Power grid power equipment anomaly prediction method, system and equipment and storage medium

Abstract

The invention relates to the technical field of power grid monitoring, and discloses a power grid power equipment anomaly prediction method, system and device, and a storage medium, which are used for solving the problem that the existing single data source and threshold alarm prediction mode cannot adapt to advanced prevention and high-precision state detection. The method comprises the following steps: acquiring real-time electrical quantity time sequence data, image data and a monitoring report of each piece of power equipment in an equipment abnormal area in a power grid through an edge intelligent gateway; respectively carrying out feature extraction on the real-time electric quantity time sequence data, the image data and the monitoring report to obtain an electric quantity feature vector, a visual feature vector and a semantic feature vector; and carrying out weight fusion on the electrical quantity feature vector, the visual feature vector and the semantic feature vector, and carrying out abnormal state prediction based on the fused feature vector and a historical feature sequence.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of power grid monitoring, and particularly relates to a power grid power equipment anomaly prediction method, system, device and storage medium. BACKGROUND

[0002] Under the background of accelerating construction of the energy internet and new power system, especially under the progress of sensing technology and artificial intelligence, power grid state monitoring has developed from the initial traditional telemetry and single data acquisition stage into a system that can perceive and collect data from multiple dimensions, and has intelligent processing and wide-area communication at the edge. However, in this system, if the current single data source and threshold alarm prediction method is still used to identify power grid equipment anomalies, it is difficult to guarantee the accuracy and high reliability of state detection, and it also cannot meet the requirements of timeliness and prevention. SUMMARY

[0003] The present application provides a power grid power equipment anomaly prediction method, system, device and storage medium, which is used to solve the problem that the existing single data source and threshold alarm prediction method cannot adapt to early prevention and high-precision state detection.

[0004] The first aspect of the present application provides a power grid power equipment anomaly prediction method, comprising: obtaining real-time electrical quantity time series data, image data and monitoring reports of each power equipment in a device anomaly area in the power grid through an edge intelligent gateway; respectively extracting features from the real-time electrical quantity time series data, the image data and the monitoring reports to obtain electrical quantity feature vectors t , visual feature vectors t and semantic feature vectors t ; performing weight fusion on the electrical quantity feature vectors t , the visual feature vectors t and the semantic feature vectors t , and performing anomaly state prediction based on the fused feature vectors and historical feature sequences.

[0005] Optionally, the respective feature extraction from the real-time electrical quantity time series data, the image data and the monitoring reports to obtain electrical quantity feature vectors t , visual feature vectors t and semantic feature vectors t comprises: extracting abnormal electrical features in the real-time electrical quantity time series data using wavelet packet transform, and converting and outputting electrical quantity feature vectors , the electrical quantity feature vectors Deepening to obtain the electrical quantity feature vector α t ; extract visual features in the image data using the improved ResNet-18 network, and convert and output a visual feature vector β t , wherein the improved ResNet-18 network is obtained by using a traditional ResNet-18 as a backbone network and replacing the classification layer in the traditional ResNet-18 with a self-defined fully connected layer; perform semantic analysis and semantic embedding on the monitoring report using the pre-trained power AI large language model to obtain a semantic feature vector ξ t .

[0006] Optionally, the wavelet packet transform is used to extract abnormal electrical features in the real-time electrical quantity time series data and convert and output an electrical quantity feature vector , the electrical quantity feature vector is deepened to obtain the electrical quantity feature vector α t , comprising: The wavelet packet transform is used to calculate the energy of all sampling points of the real-time electrical quantity time series data at each frequency band, and based on the energy, local features belonging to non-stationary signals are marked to obtain an initial frequency energy feature vector ; Based on a pre-defined key frequency band index set related to fault features, the initial frequency energy feature vector is subjected to interference suppression processing to obtain the electrical quantity feature vector ; The multi-layer perceptron (MLP) for electrical quantity feature extraction E is used to perform high-dimensional abstract processing on the electrical quantity feature vector to obtain the electrical quantity feature vector α t .

[0007] Optionally, the image data is a two-dimensional pixel matrix, and the two-dimensional pixel matrix contains infrared thermal imaging images and visible light images of power equipment. The improved ResNet-18 network is used to extract visual features in the image data and convert and output a visual feature vector β t , comprising: Based on the position information of each power equipment, the temperature pixels and gray pixels of the region of interest (ROI) in the two-dimensional pixel matrix are extracted, and the temperature pixels and gray pixels of all ROIs are converted into standardized temperature and gray distribution maps through a minimum-maximum normalization method to obtain standard image data ; The standard image data The improved ResNet-18 network is input to extract visual features, and each extracted visual feature is compressed and converted to obtain a visual feature vector β t .

[0008] Optionally, the monitoring report comprises a device historical operation and maintenance report, a real-time scheduling operation log, and environmental meteorological warning information. The pre-trained power AI large language model is used to perform semantic analysis and semantic embedding on the monitoring report to obtain a semantic feature vector ξ t , comprising: The device historical operation and maintenance report, the real-time scheduling operation log, and the environmental meteorological warning information are standardized preprocessed to obtain a text sequence , wherein the standardized preprocessing comprises removing irrelevant characters, correcting misspelled words, and standardizing device names and fault terms. The text sequence is input into the pre-trained power AI large language model for context encoding, and a deep semantic of each token in the text sequence is output, and a semantic feature vector ξ t is generated through average pooling. The pre-trained power AI large language model is obtained by fine-tuning a DeepSeek large language model using a power grid device knowledge graph and a fault case library.

