Disclosure of Invention
The embodiment of the invention provides an intelligent control method and system for fault location and quick reliability recovery of a power distribution network, which can solve the problems in the prior art.
In a first aspect of an embodiment of the present invention,
The intelligent control method for positioning the faults and quickly recovering the reliability of the power distribution network comprises the following steps:
the method comprises the steps of collecting real-time data of each node of a power distribution network, including voltage, current and power data, inputting the real-time data into a pre-trained deep learning model, rapidly identifying fault types and accurately positioning fault points by analyzing time sequence features and space distribution features of the real-time data based on a mixed architecture of a convolutional neural network and a long-short-time memory network;
Based on the generated preliminary fault evaluation report, starting a multi-objective optimization algorithm which comprehensively considers factors such as power supply reliability, economy, user importance and the like, simulating a plurality of possible fault isolation and power grid reconstruction schemes by using reinforcement learning technology, and carrying out simulation evaluation on each scheme;
The method comprises the steps of sending a formulated intelligent control instruction sequence to related intelligent terminal equipment through a safe encryption channel, gradually executing control operation according to a preset safety protocol after the intelligent terminal equipment receives the instruction, collecting state information of each node in real time in the execution process, feeding the information back to a central control system, rapidly processing feedback information by the central control system locally through an edge computing technology, evaluating the implementation effect of a scheme in real time, immediately triggering an emergency response mechanism if abnormal conditions or poor implementation effect are detected, dynamically adjusting a control strategy, continuously monitoring the recovery state of the system after fault isolation and power grid reconstruction are completed, generating detailed event reports and optimization suggestions, and providing data support for future fault prevention and system improvement.
In an alternative embodiment of the present invention,
The deep learning model is based on a mixed architecture of a convolutional neural network and a long-short-term memory network, and the rapid identification of fault types and accurate positioning of fault points by analyzing time sequence features and spatial distribution features of the real-time data comprises the following steps:
the method comprises the steps of acquiring voltage, current and power data in real time at a sampling frequency of 100 times per second through intelligent sensors deployed at key nodes of a power distribution network, preprocessing the data, including denoising, standardization and time window segmentation, to form a data format suitable for deep learning model input;
Inputting the preprocessed data into a deep learning model with a multi-channel structure, wherein the multi-channel structure respectively corresponds to voltage, current and power data, the data of each channel firstly passes through a series of convolution layers and pooling layers, the convolution layers use convolution kernels with three sizes of 3×3, 5×5 and 7×7, 64 filters are arranged in each size, and each convolution operation is followed by a maximum pooling layer;
Remodelling the feature images output by the convolution layer and the pooling layer and inputting the remodelling feature images into a long-short-time memory network layer, wherein the long-short-time memory network layer comprises 128 hidden units;
adding a self-attention layer after the long-short-term memory network layer, so that the model can adaptively pay attention to the most relevant time steps and characteristics;
the output of the self-attention layer passes through a full-connection layer, and finally the probability distribution of the fault type is output through a softmax classifier;
training the deep learning model by adopting a method of combining large-scale simulation data and actual fault data, using a cross entropy loss function and an Adam optimizer, setting the learning rate to be 0.001 initially, adopting a learning rate attenuation strategy, and simultaneously applying a dropout technology to prevent overfitting, wherein the discarding rate is set to be 0.5;
And analyzing the input real-time data by using the trained deep learning model, completing fault identification and positioning within millisecond level, and outputting fault type and fault position information, wherein the identification accuracy of the fault type is not lower than 98.5%, and the average error of fault positioning is less than 50 meters.
In an alternative embodiment of the present invention,
According to the identified fault type and fault point position, and combining preset power distribution network topological structure information, evaluating the influence range and severity of the fault, and generating a preliminary fault evaluation report comprises the following steps:
Constructing a digital twin model of a power distribution network, wherein the digital twin model of the power distribution network is based on a graph database technology, transformers, switches, lines, loads and power generation equipment in the power distribution network are represented as nodes, and connection relations among the nodes are represented as directed edges, wherein the directed edges represent power flow directions;
setting attribute information for the node, wherein the attribute information comprises equipment identification, rated capacity, current load, voltage level, position coordinates, equipment type, rated current, equipment state, line length, impedance parameters, load type and importance level;
Importing the topological structure and equipment information of the power distribution network into the graph database through a data interface with a power distribution network management system, and establishing a real-time data updating mechanism with a monitoring and data acquisition system to dynamically update the attribute information of the nodes;
Receiving fault node identification and fault type information, and positioning a fault starting point in the digital twin model of the power distribution network based on the fault node identification;
adopting a depth-first search algorithm, and gradually exploring a region affected by the fault along the power flow direction represented by the directed edge by taking the fault starting point as a starting point;
In the searching process of the depth-first searching algorithm, based on attribute information of the current access node and the adjacent node, judging whether the fault can be transmitted to the adjacent node or not by combining the fault type;
Checking whether the non-triggered protection equipment exists between the current node and the adjacent node or not when the fault type is a short circuit fault;
adding the nodes which are judged to be affected by the faults into an affected area list, and recording the father node of each affected node for reconstructing a fault propagation path;
generating a fault impact assessment result comprising an affected node list and a fault propagation order based on the affected region list and a fault propagation path;
Converting the fault influence assessment result into a visual graph by using a graph visualization technology, and intuitively displaying a fault influence range and a propagation path;
Verifying the method based on historical fault data, comparing the predicted influence range with the actual situation, and optimizing propagation judgment logic and parameters in a depth-first search algorithm according to the verification result;
And realizing a self-adaptive learning mechanism, and automatically adjusting and optimizing the judgment parameters and logic of the method according to the feedback information of each actual fault.
