Converter Station Energy Efficiency Prediction System Based on Deep Learning

Abstract

The present invention discloses an energy efficiency prediction system for a converter station based on deep learning, including: S1, collecting data of each device in the converter station to generate an original data set; S2, processing the original data set to generate a standardized data set; S3, constructing a device network diagram based on the topological relationship of the converter station devices; S4, a graph neural network module, constructing a graph neural network model based on the device network diagram, extracting feature information of nodes and edges, and generating multi-dimensional embedding vectors; S5, a meta-learning module, based on the multi-dimensional embedding vectors generated by the graph neural network module, updating the model parameters by adopting a hierarchical optimization strategy to generate a set of model parameters adaptable to different input feature distributions; S6, predicting the energy efficiency related indicators of the converter station based on the set of model parameters generated by the meta-learning module. The present invention has the advantages of strong modeling ability, high prediction accuracy, strong adaptability and good real-time performance.

Description

Technical Field

[0001] The present invention relates to the technical field of converter station energy efficiency prediction, and particularly to a converter station energy efficiency prediction system based on deep learning. Background Art

[0002] With the rapid development of energy technology and power transmission technology, converter stations have become the core equipment of high-voltage direct current transmission systems. The efficient operation of converter stations is crucial for the efficient transmission of energy and the stable operation of the system. However, due to the complex operating environment of the equipment in the converter station, accurate prediction of energy efficiency-related parameters still faces many challenges.

[0003] The purpose of converter station energy efficiency prediction is to provide a scientific basis for the operation optimization and fault prevention of converter stations by monitoring and analyzing the operating status of equipment. In the prior art, traditional energy efficiency prediction methods mainly rely on rule-based mathematical models and empirical formulas. However, with the complexity of the equipment structure in the converter station and the rapid growth of the scale of its operating data, these traditional methods have shown significant limitations:

[0004] 1. Lack of complex association modeling ability: Traditional methods usually only model a single device or independent parameter, and it is difficult to capture the complex topological relationships and dynamic interaction characteristics between the devices in the converter station.

[0005] 2. Difficulty in adapting to diverse operating environments: The operating status of converter stations is significantly affected by environmental conditions and operating modes, and existing methods lack the ability to quickly adapt to changes in the data feature distribution under different conditions.

[0006] 3. Low prediction accuracy: Traditional prediction methods based on empirical formulas have limited prediction accuracy when facing high-dimensional, non-linear, and dynamically changing operating data, and it is difficult to meet the requirements of actual operation optimization.

[0007] 4. Lack of real-time and multi-dimensional analysis ability: Existing methods mostly rely on offline analysis and cannot achieve real-time prediction of the operating status and comprehensive analysis of multi-dimensional energy efficiency parameters.

[0008] Therefore, how to provide a converter station energy efficiency prediction system based on deep learning is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0009] An object of the present invention is to propose a converter station energy efficiency prediction system based on deep learning. The present invention fully combines graph neural network and meta-learning technology, and details an algorithm for accurately predicting energy efficiency-related indicators by modeling the topological relationships of converter station equipment and quickly adapting to diverse operating environments, which has the advantages of strong modeling ability, high prediction accuracy, strong adaptability, and good real-time performance.

[0010] An energy efficiency prediction system for a converter station based on deep learning according to an embodiment of the present invention includes:

[0011] S1. A data acquisition module, configured to acquire data of each device in the converter station and generate an original data set containing the complete operating status of the devices;

[0012] S2. A data preprocessing module, configured to process the original data set, including data cleaning, missing value filling, normalization processing, and time series feature extraction, to generate a cleaned and standardized data set;

[0013] S3. A graph construction module, configured to construct a device network graph based on the topological relationship of the converter station devices, where the device network graph represents each device as a node and represents the electrical connection relationship between the devices as an edge;

[0014] S4. A graph neural network module, configured to construct a graph neural network model based on the device network graph, extract feature information of the nodes and edges, and generate multi-dimensional embedding vectors;

[0015] S5. A meta-learning module, configured to update the model parameters of the graph neural network model based on the multi-dimensional embedding vectors generated by the graph neural network module by using a hierarchical optimization strategy, and generate a set of model parameters adaptable to different input feature distributions;

[0016] S6. An energy efficiency prediction module, configured to predict the energy efficiency related indicators of the converter station based on the set of model parameters generated by the meta-learning module, in combination with the standardized data set and the multi-dimensional embedding vectors.

