of farther time steps k to the advantage estimate gradually decrease; introduce hyperparameters To adjust the weighted sum of future rewards and improve the flexibility and adaptability of the strategy; introduce hyperparameters This can enhance the control over gradient changes and thus improve the stability and accuracy of gradient estimation.

[0098] According to the initial fusion features, preset random variables and preset strategy network, an experience training set is constructed based on the Markov decision process model, and according to the experience training set and the adaptive gradient information strategy optimization algorithm, the preset strategy network is subjected to adaptive gradient strategy iterative optimization to obtain the corresponding optimal fusion weight adjustment vector; wherein, the experience training set can be understood as a vector of the form , including the state at different times, random noise, action strategy, reward value and next state data set, which can be obtained by initial fusion feature Add different random noises (preset random variables) The preset policy network in the post-input reinforcement learning model obtains actions , get rewarded And perform the action Then get the new state After obtaining the empirical training set, we can execute the adaptive gradient information policy optimization algorithm based on the empirical training set to perform deep reinforcement learning optimization based on policy gradient optimization to obtain the optimal fusion weight adjustment vector.

[0099] In this embodiment, the adaptive gradient information policy optimization algorithm is preferably set as an improved reinforcement learning algorithm that adaptively controls the policy adjustment step size based on policy gradient information. This can be understood as iterative optimization that adaptively controls the policy adjustment step size based on the policy gradient loss of the preset policy network and the second-order gradient matrix information of the preset advantage function. Specifically, the step of performing adaptive gradient policy iterative optimization on the preset policy network based on the empirical training set and the adaptive gradient information policy optimization algorithm to obtain the corresponding optimal fusion weight adjustment vector includes:

[0100] According to the initial strategy adjustment step size, the preset strategy gradient loss function and the strategy gradient information of the preset advantage function, the action mapping function of the strategy network is constructed, and the approximate optimization is performed on the preset strategy network according to the experience training set and the action mapping function; wherein the action mapping function of the strategy network can be understood as being able to be based on a given state and random variables (e.g. noise) output an action policy ,Right now The action mapping function used in this embodiment is to introduce the policy gradient information (introducing the aforementioned strategy) strategy network mapping function, that is, for -Strategy The mapping function is expressed as:

[0101]

[0102] Where,

[0103]

[0104] in, represents the policy gradient loss.

[0105] Specifically, In practical applications, given the state and random variables , by calculating the action In the advantage function Gradient on , adding gradient information To adjust the strategy, the gradient term Help improve the exploration ability of the strategy and use the strategy to adjust the step size parameter Zoom in and get new actions ; Among them, the nonlinear term It plays a greater role in the initial update and can help speed up the training. In general, the policy adjustment step size α controls the influence of gradient information in the policy update. Specifically, the larger the policy adjustment step size α, the greater the influence of gradient information on the policy. The purpose of adjusting α is to find a balance value so that the policy update can effectively utilize gradient information without causing excessive fluctuations in the policy. To achieve this, the following strategy can be adopted: for the current policy adjustment step size , define the loss function , used to minimize the current strategy and target strategy The differences between:

[0106]

[0107] That is, the optimization strategy is to minimize the loss function To update the corresponding policy network parameters , making Gradual Approximation , that is, implementation strategy Towards The approach.

[0108] According to the second-order gradient matrix information of each sample in the experience training set, the corresponding policy gradient determinant value is calculated, and the minimum determinant value and the maximum determinant value among all policy gradient determinant values ​​are obtained; wherein, the second-order gradient matrix information can be understood as the second-order gradient information obtained by the policy gradient based on the advantage function , the corresponding policy gradient determinant value is based on each and its corresponding state-action pair The variance is calculated and expressed as:

[0109]

[0110] That is, for each and its corresponding state-action pair , calculate the gradient matrix of the second-order advantage function according to the formula , for each , calculate its eigenvalue, assuming that the eigenvalue is .

[0111] Then use these eigenvalues ​​to calculate the matrix The determinant of :

[0112]

[0113] in, is the dimension of the matrix; for each sample , calculate the above determinant and collect the determinant values ​​of all samples Finally, from the collected determinant values Calculate the minimum and maximum values ​​of the samples, that is, get the required minimum and maximum determinant values.

