Power grid equipment health status prediction method and system based on multi-modal data fusion

Abstract

The present invention provides a method and system for predicting the health status of power grid equipment based on multi-modal data fusion. The method includes: obtaining in real time the multi-modal state correlation data of the target power grid equipment; extracting data features from the multi-modal state correlation data according to a preset multi-modal feature extraction model to obtain the multi-modal state features to be analyzed; dynamically obtaining the optimal fusion weight of the multi-modal state features to be analyzed according to the historical cycle health index prediction error of the target power grid equipment and the equipment operation environment data obtained; and performing feature fusion on the multi-modal state features to be analyzed according to the optimal fusion weight to obtain the corresponding predicted value of the equipment health index. By capturing rich equipment operation state features in real time through multi-source data and combining with the dynamic optimization strategy of the multi-modal feature fusion weight, the present invention effectively improves the timeliness, efficiency and accuracy of equipment health assessment, reduces power production losses and maintenance costs, and provides a reliable guarantee for the stable and safe operation of the power system.

Description

Technical Field

[0001] The present invention relates to the technical field of power grid equipment status monitoring, and particularly to a method and system for predicting the health status of power grid equipment based on multi-modal data fusion. Background Art

[0002] With the continuous expansion of the scale of the power system, the increase in power grid equipment has brought challenges to the safe operation and management of the power system. The operating stability of the power system is directly related to social production and life, and the reliability and health status maintenance of power grid equipment are the keys to ensuring stable power supply. Therefore, how to conduct a reliable and accurate health status assessment of power grid equipment is an important guarantee for maintaining the stable and safe operation of the power system.

[0003] In order to monitor the health status of power grid equipment, many physical model monitoring means and artificial intelligence fault prediction methods have been developed in the industry, but they all have certain application defects and are difficult to effectively meet the intelligent, efficient and reliable operation status assessment requirements of power grid equipment: Physical model monitoring means mainly rely on regular equipment detection and subsequent analysis, based on physical and chemical tests, combined with the experience judgment of electrical workers. For example, potential faults are detected based on the chromatographic analysis of transformer oil, or infrared thermal imaging technology is used to detect equipment overheating problems, etc. Although it can directly and specifically judge the state of a certain type of power grid equipment at a specific moment, its shutdown detection will interfere with the continuous operation of the power system, and the evaluation process has hysteresis, making it difficult to ensure the timeliness and preventability of equipment fault discovery, and the subjectivity of human evaluation makes it difficult to ensure the reliability of state analysis; Although artificial intelligence fault prediction methods can achieve fault prediction by analyzing the operation data of equipment through machine learning models, existing methods often ignore the multi-modality of power grid equipment health condition analysis, only analyze and process single-type data, and the evaluation algorithm lacks an adequate adaptive adjustment mechanism to cope with the changing power grid equipment and environmental conditions, making it difficult to adjust the model in real time to reflect the latest changes in equipment status, and it is also difficult to ensure the accuracy and reliability of equipment status prediction results. Summary of the Invention

[0004] The purpose of the present invention is to provide a method for predicting the health status of power grid equipment based on multi-modal data fusion, which can capture rich equipment operation status characteristics in real time by collecting multi-source data, and combine a dynamic optimization strategy for multi-modal feature fusion weights to conduct a scientific and reasonable real-time online assessment of the equipment health status, solve the application defects that it is difficult to ensure real-time performance and accuracy in the existing power grid equipment status prediction, effectively improve the timeliness, efficiency and accuracy of equipment health assessment, reduce power production losses and maintenance costs, and provide a reliable guarantee for the stable and safe operation of the power system.

[0005] To achieve the above object, it is necessary to provide a power grid equipment health status prediction method and system based on multi-modal data fusion for the above technical problems.

[0006] In the first aspect, an embodiment of the present invention provides a power grid equipment health status prediction method based on multi-modal data fusion. The method includes the following steps:

[0007] Obtain multi-modal state correlation data of the target power grid equipment in real time; the multi-modal state correlation data includes equipment image data, equipment operation state time series data, and equipment operation log data;

[0008] According to a preset multi-modal feature extraction model, perform data feature extraction on the multi-modal state correlation data to obtain corresponding multi-modal state features to be analyzed; the multi-modal state features to be analyzed include image features, state time series features, and log text features;

[0009] Obtain the historical cycle health index prediction error of the target power grid equipment and the equipment operation environment data, and dynamically obtain the optimal fusion weight of the multi-modal state features to be analyzed according to the historical cycle health index prediction error and the equipment operation environment data; the historical cycle health index prediction error is the prediction error generated by predicting the health index of the target power grid equipment in the previous evaluation cycle;

[0010] According to the optimal fusion weight, perform feature fusion on the multi-modal state features to be analyzed to obtain the corresponding equipment health index prediction value.

[0011] Further, the preset multi-modal feature extraction model includes a parallel preset neural network model, a long short-term memory network, and a preset natural language processing model;

[0012] The step of performing data feature extraction on the multi-modal state correlation data according to the preset multi-modal feature extraction model to obtain corresponding multi-modal state features includes:

[0013] Perform feature extraction on the equipment image data according to the preset neural network model to obtain corresponding image features;

[0014] Perform feature extraction on the equipment operation state time series data according to the long short-term memory network to obtain corresponding state time series features;

[0015] Perform feature extraction on the equipment operation log data according to the preset natural language processing model to obtain corresponding log text features.

[0016] Further, the step of obtaining the historical cycle health index prediction error of the target power grid equipment includes:

[0017] According to the preset multi-modal feature extraction model, extract features from the multi-modal state correlation data of the previous historical cycle of the target power grid equipment to obtain the corresponding multi-modal features of the historical cycle;

[0018] According to the fusion weight of the previous historical cycle, perform feature fusion on the multi-modal features of the historical cycle to obtain the predicted value of the equipment health index of the corresponding historical cycle;

[0019] Obtain the true value of the equipment health index of the previous historical cycle of the target power grid equipment, and according to the true value of the equipment health index of the previous historical cycle and the predicted value of the equipment health index of the historical cycle, obtain the corresponding prediction error of the health index of the historical cycle.

[0020] Further, the step of dynamically obtaining the optimal fusion weight of the multi-modal state features to be analyzed according to the prediction error of the health index of the historical cycle and the equipment operation environment data includes:

[0021] According to the prediction error of the health index of the historical cycle and the equipment operation environment data, based on the genetic algorithm combined with reinforcement learning, dynamically adjust the fusion weight of the previous historical cycle corresponding to the prediction error of the health index of the historical cycle to obtain the optimal fusion weight; the fusion weight is expressed as:

[0022]

[0023] Among them, and respectively represent the fusion weights of the i-th target power grid equipment in the (t + 1)-th and t-th prediction cycles; represents the prediction error of the health index of the historical cycle of the i-th target power grid equipment in the t-th prediction cycle; represents the equipment operation environment data of the i-th target power grid equipment; represents the joint execution function composed of reinforcement learning and genetic algorithm.