[0017] Optionally, the S3 specifically includes:

[0018] S31. A topology data extraction unit, configured to extract the topology structure data of the converter station devices from the standardized data set, including a node set and an edge set , where the node set represents each device in the converter station, and the edge set represents the electrical connection relationship between the devices;

[0019] S32. A feature mapping unit, configured to map the device operating status data of each node in the node set to a node feature vector , where is the dimension of the feature vector, and is a set of real numbers;

[0020] S33. An edge weight calculation unit, configured to calculate the corresponding weight value based on the attribute information of each edge in the edge set , where represents the node and the node Connection strength between:

[0021] ;

[0022] Wherein, and respectively represent the eigenvectors of nodes and node , represents the attribute vector of the edge, is the weight matrix of node features, is the weight vector of edge attributes, where represents the dimension of the edge attribute vector, represents the rectified linear unit activation function, used to map the output value to the non - negative range;

[0023] S34, the graph structure generation unit, constructs the device network graph based on the node set , the edge set and the corresponding weight values , wherein represents the weight matrix of node features, and the matrix element is the weight value.

[0024] Optionally, the S4 specifically includes:

[0025] S41, the input node feature unit, receives the device network graph generated by the graph construction module, where the node set contains the node feature vectors of each node, the edge set contains the attribute vectors of each edge, and the weight matrix contains the weight values of each edge;

[0026] S42, the feature aggregation unit, aggregates the local neighborhood information of nodes based on graph convolution operations to update the node feature vector :

[0027] ;

[0028] Wherein, represents the set of neighbor nodes of node , represents the weight value between node and node , is the weight matrix of the th layer, is the th Layer feature vector is the activation function;

[0029] S43. Feature update unit, which updates the node feature vector layer by layer based on the multi-layer graph convolutional network to generate a node embedding vector , where represents the dimension of the final embedding vector;

[0030] S44. Edge feature fusion unit, which generates an edge embedding vector based on the node embedding vector and the edge attribute vector ;

[0031] S45. Fuse the node embedding vector and the edge embedding vector to generate a multi-dimensional embedding vector.

[0032] Optionally, the S44 specifically includes:

[0033] S441. Extract the two end-node embedding vectors associated with the edge from the node embedding vector and , and extract the attribute vector corresponding to each edge from the edge attribute vector ;

[0034] S442. Perform a concatenation operation on the node embedding vector , and the edge attribute vector to generate a joint feature vector :

[0035] ;

[0036] where represents the vector concatenation operation, and the joint feature vector , is the dimension of the node embedding vector, is the dimension of the edge attribute vector;

[0037] S443. Perform a non-linear transformation on the joint feature vector to generate an edge embedding vector :

[0038] ;

[0039] where is the weight matrix, is the bias vector, is the activation function, is the generated edge embedding vector, is the dimension of the edge embedding vector.

[0040] Optionally, the S45 specifically includes:

[0041] S451. Integrate the node embedding vector and the edge embedding vector, and calculate the comprehensive embedding vector of each node according to the node embedding vector of the node and the edge embedding vector between it and its neighbor nodes: ; ;

[0042] wherein, represents the set of neighbor nodes of node , is the transformation matrix of the node embedding vector, is the transformation matrix of the edge embedding vector, represents the number of neighbor nodes of node , is the non - linear activation function;

[0043] S452. Integrate the global and local embedding features. For the comprehensive embedding vectors of all nodes, construct the global embedding feature vector , and splice it with the comprehensive embedding vector of each node to generate the final multi - dimensional embedding vector :

[0044] ;

[0045] wherein, , is the total number of the node set, represents the vector splicing operation, is the finally generated multi - dimensional embedding vector.

[0046] Optionally, the S5 specifically includes:

[0047] S51. Meta - task construction unit. Based on the multi - dimensional embedding vector generated by the graph neural network module, construct a meta - task set composed of multiple meta - tasks, wherein, represents the total number of meta - tasks, and each meta - task includes the input feature distribution and the corresponding objective function, and is used to simulate the input data feature distribution and its target mapping relationship in different operating environments;

[0048] S52. Initial parameter optimization unit. Based on the meta - task set , optimize the initial model parameters by minimizing the total loss function of all meta - tasks:

[0049] ;

[0050] wherein, Represents the loss function of the model on the meta-task ; are the model parameters to be optimized;

[0051] S53. Task Adaptation Parameter Update Unit. For each meta-task in the meta-task set , the gradient descent method is used to update the initial model parameters to task-specific parameters to adapt to the feature distribution of the meta-task ;

[0052] S54. Global Parameter Update Unit. Aggregate the task-specific parameters , and combine the weights of each meta-task to optimize the initial parameters to generate global parameters reflecting cross-tasks ;

[0053] S55. Combine the global parameters with the task-specific parameters to form a set of model parameters that can adapt to different input feature distributions .