[0009] Optionally, the electrical quantity feature vector α t , the visual feature vector β t , and the semantic feature vector ξ t are fused, and based on the fused feature vector and a historical feature sequence, an abnormal state is predicted, comprising: An adaptive weight network is used to assign weights to each feature vector, and based on the assigned weights, the electrical quantity feature vector α t , the visual feature vector β t , and the semantic feature vector ξ t are fused to obtain a multi-modal fusion feature f t w ; Based on the multi-modal fusion feature f t w and a historical feature sequence , a device abnormal state is predicted.

[0010] Optionally, in the electrical quantity feature vector α t , the visual feature vector β t , and the semantic feature vector ξ tAfter the weight fusion is performed and the abnormal state is predicted based on the fused feature vectors and the historical feature sequences, the method further includes: generating an abnormal diagnosis report and a treatment suggestion based on a result of the prediction, and notifying an operation and maintenance personnel, and generating a preliminary control instruction based on the treatment suggestion, the preliminary control instruction being used to control the power equipment to operate in a transient state.

[0011] The second aspect of the application provides an electric power grid operation and maintenance system, including: at least one edge intelligent gateway, used to acquire real-time electrical quantity time series data, image data and monitoring reports of each power equipment in an equipment abnormal area in an electric power grid; each edge intelligent gateway monitors one area, and a plurality of power equipments are arranged in each area; a server, used to respectively perform feature extraction on the real-time electrical quantity time series data, the image data and the monitoring reports, to obtain electrical quantity feature vectors α t , visual feature vectors β t and semantic feature vectors ξ t ; and perform weight fusion on the electrical quantity feature vectors α t , the visual feature vectors β t and the semantic feature vectors ξ t , and predict an abnormal state based on fused feature vectors and historical feature sequences.

[0012] The third aspect of the application provides an electronic device, including a memory and at least one processor, the memory storing instructions; the at least one processor invokes the instructions in the memory, so that the electronic device executes the power equipment abnormal prediction method of the electric power grid.

[0013] The fourth aspect of the application provides a computer readable storage medium, the computer readable storage medium storing instructions, when the instructions are run on a computer, the computer executes the power equipment abnormal prediction method of the electric power grid.

[0014] In the technical scheme provided by the application, the real-time electrical quantity time series data, the image data and the monitoring reports of each power equipment in an equipment abnormal area in an electric power grid are acquired by an edge intelligent gateway; the real-time electrical quantity time series data, the image data and the monitoring reports are respectively subjected to feature extraction, to obtain electrical quantity feature vectors α t , visual feature vectors β t and semantic feature vectors ξ t ; and the electrical quantity feature vectors α t , the visual feature vectors β t and the semantic feature vectors ξ tThe weight fusion is performed, and the abnormal state is predicted based on the fused feature vector and the historical feature sequence. The method utilizes the complementary information in the multi-modal data of electrical quantities, visual images and monitoring reports, extracts the equipment state features, and then performs fusion prediction, which significantly improves the accuracy of abnormal state prediction, greatly reduces the risk of fault false alarm and false alarm, and also adapts to the trend requirement of early prevention. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 An embodiment of the power equipment abnormality prediction method of the power grid in the present application is shown in the figure; Figure 2 A model framework of the power equipment abnormality prediction of the power grid in the present application is shown in the figure; Figure 3 A second embodiment of the power equipment abnormality prediction method of the power grid in the present application is shown in the figure; Figure 4 An embodiment of the power grid operation and maintenance system in the present application is shown in the figure; Figure 5 An embodiment of the electronic device in the present application is shown in the figure. DETAILED DESCRIPTION

[0016] The terms "first", "second", "third", "fourth" and the like in the specification and claims of the present application and in the above drawings, if any, are used for distinguishing between similar objects, and do not necessarily have a specific order or sequence. It should be understood that the data thus used can be interchanged, where appropriate, so that the embodiments described herein can be carried out in sequences other than those illustrated or described herein. Furthermore, the terms "comprise" or "have" and any variations thereof are intended to cover a non-exclusive inclusion, for example, a process, method, system, product or apparatus that comprises a list of steps or units is not necessarily limited to those steps or units that are clearly listed, but can include other steps or units that are not clearly listed or inherent to such processes, methods, products or apparatus.

[0017] In recent years, with the large-scale access of new energy and the increasingly complex load characteristics of power grid, the operating state of power equipment is more variable, and the risk of latent failure increases, which makes it difficult for the traditional monitoring method relying on a single data source (such as SCADA telemetry) and threshold alarm to meet the high reliability requirement. This prompts people to start paying attention to and using multi-modal sensor fusion and artificial intelligence technology to tap the deep value of data. Multi-modal fusion shows great potential to improve the accuracy and reliability of state perception due to its information complementarity and cross-validation. However, multi-modal fusion faces many challenges, such as data heterogeneity (different formats of time series data, images, and text), temporal and spatial asynchrony (different sampling rates and cycles of different sensors), and dynamic changes in modal reliability (such as foggy weather leading to a decrease in infrared image quality), which have an adverse effect on the accuracy and robustness of the fusion model.