[0114] Calculate the policy reward value of each sample in the experience training set, calculate the iterative expected return value based on all policy reward values, and calculate the policy update out-of-range ratio based on the preset clipping range; the policy reward value can be understood as the current policy network calculated based on the aforementioned reward function. Next state-action pair Corresponding rewards ; Based on the strategy reward values ​​of all samples obtained by sampling, the iterative expected return value (new expected return) can be calculated using the following formula:

[0115]

[0116] in, It represents the expected return value of the iteration; Represents the expected function.

[0117] At the same time, the out-of-range ratio based on the current policy network can be calculated according to the following out-of-range ratio formula:

[0118]

[0119] in, represents the estimated out-of-range ratio, Indicates the preset cropping range, which can be determined according to actual application requirements and is not specifically limited here.

[0120] In order to ensure the scientificity and accuracy of the multimodal feature fusion weight setting, this embodiment also introduces equipment operation telemetry, telesignaling, remote control, image appearance, fault diagnosis log records, and environmental data for comprehensive analysis through the following steps. Among them, environmental data can be understood as environmental factor data that will affect the operating status of the equipment, such as external variables of the power grid equipment status and performance, such as whether there is sampling anomaly causing abnormal telemetry value, whether there is loose wiring causing telesignaling jump, whether there is remote control failure, whether the equipment appearance is corroded and deformed, whether the temperature and humidity are abnormal, etc., and can also be adjusted and set according to actual application requirements.

[0121] According to the telemetry abnormal values, telesignaling jump times, remote control failure times, appearance integrity, fault time, current ambient temperature and current ambient humidity of each distribution network device, the corresponding policy gradient determinant value control parameters are obtained based on the covariance matrix adaptive evolution strategy; among them, the telemetry abnormal values, telesignaling jump times, remote control failure times, appearance integrity, fault time, current ambient temperature and current ambient humidity of the distribution network equipment can be obtained based on the distribution terminal DTU, FTU, TTU, fault indicator, image acquisition device, temperature sensor and humidity sensor, which will not be described in detail this time; the policy gradient determinant value control parameters include upper limit control parameters and lower limit control parameters, which can be understood as the consideration that the telemetry abnormal values, telesignaling jump times, remote control failure times, appearance integrity, fault time, temperature and humidity will have a significant impact on the normal operation of the distribution network equipment. The optimal design is based on the adaptive update of the feature data collected in real time, and the policy gradient determinant value boundary threshold used as the basis for adjusting the policy update step size can ensure that the key parameters in the AGIPO algorithm can be accurately updated while being consistent with the actual operating environment. In actual algorithm applications, the policy gradient determinant value control parameter is used together with the minimum determinant value and the maximum determinant value obtained above as one of the bases for subsequent strategy adjustment step size control. Specifically, the step of obtaining the corresponding policy gradient determinant value control parameter based on the covariance matrix adaptive evolution strategy according to the telemetry abnormal value, telesignaling jump number, remote control failure number, appearance integrity, fault time, current ambient temperature and current ambient humidity of each distribution network device includes:

[0122] Initialize the strategy parameters of the covariance matrix adaptive evolution strategy and generate multiple candidate solutions based on the strategy parameters; the strategy parameters include the mean of the policy gradient determinant value control parameter, the covariance matrix, and the parameter adjustment step size; wherein the number of candidate solutions can be determined according to actual needs, and the corresponding candidate solutions can be expressed as:

[0123]

[0124] In the formula, m and Respectively represent the mean and covariance matrix of the policy gradient determinant value control parameters; Represents the adjustment step size of the policy gradient determinant value control parameter; Express obedience Normally distributed random values; represents the kth candidate solution.

[0125] According to the telemetry abnormal values, telesignaling jump times, remote control failure times, appearance integrity, fault time, ambient temperature and ambient humidity of each distribution network device, the distribution network device health score corresponding to each candidate solution is calculated based on the cumulative effect, and the distribution network device health score is used as the corresponding candidate solution evaluation result; Correspondingly, the distribution network device health score can be calculated based on the acquired device feature data: First, it is necessary to introduce the cumulative effect function. During the long-term operation of the equipment, the impact of different candidate solutions on the equipment operation has a cumulative effect, which gradually affects the equipment's operating health status; define the cumulative effect function q , represents the candidate solution The cumulative effect of equipment temperature and humidity under the influence of time t, the corresponding distribution network equipment health score is expressed as:

