Substation loop phasor detection method based on lightweight artificial intelligence and related equipment

Abstract

The application provides a substation loop phasor detection method based on light artificial intelligence and related equipment, and the method comprises the following steps: obtaining a loop phasor detection task of an intelligent substation; constructing a multi-target phasor detection model; obtaining historical loop detection data of the intelligent substation and extracting static topological features of the substation loop; in response to the loop phasor detection task and controlling a current output device and a voltage output device to pressurize and flow in the substation loop to simulate a load operation process of the substation loop; in the pressurizing and flowing process, collecting device data of all loop devices in the substation loop through an intelligent sensor network; inputting the device data into the multi-target phasor detection model, and outputting a loop phasor detection result of the intelligent substation through the multi-target phasor detection model. The application has the effect of intelligently and accurately completing the substation loop phasor detection process.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of circuit phasor detection, and particularly relates to a substation circuit phasor detection method based on light artificial intelligence and related equipment. BACKGROUND

[0002] Due to the increasing demand for electricity and the diversification of energy structure, the power grid needs to support various changes in power supply and demand, which poses great challenges to the management and monitoring of substations. In this context, traditional substation detection systems usually rely on centralized management of central control stations, which is suitable for the past single power supply mode, but with the increasing complexity of power generation and electricity distribution, this mode is obviously difficult to meet the flexibility requirements of today's power system. For example, the instability of new energy such as wind and solar energy and the load fluctuation brought by distributed power generation have greatly increased the complexity of substation operation and maintenance.

[0003] The existing substation circuit detection system usually has a complex architecture, and the data processing relies on a remote large-scale data center. Such architecture design requires a large amount of data transmission between multiple links, and the reliability and efficiency of data transmission are affected by many factors such as network conditions, information transmission protocols and device performance. Therefore, in the traditional method, the process from data collection to analysis and then to result output may reduce the accuracy of detection due to delay and error accumulation. In addition, due to the diversity of devices in the substation, the state of each device and their mutual influence are complex, and the static and centralized processing mode of the traditional method is difficult to capture the real state and complex state changes of the system in real time. SUMMARY

[0004] The application provides a substation circuit phasor detection method based on light artificial intelligence and related equipment to solve the problem of poor accuracy of substation circuit detection results.

[0005] In a first aspect, the application provides a substation circuit phasor detection method based on light artificial intelligence, applied to an intelligent substation, wherein the intelligent substation is deployed with an edge computing device, the substation circuit in the intelligent substation is deployed with an intelligent sensor network, the intelligent sensor network is in communication connection with the edge computing device, and the primary circuit side of the substation circuit is further deployed with a current output device and a voltage output device.

[0006] The method comprises the following steps:

[0007] Obtaining a circuit phasor detection task of the intelligent substation;

[0008] The multi-target phasor detection model is deployed in the edge computing device, and the multi-target phasor detection model comprises a feature sharing layer and a multi-target detection task layer;

[0009] Historical loop detection data of the intelligent substation is acquired, and static topology features of the substation loop are extracted;

[0010] The feature sharing layer is trained by using the static topology features, and the multi-target detection task layer is trained by using the historical loop detection data;

[0011] The current output device and the voltage output device are controlled to pass current and pressurize in the substation loop in response to the loop phasor detection task, so as to simulate a load running process of the substation loop;

[0012] During the process of passing current and pressurizing, device data of all loop devices in the substation loop is acquired by the intelligent sensor network;

[0013] The device data is input into the multi-target phasor detection model, and a loop phasor detection result of the intelligent substation is output by the multi-target phasor detection model.

[0014] Optionally, the loop phasor detection task is any one or more of loop device phase sequence correctness detection, loop device polarity correctness detection, loop device variable ratio correctness detection, and cable connection performance detection.

[0015] Optionally, the multi-target phasor detection model is constructed based on a convolutional neural network model and according to a type of the loop phasor detection task, and comprises the following steps:

[0016] A feature sharing layer and a plurality of initial detection subnetworks are constructed based on a convolutional neural network model, the number of the initial detection subnetworks is the same as the number of types of the loop phasor detection task, and all the initial detection subnetworks have independent output layers;

[0017] The layer structure and model parameters of each initial detection subnetwork are adjusted according to the type of the loop phasor detection task, to obtain a plurality of basic detection subnetworks;

[0018] The input layers of the plurality of basic detection subnetworks are connected with the output layers of the feature sharing layer respectively, to obtain a multi-target phasor detection model, and all the basic detection subnetworks are arranged in parallel to form a multi-target detection task layer.

[0019] Optionally, the obtaining historical loop detection data of the smart substation and extracting static topology features of the substation loop comprises the following steps:

[0020] Obtaining historical loop detection data of the smart substation and device parameters of all loop devices in the substation loop;

[0021] Taking the loop devices as device nodes and the electrical connection relationship between all the loop devices in the substation loop as device node edges to construct a substation electrical connection graph, and giving all device nodes node attributes based on the device parameters;

[0022] Identifying all electrical loops in the substation electrical connection graph by using a recursive-based graph traversal algorithm;

[0023] Marking bus structure features in all the electrical loops through node aggregation analysis of all the electrical loops;

[0024] Integrating all the device parameters and all the electrical loops into static topology features of the substation loop.

[0025] Optionally, the training of the feature sharing layer using the static topology features and the training of the multi-target detection task layer using the historical loop detection data comprise the following steps:

[0026] Taking the feature sharing layer as an encoder, and mapping the static topology features to a low-dimensional space through the encoder;

[0027] Reconstructing the static topology features in the low-dimensional space through a preset decoder, and completing the training of the feature sharing layer when the reconstruction error reaches a minimum value;

[0028] Defining a subnetwork loss function for each of the basic monitoring subnetworks, and taking the weighted sum of all the subnetwork loss functions as a total loss function of the multi-target detection task layer;

[0029] Classifying the historical loop detection data according to all task types of the loop phasor detection task and inputting the classified data into all the basic monitoring subnetworks, training all the basic monitoring subnetworks by using a small batch gradient descent method, and minimizing the total loss function by using a gradient descent type algorithm to complete the training of the multi-target detection task layer.

[0030] Optionally, the inputting of the device data into the multi-target phasor detection model and the outputting of loop phasor detection results of the smart substation by the multi-target phasor detection model comprise the following steps:

[0031] Obtaining a device resource state of the edge computing device;

[0032] determining a current task set and a task priority of all tasks in the current task set by analyzing the loop phasor detection task;

[0033] constructing a computing resource allocation model of the multi-target phasor detection model in combination with the device resource state, the current task set and the task priority;

[0034] generating a reward function of the computing resource allocation model in the shortest time to complete all tasks in the current task set and the optimal resource utilization of the edge computing device;

[0035] solving the computing resource allocation model based on the reward function and using a deep Q network algorithm, and obtaining an optimal computing resource allocation strategy for the edge computing device to execute all tasks in the current task set when the reward function reaches a maximum value;

[0036] allocating the remaining device resources of the edge computing device to the feature sharing layer and a plurality of target basic detection sub-networks corresponding to the current task set according to the optimal computing resource allocation strategy;

[0037] outputting the loop phasor detection result of the smart substation through all the target basic detection sub-networks.