[0024] Further, the step of dynamically adjusting the fusion weight of the previous historical cycle corresponding to the prediction error of the health index of the historical cycle based on the genetic algorithm combined with reinforcement learning to obtain the optimal fusion weight includes:

[0025] According to the prediction error of the health index of the historical cycle and the equipment operation environment data, establish a corresponding fitness function, and with the goal of minimizing the fitness function value, optimize the fusion weight of the previous historical cycle through the genetic algorithm to obtain the initial optimized fusion weight; the fitness function is expressed as:

[0026]

[0027] Among them, represents the fitness function; represents the historical cycle health index prediction error of the i-th target power grid device in the t-th prediction cycle; represents the device operation environment data of the i-th target power grid device;

[0028] According to the historical cycle health index prediction error, construct a corresponding reward function, use the fitness function as the action value function, and perform reinforcement learning optimization on the initial optimized fusion weight according to the reward function and the action value function to obtain the optimal fusion weight; the reward function is expressed as:

[0029]

[0030] where, represents the reward function; e represents the natural exponent; represents the historical cycle health index prediction error of the i-th target power grid device in the t-th prediction cycle.

[0031] Further, the step of performing reinforcement learning optimization on the initial optimized fusion weight according to the reward function and the action value function to obtain the optimal fusion weight includes:

[0032] Use the device operation environment data and the fusion weight as the input state and output action of the intelligent agent respectively, and use the fusion weight adjustment amount as the Q value, and update the initial optimized fusion weight based on Q learning; the update rule of the initial optimized fusion weight is expressed as:

[0033]

[0034] In the formula,

[0035]

[0036] where, and represent the fusion weights in the (t + 1)-th and t-th prediction cycles respectively; represents the fusion weight adjustment amount required to update the fusion weight in the t-th prediction cycle; represents the learning rate; is the discount factor; represents the gradient of the reward function with respect to the fusion weight; represents the action value function value corresponding to the device operation environment data and the fusion weight in the t-th prediction cycle.

[0037] Further, the method further includes:

[0038] Obtain the historical health index sequence of the power system, and input the historical health index sequence of the power system and the equipment health index sequences of each target power grid device in the power system into a preset system health status prediction model for operating status evaluation to obtain the corresponding system health index prediction value.

[0039] Further, the preset system health status prediction model is a long short-term memory network; the system health index prediction value is expressed as:

[0040]

[0041] Wherein, represents the system health index prediction value output at time and are the weights and biases of the output layer of the long short-term memory network, is the hidden state of the last unit of the long short-term memory network at time point .

[0042] Further, the method further includes:

[0043] According to the preset self-evolution strategy, perform online learning and update on the preset multi-modal feature extraction model and the preset system health status prediction model.

[0044] In a second aspect, an embodiment of the present invention provides a power grid equipment health status prediction system based on multi-modal data fusion, and the system includes:

[0045] A status data acquisition module, configured to acquire multi-modal status association data of a target power grid device in real time; the multi-modal status association data includes device image data, device operation status time series data, and device operation log data;

[0046] A multi-modal feature extraction module, configured to perform data feature extraction on the multi-modal status association data according to a preset multi-modal feature extraction model to obtain corresponding multi-modal status features to be analyzed; the multi-modal status features to be analyzed include image features, status time series features, and log text features;

[0047] A fusion weight acquisition module, configured to acquire the historical cycle health index prediction error of the target power grid device and the device operation environment data, and dynamically acquire the optimal fusion weight of the multi-modal status features to be analyzed according to the historical cycle health index prediction error and the device operation environment data; the historical cycle health index prediction error is the prediction error generated by predicting the health index of the target power grid device in the previous evaluation cycle;

[0048] The device health assessment module is used to perform feature fusion on the multi-modal state features to be analyzed according to the optimal fusion weights, and obtain the corresponding predicted value of the device health index.

[0049] The present application provides a method and system for predicting the health state of power grid equipment based on multi-modal data fusion. Through the method, multi-modal state-related data including equipment image data, equipment operation state time-series data, and equipment operation log data of target power grid equipment can be obtained in real time. After data feature extraction is performed on the multi-modal state-related data according to a preset multi-modal feature extraction model, the corresponding multi-modal state features to be analyzed including image features, state time-series features, and log text features are obtained. Then, the historical cycle health index prediction error and equipment operation environment data of the target power grid equipment are obtained, and the optimal fusion weights of the multi-modal state features to be analyzed are dynamically obtained according to the historical cycle health index prediction error and the equipment operation environment data. According to the optimal fusion weights, feature fusion is performed on the multi-modal state features to be analyzed, and the corresponding technical solution of the predicted value of the device health index is obtained. Compared with the prior art, the method for predicting the health state of power grid equipment based on multi-modal data fusion can capture rich equipment operation state features in real time by collecting multi-source data, and perform scientific and reasonable real-time online evaluation of the equipment health state by combining the dynamic optimization strategy of multi-modal feature fusion weights. It can effectively improve the timeliness, efficiency, and accuracy of equipment health assessment, reduce power production losses and maintenance costs, and provide reliable guarantee for the stable and safe operation of the power system. Brief Description of the Drawings

[0050] Figure 1 is a flowchart of the method for predicting the health state of power grid equipment based on multi-modal data fusion in an embodiment of the present invention;

[0051] Figure 2 is another flowchart of the method for predicting the health state of power grid equipment based on multi-modal data fusion in an embodiment of the present invention;

[0052] Figure 3 is another flowchart of the method for predicting the health state of power grid equipment based on multi-modal data fusion in an embodiment of the present invention;

[0053] Figure 4 is a structural diagram of the system for predicting the health state of power grid equipment based on multi-modal data fusion in an embodiment of the present invention;

[0054] Figure 5 is another structural diagram of the system for predicting the health state of power grid equipment based on multi-modal data fusion in an embodiment of the present invention;

[0055] Figure 6It is another structural schematic diagram of the power grid equipment health status prediction system based on multi-modal data fusion in the embodiments of the present invention. Specific embodiments

[0056] In order to make the objectives, technical solutions, and beneficial effects of the present application clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Obviously, the following described embodiments are part of the embodiments of the present invention and are only used to illustrate the present invention, but not to limit the scope of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0057] The power grid equipment health status prediction method based on multi-modal data fusion provided by the present invention can be understood as a power grid equipment online fault detection method that extracts and integrates the characteristics of multiple data sources of power grid equipment, and constructs an equipment health index evaluation system by combining a multi-modal feature fusion weight dynamic optimization strategy that combines genetic algorithms and reinforcement learning to realize real-time, scientific, and comprehensive monitoring and evaluation of the equipment health status. It can be applied to terminals or servers. Among them, the terminal can be, but is not limited to, various personal computers, laptop computers, smart phones, tablet computers, and portable wearable devices, and the server can be implemented by an independent server or a server cluster composed of multiple servers. The server can, according to actual application requirements, use the power grid equipment health status prediction method based on multi-modal data fusion provided by the present invention to perform real-time, efficient, and accurate prediction of the power grid equipment health index, and use the obtained power grid equipment health index prediction value for subsequent research such as system health index prediction of the server, or transmit it to the terminal for the terminal user to view and analyze; the following embodiments will detail the power grid equipment health status prediction method based on multi-modal data fusion of the present invention.