[0054] Optionally, the S53 specifically includes:

[0055] S531. According to the loss function of each meta-task in the meta-task set , calculate the gradient of the initial model parameters : ;

[0056] ;

[0057] Among them, is the loss function of the meta-task , and represents the gradient of this loss function with respect to the initial model parameters ;

[0058] S532. Dynamically adjust the learning rate according to the feature distribution of the meta-task to optimize the update step size:

[0059] ;

[0060] Among them, is the initial learning rate, is the learning rate decay factor, is the number of iterations for meta-task optimization; ​

[0061] S533. Based on the learning rate and the gradient , update the initial model parameters to generate task-specific parameters :

[0062] ;

[0063] Among them, represents the task-specific parameters adapted to the feature distribution of the meta-task .

[0064] Optionally, the S54 specifically includes:

[0065] S541. Calculate the task weights for each meta-task in the meta-task set based on the feature distribution and task importance of each meta-task, and the task weights satisfy the following constraint conditions:

[0066] ;

[0067] Among them, represents the total number of meta-tasks, and the weights are allocated according to the specific importance indicators of the tasks;

[0068] S542. Perform weighted summation on the gradients of all meta-tasks to generate the global gradient :

[0069] ;

[0070] Among them, represents the gradient corresponding to the task-specific parameters of the task , and

[0071] S543. Update the global parameters based on the initial model parameters and the global gradient to generate the optimized global parameters :

[0072] ;

[0073] Among them, is the global optimization step size.

[0074] Optionally, the S6 specifically includes:

[0075] S61. A feature fusion unit constructs an input feature matrix based on the set of model parameters generated by the meta - learning module , in combination with the device operation features and multi - dimensional embedding vectors in the standardized dataset :

[0076] ;

[0077] wherein is the device operation feature of node , is the corresponding multi - dimensional embedding vector, represents the feature concatenation operation;

[0078] S62. An energy efficiency prediction model selection unit selects the corresponding global parameters and task - specific parameters from the set of model parameters to construct a task - specific prediction model , and selects the optimal task model according to the device operation features;

[0079] S63. A prediction calculation unit calculates the energy efficiency - related indicators of the converter station based on the selected task - specific prediction model and the input feature matrix :

[0080] ;

[0081] wherein represents the energy efficiency - related indicators of the converter station, represents the task - specific prediction model, is the task - specific parameter.

[0082] The beneficial effects of the present invention are:

[0083] (1) By combining the graph neural network and meta - learning technology, the present invention comprehensively models the topological relationship and dynamic interaction characteristics of the devices in the converter station, enabling the system to accurately capture complex device associations and global characteristics, thereby effectively improving the prediction accuracy of energy efficiency - related indicators, especially for the processing requirements of multi - dimensional, high - dimensional, and dynamically changing data.

[0084] (2) By introducing meta - learning technology, the model of the present invention has the ability to quickly adapt to different operating environments and changes in data feature distributions, so that it can still efficiently predict energy efficiency indicators when the operating mode or environmental conditions of the converter station change, avoiding the problem of performance degradation caused by environmental changes in traditional methods.

[0085] (3) By combining standardized data, multi-dimensional embedding vectors, and a dynamically updated set of model parameters in real time, the present invention realizes real-time monitoring and efficient calculation of the energy efficiency related indicators of the converter station. This not only significantly shortens the prediction latency time, but also supports real-time optimization of the operating state of the converter station, effectively improving the operating efficiency and safety of the system.

[0086] (4) Through the dynamic optimization of the task-specific prediction model and global parameters, the present invention provides a comprehensive analysis ability for energy efficiency related problems, enabling the system to understand the operating state of the converter station in multiple dimensions, automatically complete the prediction and analysis of energy efficiency indicators, reduce the dependence on manual intervention, and improve the intelligence and applicability of energy efficiency prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0087] The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation to the present invention. In the drawings:

[0088] Figure 1 is the overall framework diagram of a converter station energy efficiency prediction system based on deep learning proposed by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0089] Now, the present invention will be further described in detail with reference to the drawings. These drawings are all simplified schematic diagrams, only showing the basic structure of the present invention in a schematic way, so they only show the components related to the present invention.

[0090] Referring to Figure 1 , a converter station energy efficiency prediction system based on deep learning includes:

[0091] S1. A data acquisition module, which is used to collect the data of each device in the converter station and generate an original data set containing the complete operating state of the device;

[0092] In this embodiment, S1 specifically includes:

[0093] The data acquisition module is used to collect the operating state data, environmental variables, and topological relationship data from various devices in the converter station and generate an original data set containing the complete operating state of the device.