[0038] Optionally, the step of constructing the computing resource allocation model of the multi-target phasor detection model in combination with the device resource state, the current task set and the task priority comprises the following steps:

[0039] constructing a target state space based on a Markov decision model and in combination with the device resource state, the current task set and the task priority;

[0040] constructing a target action space in combination with a selection step of selecting a next task in the current task set and an initial resource allocation strategy preset by the edge computing device;

[0041] constructing the computing resource allocation model of the multi-target phasor detection model in combination with the target state space, the target action space and a preset state transition probability function.

[0042] Optionally, the expression formula of the optimal computing resource allocation strategy is as follows:

[0043]

[0044] In the formula, π * represents the optimal computing resource allocation strategy, argmax πdenotes a policy that selects a computational resource allocation strategy that maximizes the reward function, E denotes an expected cumulative reward value based on the reward function, t denotes a time step t denotes a discount factor, both alpha and beta denote weight coefficients, T(s t denotes a state in the target state space t denotes a state in the target state space t denotes a reward value in a task completion time dimension when an action a t denotes a state in the target state space t denotes a state in the target state space t denotes a state in the target state space t denotes a state in the target state space t denotes a penalty value in a resource utilization dimension.

[0045] In a second aspect, the present application further provides a circuit phasor detection device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor implements the light artificial intelligence-based substation circuit phasor detection method according to the first aspect when executing the computer program.

[0046] In a third aspect, the present application further provides a light artificial intelligence-based substation circuit phasor detection system, comprising:

[0047] The circuit phasor detection device according to the second aspect;

[0048] An edge computing device, deployed in an intelligent substation, configured to carry a multi-target phasor detection model constructed based on a convolutional neural network model;

[0049] An intelligent sensor network, deployed in a substation circuit of the intelligent substation and communicatively connected to the edge computing device, configured to collect device data of all circuit devices in the substation circuit;

[0050] A current output device, deployed on a primary circuit side of the substation circuit, configured to inject a current into the substation circuit when performing a circuit phasor detection task;

[0051] A voltage output device, deployed on a primary circuit side of the substation circuit, configured to inject a voltage into the substation circuit when performing a circuit phasor detection task.

[0052] The present application has the following beneficial effects:

[0053] Compared with the prior art, the present application has significant beneficial effects, especially in improving the real-time performance, accuracy and flexibility of the substation loop phasor detection, showing unique technical advantages. First, by deploying edge computing devices in the smart substation, the present application introduces lightweight artificial intelligence technologies such as convolutional neural networks into the phasor detection system, which can effectively reduce the delay in data processing and transmission. The adjustment of this architecture enables the detection task to be localized in the substation, greatly improving the real-time performance of the detection and enabling the actual operating state of the substation loop to be reflected more quickly. Secondly, the present application builds a multi-target phasor detection model and trains it with intelligent sensor networks and topological features, making it have higher detection accuracy and response capability. The introduction of the multi-target detection model not only enables the identification and processing of multiple types of detection tasks in complex power systems, but also enables adaptive learning and adjustment according to the dynamically changing operating environment, thereby improving the accuracy of the detection results. In addition, the present application simulates the load conditions of the substation loop, which can more realistically reflect the state and behavior of the equipment under various operating conditions, providing more reliable data support and decision-making basis for the operation and maintenance of the smart substation. In summary, the present application changes the centralized mode of traditional power system detection to a distributed stream processing mode by means of advanced artificial intelligence and edge computing technology, not only improving the response speed and detection accuracy of the system, but also improving the adaptability to complex and variable power environments, laying a solid foundation for the future management and monitoring of smart substations. BRIEF DESCRIPTION OF DRAWINGS

[0054] Figure 1 A flowchart of the substation loop phasor detection method based on lightweight artificial intelligence in one of the embodiments of the present application.

[0055] Figure 2 A model structure diagram of the multi-target phasor detection model in one of the embodiments of the present application.

[0056] Figure 3 A system structure diagram of the substation loop phasor detection system based on lightweight artificial intelligence in one of the embodiments of the present application. DETAILED DESCRIPTION

[0057] The technical solutions in the embodiments of the present application will be described in detail below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art belong to the scope of protection of the present application.

[0058] The terms "first", "second", etc. in the specification and claims of the present application are used to distinguish similar objects, and are not used to describe a particular order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments of the present application can be implemented in an order other than those illustrated or described herein, and the objects distinguished by "first", "second", etc. are generally a class, not limited to the number of objects, for example, the first object can be one or more. In addition, "and / or" in the specification and claims indicates at least one of the connected objects, and the character " / ", generally indicates that the objects before and after are in a "or" relationship.

[0059] The application discloses a substation loop phasor detection method based on light artificial intelligence technology, which is applied to intelligent substations, and provides support by deploying edge computing devices in the intelligent substations. Advanced intelligent sensor networks are installed in the substation loops in the intelligent substations, the sensors are distributed on important nodes of the substations, and operation data of various power equipment can be collected in real time. The sensor network is connected with the edge computing device through high-speed communication, so that the data is quickly transmitted and processed. The design breaks the excessive dependence of the traditional detection system on central data processing, processes the data locally in the substations, significantly reduces the delay in the data transmission process, and improves the data processing efficiency.

[0060] Current output devices and voltage output devices are specially deployed on the main loop side of the substation loop. The introduction of these devices can perform voltage boosting and current flowing operations in the loop to simulate the running state of the substation under actual load conditions. The simulation function enables the detection system to better understand and evaluate the dynamic response of the substation loop under various operating conditions, and provides more real and reliable data support for monitoring and fault diagnosis of the substation. The current output device can be a three-phase digital intelligent current source, and the parameters are as follows:

[0061] (1) three-phase output current: 0-300A; (2) output current accuracy: 1% (30-300A); (3) maximum phase current output power: 2400VA; (4) phase output range: 0-360°; (5) phase accuracy: 0.5 degrees; (6) frequency output range: 50Hz; (7) frequency accuracy: 0.01Hz.

[0062] The voltage output device can be a three-phase digital intelligent high-voltage source, and the parameters are as follows:

[0063] (1) Three-phase output voltage: 0~6000V (phase voltage), 0~10400V (line voltage); (2) Output voltage accuracy: 1% (500~6000V); (3) Phase voltage output power: 300VA; (4) Phase output range: 0~360°; (5) Phase accuracy: 0.5 degrees; (6) Frequency output range: 50Hz; (7) Frequency accuracy: 0.01Hz.

[0064] Figure 1 This is a flowchart illustrating a substation circuit phasor detection method based on lightweight artificial intelligence in one embodiment. It should be understood that, although... Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 1 At least some steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed alternately or in turn with other steps or at least a portion of the sub-steps or stages of other steps. For example Figure 1 As shown, the substation circuit phasor detection method based on lightweight artificial intelligence disclosed in this invention specifically includes the following steps:

[0065] S101. Obtain the loop phasor detection task of the intelligent substation.