[0018] The rapid change of equipment state requires the monitoring system to make judgments in time or even in advance, and the data-driven artificial intelligence method can realize effective abnormal diagnosis and prediction. Most existing methods rely on global models trained on extensive data, but the data distribution between different substations and different equipment types is significantly different, and the generalization ability of a single model is limited. In addition, most existing solutions rely on single-mode or dual-mode sensor data (such as using only electrical quantities for time series prediction or only using infrared images for target detection), which fail to provide a comprehensive portrait of equipment health status, limiting the improvement of fault warning accuracy and timeliness. In the power grid operation scene requiring high reliability and high timeliness, traditional single-mode or simple multi-modal processing methods often fail to ensure the stability of monitoring performance. However, multi-modal data contains different physical semantics and uses completely different data formats, so special data preprocessing, feature extraction, and deep fusion are needed. Therefore, in order to meet the urgent needs of new power systems for transparent perception and predictive maintenance of equipment state, and to ensure its efficiency and reliability in various complex operating environments, it is urgent to develop more advanced technical solutions.

[0019] To solve the above problems, with reference to Figure 1 The present application provides an embodiment of a power equipment anomaly prediction method for a power grid, which specifically comprises the following steps: 101. Acquire real-time electrical quantity time series data, image data, and monitoring reports of each power equipment in the equipment anomaly area of the power grid through an edge intelligent gateway.

[0020] In this embodiment, the power grid is divided into multiple areas for management and control, and an edge intelligent gateway is set up in each area to communicate data with the background of the power grid, such as a server or a cloud.

[0021] In practical application, the edge intelligent gateway is set to monitor the area where it is located in real time, and real-time data required is collected and reported to the server through the Internet. Further, the edge intelligent gateway can also be set to identify whether there is an anomaly through the collected video data, and if there is an anomaly, the video data of the abnormal power equipment is reported to the server. In addition to the original data reporting, daily monitoring reports and other information can also be reported.

[0022] In the embodiment, the server can be realized by the way of reading the data of the edge intelligent gateway at a fixed time. Specifically, at the first monitoring moment, the first edge intelligent gateway acquires real-time electrical quantity time series data, infrared thermal imaging pictures and high-definition visible light pictures, as well as operation and maintenance reports, dispatching logs, and environmental meteorological text information related to the monitoring area of the first edge intelligent gateway. t n n

[0023] 102, respectively extracting features from real-time electrical quantity time series data, image data and monitoring reports to obtain electrical quantity feature vectors, visual feature vectors and semantic feature vectors.

[0024] In this step, when extracting features, the features in real-time electrical quantity time series data, image data and monitoring reports are extracted by wavelet packet transform, improved ResNet-18 network and power AI large language model respectively, and converted and output as corresponding feature vectors.

[0025] In practical application, when extracting corresponding feature vectors, wavelet packet transform, improved ResNet-18 network and power AI large language model can be embedded into a functional module to realize, such as abstracted as electrical quantity feature extraction module, visual feature extraction module and semantic feature extraction module.

[0026] 103, weight fusion of electrical quantity feature vectors, visual feature vectors and semantic feature vectors, and prediction of abnormal state based on fused feature vectors and historical feature sequences.

[0027] In this step, an adaptive weight network is used to realize the distribution and fusion calculation of the three feature vectors, and then the fused feature vectors are combined with the historical feature sequence to predict the abnormal state.

[0028] It should be noted that the fused feature vectors are respectively predicted for real-time abnormal state, and the change rule of state is extracted based on the historical feature sequence, and then the future state of the power equipment at a certain moment or time period is predicted based on the change rule and the real-time abnormal state.

[0029] ​​​In this embodiment, the edge intelligent gateway is used to acquire real-time electrical quantity time series data, image data and monitoring reports of each power equipment in the abnormal area of the power equipment in the power grid; feature extraction is performed on the real-time electrical quantity time series data, the image data and the monitoring reports respectively to obtain electrical quantity feature vectors, visual feature vectors and semantic feature vectors; weight fusion is performed on the electrical quantity feature vectors, the visual feature vectors and the semantic feature vectors, and an abnormal state is predicted based on the fused feature vectors and historical feature sequences. This method significantly improves the accuracy of power equipment abnormal state prediction by deeply fusing multi-modal perception data and the semantic understanding ability of a large language model, and greatly reduces the risk of false negatives and false positives.

[0030] The method provided in the present application will be described in detail below taking a power equipment-edge intelligent gateway-server power grid intelligent monitoring scene as an example.

[0031] As shown in Figure 4 , a server is set as a background monitoring and analysis, an edge intelligent gateway monitors all power equipment in a region, and different power equipment is arranged in each region. The edge intelligent gateway continuously collects multi-source heterogeneous data in the process of providing state monitoring for the power equipment in the region. At each monitoring moment, the edge intelligent gateway actively predicts the abnormal state of the equipment. The edge intelligent gateway is equipped with an electrical quantity acquisition unit, an infrared thermal imager and a high-definition visible light camera. A DeepSeek large model server is deployed in the cloud, which is responsible for processing unstructured text data and providing knowledge reasoning support.

[0032] Please refer to Figure 2 and 3 , the second embodiment of the power equipment abnormality prediction method of the power grid in the present application, comprising: 301, acquiring a sample data set, acquiring real-time electrical quantity time series data, image data and monitoring reports of each power equipment in the abnormal area of the power equipment in the power grid through an edge intelligent gateway.

[0033] 302, using wavelet packet transform to extract abnormal electrical features in the real-time electrical quantity time series data, and converting and outputting electrical quantity feature vectors, deepening the electrical quantity feature vectors to obtain electrical quantity feature vectors.

[0034] Specifically, using wavelet packet transform, the energy of all sampling points of the real-time electrical quantity time series data in each frequency band is calculated, and based on the energy, the local features belonging to non-stationary signals are marked to obtain an initial frequency energy feature vector ; based on a pre-defined key frequency band index set related to fault features, interference suppression processing is performed on the initial frequency energy feature vector to obtain an initial electrical quantity feature vector ; utilize the multi-layer perceptron (MLP) extracted for electrical quantity features E , the initial electrical quantity feature vector is subjected to high-dimensional abstraction to obtain the electrical quantity feature vector α t .