[0126]

[0127] Where,

[0128]

[0129] in, 、 and They represent the cumulative effect factors of telemetry abnormal value, number of telesignaling jumps, number of remote control failures, appearance integrity, fault time, temperature and humidity respectively; 、 、 、 、 、 and Represent candidate solutions Cumulative effects of time t on equipment telemetry values, telesignaling transition times, remote control failures, appearance integrity, fault duration, temperature, and humidity; Indicates that the i-th distribution network equipment is in the k-th candidate solution The health score of the distribution network equipment calculated below; 、 、 、 、 、 and Respectively represent the i-th distribution network equipment in the Candidate solutions Average telemetry abnormal value, telesignaling jump times, remote control failure times, appearance integrity, fault time, ambient temperature and humidity of the current equipment; 、 、 、 、 、 and They represent the telemetry abnormal value, telesignaling jump number, remote control failure number, appearance integrity, fault time, ambient temperature and ambient humidity under the optimal operating state of the i-th distribution network equipment respectively; 、 、 、 、 、 and They represent the maximum telemetry abnormal value, maximum number of telesignaling transitions, maximum number of remote control failures, maximum appearance integrity, maximum fault time, maximum ambient temperature, and maximum ambient humidity of the i-th distribution network equipment respectively; 、 、 、 、 、 and represents the weight parameter;

[0130] Based on the distribution network equipment health scores corresponding to the candidate solutions obtained above, the corresponding candidate solution evaluation results can be obtained.

[0131] According to the candidate solution evaluation results, a preset number of optimal solutions are selected, and the parameter adjustment step size is updated based on all the optimal solutions; wherein the preset number of optimal solutions can be understood as the top several candidate solutions with larger candidate solution evaluation results obtained based on multiple rounds of selection and reorganization iterative optimization screening, and then the new mean and covariance matrix can be calculated by the following formula:

[0132]

[0133] in, represents the kth optimal solution; Indicates the number of optimal solutions.

[0134] The step size is adjusted according to the updated parameters, and the preset boundary control parameters are updated to obtain the policy gradient determinant value control parameters.

[0135] The current policy adjustment step size is updated according to the iterative expected reward value, the policy update out-of-range ratio, the minimum determinant value, the maximum determinant value, the policy gradient determinant value control parameter, and the preset policy adjustment step size condition; wherein the preset policy adjustment step size condition includes a preset policy adjustment step size reduction condition and an opposite adjustment step size increase condition; the specific step of updating the current policy adjustment step size includes:

[0136] determining, based on the iterative expected reward value, the policy update out-of-range ratio, the minimum determinant value, the maximum determinant value, and the policy gradient determinant value control parameter, whether any of the preset policy adjustment step size reduction conditions is satisfied; the preset policy adjustment step size reduction conditions include: the iterative expected reward value is less than the current policy expected reward value, the policy update out-of-range ratio exceeds a preset threshold, the minimum determinant value is less than the lower limit control parameter, and the maximum determinant value is greater than the upper limit control parameter;

[0137] If satisfied, the current strategy adjustment step size is updated according to the ratio of the current strategy adjustment step size to the preset strategy adjustment parameter, and the strategy adjustment step size is Updated to ,in Compare If it is large, it can be set to a constant greater than 1. Otherwise, the current strategy adjustment step size is updated according to the product of the current strategy adjustment step size and the preset strategy adjustment parameter. Updated to ; about to Limited to 0 and between.

[0138] According to the updated strategy adjustment step size and the preset proxy loss function, the network parameters of the preset strategy network are optimized and updated until the preset iteration stop condition is reached to obtain the optimal fusion weight adjustment vector; wherein the preset proxy loss function is expressed as:

[0139]

[0140] Where,

[0141]

[0142] in, is the agent loss, i.e. reinforcement learning loss; It is a constant parameter between 0 and 1, and its purpose is to prevent The special case of 0, in which case the AGIPO algorithm will not contribute to the policy update; It is a random perturbation term introduced to increase the exploratory nature of the strategy update process and avoid falling into the local optimum. represents random noise; Represents the trimming function, corresponding to and They represent the upper and lower limits of the clipping range, respectively, to prevent the strategy from changing too much in a single update, which may lead to the risk of strategy crash or failure to converge during training, thereby avoiding instability in the strategy update process. By limiting the amplitude of the policy update, the optimization process can be smoother, thereby improving the stability and convergence speed of training.