[0094] S2. A data preprocessing module, which is used to process the original data set, including data cleaning, missing value filling, normalization processing, and time series feature extraction, to generate a cleaned and standardized data set;

[0095] In this embodiment, S2 specifically includes:

[0096] The data preprocessing module is used to comprehensively process the original dataset generated by the data acquisition module. The data cleaning process includes detecting and removing outliers and noise data, filling in the missing parts of the data through interpolation, and performing linear transformation on the data through normalization to map all data values to a unified range, eliminating the influence of dimensional differences on subsequent analysis. Time series feature extraction extracts trend features, periodic features, and time series correlations from the original dataset according to the time dependence of the data, generating a standardized dataset.

[0097] S3. The graph construction module constructs a device network graph based on the topological relationship of converter station equipment. The device network graph represents each device as a node and the electrical connection relationship between devices as an edge.

[0098] In this embodiment, S3 specifically includes:

[0099] S31. The topological data extraction unit extracts the topological structure data of converter station equipment from the standardized dataset, including the node set and the edge set , where the node set represents each device in the converter station, and the edge set represents the electrical connection relationship between devices.

[0100] S32. The feature mapping unit maps the device operation status data of each node in the node set to a node feature vector , where is the dimension of the feature vector, and is a set of real numbers.

[0101] S33. The edge weight calculation unit calculates the corresponding weight value for each edge based on the attribute information of the edge set , where represents the connection strength between node and node :

[0102] ;

[0103] Among them, and respectively represent the feature vectors of node and node , represents the attribute vector of the edge, is the weight matrix of node features, is the weight vector of edge attributes, where represents the dimension of the edge attribute vector, Represents a rectified linear unit activation function for mapping output values to the non - negative range;

[0104] S34. A graph structure generation unit that constructs a device network diagram according to the node set , edge set and the corresponding weight values to construct a device network diagram , where represents the weight matrix of node features, and the matrix element is the weight value.

[0105] In this embodiment, by extracting the topological structure data of converter station devices from the standardized dataset, constructing a device network diagram, and calculating and generating complete device relationship modeling information using node feature vectors and edge weights, the electrical connection relationship and dynamic characteristics between devices in the converter station can be accurately described. This method integrates node operating states and edge attribute features, which can not only reflect local device characteristics but also capture global interaction characteristics through topological relationships, thus providing high - quality input data for subsequent deep - learning modeling, improving the system's modeling ability for the complex energy - efficiency characteristics of the converter station and the accuracy of prediction, and providing a reliable data basis and structural support for the energy - efficiency optimization of the converter station.

[0106] S4. A graph neural network module that constructs a graph neural network model based on the device network diagram, extracts the feature information of nodes and edges, and generates multi - dimensional embedding vectors;

[0107] In this embodiment, S4 specifically includes:

[0108] S41. An input node feature unit that receives the device network diagram generated by the graph construction module , where the node set contains the node feature vectors of each node , the edge set contains the attribute vectors of each edge , and the weight matrix contains the weight values of each edge ;

[0109] S42. A feature aggregation unit that aggregates the local neighborhood information of nodes based on graph convolution operations and updates the node feature vectors :

[0110] ;

[0111] Among them, represents the set of neighbor nodes of node , represents the weight value between node and node , is the weight matrix of the layer, is the -th layer feature vector of the node, is the activation function;

[0112] S43. Feature update unit, which updates the node feature vector layer by layer based on the multi-layer graph convolutional network to generate a node embedding vector , where represents the dimension of the final embedding vector;

[0113] S44. Edge feature fusion unit, which generates an edge embedding vector based on the node embedding vector and the edge attribute vector;

[0114] S45. Fuse the node embedding vector and the edge embedding vector to generate a multi-dimensional embedding vector.

[0115] S441. Extract the two end-node embedding vectors and associated with the edge from the node embedding vector, and extract the attribute vector corresponding to each edge from the edge attribute vector;

[0116] S442. Perform a concatenation operation on the node embedding vector , and the edge attribute vector to generate a joint feature vector :

[0117] ;

[0118] where represents the vector concatenation operation, the joint feature vector , is the dimension of the node embedding vector, is the dimension of the edge attribute vector;

[0119] S443. Perform a non-linear transformation on the joint feature vector to generate an edge embedding vector :

[0120] ;

[0121] where is the weight matrix, is the bias vector, is the activation function, is the generated edge embedding vector, is the dimension of the edge embedding vector.