[0035] In practical applications, first, the wavelet packet transform is used to calculate the energy of all sampling points under each frequency band to capture the local characteristics of non-stationary signals, and the initial frequency energy feature vector is obtained , the transformation can be represented as η , where B is the length of the basis function, i (i e B) is the total number of selected frequency bands, W i η is the η th wavelet packet basis function with a length of is the t th sampling point value in the electrical quantity time series data collected by the edge intelligent gateway at the monitoring time a , and T is the total number of sampling points in a single time window. Define Γ as a set of key frequency band indicators related to fault features (such as harmonics, oscillations, and transient impulses). Keep the elements corresponding to the Γ set in h t , and set other elements to 0 to highlight fault features and suppress normal component interference such as power frequency, to obtain the initial electrical quantity feature vector after key frequency band screening, which is represented by the following formula: ; Then, the high-dimensional abstract feature α t is obtained from , where MLP E represents an MLP network containing L E layers, including two hidden layers and an output layer, and the dropout rate of all dropout layers is set to 0, Θ E = {W E , b E} represents its weight and bias parameters, φ 1 E (·) represents the nonlinear function of the l th layer, and the rectified linear unit (ReLU) is used as the activation function, φ 1 E which can be represented as .

[0036] 303、Use the improved ResNet-18 network to extract visual features from image data and convert the output visual feature vector.

[0037] The improved ResNet-18 network uses a self-defined full connection layer to replace the classification layer in the traditional ResNet-18 as a backbone network.

[0038] In the embodiment, the image data is a two-dimensional pixel matrix, which contains an infrared thermal imaging image and a visible light image of the power equipment; the improved ResNet-18 network is used to extract visual features in the image data and convert and output a visual feature vector β t , comprising: Based on the position information of each power equipment, temperature pixels and gray pixels of a region of interest (ROI) in the two-dimensional pixel matrix are extracted, and the temperature pixels and the gray pixels of all ROIs are converted into normalized temperature and gray distribution maps by a min-max normalization method to obtain standard image data . The standard image data is input into the improved ResNet-18 network to extract visual features, and the extracted visual features are compressed and converted to obtain a visual feature vector β t .

[0039] The standard image data is input into the improved ResNet-18 network to extract visual features, and the extracted visual features are compressed and converted to obtain a visual feature vector β t , comprising: The standard image data is input into the convolution layer of the improved ResNet-18 network to extract low-level features, and the extracted low-level features are subjected to max-pooling; The low-level features subjected to max-pooling are input into the residual network layer of the improved ResNet-18 network to extract multi-level features to obtain a feature map; The feature map is compressed into a feature vector and input into the L V dimensional full connection layer of the improved ResNet-18 network to output a visual feature vector β t .

[0040] Exemplarily, based on the device position information, a key device region in the image is labeled as a region of interest (ROI). Pixel values (temperature values for infrared images and gray value for visible light images) in each ROI are extracted. Through minimum-maximum normalization processing, the pixel data of all ROIs are converted into standardized temperature and gray distribution maps, and the feature values are compared and aligned. The comparison and alignment refer to mapping the ROI pixel values of different pictures to the [0, 1] interval through minimum-maximum normalization, eliminating the influence caused by environmental light, camera parameter differences, and making the feature values comparable.

[0041] The pre-processed converted standardized image is represented as: , wherein I t is a device ROI region matrix extracted from the original image.

[0042] After standardization, ResNet-18 is used as the backbone network, the classification layer of the original ResNet-18 is removed and replaced with a custom fully connected layer with an output dimension of visual feature length L V . The standardized image is input into the network, low-level features are extracted through convolution layers, feature map size is reduced through max-pooling, and multi-level features are extracted through residual blocks in sequence, the feature map is compressed into a feature vector, and the final visual feature L V is output through a t dimension fully connected layer. The Kaiming initialization method is used, a batch normalization layer is added after each convolution layer to speed up training convergence, the hidden layer uses a ReLU function, the output layer does not use an activation function, and a dropout layer is added before the fully connected layer with a dropout rate.

[0043] 304. Utilize the pre-trained power AI large language model to perform semantic analysis and semantic embedding on the monitoring report to obtain a semantic feature vector.

[0044] Specifically, the monitoring report includes: a device historical operation and maintenance report, a real-time scheduling operation log, and environmental meteorological warning information; the pre-trained power AI large language model is used to perform semantic analysis and semantic embedding on the monitoring report to obtain a semantic feature vector t , including: The device historical operation and maintenance report, the real-time scheduling operation log, and the environmental meteorological warning information are standardized pre-processed to obtain a text sequence , wherein the standardized pre-processing includes removing irrelevant characters, correcting misspelled words, and standardizing device names and fault terminology. The text sequence is input into the pre-trained power AI large language model for context encoding, and a text sequence deep semantic of each token in the text sequence and generate a semantic feature vector ξ t , wherein the pre-trained power AI large language model is obtained by fine-tuning a DeepSeek large language model using a power grid equipment knowledge graph and a fault case library. For example, the model is trained using device historical fault reports, maintenance work orders, power grid safety procedure texts, and a power equipment knowledge graph.

[0045] wherein the text sequence is input into the pre-trained power AI large language model for context encoding, and the text sequence is output, wherein the deep semantic of each token in the text sequence is obtained, and a semantic feature vector ξ t is generated by average pooling. t wherein, d is the dimension of the embedding vector.