[0143] In practical applications, after obtaining the updated strategy adjustment step size, we can use the above action mapping function The latest action mapping function is obtained, that is, the preset strategy network is updated, and then the network parameters of the updated preset strategy network are optimized and updated according to the preset proxy loss function until the preset stop condition is reached. For example, the preset number of iterations is reached or the loss converges, and the optimization iteration is stopped. The corresponding optimal fusion weight adjustment vector is obtained based on the strategy network obtained at this time and the current multi-mode operation characteristics to be fused.

[0144] In this embodiment, the step size parameter is updated by introducing a strategy To adjust the action selection of the current strategy, and use the policy gradient information of the advantage function to make the strategy perform better in terms of expected returns, this strategy fine-tuning method can not only ensure the correct direction of strategy improvement, but also ensure the stability of strategy improvement. It can also provide a more accurate strategy improvement direction based on the use of the second-order gradient information of the advantage function, thereby helping to break through the local optimal point, and then ensure the efficiency and reliability of the optimal fusion weight adjustment vector obtained by dynamic optimization of the AGIPO algorithm.

[0145] The initial fusion weight vector is adjusted according to the optimal fusion weight adjustment vector to obtain an optimal fusion weight vector, and the multi-mode operation features to be fused are weightedly fused according to the optimal fusion weight vector to obtain the device state fusion feature.

[0146] This embodiment adopts a fusion weight vector adaptive optimization method that combines PSO preliminary optimization with AGIPO dynamic optimization based on historical cycle health index prediction error and equipment operating environment data. It can adaptively adjust the fusion weight of each sample to adapt to the characteristics of different input samples, enhance the flexibility, adaptability and reliability of the fusion features, and thus provide reliable guarantees for the accuracy of subsequent equipment operation status assessments.

[0147] S14. According to the device status fusion characteristics of each distribution network device, a distribution network status topology map is constructed based on the network topology of the target distribution network, and the distribution network status topology map is analyzed according to a preset graph neural network model to obtain a corresponding distribution network operation status assessment result; wherein, the network topology of the target distribution network can be obtained according to the system networking architecture of the actual distribution network, which will not be described in detail here; the corresponding constructed distribution network status topology map can be understood as a map obtained by taking each distribution network device in the distribution network as a node, the corresponding device status fusion characteristics as the node value, and constructing the node connection edge according to the association relationship between the devices, so as to better capture the correlation and overall topological structure between the quantitative devices, thereby providing a more accurate and comprehensive health status assessment.

[0148] The preset graph neural network model in this embodiment can be built using graph neural networks (GNNs), including graph convolution layers, graph attention layers, and Transformer layers. It captures deeper correlations between devices that are easily overlooked by modeling high-order relationships.

[0149] Specifically, the step of analyzing the distribution network state topology map according to the preset graph neural network model to obtain the corresponding distribution network operation state evaluation result includes:

[0150] Analyze the distribution network state topology through the graph convolution layer to extract the features of each device node;

[0151] Input all device node features into the graph attention layer for feature attention fusion to obtain the corresponding distribution network fusion feature graph;

[0152] Input the distribution network fusion feature map into the Transformer layer for calculation to obtain the distribution network deep feature;

[0153] The distribution network deep features are processed in sequence by the linear change layer and the Softmax classifier to obtain the distribution network operation status evaluation result.

[0154] In practical applications, the process of using the preset graph neural network model to analyze the distribution network state topology map is as follows:

[0155] 1) Use the graph convolution layer to extract device node features:

[0156]

[0157] in, is the normalized graph adjacency matrix; is the degree matrix; It is Node features of the layer; It is The trainable weight matrix of the layer, is the activation function.

[0158] 2) Based on the graph attention mechanism, calculate the attention coefficient of the node pair:

[0159]

[0160] in, is the attention weight vector, is the weight matrix, and Is a device node and The eigenvector of Represents vector concatenation operation;

[0161] Normalized attention coefficient:

[0162]

[0163] Weighted sum:

[0164]

[0165] in, is the updated feature vector of node i, is the attention coefficient, is a trainable weight matrix, is the activation function (such as ReLU); then introduce the position encoding matrix for the device node , combining the outputs of the graph convolution and attention layers:

[0166]

[0167] Here is the node feature matrix after attention mechanism and graph convolution processing, It is the node feature matrix after adding position encoding.