[0066] The loop phasor detection task is any one or more of loop device phase sequence correctness detection, loop device polarity correctness detection, loop device ratio correctness detection, and cable connection performance detection. Such a wide range of detection aims to ensure that each loop component of the substation is in the best working state, thereby ensuring the normal operation of the power grid. Phase sequence correctness detection is used to confirm whether the order of each phase signal meets the predetermined standard, because phase sequence errors can cause the device to fail to start normally or run unstably. Polarity correctness detection ensures that parameters such as current and voltage are measured and transmitted in the correct direction, and deviations can cause system failure. Ratio correctness detection evaluates whether the transformer ratio operates according to the design parameters, which is particularly critical for the effective operation of the transformer. Cable connection performance detection checks the connection quality in the line, thereby ensuring the safety of electrical transmission. In implementation, the smart substation automatically captures any one or more of the above detection tasks through the deployed sensors and edge computing devices. This is achieved by configuring the system intelligent control strategy to allocate and prioritize multiple tasks. The sensors monitor various parameters in the loop in real time and send them to the edge device for processing. Without human intervention, the intelligent system automatically defines the most urgent detection task in combination with historical data and actual conditions, ensuring continuous and stable power transmission of the substation.

[0067] S102. Based on the convolutional neural network model, a multi-target phasor detection model is constructed according to the type of the loop phasor detection task, and the multi-target phasor detection model is deployed in the edge computing device.

[0068] The multi-target phasor detection model includes a feature sharing layer and a multi-target detection task layer. Building an efficient multi-target phasor detection model requires the introduction of convolutional neural network (CNN) technology. CNN is widely used in computer vision and signal processing due to its powerful feature extraction capability. By using CNN to build a multi-target phasor detection model, the detection accuracy can be significantly improved. First, according to different loop phasor detection tasks, the model needs to have the ability to handle multiple detection tasks, such as the aforementioned phase sequence, polarity, ratio, and cable connection functions. The core of the multi-task model is its feature sharing layer, which fully utilizes the common features required in phasor detection tasks to extract multi-level features from input data through deep learning. On this basis, the feature sharing layer can reduce feature redundancy and improve detection speed and efficiency. Next, the multi-target detection task layer optimizes the output of each detection task with different task loss functions based on the features extracted by the feature sharing layer. This multi-task collaborative optimization approach takes into account the possible interdependence between detection tasks. Through the training of various historical task data and specific types of current and voltage data obtained in advance, the multi-target phasor detection model is deployed in real-time on edge computing devices. During practical use, the model continuously processes real-time data and adjusts itself to adapt to actual detection conditions and environmental changes. Therefore, efficient and accurate substation loop multi-task phasor detection can be achieved.

[0069] S103. Obtain historical loop detection data of the intelligent substation and extract static topology features of the substation loop.

[0070] Before conducting intelligent detection, historical loop detection data of the substation must be obtained. These data provide the operating conditions of the substation loop at different time nodes, including electrical parameters under various load conditions. These data are accumulated over a long period of time through sensor networks and edge computing devices, and truly reflect the operating mode and potential problems of the substation. After obtaining these data, the next step is to extract the static topology features of the substation loop. These features include the connection mode of the loop, the rated parameters of the transformer, the layout of the switch, and the load distribution condition, etc. These topology information runs through the entire physical connection structure and is crucial for system analysis. Through this technical solution, the algorithm directly extracts the design parameters from the data, enabling the model to more accurately depict the respective positions of each part of the loop and their relationships. Combined with historical data and topology features, the system organizes and encodes the obtained data, laying a foundation for the optimization of the feature sharing and task detection layers. Analyzing and extracting these features not only provides support for improving the efficiency of model training, but also plays a clear guiding role in subsequent intelligent detection.

[0071] S104. Train the feature sharing layer using the static topology features and train the multi-object detection task layer using the historical loop detection data.

[0072] Where, after the extraction of topology features and historical data, the next step is how to effectively apply them to the training of the model. This process contains two main parts: the training of the feature sharing layer and the training of the multi-object detection task layer. In the training process of the feature sharing layer, with the help of the previously extracted static topology features, the adaptive learning algorithm in deep learning is used to iteratively optimize the network parameters. Since these features provide static information of the substation loop, which is closely related to the actual power system running state, therefore, in the training, the feature sharing layer uses these information to capture the basic features shared in each task, which not only improves the performance of the model, but also avoids repeated calculation. After the training of the feature sharing layer is completed, the training of the multi-object detection task layer in the model is particularly important. Its training is mainly based on historical loop detection data, using labeled phase sequence, polarity, and transformer ratio label data, to optimize the loss function of the task layer specifically, in order to enhance its recognition ability and accuracy in processing specific tasks. The whole training process is iterated, and the task layer is updated constantly to adapt to the needs of multi-task processing and maximize the collaborative efficiency between tasks. This training method makes the model more sensitive to the discovery of potential problems in the substation loop, and realizes the rapid response and positioning of the abnormality.

[0073] S105. In response to the loop phasor detection task and control the current output device and voltage output device to pressurize and flow in the substation loop to simulate the process of the substation loop under load operation.

[0074] Where, by controlling the current output device and voltage output device to pressurize and flow in the loop. This process simulates the running state of the substation under real load conditions. First, the system configures the output parameters of the current and voltage devices according to the preset load curve or power demand at different times, and gradually pressurizes the power load to the target point through accurate control. This not only allows the detection system to evaluate the behavior parameters of the loop under overload or underload conditions, but also provides a risk-free environment to identify potential hidden information. Through load testing under stable and dynamic conditions, the system can verify the correlation between its performance in actual working conditions and model prediction, so as to evaluate or adjust the parameter settings of the detection model.

[0075] S106. During the pressurized flow process, collect device data of all loop devices in the substation loop through the intelligent sensor network.

[0076] The loop equipment includes transformers, circuit breakers, disconnectors, current transformers, voltage transformers, arresters, busbars, cables and overhead lines, protection devices (relay protection devices), etc. The sensor can capture real-time current, voltage, phase, frequency and temperature and other key parameters. With the help of high-sensitivity and high-precision measurement technology, a large amount of real-time data collected is transmitted to the edge computing device. This connection provides an accurate and timely information source for subsequent intelligent analysis, ensuring the continuity and consistency of the data from collection to processing. In order to ensure the effectiveness of the data, the sensor network converts the data stream into a structured power grid feature set, and through the priority processing mechanism and data integrity protocol, avoids information loss or error accumulation caused by the transmission process. All collected data is integrated and converted into effective data signals for model analysis to achieve comprehensive and accurate monitoring of the loop equipment state of the substation.

[0077] S107. Input the device data into the multi-target phasor detection model, and output the loop phasor detection result of the smart substation through the multi-target phasor detection model.