[0122] S451. Integrate the node embedding vector and the edge embedding vector, and calculate the comprehensive embedding vector of the node according to the node embedding vector of each node and the edge embedding vector between it and its neighbor nodes : : ;

[0123] wherein, represents the set of neighbor nodes of node , is the transformation matrix of the node embedding vector, is the transformation matrix of the edge embedding vector, represents the number of neighbor nodes of node , is the non-linear activation function;

[0124] S452. Integrate the global and local embedding features. For the comprehensive embedding vectors of all nodes, construct the global embedding feature vector , and splice it with the comprehensive embedding vector of each node to generate the final multi-dimensional embedding vector :

[0125] ;

[0126] wherein, , is the total number of the node set, represents the vector splicing operation, is the finally generated multi-dimensional embedding vector.

[0127] In this embodiment, through the fusion processing of the node embedding vector and the edge embedding vector in the device network diagram, the local topological characteristics and the global features are effectively combined to generate a multi-dimensional embedding vector. This method can not only accurately capture the complex interaction relationships between converter station devices, but also enhance the understanding of the overall network dynamics through the global features, making the expression of the information related to the energy efficiency of the converter station more comprehensive and accurate. Through the comprehensive modeling of node and edge features, the representation ability of the energy efficiency prediction model is further improved, providing an efficient and reliable data basis support for the energy efficiency optimization in the complex converter station environment.

[0128] S5. Meta-learning module. Based on the multi-dimensional embedding vector generated by the graph neural network module, adopt a hierarchical optimization strategy to update the model parameters of the graph neural network model, and generate a set of model parameters that can adapt to different input feature distributions;

[0129] In this embodiment, S5 specifically includes:

[0130] S51. Meta-task construction unit, which constructs a meta-task set composed of multiple meta-tasks based on the multi-dimensional embedding vectors generated by the graph neural network module. , where represents the total number of meta-tasks, and each meta-task includes an input feature distribution and a corresponding objective function, and is used to simulate the input data feature distribution and its target mapping relationship in different operating environments.

[0131] S52. Initial parameter optimization unit, based on the meta-task set , optimizes the initial model parameters by minimizing the total loss function of all meta-tasks:

[0132] ;

[0133] where represents the loss function of the model on the meta-task , and are the model parameters to be optimized;

[0134] S53. Task adaptation parameter update unit, for each meta-task in the meta-task set, uses the gradient descent method to update from the initial model parameters to task-specific parameters to to adapt to the feature distribution of the meta-task ;

[0135] S54. Global parameter update unit, aggregates the task-specific parameters , combines the weights of each meta-task to optimize the initial parameters to generate global parameters that reflect cross-tasks;

[0136] S55. Combines the global parameters with the task-specific parameters to form a set of model parameters that can adapt to different input feature distributions.

[0137] S531. According to the loss function of each meta-task in the meta-task set, calculates the gradient of the initial model parameters : :

[0138] ;

[0139] where is the meta-task The loss function, represents the gradient of the loss function with respect to the initial model parameters ;

[0140] S532. Dynamically adjust the learning rate according to the feature distribution of the meta-task to optimize the update step size:

[0141] ;

[0142] wherein, is the initial learning rate, is the learning rate decay factor, is the number of iterations for meta-task optimization;

[0143] S533. Based on the learning rate and the gradient , update the initial model parameters to generate task-specific parameters :

[0144] ;

[0145] wherein, represents the task-specific parameters adapted to the feature distribution of the meta-task .

[0146] S541. Calculate the task weights for each meta-task in the meta-task setbased on the feature distribution and task importance of each meta-task, and the task weights satisfy the following constraint conditions:

[0147] ;

[0148] wherein, represents the total number of meta-tasks, and the weight is allocated according to the specific importance index of the task;

[0149] S542. Perform weighted summation on the gradients of all meta-tasks to generate the global gradient :

[0150] ;

[0151] wherein, represents the gradient corresponding to the task-specific parameters of task , the corresponding gradient, and the global gradient , update the global parameters to generate optimized global parameters : ;

[0152] Among them, is the global optimization step size.

[0153] In this embodiment, a hierarchical optimization strategy is introduced through the meta-learning module, and the learning rate and the model parameter update path are dynamically adjusted in combination with the task characteristics in the meta-task set, enabling the model to quickly adapt to different input feature distributions. It can not only capture the global change characteristics of the converter station operation environment but also refine to the local feature distribution of specific tasks, generating a more accurate task-specific parameter set. In addition, the effective combination of global parameters and task-specific parameters combines the generality across tasks and the personalized adaptation ability, significantly improving the generalization performance and prediction accuracy of the converter station energy efficiency prediction system, providing efficient and reliable support for energy efficiency optimization decisions in complex operation scenarios.

[0154] S6. Energy efficiency prediction module, based on the model parameter set generated by the meta-learning module, combines the standardized data set and the multi-dimensional embedding vector to predict the energy efficiency related indicators of the converter station.