[0058] In one embodiment, as Figure 1 shown, a power grid equipment health status prediction method based on multi-modal data fusion is provided, including the following steps:

[0059] S11. Obtain the multi-modal state correlation data of the target power grid equipment in real time; among them, the target power grid equipment can be understood as various distribution equipment used in the power system, such as transformers, relays, and circuit breakers, etc., which are not specifically limited here; the corresponding multi-modal state correlation data can be understood as multi-source data collected related to the health status of the target power grid equipment. Considering the physical state of the power grid equipment, such as equipment wear, damage, or other signs of failure, etc., can all be reflected on the equipment image, and the relevant data such as voltage, current, and fault codes that change over time during the operation of the power grid equipment, as well as the operation log information recorded during the operation of the equipment, etc., can all directly or indirectly reflect the health status of the power grid equipment. In order to ensure the comprehensiveness and reliability of equipment state analysis, in this embodiment, preferably, multi-modal state correlation data including equipment image data, equipment operation state time-series data, and equipment operation log data, etc., are collected in each health state assessment cycle of the target power grid equipment as the basic data for operation state analysis; it should be noted that the specific data content involved in the multi-modal state correlation data varies according to the type of the target power grid equipment, which will not be elaborated here.

[0060] S12. According to the preset multi-modal feature extraction model, perform data feature extraction on the multi-modal state correlation data to obtain the corresponding multi-modal state features to be analyzed; the multi-modal state features to be analyzed include image features, state time-series features, and log text features; among them, the preset multi-modal feature extraction model can be understood as a machine learning model that can effectively extract various source data features in the collected multi-modal state correlation data. In order to ensure the efficiency and reliability of various source data feature extraction, in this embodiment, preferably, a preset multi-modal feature extraction model including a parallel preset neural network model, a long short-term memory network, and a preset natural language processing model is used for multi-modal data feature extraction.

[0061] Specifically, the step of performing data feature extraction on the multi-modal state correlation data according to the preset multi-modal feature extraction model to obtain the corresponding multi-modal state features includes:

[0062] Perform feature extraction on the equipment image data according to the preset neural network model to obtain the corresponding image features; among them, the preset neural network model can be understood as a neural network that can extract image features, but in order to ensure the efficiency and reliability of image feature extraction, in this embodiment, preferably, a convolutional neural network model (Convolutional Neural Networks, CNN) is used as the preset neural network model, and the specific structure and parameters of the convolutional neural network model can be adjusted and set through training according to actual application requirements;

[0063] Suppose the power grid equipment image set obtained by image acquisition of the target power grid equipment in the power system , where n represents the total number of grid devices in the power system, represents the device image data of the i-th grid device, which can describe the physical state of transformers, relays, circuit breakers, or other power distribution facilities, and is the feature map processed by the -th convolutional network layer, represents the total number of convolutional layers of the convolutional neural network model. Then, the specific process of extracting image features from the device image data through the convolutional neural network model is as follows:

[0064] Each convolutional layer sequentially includes a convolution module, an activation module, and a pooling module. For each layer :

[0065] 1) The convolution module performs a convolution operation: A set of convolution kernels are used to perform a convolution operation on the input feature map (for , that is, the original image). The mathematical representation of the convolution is:

[0066]

[0067] where * represents the convolution operation, is the feature map corresponding to the -th convolutional kernel of the -th layer;

[0068] 2) The activation module performs an activation operation: The activation function (such as the ReLU activation function) performs a non-linear transformation on to increase the expressive power of the model. The activation output of each convolution kernel is:

[0069]

[0070] 3) The pooling module performs a pooling operation: Pooling (such as max pooling) is performed on to reduce the feature dimension and increase the robustness of the model. The pooling result is:

[0071]

[0072] Then, after being sequentially processed by the above three modules, the output of the -th layer is composed of all the pooled feature maps; After being processed by each convolutional layer in the convolutional neural network model, the final image feature can be obtained.

[0073] Extract the state time series features from the time series data of the device operating state according to the long short-term memory network, and obtain the corresponding state time series features; among them, the long short-term memory network can be understood as directly using the existing LSTM (Long Short-Term Memory) network or an improved long short-term memory network trained and constructed based on the LSTM (Long Short-Term Memory) network, as long as it can accurately and reliably extract the time series features of the voltage, current, fault code, etc. that change over time collected during the operation of the power grid equipment. There is no specific limitation here; assume that the time series data of the device operating state of the target power grid equipment in the power system is collected as , where n represents the total number of power grid equipment in the power system, represents the th power grid equipment at the th moment of the device operating state time series data. Then, the processing process of the LSTM unit of the th sequence in the state time series features extracted by the LSTM network at the th moment can be expressed as:

[0074] 1) Forget gate, determining which information will be discarded:

[0075]

[0076] 2) Input gate, determining which new information will be stored in the cell state :

[0077]

[0078]

[0079] 3) Cell state update, combining the information of the forget gate and the input gate to update the cell state:

[0080]

[0081] 4) Output gate, determining the output value:

[0082]

[0083]

[0084] Among them, represents the sigmoid function, represents element-wise multiplication, and are the LSTM network parameters.

[0085] Extract features from the device operation log data according to the preset natural language processing model to obtain corresponding log text features; among them, in principle, any model that can implement text data feature extraction can be used as the preset natural language processing model. To ensure the efficiency and timeliness of device health status evaluation, this embodiment preferably uses TF-IDF (term frequency–inverse document frequency) to reliably extract log file features; assume that the device operation log data set of the target grid device in the power system collected is , where n represents the total number of grid devices in the power system, represents the -th device operation log data of the grid device. Since the device operation log data is file data, it needs to be preprocessed including word segmentation, removing useless words, etc. before using TF-IDF for text feature extraction; specifically, the weight calculation of a specific word in the log can be expressed as:

[0086]

[0087] Among them, is the word frequency, is the inverse document frequency, calculating the importance of each word in all log documents; combining the weights of each word, or performing preset optimization processing on the combined vector according to actual application requirements, the required log text features can be obtained.