[0078] After the multi-target phasor detection model training and data collection are completed, all collected device data is input into the model for real-time calculation and analysis. According to the parameters obtained by learning and training, the multi-target phasor detection model outputs a series of phasor detection results for the loop of the smart substation through its feature sharing and multi-task optimization mechanism. In specific implementation, the model will comprehensively evaluate different categories of data, such as whether the phase sequence is normal, whether the polarity is accurate, whether the transformation ratio is consistent with the expected value, and whether the connection performance of the cable meets the standard. After rigorous algorithm analysis, these results are presented in various forms of tables or graphs, providing a humanized interface interaction, so that technical personnel can intuitively understand and make decisions. The output detection results usually come with problem diagnosis prompts, helping technical personnel quickly identify and solve potential problems, ensuring the normal operation of the substation and continuously optimizing the reliability and safety of the power system. At the same time, the historical archive of the detection results will become a strong basis for subsequent improvement and analysis, supporting further optimization of power grid management strategies.

[0079] In one embodiment, based on a convolutional neural network model, a multi-target phasor detection model is constructed according to the types of tasks in the loop phasor detection task, including the following steps:

[0080] A feature sharing layer and a plurality of initial detection sub-networks are constructed based on the convolutional neural network model. The number of initial detection sub-networks is the same as the number of types of tasks in the loop phasor detection task, and all initial detection sub-networks have independent output layers;

[0081] Adjust the layer structure and model parameters of each initial detection subnetwork according to the task type in the circuit phasor detection task to obtain a plurality of basic detection subnetworks;

[0082] Connect the input layer in the plurality of basic detection subnetworks with the output layer in the feature sharing layer to obtain a multi-target phasor detection model.

[0083] In this embodiment, with reference to Figure 2 In this process, first, a feature sharing layer needs to be created, which is designed to extract shared low-level features from input data. Convolutional neural networks are widely used due to their powerful feature extraction capabilities, especially in computer vision and signal processing. The feature sharing layer can be considered as a set of convolutional layers and pooling layers combined together to effectively capture local characteristics and cross-feature correlations in input data. Convolutional layers scan input data through a set of convolutional kernels (or filters) to extract various detailed features, while pooling layers perform downsampling operations on input feature maps to compress their size and reduce computational burden. The feature sharing layer uses a combination of multiple convolutional and pooling operations to adapt to a wide range of phasor detection task scenarios, providing rich basic features for subsequent detection subnetworks. The key to this step is to design appropriate numbers of convolutional kernels, sizes, and numbers of layers to maximize feature extraction versatility in various detection tasks.

[0084] After the feature sharing layer is constructed, the next step is to design the initial detection subnetworks. The initial detection subnetworks are multiple independent basic networks, each corresponding to a specific detection task type. The main role of these subnetworks is to receive the basic features extracted by the feature sharing layer to further optimize specific types of detection tasks. The initial detection subnetworks include a set of specific convolutional layers or fully connected layers, and the structure composed of these layers and the parameters therein need to be adjusted according to the characteristics of the target detection task. Each detection subnetwork contains its own output layer, which is used to pass the features processed by the specific layers to the final output of the model. This ensures that even under the same input, each subnetwork can generate output results independently according to its specific training goal to adapt to different types of detection tasks, such as phase sequence detection, polarity check, variable ratio verification, and cable connection performance testing, etc. The implementation effect of this step is to lay the basic framework for subsequent network adjustment and optimization through the diversification of the initial network structure, and to ensure that each subnetwork has an independent learning channel to adjust itself.

[0085] Based on the construction of the initial detection sub-network, the next step is to individualize the adjustment of these initial sub-networks for different detection task types. This process involves optimizing the layer structure and model parameters of the detection sub-network one by one to meet the specific needs of different tasks. The objects of adjustment include the depth of the layer, the size of the convolution kernel, the selection of the pooling strategy, and the configuration of the activation function and the loss function, etc. Different task types may require different feature extraction depth and pattern recognition ability, so detailed parameter adjustment needs to be performed on each sub-network. For example, for phase sequence detection, attention is paid to the relative positional relationship of the three-phase signals, so the network may focus on the capture of periodic signal features and the amplification of phase differences. For polarity detection, the network needs to have stronger direction sensitivity to ensure the correctness of the current detection signal direction. After these individualized adjustments, each initial detection sub-network is transformed into a highly optimized basic detection sub-network for a specific task. Through detailed optimization and parameter adjustment, the construction of the basic detection sub-network ensures efficient recognition and accuracy improvement of multiple task types, and can accurately respond to the complex signal processing needs in actual application scenarios.

[0086] After the construction of the basic detection sub-network is completed, the input layers of these networks need to be connected with the output layers of the feature sharing layer respectively to form a whole multi-target phasor detection model. The feature sharing layer has extracted rich shared features, while the basic detection sub-network has the processing ability for specific detection tasks. By combining the basic detection sub-network with the feature sharing layer, the goal of simultaneous task processing can be achieved. The structure of the multi-target phasor detection model is equivalent to stacking each optimized basic detection sub-network in parallel on top of the feature sharing layer, and these sub-networks together constitute the multi-target detection task layer. The structure design aims to realize the simultaneous detection and evaluation of multiple operating parameters in the power grid through the unified feature extraction of the sharing layer and the specific task recognition of the basic sub-network. The connected multi-target phasor detection model not only can greatly improve the recognition efficiency by using shared features, but also can form a collaborative processing relationship between multiple tasks, speed up the processing speed of detection data through parallel computing, and provide a comprehensive phasor detection report at the task output layer. The implementation effect of the whole model is to greatly improve the response speed and accuracy of the detection task, making real-time monitoring and fault diagnosis in complex power grid environment more efficient and reliable.

[0087] In one embodiment, obtaining historical loop detection data of the smart substation and extracting static topology features of the substation loop includes the following steps:

[0088] Obtaining historical loop detection data of the smart substation and device parameters of all loop devices in the substation loop;

[0089] The circuit equipment is taken as the equipment node, and the electrical connection relationship between all the circuit equipment in the substation circuit is taken as the equipment node edge to construct the substation electrical connection graph, and all the equipment nodes are given node attributes based on the equipment parameters;

[0090] All the electrical circuits in the substation electrical connection graph are identified by using a recursive-based graph traversal algorithm;

[0091] The bus structure features in all the electrical circuits are marked through the node aggregation analysis of all the electrical circuits;

[0092] All the equipment parameters and all the electrical circuits are integrated as the static topological features of the substation circuit.

[0093] In the embodiment, first, it is crucial to collect the historical circuit detection data of the smart substation. These data can provide information about the electrical system performance and operation history, which helps to analyze the health condition and failure mode of the power grid. The historical circuit detection data includes data records of electrical parameters such as voltage, current, power, frequency, etc., which are accumulated during long-term operation and can help analysts identify potential abnormalities or instabilities in the electrical system. At the same time, the equipment parameters of all the circuit equipment in the substation circuit need to be obtained, which usually include equipment type, rated capacity, operating voltage, rated current, short-circuit current capacity, manufacturer information, installation date, operating status, etc. The equipment parameters provide detailed static information, making the role and characteristics of the equipment in the electrical connection graph more explicit. The collection of these data can be achieved through various channels such as the monitoring system of the substation, maintenance logs, operation records, etc. By comprehensively obtaining the static parameters of the equipment and the dynamic detection data of the circuit, detailed and accurate basic information is provided for subsequent analysis and model construction. Finally, the collected and sorted data will provide effective support for establishing an accurate electrical connection graph and analyzing the behavior of the electrical circuit.