[0168] 3) Distribution network fusion feature map Use the Transformer layer to process and obtain the deep features of the distribution network , classify using linear transformation and Softmax function:

[0169]

[0170] in, It represents the final evaluation result of the distribution network operation status.

[0171] The embodiment of the present application provides a method for obtaining the operating status data of each distribution network device in the target distribution network in real time, including device sensor monitoring data, device image data and device log text data, and extracting features of the operating status data of each distribution network device according to a preset operating status feature extraction model to obtain corresponding device multi-mode operating features including device operating data features, device image data features and device log text features. Then, according to an adaptive gradient information strategy optimization algorithm, weighted fusion is performed on the device multi-mode operating features of each distribution network device to obtain corresponding device status fusion features, and according to the device status fusion features of each distribution network device, a distribution network status topology map is constructed based on the network topology of the target distribution network, and the distribution network status topology map is analyzed according to a preset graph neural network model to obtain The corresponding scheme for the distribution network operation status assessment results can not only capture rich equipment operation status characteristics in real time by collecting multi-source data, and combine the multi-modal feature fusion weight dynamic optimization strategy to conduct real-time online scientific and reasonable assessment of equipment health status, effectively improving the timeliness, efficiency and accuracy of equipment health assessment, facilitating timely perception and prevention of power grid equipment failures, thereby effectively reducing power production losses and maintenance costs, but also can be based on the topological correlation analysis between devices. On the basis of improving the timeliness, comprehensiveness and reliability of distribution network equipment status assessment, it can further conduct real-time and efficient operation status prediction analysis of the entire power system, effectively improve the comprehensiveness and accuracy of the automatic assessment of the distribution network status, provide reliable data support for the formulation of stable operation strategies for the power system, and thus provide effective guarantees for the stable and safe operation of the power system.

[0172] It should be noted that although the steps in the above flowchart are shown in sequence as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders.

[0173] In one embodiment, Figure 2 As shown, a distribution network status assessment system based on artificial intelligence is provided, and the system includes:

[0174] Data acquisition module 1 is used to obtain the operating status data of each distribution network device in the target distribution network in real time; the operating status data includes device sensor monitoring data, device image data and device log text data; the device sensor monitoring data includes the corresponding telemetry data, telesignaling data, remote control data and environmental data of the device;

[0175] Feature extraction module 2 is used to extract features from the operating status data of each distribution network device according to a preset operating status feature extraction model to obtain corresponding multi-mode operating features of the device; the device operating data features include telemetry features, telesignaling features, remote control features, and environmental features; the multi-mode operating features of the device include device operating data features, device image data features, and device log text features;

[0176] Feature fusion module 3 is used to perform weighted fusion on the multi-mode operation features of each distribution network device according to the adaptive gradient information strategy optimization algorithm to obtain the corresponding device state fusion features;

[0177] The status assessment module 4 is used to construct a distribution network status topology map based on the network topology of the target distribution network according to the device status fusion characteristics of each distribution network device, and analyze the distribution network status topology map according to the preset graph neural network model to obtain the corresponding distribution network operation status assessment result.