[0088] Through the above method steps, data feature extraction can be performed on the multi-modal state correlation data of each target grid device to be evaluated, and the corresponding multi-modal state features to be analyzed can be obtained, where , and respectively represent the image features, state time series features and log text features of each power grid device; in principle, by directly splicing and fusing the multimodal state features corresponding to each power grid device, an indicator that can characterize the health status of the power grid equipment can be obtained, but considering that there are certain differences in the descriptions of the operating status of the power grid equipment by different types of data features, direct fusion may result in the final health index not being in line with reality and having no specific application value. In order to further ensure that the health index constructed based on the fusion of multimodal data features is closer to reality, more accurate and reliable, this embodiment preferably introduces the fusion weight of multimodal features to control the importance of different features to the modeling of the power grid equipment state, and adopts the following recent historical health index evaluation deviation based on the target power grid equipment combined with the actual operating environment of the power grid equipment to dynamically optimize and update the optimal fusion weight of the multimodal features; for the specific method of obtaining the optimal fusion weight of the multimodal features used in each health status evaluation cycle, refer to the following steps.

[0089] S13, obtaining the historical cycle health index prediction error and the equipment operating environment data of the target power grid equipment, and dynamically obtaining the optimal fusion weight of the multimodal state feature to be analyzed based on the historical cycle health index prediction error and the equipment operating environment data; wherein the historical cycle health index prediction error can be understood as the prediction error generated by predicting the health state of the target power grid equipment using the same health state prediction method in the previous evaluation cycle; specifically, the step of obtaining the historical cycle health index prediction error of the target power grid equipment includes:

[0090] According to the preset multimodal feature extraction model, feature extraction is performed on the multimodal state association data of the previous historical period of the target power grid equipment to obtain the corresponding historical period multimodal features; wherein, the specific extraction process of the historical period multimodal features can refer to the acquisition process of the multimodal state features to be analyzed in the previous text, which will not be repeated here;

[0091] According to the fusion weight of the previous historical cycle, the multimodal features of the historical cycle are feature fused to obtain the corresponding historical cycle equipment health index prediction value; wherein, the fusion weight of the previous historical cycle can be understood as the multimodal feature fusion weight used for equipment health status prediction in the previous cycle of the current evaluation cycle. If the previous historical cycle is the first evaluation cycle, the corresponding fusion weight can be set according to actual application requirements according to human experience or other methods. If it corresponds to the second and subsequent evaluation cycles, the optimal fusion weight acquisition method given in this embodiment can be used to obtain and update the corresponding fusion weight, which will not be described in detail here; correspondingly, the historical cycle equipment health index prediction value can be understood as the fusion feature obtained by weighted fusion of the multimodal features of each historical cycle according to the fusion weight of the previous historical cycle;

[0092] Obtain the true value of the equipment health index of the target power grid equipment in the previous historical period, and obtain the corresponding historical period health index prediction error according to the true value of the equipment health index in the previous historical period and the predicted value of the equipment health index in the historical period.

[0093] To ensure the scientificity and accuracy of the multi-modal feature fusion weight setting, in this embodiment, on the basis of considering the preset error in the historical period, data on the equipment operating environment is also introduced for comprehensive analysis; the data on the equipment operating environment in this embodiment can be understood as other environmental factor data that will affect the operating state of the equipment. For example, external variable data of the power grid equipment state and performance, such as humidity, temperature, whether there is vibration, electromagnetic interference, etc., can be set according to actual application requirements and are not specifically limited here; specifically, the step of dynamically obtaining the optimal fusion weight of the multi-modal state features to be analyzed according to the historical period health index prediction error and the equipment operating environment data includes:

[0094] Based on the historical period health index prediction error and the equipment operating environment data, and combining genetic algorithms with reinforcement learning, dynamically adjust the previous historical period fusion weight corresponding to the historical period health index prediction error to obtain the optimal fusion weight; the fusion weight is expressed as:

[0095]

[0096] In the formula,

[0097]

[0098]

[0099] Among them, and respectively represent the fusion weights of the i-th target power grid equipment in the (t + 1)-th and t-th prediction periods; represents the historical period health index prediction error of the i-th target power grid equipment in the t-th prediction period; and respectively represent the predicted value of the historical period equipment health index and the true value of the equipment health index in the previous historical period of the i-th target power grid equipment in the t-th prediction period; , and respectively represent the image feature, state time series feature, and log text feature of the -th power grid equipment; represents the splicing operation of features; is the dot product operation; represents the , and Weight coefficients of three features; Denote the equipment operation environment data of the i-th target power grid equipment; Denote the joint execution function composed of reinforcement learning and genetic algorithm. Preferably, the genetic algorithm is responsible for initially obtaining the optimized weights under the action of environmental factors, and then using the reinforcement learning algorithm to perform secondary optimization on the fusion weights obtained by the genetic algorithm, and finally obtaining the required optimal fusion weights through cyclic optimization of a preset number of times.

[0100] Specifically, the steps of dynamically adjusting the fusion weights of the previous historical cycle corresponding to the prediction error of the historical cycle health index based on the genetic algorithm combined with reinforcement learning to obtain the optimal fusion weights include:

[0101] According to the prediction error of the historical cycle health index and the equipment operation environment data, establish a corresponding fitness function, and with the goal of minimizing the fitness function value, optimize the fusion weights of the previous historical cycle through the genetic algorithm to obtain the initial optimized fusion weights; the fitness function is expressed as:

[0102]

[0103] Among them, Denote the fitness function; Denote the prediction error of the historical cycle health index of the i-th target power grid equipment in the t-th prediction cycle; Denote the equipment operation environment data of the i-th target power grid equipment;

[0104] In practical applications, the process of optimizing the fusion weights of the previous historical cycle through the genetic algorithm to obtain the initial optimized fusion weights is as follows:

[0105] 1) Establish an optimization objective with the goal of minimizing the fitness function:

[0106]

[0107] Among them, Denote the , and weight coefficients of three features corresponding to the i-th power grid equipment when the fitness function value is the smallest;

[0108] 2) Initialize and define the initial population of the feature weights ; each individual represents a group of possible weight configurations;

[0109] 3) Genetic algorithm stage:

[0110] Select: Apply the fitness function to select individuals with higher fitness; a higher fitness means that the device health index under the weight configuration is closer to the target health index, and these individuals are more likely to be selected to enter the next generation;

[0111] Crossover: Randomly select a pair of individuals for crossover operation, that is, exchange part of their genes (weights) at one or more points, so as to generate new offspring, which can introduce new combinations of weight configurations and increase the diversity of the search solution space;

[0112] Mutate: Modify part of the genes (weights) of individuals with a certain probability, further increasing the diversity of the population, helping the algorithm to jump out of the local optimal solution and explore a wider solution space;

[0113] Update and iteration: Repeat the above steps until a certain number of iterations or performance criteria are reached. At this time, a set of relatively optimal weight configurations preliminarily screened by the genetic algorithm will be obtained, that is, the corresponding initial optimized fusion weights will be obtained.