[0094] The electrical connection graph of a substation is constructed by taking each circuit device as a node and the electrical connection relationship between all circuit devices in the substation circuit as the edges of the device nodes. During this process, the electrical equipment and its interconnection relationship are transformed into a graph theory model. First, each circuit device is considered as a node, including transformers, circuit breakers, transformers, etc., which play different roles in the electrical system. The definition of node attributes is based on device parameters, which may include important information such as device specifications and location. The connection relationship of device nodes is represented as edges in the graph, and the construction of edges reflects the actual electrical connection between devices in the circuit, such as cable lines, wires, etc. The construction of this electrical connection graph needs to make full use of the device parameters and historical detection data collected in the previous step to ensure the integrity and accuracy of the graph. In order to realize this process, the nodes and edges in graph theory are usually used to build an abstract model of the entire electrical system. In addition, this connection graph is not only a visual structural representation, but also needs to be continuously updated and maintained through a digital information system, and the structure of the graph also needs to be dynamically adjusted with the addition or modification of devices.

[0095] Using a recursive graph traversal algorithm to identify all electrical circuits in the substation electrical connection graph is an important step in analyzing the complex relationships of the electrical system. Recursive algorithm is a computational method suitable for solving similar structure problems, which divides and solves sub-problems to achieve the final solution of the whole problem. For the electrical connection graph, recursive graph traversal mainly refers to starting from a node in the graph, following the connection path of the edge to visit each node in turn, and returning to the starting node to form a complete electrical circuit. Graph traversal algorithms mainly include depth-first search (DFS) and breadth-first search (BFS), etc., and the application of each method depends on the focus of the demand, such as finding the shortest path or complete circuit. During the traversal process, the algorithm visits each node recursively and tracks the path of nodes that have been visited to avoid repeated visits and fall into a dead loop. To identify electrical circuits, each possible path needs to be checked to confirm whether they constitute a closed loop. The efficiency of the recursive traversal algorithm lies in its ability to adaptively handle a large number of nodes and complex connection structures, which is particularly suitable for substation electrical graphs with dense connections and numerous nodes.

[0096] The node aggregation analysis of all electrical circuits and the marking of busbar structure features in all electrical circuits are important steps for fine identification of the topological characteristics of the electrical network. The node aggregation analysis involves in-depth analysis of the identified electrical circuits, focusing on identifying the aggregation (i.e. busbar) between key nodes and its importance in the entire cycle. The busbar is a key location where multiple devices in the electrical circuit converge, and is usually used to converge and distribute power, so its role in the electrical system is crucial. In the aggregation analysis, the nodes with high connectivity are identified as potential busbar structures by examining the connectivity of each node in the circuit. Connectivity refers to the number of connections a node has with other nodes, and a higher connectivity means that the node plays a role in converging in multiple circuits. In order to mark these busbar structures, the connection density and distribution characteristics between nodes need to be analyzed, and the physical properties and functions of the devices need to be identified. Once the key busbar structures are marked, they can be visualized in the form of topological views and data tables, presenting the important connection locations in the electrical circuit. Finally, all device parameters and all electrical circuits are integrated into the static topological characteristics of the substation circuit. The determination of static topological characteristics is based on the results of the previous steps of identification and analysis, and integrates device parameters, circuit structure and busbar characteristics in a systematic way. This integration process is not just a mechanical superposition of elements, but also takes into account the functional coupling and influence between devices, forming a logically connected and practically valuable overall model.

[0097] In one embodiment, the training of the feature sharing layer using the static topological characteristics and the training of the multi-objective detection task layer using the historical circuit detection data include the following steps:

[0098] The feature sharing layer is used as an encoder, and the static topological characteristics are mapped to a low-dimensional space through the encoder;

[0099] The static topological characteristics are reconstructed in the low-dimensional space through a pre-set decoder, and the training of the feature sharing layer is completed when the reconstruction error reaches a minimum value;

[0100] Define a sub-network loss function for each basic monitoring sub-network, and use the weighted sum of all sub-network loss functions as the total loss function of the multi-objective detection task layer;

[0101] According to all task types of the circuit phasor detection task, the historical circuit detection data is classified and input into all basic monitoring sub-networks, and the small batch gradient descent method is used to train all basic monitoring sub-networks, and the gradient descent algorithm is used to minimize the total loss function, and the training of the multi-objective detection task layer is completed.

[0102] In this embodiment, the encoder is a neural network model whose task is to compress the input data into a lower-dimensional latent space through a series of nonlinear transformations. Static topology features refer to the structural and connectivity information in the power system, which are usually high-dimensional and complex. Through the processing of the encoder, these high-dimensional data are transformed into a low-dimensional form that can effectively express their essential information. To achieve this, a suitable encoder network structure is first constructed, which includes an input layer, several hidden layers, and an output layer. The input layer is responsible for receiving the static topology feature data, the hidden layers extract features through convolution, activation, and pooling operations, and the output layer maps the data to a lower-dimensional space. At this time, by appropriately selecting the depth and width of the feature sharing layer, the original information of the input can be maximally preserved while the data is compressed. In addition, parameter optimization during the training process can further improve the encoding effect. After the features are mapped to the low-dimensional space, the computational efficiency of the data is significantly improved, laying a foundation for subsequent analysis and processing. In addition, low-dimensional feature representation also reduces the burden of data storage and transmission, improving the response speed of the model.

[0103] The reconstruction of static topology features in the low-dimensional space through the preset decoder is completed when the reconstruction error reaches a minimum value, which describes the reverse operation process of the autoencoder. The decoder is usually a symmetric neural network structure with the encoder, and its role is to map the features compressed into the low-dimensional space back to the high-dimensional space, i.e., the original data space. The design of the decoder generally follows the same principles: its input is the output of the encoder, and through training processes such as backpropagation and gradient descent, the weights in the network are gradually adjusted so that the output of the decoder is as close to the original input data as possible. The evaluation criterion for training is the reconstruction error, which is usually measured by a certain loss function, such as mean square error (MSE), which calculates the difference between the input data and the decoder reconstruction output. When the value of the loss function decreases to a certain set threshold or reaches a local minimum, the training process is considered to have converged. At this time, the feature sharing layer has been fully trained, and it can ensure that the low-dimensional features output by the encoder can be accurately reconstructed into the original form in the decoder. This adaptive training mechanism enables the feature sharing layer to retain key information in a limited data representation space while accurately reconstructing the original data when necessary.