[0178] For the specific limitations of the distribution network status assessment system based on artificial intelligence, please refer to the limitations of the distribution network status assessment method based on artificial intelligence above. The corresponding technical effects can also be obtained equivalently, so they will not be repeated here. The various modules in the above-mentioned distribution network status assessment system based on artificial intelligence can be implemented in whole or in part through software, hardware and their combination. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0179] In summary, an embodiment of the present invention provides an artificial intelligence-based distribution network state assessment method and system, wherein the artificial intelligence-based distribution network state assessment method realizes real-time acquisition of the operating state data of each distribution network device in the target distribution network, including device sensor monitoring data, device image data, and device log text data; according to a preset operating state feature extraction model, feature extraction is performed on the operating state data of each distribution network device respectively, and after obtaining the corresponding device multi-mode operating features including device operating data features, device image data features, and device log text features, weighted fusion is performed on the device multi-mode operating features of each distribution network device respectively according to an adaptive gradient information strategy optimization algorithm to obtain corresponding device state fusion features; and according to the device state fusion features of each distribution network device, a distribution network state topology map is constructed based on the network topology of the target distribution network, and according to the preset graph neural network model, a distribution network state topology map is constructed based on the network topology of the target distribution network. The invention discloses a technical solution for analyzing the distribution network state topology diagram to obtain the corresponding distribution network operation state evaluation result. The method can not only capture rich equipment operation state characteristics in real time by collecting multi-source data, and combine the multi-modal feature fusion weight dynamic optimization strategy to conduct real-time online scientific and reasonable evaluation of equipment health state, effectively improve the timeliness, efficiency and accuracy of equipment health evaluation, facilitate timely perception and prevention of power grid equipment failure, and thus effectively reduce power production losses and maintenance costs, but also can be based on the topological correlation analysis between devices. On the basis of improving the timeliness, comprehensiveness and reliability of distribution network equipment state evaluation, further real-time and efficient operation state prediction analysis of the entire power system can be carried out, effectively improving the comprehensiveness and accuracy of automatic evaluation of distribution network state, providing reliable data support for the formulation of stable operation strategy of power system, and thus providing effective guarantee for stable and safe operation of power system.

[0180] Each embodiment in this specification is described in a progressive manner, and the same or similar parts of each embodiment can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment. It should be noted that the various technical features of the above embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the various technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0181] The above-described embodiments merely represent several preferred implementations of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent. It should be noted that a person skilled in the art could make several improvements and substitutions without departing from the technical principles of the present invention, and these improvements and substitutions should also be considered within the scope of protection of the present application. Therefore, the scope of protection of the present patent application shall be based on the scope of protection of the claims.

Claims

1. A distribution network status assessment method based on artificial intelligence, characterized in that: The method comprises the following steps: Real-time acquisition of operating status data of each distribution network device within the target distribution network; the operating status data includes device sensor monitoring data, device image data, and device log text data; the device sensor monitoring data includes the device's corresponding telemetry data, telesignaling data, remote control data, and environmental data; According to the preset operation status feature extraction model, feature extraction is performed on the operation status data of each distribution network device to obtain the corresponding multi-mode operation features of the device; the multi-mode operation features of the device include device operation data features, device image data features and device log text features; the device operation data features include telemetry features, telesignaling features, remote control features and environmental features; According to the adaptive gradient information strategy optimization algorithm, the multi-mode operation characteristics of each distribution network device are weighted and fused to obtain the corresponding device status fusion characteristics, including: Standardizing the multi-mode operation characteristics of the device to obtain corresponding multi-mode operation characteristics to be fused; Establishing a device state classification task based on the multi-mode operation characteristics to be fused, and taking maximizing the device state classification prediction accuracy as the optimization goal, iteratively solving the device state classification task through the particle swarm optimization algorithm to obtain the corresponding initial fusion weight vector; Defining a state space and an action space respectively according to the device state fusion feature and the feature fusion weight adjustment vector to construct a Markov decision process model, and performing weighted fusion on the multi-mode operation features of the device according to the initial fusion weight vector to obtain a corresponding initial fusion feature; According to the initial fusion features, the preset random variables and the preset policy network, an experience training set is constructed based on the Markov decision process model, and according to the experience training set and the adaptive gradient information policy optimization algorithm, the preset policy network is subjected to adaptive gradient policy iterative optimization to obtain a corresponding optimal fusion weight adjustment vector; the adaptive gradient information policy optimization algorithm is an improved reinforcement learning algorithm that adaptively controls the policy adjustment step size based on policy gradient information; Adjusting the initial fusion weight vector according to the optimal fusion weight adjustment vector to obtain an optimal fusion weight vector, and performing weighted fusion on the multi-mode operation features to be fused according to the optimal fusion weight vector to obtain the device state fusion feature; According to the equipment status fusion characteristics of each distribution network equipment, a distribution network status topology map is constructed based on the network topology of the target distribution network, and the distribution network status topology map is analyzed according to a preset graph neural network model to obtain the corresponding distribution network operation status evaluation result.