[0114] Construct a corresponding reward function according to the historical cycle health index prediction error, and use the fitness function as the action value function, and optimize the initial optimized fusion weights according to the reward function and the action value function to obtain the optimal fusion weights; among them, reinforcement learning optimization can be understood as defining the performance of the reward function to quantify the weight configuration, and using Q-learning to fine-tune the weight configuration based on the reward feedback; in this embodiment, the reward function is defined as the closer the predicted value of the device health index is to the true value of the device health index, the larger the value of the reward function, and vice versa, so that the optimization process will be more inclined to find those weight configurations that can make the predicted value of the device health index close to the true value of the device group health. Preferably, the reward function is defined as the negative exponential function of the opposite number of the difference (historical cycle health index prediction error) between the two, expressed as:

[0115]

[0116] where e represents the natural exponent; represents the historical cycle health index prediction error of the i-th target power grid device in the t-th prediction cycle; represents the reward function, when is smaller, is closer to 1, the reward is negative, and then taking the opposite number gets a larger reward, and vice versa, the reward is smaller; the reward function design adopted in this embodiment can motivate the model to make the predicted value of the device health index close to the true value of the device health index through the optimization process, that is, it is inclined to select those weight configurations that make the device operating state healthier.

[0117] Q-Learning is a very famous model-free learning algorithm in reinforcement learning. It aims to find the expected utility of taking a certain action in a given state, that is, the so-called Q value (Quality value); the Q-learning algorithm does not need to model the dynamics of the environment, and it directly learns the optimal policy through interaction with the environment; in this embodiment, preferably, the fusion weight adjustment amount is set as the Q value, and the device operating environment data is set as the environment data obtained by the agent, and the fusion weight is the action that the agent needs to take, that is, combined with Q-learning, the update rule of the weight is as follows. The time t is used to distinguish the current value and the value at the next moment, and it is defined The calculation expression is the fitness function defined by the above genetic algorithm:

[0118]

[0119]

[0120]

[0121] where and represent the fusion weights in the (t + 1)-th and t-th prediction periods respectively; represents the fusion weight adjustment amount required to update the fusion weight in the t-th prediction period; is the learning rate, and the initial value is set to 0.00001; is the discount factor, which is set to 0.4; represents the gradient of the reward function with respect to the weight ; represents the action value function value corresponding to the device operating environment data and the fusion weight in the t-th prediction period.

[0122] The above process of optimizing the fusion weight can actually be understood as continuously jointly executing the genetic algorithm and the reinforcement learning algorithm, and adjusting the weight configuration according to the historical cycle health index prediction error and the device operating environment data feedback until the algorithm converges (the weight tends to a stable value) or reaches the preset number of iterations, and finally obtaining the weight ; In this embodiment, by adopting a weight dynamic optimization strategy that combines genetic algorithms and reinforcement learning, it is possible to search for a reasonable weight configuration globally using a genetic algorithm, and then fine-tune and locally optimize it based on this using a reinforcement learning algorithm, effectively optimizing the weights of multi-modal feature fusion. Furthermore, while providing a reliable guarantee for obtaining more accurate equipment health status prediction results, it can also effectively improve the optimization efficiency of the fusion weights, enhancing the robustness and adaptability of the fusion weight dynamic optimization algorithm.

[0123] S14. According to the optimal fusion weights, perform feature fusion on the multi-modal state features to be analyzed, and obtain the corresponding predicted value of the equipment health index; among them, the predicted value of the equipment health index can be understood as the multi-modal fusion feature obtained by weighted fusion of the multi-modal state features to be analyzed in the current cycle based on the optimal fusion weights obtained through the above method steps. For the specific calculation, reference can be made to the specific expression about here, which will not be elaborated further.

[0124] In this embodiment, by obtaining in real time the multi-modal state correlation data of the target power grid equipment, including equipment image data, equipment operation status time-series data, and equipment operation log data, extracting data features from the multi-modal state correlation data according to a preset multi-modal feature extraction model, obtaining the corresponding multi-modal state features to be analyzed including image features, state time-series features, and log text features, and then obtaining the historical cycle health index prediction error and equipment operation environment data of the target power grid equipment, and dynamically obtaining the optimal fusion weights of the multi-modal state features to be analyzed according to the historical cycle health index prediction error and equipment operation environment data, and performing feature fusion on the multi-modal state features to be analyzed according to the optimal fusion weights to obtain the corresponding predicted value of the equipment health index, the technical solution can capture rich equipment operation state features in real time by collecting multi-source data, and combine the multi-modal feature fusion weight dynamic optimization strategy to conduct a scientific and reasonable real-time online assessment of the equipment health state, effectively improving the timeliness, efficiency, and accuracy of equipment health assessment, facilitating the timely perception and prevention of power grid equipment failures, and thus effectively reducing power production losses and maintenance costs.

[0125] At the same time, considering that the health status of each power grid equipment in the power system is directly related to the normal operation of the entire power system, preferably in this embodiment, on the basis of realizing a reliable health state assessment of each power grid equipment through the above method steps, further conduct real-time and efficient operation state prediction and analysis on the entire power system, and thus provide reliable data support for the formulation of power system stable operation strategies; specifically, as Figure 2 shown, a method for predicting the health status of power grid equipment based on multi-modal data fusion is provided, and the method further includes:

[0126] S15. Obtain the historical health index sequence of the power system, and input the historical health index sequence of the power system and the equipment health index sequences of each target grid equipment in the power system into a preset system health status prediction model for operation status evaluation to obtain the corresponding system health index prediction value; among them, the historical health index sequence of the power system can be understood as the historical time series data of the operation health index of the power system, and the specific acquisition method can refer to the corresponding existing technology for implementation; correspondingly, the equipment health index sequences of each target grid equipment can be understood as data sequences composed of the equipment health indices (true health index values) of the corresponding historical cycles and the equipment health index prediction values of the current cycle obtained through the aforementioned step S14 in chronological order , where represents the equipment health index prediction value of the i-th target grid equipment in the t-th prediction cycle; represents the true health index value of the i-th target grid equipment in the (t - m)-th historical cycle, and m represents the time step (total number of historical evaluation cycles).