[0104] A sub-network loss function is defined for each base monitoring sub-network, and the weighted sum of all sub-network loss functions is taken as the total loss function of the multi-objective detection task layer, i.e. the target optimization is implemented in the multi-task deep learning framework, which emphasizes the importance of multi-objective optimization. Each base monitoring sub-network represents a different monitoring task, such as voltage monitoring, frequency monitoring, etc., each with its own learning goal and scope, corresponding to an independent task function. Each task function is embodied as a loss function that measures the difference between the predicted output and the true value of the corresponding task. These sub-network loss functions can be mean square error, cross entropy, etc., selected according to the specific task requirements. Then, the loss functions of all sub-networks are weighted to reflect the importance and priority of different tasks, forming a comprehensive total loss function. The weighted sum is usually pre-set with a weight coefficient, or dynamically adjusted according to the real-time performance of the task, to better coordinate the learning between tasks. Through the optimization of the total loss function, the base monitoring sub-networks can work together in the overall network architecture to achieve unified detection and optimization of multiple objectives. This multi-objective optimization strategy helps to improve the generalization ability of the entire model, ensuring that the performance is effectively balanced when dealing with different types of tasks, and thus obtaining more robust prediction results.

[0105] The historical loop detection data is classified according to all task types of the loop phasor detection task and input into all base monitoring sub-networks, the small batch gradient descent method is used to train all base monitoring sub-networks, and the gradient descent algorithm is used to minimize the total loss function, completing the training of the multi-objective detection task layer, which details the training mechanism and optimization strategy of deep learning for power system detection. First, the historical loop detection data involves a large number of records, including current, voltage, phase angle, etc. The data is classified and sorted according to the different needs of the phasor detection task and distributed to the corresponding sub-network modules. Through this targeted input classification, it is ensured that each sub-network can focus on its specific task data, thereby improving the learning efficiency. Then, in the model training process, the small batch gradient descent method is used for parameter update. Small batch gradient descent is an algorithm that balances computational efficiency and model stability, compared to gradient descent of the entire data, it can reduce computational overhead while preventing falling into local minimum. For all base monitoring sub-networks, the goal is to reduce the prediction error of each sub-network by adjusting the weights, thereby gradually reducing the value of the total loss function. In order to achieve this, advanced gradient descent algorithms such as Adam, RMSProp, etc. are used to further improve the optimization efficiency and convergence speed. Finally, after multiple iterations, when the total loss function reaches the preset minimization target, it indicates that the sub-networks of the model have been successfully trained, providing a reliable foundation for subsequent practical applications.

[0106] In one embodiment, the device data is input into the multi-target phasor detection model, and the loop phasor detection result of the smart substation is output by the multi-target phasor detection model, including the following steps:

[0107] Obtaining the device resource state of the edge computing device;

[0108] Determining the current task set and the task priority of all tasks in the current task set by analyzing the loop phasor detection task;

[0109] Combining the device resource state, the current task set and the task priority to construct a computing resource allocation model of the multi-target phasor detection model;

[0110] Generating a reward function of the computing resource allocation model in the shortest time to complete all tasks in the current task set and the optimal resource utilization of the edge computing device;

[0111] Solving the computing resource allocation model based on the reward function and using a deep Q network algorithm, and obtaining the optimal computing resource allocation strategy of the edge computing device for executing all tasks in the current task set when the reward function reaches the maximum value;

[0112] Allocating the remaining device resources of the edge computing device to the feature sharing layer and the plurality of target basic detection sub-networks corresponding to the current task set according to the optimal computing resource allocation strategy;

[0113] Outputting the loop phasor detection result of the smart substation by all target basic detection sub-networks.

[0114] In this embodiment, the state of the resource generally includes computing capability (such as the use of CPU and GPU), storage remaining space, network bandwidth, and memory usage, etc. This information can be obtained by inquiring the operating system of the device, or by a specific monitoring tool. Real-time monitoring and collection of these data can adopt a proxy-based monitoring scheme, which can deploy a lightweight proxy process on the device, and the process is responsible for collecting and reporting the state of resource usage. These state information generally needs to be recorded and analyzed to provide data support for subsequent resource scheduling decisions. The collection of resource state information needs to be formatted, such as through data serialization technology (such as JSON or XML) to ensure the readability of data and the simplicity of transmission. When storing these information, a time series database can also be selected, which can efficiently process real-time data streams.

[0115] The task scheduling system classifies and sorts tasks according to their properties. In this process, the priority needs to be set according to the urgency of the task, the scope of the impact, and the significance to the stability and security of the system. This can be achieved through pre-set task attribute tags and weight parameters, where task importance and the amount of resources required for execution are common evaluation criteria. Through the division of priorities, the most important and urgent tasks can be given priority, avoiding any waste of system resources and task delays. Reasonable allocation of priorities can improve the overall task processing capacity, optimize system response time, and ensure the real-time and effectiveness of power system monitoring.

[0116] Building a computing resource allocation model requires comprehensive consideration of the resource status of the device and the current task requirements, including the computing power of the edge computing device, the complexity of the task, and the data processing requirements. First, resource allocation equations can be developed through linear programming or integer programming methods, which define the task's resource requirements and the device's resource limitations. In a multi-objective context, a balance equation is often needed to ensure that the basic resource requirements of each task are met while improving the resource utilization of the entire computing system. Specifically, a target function can be set to complete the task set in the shortest time while considering the efficiency of resource use, and solved in a multi-variable optimization framework. By determining the unit resource benefit of different tasks (such as the amount of tasks that can be completed per unit of computing resources), dynamic adjustment of the model can be achieved to adapt to different tasks and resource changes.

[0117] The reward function of the computing resource allocation model is generated based on the shortest time to complete all tasks in the current task set and the optimal resource utilization of the edge computing device, which is part of establishing a feedback mechanism. The reward function is used in the computing field to guide and optimize the decision-making process, and its core idea is to seek the solution that makes all tasks complete in the shortest time and uses the highest resource rate under a specific computing resource and task allocation strategy. When building the reward function, evaluation indicators such as task completion time, resource usage ratio, and utility need to be defined first. These indicators need to be combined through reasonable mathematical formula relationships to form a comprehensive reward value. For time and resource rate measurement, the commonly used method is the weighted average of unit task completion time and device utilization rate. Completion time can be calculated by counting the start and end times of each task, and resource rate can be calculated by the ratio of available resources to allocated resources. To balance the calculation results, a penalty parameter can be introduced to negatively reward situations where resource utilization is below a certain threshold or tasks are completed overtime. Through such reward function design, the resource allocation algorithm can be guided to converge towards the optimal direction without sacrificing performance requirements, thereby achieving efficient resource allocation and reasonable strategy in complex computing environments.