[0155] In this embodiment, S6 specifically includes:

[0156] S61. Feature fusion unit, based on the model parameter set generated by the meta-learning module , combines the device operation characteristics in the standardized data set and the multi-dimensional embedding vector to construct an input feature matrix :

[0157] ;

[0158] Among them, is the device operation characteristic of node , is the corresponding multi-dimensional embedding vector, represents the feature concatenation operation;

[0159] S62. Energy efficiency prediction model selection unit, selects the corresponding global parameter and task-specific parameter from the model parameter set to construct a task-specific prediction model , and selects the optimal task model according to the device operation characteristics;

[0160] S63. Prediction calculation unit, based on the selected task-specific prediction model and the input feature matrix , calculates the energy efficiency related indicators of the converter station:

[0161] ​​ ;

[0162] Among them, represents the index related to the energy efficiency of the converter station, represents the task-specific prediction model, is the task-specific parameter.

[0163] In this embodiment, the feature fusion unit combines the device operation data and the multi-dimensional embedding vectors to construct an input feature matrix, comprehensively integrating the operation state of the converter station and the global topology features, and capable of capturing complex system dynamic relationships; by selecting the optimal task-specific prediction model from the model parameter set and dynamically adapting to different operation data features, the accuracy of the prediction model is ensured; the energy efficiency related indicators are calculated by the prediction calculation unit, such as the top oil temperature of the converter transformer and the system energy loss, to achieve high-precision modeling of the energy efficiency features. This embodiment combines the global feature modeling ability of deep learning and the fast adaptation ability of meta-learning, providing an accurate and efficient solution for the energy efficiency prediction of the converter station, and at the same time providing reliable data support for energy efficiency optimization and operation decision-making. Embodiment 1:

[0164] To verify the feasibility of the present invention, the present invention is applied to the high-voltage DC converter station system of a large power grid operation company. The company uses multiple converter stations in power transmission. As the core nodes, the energy efficiency management of the converter stations directly affects the operation efficiency and stability of the entire power grid. However, the company has found the following problems in energy efficiency management in recent years: the prediction deviation of the energy efficiency related indicators of some converter stations is relatively large, and it cannot effectively support operation optimization and early fault warning. To solve this problem, the company decides to introduce the converter station energy efficiency prediction system based on deep learning of the present invention.

[0165] The deployment process of this system includes data collection, graph construction and modeling, training of the meta-learning module, and real-time application of the energy efficiency prediction module. The operation data in the converter station is first collected by the collection device, including the converter station equipment topology, electrical connection data, operation state parameters, and environmental variables. Through the graph construction module, these data are mapped into a device network graph, where the nodes represent the converter station equipment and the edges represent the electrical relationships between the equipment. On this basis, the node features are extracted through the graph neural network module, and the meta-learning module generates an adaptation parameter set for different operation environments. Finally, the energy efficiency prediction module makes real-time predictions based on these parameters.

[0166] A certain converter station is located in a coastal area. Due to the long-term exposure of its converter transformers to high humidity and high temperature environments, the abnormal increase in the top oil temperature occurs frequently. However, the traditional empirical model cannot accurately predict the oil temperature, resulting in several equipment failures caused by overheating. After the energy efficiency prediction system of the present invention is deployed in this converter station, it constructs a device network diagram using the real-time collected device operation data and environmental data, and quickly adapts to the characteristic distribution of this coastal environment through the meta-learning module. In the energy efficiency prediction module, by combining the standardized data and multi-dimensional embedding vectors, it makes real-time predictions on the top oil temperature and energy loss of the converter transformer, providing data support for the operation optimization of the converter station.

[0167] In practical applications, the system first extracts the complex interaction characteristics between devices through a graph neural network to generate the embedding vectors of the converter station. After these embedding vectors are fused with the standardized operation data, they are input into the energy efficiency prediction module, and a task-specific model is selected for real-time prediction. For example, in a scenario where the load current suddenly increases, the system accurately predicts that the top oil temperature will reach the alarm threshold after 45 minutes, providing sufficient time for the operator to take load reduction measures in advance. To evaluate the system performance, the company compared the energy efficiency prediction results and operation optimization effects before and after the system deployment, and the specific data records are shown in Table 1.

[0168] Table 1 Performance Verification Report of the Energy Efficiency Prediction System of the Converter Station

[0169] ;

[0170] As can be seen from Table 1, after the system of the present invention is deployed, the error of the energy efficiency prediction of the converter station is significantly reduced, especially the performance improvement is most significant under complex operation environments and high dynamic loads. Specifically, the average prediction error of the top oil temperature drops from 7.3 °C of the traditional method to 2.1 °C, greatly improving the prediction accuracy and avoiding the equipment operation risks caused by inaccurate predictions. For the energy loss index, the prediction error drops from the original 15.4% to 3.8%, which provides more reliable data support for the operation department to optimize energy management and formulate energy-saving strategies. In addition, the system prediction delay is shortened from 45 seconds to 12 seconds, achieving a significant improvement in real-time performance.