[0127] In principle, any machine learning model capable of making predictions based on time series data can be used for the preset system health status prediction model in this embodiment. However, to ensure the real-time performance and efficiency of the power system health status evaluation, this embodiment preferably uses a long short-term memory (LSTM) network that integrates a convolutional structure and time series analysis and is committed to continuous learning and updating to quickly adapt to new data input and efficiently and dynamically predict the future system health status using the existing equipment health index data; assume that the health data of N target devices for m time steps have been collected; for the system health index, an input feature vector can be defined represents the time point containing the health index sequences of all devices; the mathematical model of the LSTM unit specifically used for system health status prediction is as follows:

[0128] 1) Forget gate:

[0129]

[0130] where represents the sigmoid activation function, and represent the weights and biases, is the concatenation of the previous hidden state and the current input;

[0131] 2) Input gate:

[0132]

[0133]

[0134] Among them, is the candidate cell state, and are the weights and biases of the input gate;

[0135] 3) Cell state update:

[0136]

[0137] Here, multiplication operations are used to implement the process of forgetting and adding new information;

[0138] 4) Output gate:

[0139]

[0140]

[0141] Among them, is the hidden state at the current time point, and are the weights and biases of the output gate;

[0142] In practical applications, the entire long short-term memory (LSTM) network is composed of multiple such LSTM units, and each unit at time point generates a hidden state , and the final network output, the final system health index is expressed as:

[0143]

[0144] Among them, and are the weights and biases of the output layer, is the hidden state of the last unit of the LSTM network at time point .

[0145] In this embodiment, by obtaining the multi-modal state correlation data of the target power grid equipment in real time, extracting the data features of the multi-modal state correlation data according to the preset multi-modal feature extraction model, obtaining the corresponding multi-modal state features to be analyzed, dynamically obtaining the optimal fusion weights of the multi-modal state features to be analyzed according to the historical cycle health index prediction error of the target power grid equipment and the equipment operation environment data obtained, and performing feature fusion on the multi-modal state features to be analyzed according to the optimal fusion weights to obtain the corresponding equipment health index prediction value, and inputting the obtained historical health index sequence of the power system and the equipment health index sequences of each target power grid equipment in the power system into the preset system health state prediction model for operation state evaluation to obtain the corresponding system health index prediction value, the technical solution can not only effectively improve the timeliness, efficiency and accuracy of equipment health assessment, facilitate the timely perception and prevention of power grid equipment failures, and effectively reduce power production losses and maintenance costs, but also can perform real-time and efficient operation state prediction and analysis on the entire power system based on the health state prediction results of each power grid equipment, thereby providing reliable data support for the formulation of power system stable operation strategies.

[0146] In addition, considering that in practical applications, the operation of power grid equipment and the operation environment will constantly change, in order to ensure the long-term effectiveness and robustness of the power grid equipment health state prediction model and the power system health state prediction model, and further ensure the prediction accuracy and efficiency of the power grid equipment health state and the power system health state in each evaluation cycle, this embodiment preferably adopts an online learning method to adaptively update and adjust the parameters of the health state evaluation and prediction models in real time to ensure that each model can continuously adapt to and optimize its own prediction ability; specifically, as Figure 3 shown, a method for predicting the health state of power grid equipment based on multi-modal data fusion is provided, and the method further includes:

[0147] S16. Perform online learning and update on the preset multi-modal feature extraction model and the preset system health state prediction model according to the preset self-evolution strategy;

[0148] The preset self-evolution strategy in this embodiment can be understood as an online learning strategy, which can update the model in real time based on the gradually received new data during operation, and independently improve the operation performance of the preset multi-modal feature extraction model and the preset system health state prediction model to create an intelligent agent system that can quickly adapt to new situations and continuously improve. During the process of the intelligent agent interacting with the environment, it can update its strategy in real time to adapt to the changes in the environment; the specific process of performing online learning and update based on the preset self-evolution strategy is as follows:

[0149] At each time point , assume a corresponding set of model parameters It contains the weights and biases of all models involved in the evaluation and prediction processes; as new data of each model is received in real time, it is necessary to continuously modify the model parameters , so that each model can better adapt to the new situation and improve the prediction accuracy;

[0150] A self-evolution strategy is set that can, based on the current model parameters (the preset multi-modal feature extraction model and the preset system health status prediction model involved in steps S11 - S15) and newly collected data (device image data, device operation status time series data, device operation log data, device operation environment data, etc. of each target power grid device), as well as a factor called "learning rate" (this factor can control the update rate (update amplitude) of the parameters, which determines the step size in the gradient descent process and must be just right; too large may lead to unstable convergence, and too small will result in a slow convergence rate) to help update the model parameters; the specific formula is expressed as:

[0151]

[0152] In the formula,

[0153]

[0154]

[0155] where, and respectively represent the time moment and the self-evolution strategy parameters at the +1 moment, corresponding to the model parameters at time and +1, including all the weights and biases of the preset multi-modal feature extraction model and the preset system health status prediction model, that is, the update objects that need to be continuously updated and optimized based on the newly collected data to make the model more adaptable and have better prediction ability; The function uses the historical health index of the power system and the current parameters to predict the system health status at the next moment and update the model parameters accordingly; represents the loss function, which is used to measure the performance of the model on the data , represents the loss gradient with respect to the parameter ; is the number of newly collected samples at time ; and are respectively the model parameters Under the condition, corresponding to the predicted value of the system health index and the actually measured value of the system health index for the

[0156] In practical applications, assume that the health status of an important power grid transformer is being monitored, and the data collection device will regularly transmit the collected data back to the central system; when new data at time is transmitted to the central system, update the parameters of the evaluation and prediction model according to the following process:

[0157] 1) Use the current model and the current parameters to perform forward calculation on the new data to obtain the predicted value of the current device health status;

[0158] 2) Calculate the value of the loss function and calculate the gradient using backpropagation;

[0159] 3) Use the above function to update the parameters: , and replace the old parameters with the new parameters to update the model to a new version and continue to put it into use.