[0118] The optimal computing resource allocation strategy is obtained when the reward function reaches the maximum value, which involves the application of deep reinforcement learning. The Deep Q-Network (DQN) algorithm is a reinforcement learning method that combines deep learning and Q-learning. It uses a neural network to estimate the Q value of each state-action pair, which evaluates the long-term benefits of taking a certain action in a specific state. To optimize the computing resource allocation model, DQN uses the device resource state and task set as input to predict the reward value of each possible resource allocation strategy. By continuously updating the weights of the neural network, DQN can gradually approach the optimal strategy, i.e., taking the most correct resource allocation action in each task state. The reward function plays a role in evaluating the decision output in this process, guiding the model to continuously adjust itself to achieve the optimization of the resource allocation strategy. Through iterative experiments and the experience replay mechanism, the DQN algorithm can quickly and effectively converge in a dynamic and complex computing environment and find a strategy with high rewards. Under the guidance of this strategy, the edge computing device can efficiently allocate task loads, maximizing the utilization of computing resources.

[0119] Allocating the remaining device resources of the edge computing device to the feature sharing layer and multiple target basic detection sub-networks corresponding to the current task set according to the optimal computing resource allocation strategy is a key measure to effectively implement multi-task processing. In this implementation process, the feature sharing layer is used to extract and learn common features in task data, so it needs certain computing resource support. In addition, multiple basic detection sub-networks perform in-depth analysis and processing for specific tasks, such as voltage detection, phasor calculation, etc. According to the optimal resource allocation strategy obtained, the remaining resources will be reasonably allocated to different sub-networks and the feature sharing layer to ensure that all tasks are processed most effectively with appropriate computing resource support. This allocation process can be achieved by adjusting the resource allocation ratio to ensure that resources are moderately tilted according to priority and task workload requirements, improving the satisfaction of specific task requirements. At the same time, a dynamic resource scheduling mechanism can be introduced to continuously adjust the device resource allocation strategy based on real-time feedback of task execution to adapt to changes in task load. Under this resource optimization allocation, the device can operate more flexibly and efficiently, greatly improving its ability to handle large multi-task and complex computing.

[0120] Each target-based detection sub-network is committed to solving specific power system detection problems and achieving accurate detection of phasor data through an efficient computing resource allocation strategy. The final detection results include the measurement of various power parameters and their certainty analysis, which are published through the output interface and used for further operation and technical decision-making. Through the verification and feedback of the detection results, the shortcomings in the previous resource configuration and computing model can be found and optimized, forming a complete closed-loop control to continuously improve in future task processing. In addition, the detection results can also be updated to the power grid monitoring platform through the information sharing interface to realize more intelligent analysis and management.

[0121] In the embodiment, the expression formula of the optimal computing resource allocation strategy is as follows:

[0122]

[0123] In the formula, π * represents the optimal computing resource allocation strategy, argmax π represents selecting the computing resource allocation strategy that can maximize the reward function, E represents the cumulative reward expectation value based on the reward function, t represents the time step, γ t represents the discount factor, α and β both represent the weight coefficient, T(s t , a t ) represents the reward value in the task completion time dimension when the action a t is taken in the state s t in the target state space, R(s t , a t ) represents the penalty value in the resource utilization dimension when the action a t is taken in the state s t in the target state space.

[0124] In one embodiment, the computing resource allocation model of the multi-target phasor detection model is constructed in combination with the device resource state, the current task set, and the task priority, and includes the following steps:

[0125] A target state space is constructed based on a Markov decision model and in combination with the device resource state, the current task set, and the task priority;

[0126] A target action space is constructed with a selection step of selecting the next task in the current task set and an initial resource allocation strategy preset by the edge computing device;

[0127] A computing resource allocation model of a multi-target phasor detection model is constructed in combination with the target state space, the target action space, and a preset state transition probability function.

[0128] In this embodiment, the Markov decision model serves as a dynamic decision-making tool that effectively describes decision-making problems with randomness and temporal correlation. When defining the state space, the device's resource status needs to be considered, including the current available CPU, storage, and other hardware resource information. The characteristics of the device task, such as the size and complexity of the current task set, also need to be considered. Each state can be regarded as a specific configuration and load condition of the device at a certain time. The priority of the task is an important reference factor for the construction of the state space, and can affect the value and transition path of the state by setting the priority weight. Combining these information, the state space will cover all potential configuration states and have the ability to abstract complex system conditions. The state transition under different behavior strategies needs to be analyzed in detail to form the state transition matrix, which describes how the system transitions between states under different action choices.

[0129] The target action space is constructed by selecting the next task in the current task set and the initial resource allocation strategy preset by the edge computing device. In this process, the action space contains all possible behaviors that can be taken in a given state, determining how the current selection affects the future state. First, the selection of each task in the current task set is regarded as a scheduling behavior, and each selection step corresponds to a decision action. This step is not just picking a task from the task set, but also involves evaluating the importance, urgency, and impact of the task on the overall system, and then selecting the task that deserves priority execution. In addition, the initial resource allocation strategy limits the initial allocation of computing resources under the current device resource conditions to achieve the best performance. Combining these elements, the construction of the action space needs to consider the differentiated effects of different task combination selections and resource allocation strategies. In this process, detailed action descriptions need to be developed, and possible combination schemes can be displayed through charts or matrices. Reasonable action space construction helps to include potential strategies and ensures that an optimal solution that meets the conditions can always be found under different strategy paths.

[0130] The computing resource allocation model of the multi-objective phasor detection model is constructed by combining the target state space, the target action space and the preset state transition probability function. The computing resource allocation model is formed by integrating the state space and the action space constructed in the early stage, combining the state transition probability, and forming a comprehensive decision model. The role of the state transition probability function in this process is to describe the possibility of a state transitioning to another state after taking a certain action. These probabilities can be estimated based on historical data statistics, or designed according to system simulation or expert knowledge. During the model construction process, the probability distribution of each state transition process needs to be defined and calibrated to ensure that the model can accurately define the behavior response under various possible situations. By constructing the multi-objective phasor detection model, the model aims to solve the simultaneous satisfaction of multiple resource optimization objectives, such as maximizing resource utilization, minimizing task response time and satisfying task priority, etc. In the resource allocation model, by analyzing each strategy, the optimal order of task execution and resource allocation can be identified, so that all objectives can be effectively supported by decision-making under various strategies. Finally, through iterative optimization of the model, a dynamic resource allocation strategy framework is formed that conforms to the current device resource state and task demand, providing an intelligent decision support system for edge computing devices when executing complex tasks.

[0131] The application further discloses a loop phasor detection device, which comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements the substation loop phasor detection method based on light artificial intelligence in any one of the above embodiments when executing the computer program.

[0132] The processor can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc. The application does not limit this.

[0133] The memory can be an internal storage unit of the computer device, such as a hard disk or a memory of the computer device, or an external storage device of the computer device, such as a plug-in hard disk, a smart memory card (SMC), a secure digital card (SD) or a flash memory card (FC) equipped on the computer device, etc. The memory can also be a combination of the internal storage unit and the external storage device of the computer device. The memory is used to store computer programs and other programs and data required by the computer device. The memory can also be used to temporarily store data that has been output or will be output. The application does not limit this.