In an alternative embodiment of the present invention,
Based on the generated preliminary fault assessment report, a multi-objective optimization algorithm is started, wherein the algorithm comprehensively considers factors of power supply reliability, economy and user importance, and the factors comprise:
receiving a preliminary fault assessment report, wherein the preliminary fault assessment report comprises an affected area, a fault type, a fault position and an estimated influence degree;
Starting a multi-objective optimization algorithm based on the preliminary fault evaluation report, wherein the multi-objective optimization algorithm adopts an improved non-dominant ranking genetic algorithm II so as to balance three objectives of power supply reliability, economy and user importance;
For the power supply reliability target, calculating a system average interrupt frequency index and a system average interrupt duration index, and minimizing the two indexes; for the economic goal, consider the equipment replacement cost, cost of labor and cost of electric energy consumption, and minimize the total cost;
setting the population scale as 100, the evolution algebra as 500, adopting single-point cross operation, the cross probability as 0.8, adopting uniform mutation operation and the mutation probability as 0.1;
Introducing a self-adaptive crossover and mutation probability adjustment mechanism, calculating an average fitness value of a population every 10 generations, and increasing crossover probability by 0.05 and mutation probability by 0.01 when the average fitness value calculated for three continuous times changes by less than 1 percent until the average fitness value reaches a preset upper limit;
Setting network topology constraint, voltage constraint and current constraint, wherein the voltage constraint requires that the voltage deviation of each node is not more than +/-7% of rated voltage, the current constraint requires that the line current is not more than 80% of rated capacity, punishing individuals violating constraint by adopting a penalty function method, layering the individuals by adopting a rapid non-dominant sorting method in each generation of evolution process, and using crowding degree distance calculation to maintain population diversity, and outputting a group of non-dominant solutions as a pareto optimal solution set for a decision maker to select the most suitable fault recovery scheme.
In an alternative embodiment of the present invention,
Simulating a plurality of possible fault isolation and power grid reconstruction schemes by using reinforcement learning technology, and carrying out simulation evaluation on each scheme, wherein the selection of the optimal fault isolation and power grid reconstruction scheme through iterative optimization comprises the following steps:
The method comprises the steps of constructing a power distribution network environment model, defining an environment state, setting an action space, defining a specific fault isolation and network reconstruction scheme for each action, designing a reward function, wherein the environment state is represented by a high-dimensional vector and comprises voltage, current and power factor parameters of each node and the on-off state of a switch;
constructing a deep Q network, wherein the deep Q network comprises a plurality of layers of convolutional neural networks and a full-connection layer, an input layer receives the environment state, and an output layer corresponds to the Q value of each possible action;
Setting an experience playback mechanism, storing a transfer sample by using an experience pool with fixed capacity, setting a target network updating mechanism, updating target network parameters once every fixed training step, adopting an epsilon-greedy strategy to perform action selection, setting the initial epsilon value of the epsilon-greedy strategy to be 1, attenuating fixed proportion in each round until reaching the minimum value, and executing a training process, wherein the method comprises the following substeps:
Resetting an environment model of the power distribution network to obtain an initial state, selecting actions according to an epsilon-greedy strategy, executing the selected actions to obtain rewards and a next state, storing a transfer sample into an experience pool, randomly extracting a batch of samples from the experience pool for training, updating a depth Q network parameter by using a mean square error loss function and an Adam optimizer, and periodically updating a target network;
Performing test evaluation on the trained deep Q network, wherein the test evaluation comprises single fault, multiple faults and cascading failure scenes; and selecting an optimal fault isolation and power grid reconstruction scheme based on the test evaluation result.