[0171] The early warning ability of the system provides an important guarantee for the operation safety of the converter station. In a scenario of sudden load change, the system successfully predicts that the top oil temperature will reach the alarm threshold within 45 minutes, providing sufficient time for the operation and maintenance team to take countermeasures. Compared with the limitation of the traditional method that cannot provide effective early warnings, the present invention can provide at least a 30-minute early warning time window, greatly reducing the possibility of equipment failures caused by overheating. In practical applications, the number of equipment failures caused by abnormal oil temperature drops from 3 times per month before deployment to 0 times after deployment, significantly improving the reliability of the equipment and the stability of the power grid operation.

[0172] In addition, the energy efficiency prediction system of the present invention provides a comprehensive perspective for energy efficiency analysis through the efficient integration of multi-dimensional embedded vectors and standardized operation data. In actual operation, the system can not only accurately predict a single indicator, but also comprehensively analyze multi-dimensional data related to energy efficiency, providing more insightful decision-making support for the operation team. For example, through the correlation analysis between the top oil temperature and energy loss, the system can provide optimization suggestions for load distribution and equipment scheduling under different operating conditions. This prediction method based on multi-dimensional data fusion not only reduces the limitations of relying on manual experience in operation and maintenance, but also significantly improves the intelligent and refined level of energy efficiency management in the converter station.

[0173] The above are only the preferred specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, making equivalent substitutions or changes should be covered within the protection scope of the present invention.

Claims

1. A converter station energy efficiency prediction system based on deep learning, characterized in that: include: S1, data acquisition module, used to collect data of various devices in the converter station and generate an original data set containing the complete equipment operating status; S2, data preprocessing module, used to process the original data set, including data cleaning, missing value filling, normalization and time series feature extraction, to generate a cleaned standardized data set; S3, a graph construction module, which constructs a device network graph based on the topological relationship of the converter station devices, wherein the device network graph represents each device as a node and the electrical connection relationship between the devices as an edge; S4, graph neural network module, builds a graph neural network model based on the device network graph, extracts feature information of nodes and edges, and generates a multi-dimensional embedding vector; S5, the meta-learning module, based on the multi-dimensional embedding vector generated by the graph neural network module, uses a hierarchical optimization strategy to update the model parameters of the graph neural network model and generate a set of model parameters that can adapt to different input feature distributions; S6, energy efficiency prediction module, based on the model parameter set generated by the meta-learning module, combined with the standardized data set and multi-dimensional embedding vector, predicts the energy efficiency related indicators of the converter station; The S3 specifically includes: S31, topology data extraction unit, extracts the topology data of converter station equipment from the standardized data set, including node set and edge set , where the node set Indicates the various devices in the converter station, and the edge collection Indicates the electrical connection relationship between devices; S32, feature mapping unit, node set The device operation status data of each node in is mapped into the node feature vector ,in is the feature vector dimension, is the set of real numbers; S33, edge weight calculation unit, based on edge set The attribute information of each edge in the calculation corresponds to the weight value ,in Representation Node With Node The connection strength between: ； in, and Respectively represent nodes and nodes The characteristic vector of represents the attribute vector of the edge, is the weight matrix of node features, is the weight vector of edge attributes, where represents the dimension of the edge attribute vector, Represents the rectified linear unit activation function, which is used to map the output value to a non-negative range; S34, graph structure generation unit, based on the node set , edge set And the corresponding weight value Build a device network diagram ,in The weight matrix representing the node features, the matrix elements is the weight value; The S4 specifically includes: S41, input node feature unit, receive the device network graph generated by the graph construction module , where the node set Contains the node feature vector for each node , edge set A vector containing the attributes of each edge , the weight matrix Contains the weight of each edge ; S42, feature aggregation unit, aggregates the local neighborhood information of the node based on the graph convolution operation and updates the node feature vector : ； in, Representation Node The set of neighbor nodes of Representation Node and nodes The weight value between For the The weight matrix of the layer, For Node No. Layer feature vector, is the activation function; S43, feature update unit, updates the node feature vector layer by layer based on the multi-layer graph convolutional network to generate the node embedding vector ,in represents the dimension of the final embedding vector; S44, edge feature fusion unit, generates edge embedding vector based on node embedding vector and edge attribute vector ; S45, fusing the node embedding vector with the edge embedding vector to generate a multi-dimensional embedding vector; The S5 specifically includes: S51, meta-task construction unit, constructs a meta-task set consisting of multiple meta-tasks based on the multi-dimensional embedding vector generated by the graph neural network module ,in, Represents the total number of meta-tasks, each meta-task It includes input feature distribution and corresponding objective function, which is used to simulate the input data feature distribution and its target mapping relationship in different operating environments; S52, initial parameter optimization unit, based on meta-task set , by minimizing the total loss function of all meta-tasks on the initial model parameters To optimize: ； in, Representation model in meta-task The loss function on are the model parameters to be optimized; S53, task adaptation parameter updating unit, for the meta-task set Each meta-task in , using the gradient descent method from the initial model parameters Updated to task specific parameters , to adapt the meta-task The characteristic distribution of S54, global parameter update unit, updates task specific parameters Aggregation, combining the weights of each meta-task For initial parameters Optimize and generate global parameters that reflect cross-task ; S55, global parameters With task-specific parameters Combining to form a set of model parameters that can adapt to different input feature distributions ; The S6 specifically includes: S61, feature fusion unit, a set of model parameters generated based on the meta-learning module , combining the equipment operation features and multi-dimensional embedding vectors in the standardized dataset to construct the input feature matrix : ； in, For Node The equipment operating characteristics, is the corresponding multi-dimensional embedding vector, Represents feature concatenation operation; S62, energy efficiency prediction model selection unit, from the model parameter set Select the corresponding global parameters and task-specific parameters Building task-specific prediction models , select the optimal task model according to the equipment operation characteristics; S63, prediction calculation unit, based on the selected task-specific prediction model And the input feature matrix Calculate the indicators related to converter station energy efficiency: ； in, Indicates the energy efficiency related indicators of the converter station, represents a task-specific prediction model, are task-specific parameters.