[0046] 305、weight fusion is performed on the electrical quantity feature vector, the visual feature vector, and the semantic feature vector, and based on the fused feature vector and the historical feature sequence, an abnormal state is predicted.

[0047] Specifically, an adaptive weight network is used to assign weights to each feature vector, and based on the assigned weights, the electrical quantity feature vector α t , the visual feature vector β t , and the semantic feature vector ξ t are fused to obtain a multi-modal fusion feature f t W ; based on the multi-modal fusion feature f t W and the historical feature sequence , an equipment abnormal state is predicted.

[0048] For example, in the adaptive weight network, the unweighted multi-modal feature f' t =α t ⊕β t ⊕ξ t is input, φ 1 W (·) is a nonlinear function of the i-th layer in the adaptive weight network, and ReLU is used as the activation function, l φ ​​1 W may be expressed as and then the adaptive weight network outputs a weight vector where W L and b L are the weights and bias of the last fully connected layer in the adaptive weight network, The function is a sigmoid function, expressed as The average values of the electrical quantity feature weight, visual feature weight and semantic feature weight are calculated by W' t E V L , respectively, which are calculated by and The multi-modal fusion features f w t are weighted by element-wise multiplication, and the weighted multi-modal fusion features f E α t V β t L ξ t .

[0049] For example, define the electrical quantity feature extraction module to output a high-dimensional abstract electrical quantity feature vector α t with a dimension of L E , the visual feature extraction module to output a visual feature vector β t with a dimension of L V , and the DeepSeek semantic feature extraction module to output a semantic feature vector ξ t with a dimension of L L .

[0050] In the fusion, first, the unweighted multi-modal fusion feature f' t is generated by vector splicing, which is expressed as t t t , and the total dimension is L E + L V + L L , which is taken as the input of the adaptive weight network; Then, a multi-layer perceptron (MLP) is used for nonlinear transformation, and the ReLU activation function is used in each layer of the MLP to gradually extract high-order interaction relationships between features. ​​​​​​​​

[0051] Further, the output layer is passed through a Sigmoid function to generate the original weight vector where W AWN and b AWN are the weights and bias of the last fully connected layer in the adaptive weight network.

[0052] Finally, the average of the electrical feature weight, visual feature weight and semantic feature weight is calculated to simplify the high-dimensional weight vector into a scalar weight, reduce the parameter complexity and modal level dynamic adaptability, denoted as , and respectively. Element-wise scalar multiplication (i.e. feature multiplied by weight) is performed on the feature vector of each modal, denoted as w E α t , w V β t and w L ξ t . Subsequently, the final weighted multi-modal feature vector f w t is integrated by concatenation operation = w E α t ⊕ w V β t ⊕ w L ξ t .

[0053] In the present embodiment, the prediction of the device abnormal state based on the multi-modal fusion feature f w t and the historical feature sequence includes: inputting the multi-modal fusion feature f w t to a preset abnormal classification network for abnormal identification to obtain the current power device abnormal state of the device abnormal area; obtaining the historical feature sequence of the device abnormal area and inputting the historical feature sequence to a preset time series prediction network for abnormal state change rule identification to obtain the power device state change sequence of the device abnormal area; based on the current power device abnormal state and the power device state change sequence, the future power device abnormal state of the device abnormal area is obtained.

[0054] Specifically, the anomaly classification network is composed of a four-layer MLP, including three hidden layers and an output layer, and the dropout rate of all Dropout layers is 0. The number of output layer neurons of this module is set to K , corresponding to K a preset abnormal state category, and a Softmax function is used to output the probability of belonging to each category. The weighted multi-modal feature f w t is input into the anomaly classification network, and the output probability vector P is p 1 ,p 2 ,…p k , wherein p k represents the probability that the device belongs to the i-th k abnormal state category, and the category corresponding to the maximum probability is taken as the predicted abnormal state label at the current time t, denoted as , as the abnormal state judgment.

[0055] The time series prediction network uses a gated recurrent unit (GRU) network to extract the time features of the device state evolution and make trend prediction, including two layers of GRU, each layer being configured with G t GRU units, G t , representing the historical window size. The weighted multi-modal fusion feature historical sequence is input , which contains all the fusion features in the window period from time t-G t to the current time t . Through the GRU network, each feature vector in the sequence is processed step by step, and the time dependence is extracted using the gating mechanism. The final hidden state is input into the MLP containing two hidden layers and an output layer, and the output layer uses Z = M x D neurons, wherein M is the prediction step, D is the state dimension, and the output is a device state prediction value sequence of the future M time steps, denoted as , which predicts the device state sequence from t+1 to t + M , a total of M time points, and the process can be represented as .

[0056] 306、Based on the predicted results, an abnormal diagnosis report and a treatment suggestion are generated, and an operation and maintenance personnel is notified, and a preliminary control instruction is generated based on the treatment suggestion to control the power equipment to operate in a transient state.

[0057] In this embodiment, the server proactively predicts the abnormal state of the equipment in the next moment and generates a handling strategy to ensure intervention before the anomaly occurs or worsens. An ideal proactive early warning system needs to predict future equipment states as accurately as possible and maintain stable performance in the face of various uncertainties in actual operation, thereby maximizing operational efficiency and reducing the risk of power outages and operational costs in the power grid system. This objective can be expressed as: , in, U(Θ) It is the overall utility function of the system. P k This represents the probability of accurate detection of the k-th type of anomaly. W k It is the first k Weighting factors for class anomalies C FP The costs associated with false alarms include unnecessary downtime for inspections and manpower consumption. λ It is a penalty factor for the cost of false alarms. f Θ This refers to the multimodal sensing data feature extraction and fusion model, which includes electrical quantity, visual, and semantic feature extraction modules, an adaptive weight network (AWN) module, an anomaly classification module, and a time series prediction module. Θ is the model parameter set. , , These are the preprocessed electrical quantity data, images, and semantic information at time (t-1).