[0160] In the embodiment of the present application, by obtaining the multi-modal state correlation data of the target power grid equipment in real time, extracting the data features of the multi-modal state correlation data according to the preset multi-modal feature extraction model to obtain the corresponding multi-modal state features to be analyzed, and according to the optimal fusion weight of the multi-modal state features to be analyzed dynamically obtained based on the historical cycle health index prediction error of the obtained target power grid equipment and the equipment operation environment data, fusing the multi-modal state features to be analyzed to obtain the corresponding equipment health index prediction value, then inputting the obtained power system historical health index sequence and the equipment health index sequences of each target power grid equipment in the power system into the preset system health state prediction model to perform the operation state evaluation to obtain the corresponding system health index prediction value, and according to the preset self-evolution strategy, online learning and updating the preset multi-modal feature extraction model and the preset system health state prediction model, the technical solution can not only effectively improve the timeliness, efficiency and accuracy of equipment health assessment, facilitate the timely perception and prevention of power grid equipment failures, effectively reduce power production losses and maintenance costs, but also can perform real-time and efficient operation state prediction and analysis on the entire power system based on the prediction results of the health states of each power grid equipment, and further provide reliable data support for the formulation of the power system stable operation strategy. It can also ensure that each model can continuously adapt to new data and optimize its own prediction ability, ensure the long-term effectiveness and robustness of the model, and then ensure the prediction accuracy and efficiency of the power grid equipment health state and the power system health state, providing reliable guarantee for the stable and safe operation of the entire power system.

[0161] It should be noted that although each step in the above flowchart is shown in sequence according to the indication of the arrow, these steps are not necessarily executed in the order indicated by the arrow. Unless otherwise clearly stated in this article, the execution of these steps has no strict order limit, and these steps can be executed in other orders.

[0162] In one embodiment, as Figure 4 shown, a power grid equipment health state prediction system based on multi-modal data fusion is provided, and the system includes:

[0163] A state data acquisition module 1, configured to acquire the multi-modal state correlation data of the target power grid equipment in real time; the multi-modal state correlation data includes equipment image data, equipment operation state time series data, and equipment operation log data;

[0164] A multi-modal feature extraction module 2, configured to extract the data features of the multi-modal state correlation data according to the preset multi-modal feature extraction model to obtain the corresponding multi-modal state features to be analyzed; the multi-modal state features to be analyzed include image features, state time series features, and log text features;

[0165] The fusion weight acquisition module 3 is configured to obtain the historical cycle health index prediction error of the target power grid device and the device operation environment data, and dynamically obtain the optimal fusion weight of the multi-modal state features to be analyzed according to the historical cycle health index prediction error and the device operation environment data; the historical cycle health index prediction error is the prediction error generated by predicting the health index of the target power grid device in the previous evaluation cycle.

[0166] The device health assessment module 4 is configured to perform feature fusion on the multi-modal state features to be analyzed according to the optimal fusion weight to obtain the corresponding device health index prediction value.

[0167] In one embodiment, as Figure 5 shown, a power grid device health status prediction system based on multi-modal data fusion is provided. The system further includes:

[0168] The system health assessment module 5 is configured to obtain the historical health index sequence of the power system, and input the historical health index sequence of the power system and the device health index sequences of each target power grid device in the power system into a preset system health status prediction model for operation status assessment to obtain the corresponding system health index prediction value.

[0169] In one embodiment, as Figure 6 shown, a power grid device health status prediction system based on multi-modal data fusion is provided. The system further includes:

[0170] The model online update module 6 is configured to perform online learning and update on the preset multi-modal feature extraction model and the preset system health status prediction model according to a preset self-evolution strategy.

[0171] For the specific limitations of the power grid device health status prediction system based on multi-modal data fusion, reference can be made to the limitations of the power grid device health status prediction method based on multi-modal data fusion in the above text, and the corresponding technical effects can also be equivalently obtained, which will not be elaborated here. Each module in the above power grid device health status prediction system based on multi-modal data fusion can be implemented in whole or in part by software, hardware, and their combinations. The above modules can be embedded in the processor of the computer device in hardware form or be independent of it, or can be stored in the memory of the computer device in software form, so that the processor can call and execute the operations corresponding to the above modules.

[0172] In summary, the power grid equipment health status prediction method and system provided by the embodiments of the present invention, the power grid equipment health status prediction method based on multi-modal data fusion realizes real-time acquisition of multi-modal state correlation data of target power grid equipment, extracts data features from the multi-modal state correlation data according to a preset multi-modal feature extraction model to obtain corresponding multi-modal state features to be analyzed, and according to the optimal fusion weight of the multi-modal state features to be analyzed dynamically obtained based on the historical cycle health index prediction error of the acquired target power grid equipment and equipment operation environment data, performs feature fusion on the multi-modal state features to be analyzed to obtain a corresponding equipment health index prediction value. Then, according to the acquired historical health index sequence of the power system and the equipment health index sequences of each target power grid equipment in the power system, inputs them into a preset system health status prediction model to perform operation state evaluation to obtain a corresponding system health index prediction value, and according to a preset self-evolution strategy, performs online learning and updating on the preset multi-modal feature extraction model and the preset system health status prediction model. The method can not only effectively improve the timeliness, efficiency and accuracy of equipment health assessment, facilitate timely perception and prevention of power grid equipment failures, and effectively reduce power production losses and maintenance costs, but also perform real-time and efficient operation state prediction and analysis on the entire power system based on the health status prediction results of each power grid equipment, thereby providing reliable data support for the formulation of power system stable operation strategies. It can also ensure that each model can continuously adapt to new data and optimize its own prediction ability, ensure the long-term effectiveness and robustness of the model, and then ensure the prediction accuracy and efficiency of the power grid equipment health status and the power system health status, providing reliable guarantee for the stable and safe operation of the entire power system.

[0173] Each embodiment in this specification is described in a progressive manner. For parts that are the same or similar in each embodiment, reference can be made to each other. Each embodiment focuses on the differences from other embodiments. In particular, for the system embodiment, since it is basically similar to the method embodiment, it is described relatively simply. For the relevant parts, reference can be made to the partial description of the method embodiment. It should be noted that the technical features of the above embodiments can be combined arbitrarily. For the sake of concise description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope recorded in this specification.

[0174] The above-described embodiments merely represent several preferred embodiments of the present application. The description thereof is relatively specific and detailed, but it should not be construed as a limitation on the scope of the invention patent. It should be noted that for those of ordinary skill in the art, without departing from the technical principle of the present invention, several improvements and substitutions can be made, and these improvements and substitutions should also be regarded as the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the protection scope of the claims described above.