2. According to claim 1, a converter station energy efficiency prediction system based on deep learning is characterized in that: The S44 specifically includes: S441. Extract the embedding vectors of the two end nodes associated with the edge from the node embedding vector and , extract the attribute vector corresponding to each edge from the attribute vector of the edge ; S442, embedding vector for nodes , The attribute vector of the edge Perform concatenation operation to generate joint feature vector : ； in, Represents vector concatenation operation, joint feature vector , is the dimension of the node embedding vector, is the dimension of the edge attribute vector; S443, joint feature vector Perform nonlinear transformation to generate edge embedding vector : ； in, is the weight matrix, is the bias vector, is the activation function, is the generated edge embedding vector, is the dimension of the edge embedding vector.

3. The converter station energy efficiency prediction system based on deep learning according to claim 1, characterized in that: The S45 specifically includes: S451, merge node embedding vector and edge embedding vector, according to each node The node embedding vector The edge embedding vector between the node and its neighbor , calculate the comprehensive embedding vector of the node : ; in, Representation Node The set of neighbor nodes of is the transformation matrix of the node embedding vector, is the transformation matrix of the edge embedding vector, Representation Node The number of neighbor nodes, is a nonlinear activation function; S452, fusion of global and local embedding features, comprehensive embedding vector for all nodes , construct the global embedding feature vector , and concatenated with the comprehensive embedding vector of each node to generate the final multidimensional embedding vector : ； in, , is the total number of node sets, represents the vector concatenation operation, is the final generated multi-dimensional embedding vector.

4. The converter station energy efficiency prediction system based on deep learning according to claim 1, characterized in that: The S53 specifically includes: S531. According to the meta-task set Each meta-task The loss function , calculate the initial model parameters Gradient : ； in, Meta-task The loss function is Represents the loss function with respect to the initial model parameters The gradient of S532. According to the meta-task Dynamically adjust the learning rate based on the feature distribution Optimize update step size: ； in, is the initial learning rate, is the learning rate decay factor, The number of iterations optimized for the meta-task; S533, based on learning rate and gradient , for the initial model parameters Update to generate task specific parameters : ； in, Representation and Meta-Tasks Task-specific parameters adapted to feature distribution.

5. The converter station energy efficiency prediction system based on deep learning according to claim 1, characterized in that: The S54 specifically includes: S541, based on meta-task set Each meta-task The feature distribution and task importance of , calculate the task weight , the task weights satisfy the following constraints: ; in, Represents the total number of meta-tasks, weight Assign tasks based on their specific importance indicators; S542. Gradients of all meta-tasks Perform weighted summation to generate global gradient : ； in, Indicates the task Task-specific parameters The corresponding gradient, is the task weight; S543, based on initial model parameters and the global gradient , update the global parameters and generate optimized global parameters : ； in, is the global optimization step size.

Images