[0058] like Figure 2 As shown, the prediction method here actually includes modules for extracting electrical quantities, visual and semantic features, an adaptive weight network, an anomaly classification network, and a temporal prediction network. Θ is the model parameter set. , , These are the preprocessed electrical quantity data, images, and semantic information at time (t-1).

[0059] Specifically, based on the obtained anomaly prediction and classification results S t-1 , generate the first n Corresponding handling strategies for each edge smart gateway A n The strategy, after undergoing compliance and rationality verification by the DeepSeek large language model, is transformed into specific control instructions or operational suggestions and issued to the execution unit. a. If an emergency anomaly is predicted, the system sends instructions to the automation device via the control interface. A n(e.g., trip command, regulating the tap of on-load tap-changing transformer, switching capacitor reactor) to isolate the fault or suppress the abnormal development.

[0060] b. If it is predicted as a general abnormality or potential risk, the system sends an inspection work order to the operation and maintenance personnel through the production management system (such as a mobile terminal for distribution network operation and maintenance), which contains the device location, abnormal type, and treatment suggestions generated by DeepSeek. A n

[0061] In summary, by fully utilizing the complementary information in multi-modal data such as electrical quantities, visual images, and text reports, and deeply mining device state features, the accuracy of abnormal state prediction and classification is significantly improved, and the risk of fault false alarm and false alarm is greatly reduced.

[0062] Further, from the heterogeneous multi-modal data, a variety of complementary deep features that can more comprehensively depict the health status of the device are efficiently extracted, which together constitute a complete portrait of the device state, providing a reliable basis for accurate diagnosis. Further, an adaptive weight allocation mechanism is introduced, which can dynamically adjust the fusion weights of electrical, visual, and semantic features according to the reliability of each modal data in different environments, effectively reducing the performance loss caused by the failure of a single modal data, and enhancing the robustness and adaptability of the overall monitoring system.

[0063] The above describes the power equipment anomaly prediction method in the power grid in the present application, and the power grid operation and maintenance system in the present application is described below. Please refer to Figure 4 One embodiment of the power grid operation and maintenance system in the present embodiment includes: At least one edge intelligent gateway 410 is used to acquire real-time electrical quantity time series data, image data, and monitoring reports of each power equipment in the device anomaly area in the power grid; wherein each edge intelligent gateway 410 monitors one area, and a plurality of power equipment 430 are arranged in each area; A server 420 is used to respectively extract features from the real-time electrical quantity time series data, the image data, and the monitoring reports to obtain electrical quantity feature vectors α t , visual feature vectors β t , and semantic feature vectors ξ t ; weight fusion is performed on the electrical quantity feature vectors α t , the visual feature vectors β t , and the semantic feature vectors ξ t , and based on the fused feature vectors and historical feature sequences, the abnormal state is predicted.

[0064] In the present embodiment, the server 420 is specifically used for: ​Wavelet packet transform is used to extract abnormal electrical features from the real-time electrical quantity time-series data, and the output electrical quantity feature vector is transformed. For the electrical quantity feature vector Further refinement yields the electrical quantity characteristic vector α. t ; Using an improved ResNet-18 network, visual features are extracted from the image data, and the output visual feature vector β is transformed. t The improved ResNet-18 network is obtained by using the traditional ResNet-18 as the backbone network and replacing the classification layer in the traditional ResNet-18 with a custom fully connected layer. Using a pre-trained power AI large language model, semantic parsing and semantic embedding are performed on the monitoring report to obtain the semantic feature vector ξ. t .

[0065] In this embodiment, the server 420 is specifically used for: Using wavelet packet transform, the energy of all sampling points in each frequency band of the real-time electrical quantity time-series data is calculated, and based on the energy, the local features of non-stationary signals are marked to obtain an initial frequency domain energy feature vector. ; Based on a predefined set of key frequency band indicators related to fault characteristics, the initial frequency domain energy feature vector is... Interference suppression processing is performed to obtain the electrical quantity feature vector. ; Multilayer perceptron (MLP) for electrical quantity feature extraction E For electrical quantity feature vectors By performing high-dimensional abstraction, we obtain the electrical quantity feature vector ξ. t .

[0066] In this embodiment, the image data is a two-dimensional pixel matrix, which contains infrared thermal imaging images and visible light images of the power equipment; The server 420 is specifically used for: Based on the location information of each power device, temperature and grayscale pixels of the Region of Interest (ROI) in the two-dimensional pixel matrix are extracted. Then, the temperature and grayscale pixels of all ROIs are transformed into standardized temperature and grayscale distribution maps using a min-max normalization method to obtain standard image data. ; The standard image data The input is fed into the improved ResNet-18 network to extract visual features, and the extracted visual features are compressed and transformed to obtain the visual feature vector β. t .

[0067] In the embodiment, the monitoring report includes a device historical operation and maintenance report, a real-time scheduling operation log, and environment meteorological warning information. The server 420 is specifically configured to: standardize and preprocess the device historical operation and maintenance report, the real-time scheduling operation log, and the environment meteorological warning information to obtain a text sequence , wherein the standardization preprocessing includes removing irrelevant characters, correcting misspelled words, and standardizing device names and fault terms. The text sequence is input into a pre-trained power AI large language model to perform context encoding, and a text sequence is output, in which a deep semantic of each token is obtained, and a semantic feature vector ξ t is generated through average pooling. The pre-trained power AI large language model is obtained by fine-tuning a DeepSeek large language model using a power grid device knowledge graph and a fault case library.