Claims

1. A method for predicting the health status of power grid equipment based on multi-modal data fusion, characterized in that The method includes the following steps: Obtain multi-modal status correlation data of the target power grid device in real time; the multi-modal status correlation data includes device image data, device operation status time-series data, and device operation log data; According to a preset multi-modal feature extraction model, perform data feature extraction on the multi-modal status correlation data to obtain corresponding multi-modal status features to be analyzed; the multi-modal status features to be analyzed include image features, status time-series features, and log text features; Obtain the historical cycle health index prediction error and device operation environment data of the target power grid device, and dynamically obtain the optimal fusion weight of the multi-modal status features to be analyzed according to the historical cycle health index prediction error and the device operation environment data; the historical cycle health index prediction error is the prediction error generated by predicting the health index of the target power grid device in the previous evaluation cycle; According to the optimal fusion weight, perform feature fusion on the multi-modal status features to be analyzed to obtain the corresponding device health index prediction value; Among them, the step of dynamically obtaining the optimal fusion weight of the multi-modal status features to be analyzed according to the historical cycle health index prediction error and the device operation environment data includes: According to the historical cycle health index prediction error and the device operation environment data, based on the genetic algorithm combined with reinforcement learning, dynamically adjust the previous historical cycle fusion weight corresponding to the historical cycle health index prediction error to obtain the optimal fusion weight; the fusion weight is expressed as: Among them, and respectively represent the fusion weights of the i-th target power grid device in the (t + 1)-th and t-th prediction cycles; represents the historical cycle health index prediction error of the i-th target power grid device in the t-th prediction cycle; represents the device operation environment data of the i-th target power grid device; represents the joint execution function composed of reinforcement learning and genetic algorithm.

2. The method for predicting the health status of power grid equipment based on multi-modal data fusion according to claim 1, wherein The preset multi-modal feature extraction model includes a preset neural network model, a long short-term memory network, and a preset natural language processing model in parallel; The step of performing data feature extraction on the multi-modal status correlation data according to the preset multi-modal feature extraction model to obtain corresponding multi-modal status features includes: Perform feature extraction on the device image data according to the preset neural network model to obtain corresponding image features; Perform feature extraction on the device operation status time-series data according to the long short-term memory network to obtain corresponding status time-series features; Perform feature extraction on the device operation log data according to the preset natural language processing model to obtain corresponding log text features.

3. The method for predicting the health state of power grid equipment based on multimodal data fusion according to claim 1, wherein, The step of obtaining the historical cycle health index prediction error of the target power grid device includes: According to the preset multi-modal feature extraction model, perform feature extraction on the multi-modal status correlation data of the previous historical cycle of the target power grid device to obtain corresponding historical cycle multi-modal features; According to the previous historical cycle fusion weight, perform feature fusion on the historical cycle multi-modal features to obtain the corresponding historical cycle device health index prediction value; Obtain the true value of the device health index of the previous historical cycle of the target power grid device, and obtain the corresponding historical cycle health index prediction error according to the true value of the device health index of the previous historical cycle and the historical cycle device health index prediction value.

4. The power grid equipment health status prediction method based on multi-modal data fusion according to claim 1, wherein The step of dynamically adjusting the fusion weight of the previous historical period corresponding to the prediction error of the historical period health index based on the genetic algorithm combined with reinforcement learning to obtain the optimal fusion weight includes: According to the prediction error of the historical period health index and the device operation environment data, a corresponding fitness function is established, and with the goal of minimizing the fitness function value, the fusion weight of the previous historical period is optimized by the genetic algorithm to obtain an initial optimized fusion weight; the fitness function is expressed as: Among them, represents the fitness function; represents the historical cycle health index prediction error of the i-th target power grid device in the t-th prediction cycle; represents the device operation environment data of the i-th target power grid device; According to the prediction error of the historical period health index, a corresponding reward function is constructed, and the fitness function is used as the action value function, and the initial optimized fusion weight is optimized by reinforcement learning according to the reward function and the action value function to obtain the optimal fusion weight; the reward function is expressed as: Among them, represents the reward function; e represents the natural exponent; represents the historical cycle health index prediction error of the i-th target grid device in the t-th prediction cycle.

5. The method for predicting the health state of power grid equipment based on multi-modal data fusion according to claim 4, characterized in that, The step of optimizing the initial optimized fusion weight by reinforcement learning according to the reward function and the action value function to obtain the optimal fusion weight includes: The device operation environment data and the fusion weight are respectively used as the input state and output action of the intelligent agent, and the fusion weight adjustment amount is used as the Q value, and the initial optimized fusion weight is updated based on Q learning; the update rule of the initial optimized fusion weight is expressed as: In the formula, Among them, and represent the fusion weights for the (t + 1)-th and t-th prediction cycles, respectively; represents the fusion weight adjustment amount required to update the fusion weight for the t-th prediction cycle; represents the learning rate; is the discount factor; represents the gradient of the reward function with respect to the fusion weight; represents the action value function value corresponding to the device operating environment data and the fusion weight for the t-th prediction cycle.

6. The method for predicting the health state of power grid equipment based on multi-modal data fusion according to claim 1, characterized in that The method further includes: Obtain the historical health index sequence of the power system, and input the historical health index sequence of the power system and the device health index sequences of each target grid device in the power system into a preset system health state prediction model for operation state evaluation to obtain the corresponding system health index prediction value.

7. The method for predicting the health state of power grid equipment based on multi-modal data fusion according to claim 6, characterized in that, The preset system health state prediction model is a long short-term memory network; the system health index prediction value is expressed as: Among them, represents the predicted value of the system health index output at the moment; and are the weights and biases of the output layer of the long short-term memory network, is the hidden state of the last unit of the long short-term memory network at the time point .

8. The method for predicting the health state of power grid equipment based on multi-modal data fusion according to claim 6, wherein The method further includes: According to the preset self-evolution strategy, online learning and updating are performed on the preset multi-modal feature extraction model and the preset system health state prediction model.

9. A power grid equipment health status prediction system based on multi-modal data fusion, characterized in that, Applying the power grid equipment health state prediction method based on multi-modal data fusion as described in claim 1, the system includes: A state data acquisition module, configured to acquire multi-modal state correlation data of target grid equipment in real time; the multi-modal state correlation data includes equipment image data, equipment operation state time series data, and equipment operation log data; A multi-modal feature extraction module, configured to perform data feature extraction on the multi-modal state correlation data according to a preset multi-modal feature extraction model to obtain corresponding multi-modal state features to be analyzed; the multi-modal state features to be analyzed include image features, state time series features, and log text features; A fusion weight acquisition module, configured to acquire the prediction error of the historical period health index of the target grid equipment and the device operation environment data, and dynamically acquire the optimal fusion weight of the multi-modal state features to be analyzed according to the prediction error of the historical period health index and the device operation environment data; the prediction error of the historical period health index is the prediction error generated by predicting the health index of the target grid equipment in the previous evaluation period. The device health assessment module is used to perform feature fusion on the multi-modal state features to be analyzed according to the optimal fusion weights, so as to obtain the corresponding predicted value of the device health index.

Images