2. The distribution network status assessment method based on artificial intelligence according to claim 1, characterized in that: The preset operating state feature extraction model includes a first feature extraction model, a second feature extraction model and a third feature extraction model in parallel; The step of extracting features from the operating status data of each distribution network device according to a preset operating status feature extraction model to obtain corresponding multi-mode operating features of the device includes: According to the first feature extraction model, feature extraction is performed on the device sensor monitoring data of each distribution network device to obtain corresponding device operation data features; the first feature extraction model includes a preset neural network and an operation data feature fusion module connected in sequence; According to the second feature extraction model, feature extraction is performed on the device image data of each distribution network device to obtain corresponding device image data features; the device image data features include operating status indicator light features, alarm indicator light features, device appearance integrity features, corrosion features, sealing features, and connection tightness features; the second feature extraction model includes a convolutional neural network, a topological data analysis module, a sparse coding module, and an image feature fusion module; According to the third feature extraction model, feature extraction is performed on the equipment log text data of each distribution network equipment to obtain corresponding equipment log text features; the equipment log text features include system log features, fault diagnosis log features, maintenance record features and operation record features; the second feature extraction model includes a popular learning module, a hypergraph neural network and a text feature fusion module.

3. The distribution network status assessment method based on artificial intelligence according to claim 2, characterized in that: The step of extracting features from the device image data of each distribution network device according to the second feature extraction model to obtain corresponding device image data features includes: Preprocessing the device image data to obtain corresponding preprocessed device image data; Inputting the pre-processing device image data into the convolutional neural network for feature extraction to obtain corresponding first image features; Inputting the pre-processed device image data into the topology data analysis module for topology extraction to obtain corresponding second image features; the topology data analysis module extracts topology features based on a persistent homology algorithm; The first image feature and the second image feature are input into the image feature fusion module for splicing and fusion to obtain the device image data feature.

4. The distribution network status assessment method based on artificial intelligence according to claim 2, characterized in that: The steps of performing feature extraction on the device log text data of each distribution network device according to the third feature extraction model to obtain corresponding device log text features include: Performing word vector conversion on the device log text data to obtain corresponding text word vector embedding, and adding position encoding to the text word vector embedding to obtain reconstructed text word vector embedding data; Inputting the reconstructed text word vector embedding data into the popular learning module for dimensionality reduction processing to obtain the corresponding first text feature; According to the reconstructed text word vector embedding data, a log text hypergraph is constructed based on the context information of the device log text data, and the log text hypergraph is input into the hypergraph neural network for feature extraction to obtain corresponding second text features; The first text feature and the second text feature are input into the text feature fusion module for splicing and fusion to obtain the device log text feature.

5. The distribution network status assessment method based on artificial intelligence according to claim 1, characterized in that: The step of performing adaptive gradient strategy iterative optimization on the preset strategy network according to the experience training set and the adaptive gradient information strategy optimization algorithm to obtain the corresponding optimal fusion weight adjustment vector includes: Constructing an action mapping function of a policy network according to the initial policy adjustment step size, a preset policy gradient loss function, and policy gradient information of a preset advantage function, and performing approximation optimization on the preset policy network according to the experience training set and the action mapping function; Calculate the corresponding policy gradient determinant value according to the second-order gradient matrix information of each sample in the experience training set, and obtain the minimum determinant value and the maximum determinant value among all policy gradient determinant values; Calculate the policy reward value of each sample in the experience training set, calculate the iterative expected return value based on all policy reward values, and calculate the policy update out-of-range ratio based on the preset clipping range; According to the telemetry abnormal values, telesignaling jump times, remote control failure times, appearance integrity, fault time, current ambient temperature and current ambient humidity of each distribution network device, the corresponding policy gradient determinant value control parameters are obtained based on the covariance matrix adaptive evolution strategy; the policy gradient determinant value control parameters include upper limit control parameters and lower limit control parameters; Update the current policy adjustment step size according to the iterative expected reward value, the policy update out-of-range ratio, the minimum determinant value, the maximum determinant value, the policy gradient determinant value control parameter, and a preset policy adjustment step size condition; According to the updated strategy adjustment step size and the preset proxy loss function, the network parameters of the preset strategy network are optimized and updated until the preset iteration stop condition is reached to obtain the optimal fusion weight adjustment vector.