[0068] In the embodiment, the server 420 is specifically configured to: distribute weights of each feature vector using an adaptive weight network, and fuse the electrical quantity feature vector α t , the visual feature vector β t , and the semantic feature vector ξ t based on the distributed weights to obtain a multi-modal fusion feature f w t . Based on the multi-modal fusion feature f w t and a historical feature sequence , a device abnormal state is predicted.

[0069] In the embodiment, the server 420 is specifically configured to: generate an abnormal diagnosis report and a treatment suggestion based on the prediction result, and notify an operation and maintenance personnel, and generate a preliminary control instruction based on the treatment suggestion, the preliminary control instruction being used to control the power device to operate in a transient state.

[0070] In the embodiment, the edge intelligent gateway obtains real-time electrical quantity time series data, image data, and monitoring reports of each power device in a device abnormal area in a power grid; feature extraction is performed on the real-time electrical quantity time series data, the image data, and the monitoring reports to obtain an electrical quantity feature vector α t , a visual feature vector β t , and a semantic feature vector ξ t ; and the electrical quantity feature vector αt the visual feature vector β t and the semantic feature vector ξ t weight fusion is performed, and based on the fused feature vector and the historical feature sequence, an abnormal state is predicted. This method utilizes the complementary information in the multi-modal data of electrical quantities, visual images and monitoring reports, extracts device state features and performs fusion prediction, which significantly improves the accuracy of abnormal state prediction, greatly reduces the risk of fault false alarm and false alarm, and also adapts to the trend of early prevention.

[0071] The above Figure 4 The power grid operation and maintenance system of the present application is described in detail from the perspective of modular functional entities, and the electronic device of the present application is described in detail from the perspective of hardware processing.

[0072] Referring to Figure 5 The electronic device includes a processor 500 and a memory 501, and the memory 501 stores machine executable instructions that can be executed by the processor 500, and the processor 500 executes the machine executable instructions to implement the power equipment abnormality prediction method of the power grid described above.

[0073] Further, Figure 5 The electronic device shown in the figure also includes a bus 502 and a communication interface 503, and the processor 500, the communication interface 503 and the memory 501 are connected through the bus 502.

[0074] The memory 501 can include a high-speed random access memory (RAM), and can also include a non-volatile memory, for example, at least one disk memory. The communication connection between the system network element and at least one other network element is realized through at least one communication interface 503 (which can be wired or wireless), and the Internet, a wide area network, a local area network, a metropolitan area network, etc. can be used. The bus 502 can be an ISA bus, a PCI bus or an EISA bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. For the sake of expression, Figure 5 In the figure, only one bidirectional arrow is used to represent, but it does not mean that there is only one bus or one type of bus.

[0075] The processor 500 can be an integrated circuit chip having a processing capability of signals. In the implementation process, each step of the above method can be completed by the integrated logic circuit of hardware in the processor 500 or the instruction in the form of software. The processor 500 described above can be a general processor, including a central processing unit (CPU), a network processor (NP), etc.; can also be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic device, a discrete gate or transistor logic device, a discrete hardware component. Each method, step and logic block diagram disclosed in the embodiments of the present disclosure can be implemented or executed. The general processor can be a microprocessor or the processor can also be any conventional processor. The steps of the method disclosed in combination with the embodiments of the present disclosure can be directly embodied as a hardware code processor for execution, or a combination of hardware and software modules in the code processor for execution. The software module can be located in a random access memory, a flash memory, a read only memory, a programmable read only memory or an electrically erasable programmable memory, a register, etc. The storage medium in the art. The storage medium is located in the memory 501, and the processor 500 reads the information in the memory 501, and combines the hardware to complete the method steps of the above embodiments.

[0076] The present application also provides a computer readable storage medium, which can be a non-volatile computer readable storage medium, and can also be a volatile computer readable storage medium, and the computer readable storage medium stores instructions, when the instructions run on the computer, so that the computer executes the steps of the power equipment anomaly prediction method of the power grid.

[0077] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the above described system, device and unit can refer to the corresponding process in the foregoing method embodiment, which will not be repeated here.

[0078] The integrated unit, if implemented in the form of a software function unit and sold or used as an independent product, can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application or the part of the prior art that essentially contributes or the whole or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various media that can store program codes.

[0079] The above, the above embodiments are only used to illustrate the technical solutions of the present application, but not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for predicting power equipment anomalies in a power grid, characterized in that, The method includes: Real-time electrical quantity time-series data, image data, and monitoring reports of each power device in the abnormal equipment area of ​​the power grid are obtained through the edge intelligent gateway; Feature extraction is performed on the real-time electrical quantity time-series data, the image data, and the monitoring report to obtain the electrical quantity feature vector α. t Visual feature vector β t and semantic feature vector ξ t ; For the electrical quantity feature vector α t The visual feature vector β t and the semantic feature vector ξ t Weight fusion is performed, and abnormal states are predicted based on the fused feature vector and historical feature sequence.