[0134] The application also discloses a substation loop phasor detection system based on light artificial intelligence, referring to Figure 3 , the system comprises:

[0135] a loop phasor detection device;

[0136] an edge computing device, which is arranged in the intelligent substation and is used for carrying a multi-target phasor detection model constructed based on a convolutional neural network model;

[0137] an intelligent sensor network, which is arranged in a substation loop of the intelligent substation and is in communication connection with the edge computing device, and is used for collecting device data of all loop devices in the substation loop;

[0138] a current output device, which is arranged on a primary loop side of the substation loop, and is used for injecting current into the substation loop when performing a loop phasor detection task;

[0139] a voltage output device, which is arranged on the primary loop side of the substation loop, and is used for injecting voltage into the substation loop when performing the loop phasor detection task.

[0140] It should be understood by those skilled in the art that the above discussion of any embodiment is only exemplary and is not intended to imply that the protection scope of the application is limited to these examples; under the idea of the application, the above embodiments or technical features in different embodiments can also be combined, the steps can be implemented in any order, and there are many other changes of different aspects of one or more embodiments of the application as above. In order to be brief, they are not provided in details.

[0141] One or more embodiments of the application are intended to cover all such replacements, modifications and variations falling within the broad scope of the application. Therefore, any omission, modification, equivalent replacement, improvement, etc. made in the spirit and principles of one or more embodiments of the application should be included in the protection scope of the application.

Claims

1. A lightweight artificial intelligence-based substation loop phasor detection method, characterized in that, The application is applied to an intelligent substation, wherein an edge computing device is deployed in the intelligent substation, an intelligent sensor network is deployed in a substation loop in the intelligent substation, the intelligent sensor network is in communication connection with the edge computing device, and a current output device and a voltage output device are further deployed on a primary loop side of the substation loop. The method comprises the following steps: obtaining a loop phasor detection task of the intelligent substation, wherein the loop phasor detection task is any one or more of loop device phase sequence correctness detection, loop device polarity correctness detection, loop device variable ratio correctness detection and cable connection performance detection; constructing a feature sharing layer and a plurality of initial detection subnetworks based on a convolutional neural network model, wherein the number of the initial detection subnetworks is the same as the type number of task types in the loop phasor detection task, and all the initial detection subnetworks have independent output layers; adjusting the layer structure and model parameters of each initial detection subnetwork according to the task types in the loop phasor detection task to obtain a plurality of basic detection subnetworks; connecting the input layers in the plurality of basic detection subnetworks with the output layers in the feature sharing layer to obtain a multi-target phasor detection model, and all the basic detection subnetworks are arranged in parallel to form a multi-target detection task layer; deploying the multi-target phasor detection model in the edge computing device, wherein the multi-target phasor detection model comprises the feature sharing layer and the multi-target detection task layer; obtaining historical loop detection data of the intelligent substation and extracting static topology features of the substation loop; training the feature sharing layer by using the static topology features and training the multi-target detection task layer by using the historical loop detection data; in response to the loop phasor detection task and controlling the current output device and the voltage output device to pass through the substation loop under pressure to simulate the load running process of the substation loop; during the process of passing through the substation loop under pressure, collecting device data of all loop devices in the substation loop through the intelligent sensor network; inputting the device data into the multi-target phasor detection model to output loop phasor detection results of the intelligent substation through the multi-target phasor detection model.

2. The lightweight artificial intelligence based substation loop phasor detection method of claim 1, wherein, The method comprises the following steps: obtaining historical loop detection data of the intelligent substation and device parameters of all loop devices in the substation loop; taking the loop devices as device nodes and taking the electrical connection relationship between all the loop devices in the substation loop as device node edges to construct a substation electrical connection graph, and giving all the device nodes node attributes based on the device parameters; identifying all electrical loops in the substation electrical connection graph by using a recursive graph traversal algorithm; marking bus structure features in all the electrical loops through node aggregation analysis of all the electrical loops; integrating all the device parameters and all the electrical loops into static topology features of the substation loop.

3. The lightweight artificial intelligence based substation loop phasor detection method of claim 2, wherein, The training of the feature sharing layer is completed by using the static topology features, and the training of the multi-target detection task layer is completed by using the historical loop detection data, including the following steps: The feature sharing layer is used as an encoder, and the static topology features are mapped to a low-dimensional space through the encoder; The static topology features are reconstructed in the low-dimensional space through a preset decoder, and the training of the feature sharing layer is completed when the reconstruction error reaches a minimum value; A sub-network loss function is defined for each of the basic monitoring sub-networks, and a weighted sum of all the sub-network loss functions is used as a total loss function of the multi-target detection task layer; The historical loop detection data is classified according to all task types of the loop phasor detection task and input into all the basic monitoring sub-networks, and the small batch gradient descent method is used to train all the basic monitoring sub-networks, and the gradient descent algorithm is used to minimize the total loss function, thereby completing the training of the multi-target detection task layer.

4. The lightweight artificial intelligence based substation loop phasor detection method of claim 1, wherein, The device data is input into the multi-target phasor detection model, and the loop phasor detection result of the smart substation is output through the multi-target phasor detection model, including the following steps: Obtaining the device resource state of the edge computing device; Determining a current task set and task priorities of all tasks in the current task set by analyzing the loop phasor detection task; Constructing a computing resource allocation model of the multi-target phasor detection model in combination with the device resource state, the current task set and the task priorities; Generating a reward function of the computing resource allocation model in order to complete all tasks in the current task set in the shortest time and achieve optimal resource utilization of the edge computing device; Solving the computing resource allocation model based on the reward function and using a deep Q network algorithm, and obtaining an optimal computing resource allocation strategy for the edge computing device to execute all tasks in the current task set when the reward function reaches a maximum value; Allocating the remaining device resources of the edge computing device to the feature sharing layer and multiple target basic detection sub-networks corresponding to the current task set according to the optimal computing resource allocation strategy; Outputting the loop phasor detection result of the smart substation through all the target basic detection sub-networks.

5. The lightweight artificial intelligence based substation loop phasor detection method of claim 4, wherein, The construction of the computing resource allocation model of the multi-target phasor detection model in combination with the device resource state, the current task set and the task priorities includes the following steps: Constructing a target state space based on a Markov decision model and in combination with the device resource state, the current task set and the task priorities; Constructing a target action space with a selection step of selecting a next task in the current task set and a preset initial resource allocation strategy of the edge computing device; Constructing the computing resource allocation model of the multi-target phasor detection model in combination with the target state space, the target action space and a preset state transition probability function.

6. The lightweight artificial intelligence based substation loop phasor detection method of claim 5, wherein, The expression formula of the optimal computing resource allocation strategy is as follows: ， wherein: represents the optimal computing resource allocation strategy, represents selecting a computing resource allocation strategy that maximizes the reward function, represents the cumulative reward expectation value based on the reward function, represents a time step, represents a discount factor, and both represent a weight coefficient, represents a state in the target state space under taking an action in the target action space at a reward value in a task completion time dimension, represents a state in the target state space under taking an action in the target action space at a penalty value in a resource utilization dimension.

7. A loop phasor detection device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein, The processor implements the light artificial intelligence-based substation loop phasor detection method according to any one of claims 1-6 when executing the computer program.

Images