6. The distribution network status assessment method based on artificial intelligence according to claim 5, characterized in that: The step of obtaining corresponding policy gradient determinant value control parameters based on the covariance matrix adaptive evolution strategy according to the telemetry abnormal values, telesignaling jump times, remote control failure times, appearance integrity, fault time, current ambient temperature and current ambient humidity of each distribution network device includes: Initializing the strategy parameters of the covariance matrix adaptive evolution strategy and generating multiple candidate solutions based on the strategy parameters; the strategy parameters include the mean value of the strategy gradient determinant value control parameter, the covariance matrix and the parameter adjustment step size; Based on the cumulative effect of the telemetry abnormal values, telesignaling jump times, remote control failure times, appearance integrity, fault time, ambient temperature, and ambient humidity of each distribution network device, the distribution network device health score corresponding to each candidate solution is calculated and used as the evaluation result of the corresponding candidate solution. The distribution network device health score is expressed as: in, Indicates that the i-th distribution network equipment is in the Candidate solutions The health score of the distribution network equipment calculated below; 、 、 、 、 、 and Respectively represent the i-th distribution network equipment in the Candidate solutions Average telemetry abnormal value, telesignaling jump times, remote control failure times, appearance integrity, fault time, ambient temperature and humidity of the current equipment; 、 、 、 、 、 and They represent the telemetry abnormal value, telesignaling jump number, remote control failure number, appearance integrity, fault time, ambient temperature and ambient humidity under the optimal operating state of the i-th distribution network equipment respectively; 、 、 、 、 、 and They represent the maximum telemetry abnormal value, maximum number of telesignaling transitions, maximum number of remote control failures, maximum appearance integrity, maximum fault time, maximum ambient temperature, and maximum ambient humidity of the i-th distribution network equipment respectively; 、 、 、 、 、 and represents the weight parameter; Selecting a preset number of optimal solutions based on the candidate solution evaluation results, and updating the parameter adjustment step size based on all the optimal solutions; The step size is adjusted according to the updated parameters, and the preset boundary control parameters are updated to obtain the policy gradient determinant value control parameters.

7. The distribution network status assessment method based on artificial intelligence according to claim 5, characterized in that: The preset proxy loss function is expressed as: Where, in, represents the loss of the agent; is a constant parameter; represents the introduced random disturbance term; represents the trimming function; 、 and Respectively indicate status Next select action strategy The corresponding past policy network, current policy network and strategic networks; Representation Policy Network and strategic networks Weighted ensemble strategy network; Represents The advantage function corresponding to the policy network; Indicates the preset cropping range; Represents the total number of training set samples.

8. The distribution network status assessment method based on artificial intelligence according to claim 1, characterized in that: The preset graph neural network model includes a graph convolution layer, a graph attention layer, a Transformer layer, a linear change layer and a Softmax classifier; The step of analyzing the distribution network state topology map according to the preset graph neural network model to obtain the corresponding distribution network operation state evaluation result includes: Analyze the distribution network state topology through the graph convolution layer to extract the features of each device node; Input all device node features into the graph attention layer for feature attention fusion to obtain the corresponding distribution network fusion feature graph; Input the distribution network fusion feature map into the Transformer layer for calculation to obtain the distribution network deep feature; The distribution network deep features are processed in sequence by the linear change layer and the Softmax classifier to obtain the distribution network operation status evaluation result.

9. A distribution network status assessment system based on artificial intelligence, characterized in that: The method for evaluating the state of a distribution network based on artificial intelligence according to claim 1 is applied, wherein the system comprises: A data acquisition module is used to obtain the operating status data of each distribution network device in the target distribution network in real time; the operating status data includes device sensor monitoring data, device image data and device log text data; the device sensor monitoring data includes the corresponding telemetry data, telesignaling data, remote control data and environmental data of the device; A feature extraction module is used to extract features from the operating status data of each distribution network device according to a preset operating status feature extraction model to obtain corresponding multi-mode operating features of the device; the multi-mode operating features of the device include device operating data features, device image data features, and device log text features; the device operating data features include telemetry features, telesignaling features, remote control features, and environmental features; The feature fusion module is used to perform weighted fusion on the multi-mode operation features of each distribution network device based on the adaptive gradient information strategy optimization algorithm to obtain the corresponding device state fusion features; The status assessment module is used to construct a distribution network status topology map based on the network topology of the target distribution network according to the device status fusion characteristics of each distribution network device, and analyze the distribution network status topology map according to the preset graph neural network model to obtain the corresponding distribution network operation status assessment result.

Images