2. The method for predicting power equipment anomalies in a power grid according to claim 1, characterized in that, The process involves extracting features from the real-time electrical quantity time-series data, the image data, and the monitoring report to obtain the electrical quantity feature vector α. t Visual feature vector β t and semantic feature vector ξ t ,include: Wavelet packet transform is used to extract abnormal electrical features from the real-time electrical quantity time-series data, and the output electrical quantity feature vector is transformed. For the electrical quantity feature vector Further refinement yields the electrical quantity characteristic vector α. t ; Using an improved ResNet-18 network, visual features are extracted from the image data, and the output visual feature vector β is transformed. t The improved ResNet-18 network is obtained by using the traditional ResNet-18 as the backbone network and replacing the classification layer in the traditional ResNet-18 with a custom fully connected layer. Using a pre-trained power AI large language model, semantic parsing and semantic embedding are performed on the monitoring report to obtain the semantic feature vector ξ. t .

3. The method for predicting power equipment anomalies in a power grid according to claim 2, characterized in that, The process involves using wavelet packet transform to extract abnormal electrical features from the real-time electrical quantity time-series data and then transforming and outputting an electrical quantity feature vector. For the electrical quantity feature vector Further refinement yields the electrical quantity characteristic vector α. t ,include: Using wavelet packet transform, the energy of all sampling points in each frequency band of the real-time electrical quantity time-series data is calculated, and based on the energy, the local features of non-stationary signals are marked to obtain an initial frequency domain energy feature vector. ; Based on a predefined set of key frequency band indicators related to fault characteristics, the initial frequency domain energy feature vector is... Interference suppression processing is performed to obtain the electrical quantity feature vector. ; Multilayer perceptron (MLP) for electrical quantity feature extraction E For electrical quantity feature vectors By performing high-dimensional abstraction, we obtain the electrical quantity feature vector α. t .

4. The method for predicting power equipment anomalies in a power grid according to claim 2, characterized in that, The image data is a two-dimensional pixel matrix, which contains infrared thermal imaging images and visible light images of the power equipment; The improved ResNet-18 network is used to extract visual features from the image data and transform the output visual feature vector β. t ,include: Based on the location information of each power device, temperature and grayscale pixels of the Region of Interest (ROI) in the two-dimensional pixel matrix are extracted. Then, the temperature and grayscale pixels of all ROIs are transformed into standardized temperature and grayscale distribution maps using a min-max normalization method to obtain standard image data. ; The standard image data The input is fed into the improved ResNet-18 network to extract visual features, and the extracted visual features are compressed and transformed to obtain the visual feature vector β. t .

5. The method for predicting power equipment anomalies in a power grid according to claim 2, characterized in that, The monitoring report includes: historical equipment operation and maintenance reports, real-time dispatch operation logs, and environmental and meteorological early warning information; The pre-trained power AI large language model is used to perform semantic parsing and semantic embedding on the monitoring report to obtain the semantic feature vector ξ. t ,include: The historical operation and maintenance reports of the equipment, the real-time scheduling operation logs, and the environmental and meteorological early warning information are standardized and preprocessed to obtain a text sequence. The standardization preprocessing includes removing irrelevant characters, correcting typos, and standardizing equipment names and fault terms. The text sequence The input is fed into a pre-trained power AI large language model for context encoding, and the output is a text sequence. The deep semantics of each token are analyzed, and a semantic feature vector ξ is generated through average pooling. t The pre-trained power AI language model is obtained by performing domain-adaptive fine-tuning on the DeepSeek language model using a power grid equipment knowledge graph and a fault case library.

6. The method for predicting power equipment anomalies in a power grid according to any one of claims 1-5, characterized in that, The electrical quantity feature vector α t The visual feature vector β t and the semantic feature vector ξ t Weight fusion is performed, and abnormal states are predicted based on the fused feature vector and historical feature sequences, including: An adaptive weighting network is used to assign weights to each feature vector, and the electrical quantity feature vector α is then adjusted based on the assigned weights. t The visual feature vector β t and the semantic feature vector ξ t The fusion is performed to obtain the multimodal fusion feature f. t w ; Based on the multimodal fusion feature f t w and historical feature sequences Predict abnormal equipment conditions.

7. The method for predicting power equipment anomalies in a power grid according to claim 6, characterized in that, In the pair of electrical quantity feature vectors α t The visual feature vector β t and the semantic feature vector ξ t After performing weight fusion and predicting abnormal states based on the fused feature vector and historical feature sequences, the process also includes: Based on the prediction results, an anomaly diagnosis report and handling suggestions are generated and notified to the operation and maintenance personnel. Based on the handling suggestions, preliminary control instructions are generated to control the power equipment to operate in a transient state.

8. A power grid operation and maintenance system, characterized in that, The power grid operation and maintenance system includes: At least one edge smart gateway is used to acquire real-time electrical quantity time-series data, image data and monitoring reports of each power device in the abnormal equipment area of ​​the power grid; wherein, each edge smart gateway monitors one area, and each area has multiple power devices; The server is used to extract features from the real-time electrical quantity time-series data, the image data, and the monitoring report to obtain an electrical quantity feature vector α. t Visual feature vector β t and semantic feature vector ξ t ; for the electrical quantity feature vector α t The visual feature vector β t and the semantic feature vector ξ t Weight fusion is performed, and abnormal states are predicted based on the fused feature vector and historical feature sequence.

9. An electronic device, characterized in that, The electronic device includes: a memory and at least one processor, wherein the memory stores instructions; The at least one processor invokes the instructions in the memory to cause the electronic device to execute the power equipment anomaly prediction method for the power grid as described in any one of claims 1-7.

10. A computer-readable storage medium storing instructions thereon, characterized in that, When the instruction is read and executed, it performs the power equipment anomaly prediction method for the power grid as described in any one of claims 1-7.

Images