In an alternative embodiment of the present invention,
The method comprises the steps of sending a formulated intelligent control instruction sequence to related intelligent terminal equipment through a secure encryption channel, gradually executing control operation according to a preset security protocol after the intelligent terminal equipment receives the instruction, collecting state information of each node in real time in the execution process, feeding the information back to a central control system, rapidly processing feedback information by the central control system locally by using an edge computing technology, evaluating an executing effect of a scheme in real time, and immediately triggering an emergency response mechanism if abnormal conditions or poor executing effect are detected, wherein the dynamic adjustment control strategy comprises the following steps:
performing test evaluation on the trained deep Q network, wherein the test evaluation comprises single fault, multiple faults and cascading failure scenes;
Based on the test evaluation result, selecting an optimal fault isolation and power grid reconstruction scheme;
Converting the selected optimal scheme into an intelligent control instruction sequence, and transmitting the intelligent control instruction sequence to related intelligent terminal equipment through a secure encryption channel based on elliptic curve cryptography;
After receiving the instruction, the intelligent terminal device gradually executes control operation according to a preset safety protocol, wherein the safety protocol comprises environment inspection before operation, operation interval time control and operation sequence forced execution;
in the execution process, the state information of each node is collected in real time through a high-precision sensor network, and the information is fed back to a central control system through a multi-level data convergence structure;
the central control system utilizes a layered edge computing architecture to rapidly process feedback information locally and evaluate the scheme execution effect in real time;
if abnormal conditions or poor execution effects are detected, an emergency response mechanism is immediately triggered, and a multi-objective optimization algorithm is adopted to dynamically adjust a control strategy.
In a second aspect of an embodiment of the present invention,
An intelligent control system for providing fault location and fast reliability recovery of a power distribution network, comprising:
The first unit is used for collecting real-time data of each node of the power distribution network, including voltage, current and power data, and inputting the real-time data into a pre-trained deep learning model; the deep learning model is based on a mixed architecture of a convolutional neural network and a long-short-time memory network, and rapidly identifies fault types and accurately positions fault points by analyzing time sequence features and space distribution features of the real-time data;
The second unit is used for starting a multi-objective optimization algorithm based on the generated preliminary fault evaluation report, wherein the algorithm comprehensively considers factors such as power supply reliability, economy, user importance and the like, simulating a plurality of possible fault isolation and power grid reconstruction schemes by using a reinforcement learning technology, and carrying out simulation evaluation on each scheme;
The third unit is used for sending the formulated intelligent control instruction sequence to the related intelligent terminal equipment through the safe encryption channel, the intelligent terminal equipment gradually executes control operation according to a preset safe protocol after receiving the instruction, collects the state information of each node in real time in the execution process and feeds the information back to the central control system, the central control system rapidly processes the feedback information locally by utilizing the edge computing technology, evaluates the execution effect of the scheme in real time, immediately triggers an emergency response mechanism if abnormal conditions or poor execution effect are detected, dynamically adjusts the control strategy, continuously monitors the recovery state of the system after fault isolation and power grid reconstruction are completed, and generates detailed event reports and optimization suggestions to provide data support for future fault prevention and system improvement.
In a third aspect of an embodiment of the present invention,
There is provided an electronic device including:
A processor;
a memory for storing processor-executable instructions;
wherein the processor is configured to invoke the instructions stored in the memory to perform the method described previously.
In a fourth aspect of an embodiment of the present invention,
There is provided a computer readable storage medium having stored thereon computer program instructions which, when executed by a processor, implement the method as described above.
The method can complete influence evaluation within a few seconds after the fault occurs, and provides accurate decision basis for fault isolation and system recovery. For a typical urban distribution network, which comprises 5000 nodes and 7500 edges, the method can complete one-time complete fault influence evaluation within 3 seconds, and the accuracy rate is more than 95%. The high-efficiency and accurate evaluation capability has important significance for quickly responding to faults of the power distribution network, minimizing the influence range of the faults and improving the reliability and toughness of the power distribution network.
Detailed Description
For the purpose of making the objects, technical solutions and advantages of the embodiments of the present invention more apparent, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention, and it is apparent that the described embodiments are only some embodiments of the present invention, not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
The technical scheme of the invention is described in detail below by specific examples. The following embodiments may be combined with each other, and some embodiments may not be repeated for the same or similar concepts or processes.
Fig. 1 is a schematic flow chart of an intelligent control method for fault location and reliability quick recovery of a power distribution network according to an embodiment of the present invention, as shown in fig. 1, the method includes:
S101, collecting real-time data of each node of a power distribution network, including voltage, current and power data, and inputting the real-time data into a pre-trained deep learning model, wherein the deep learning model is based on a mixed architecture of a convolutional neural network and a long-short-time memory network, and rapidly identifies fault types and accurately positions fault points by analyzing time sequence features and space distribution features of the real-time data;
S102, starting a multi-objective optimization algorithm based on the generated preliminary fault evaluation report, wherein the algorithm comprehensively considers factors such as power supply reliability, economy, user importance and the like, simulating a plurality of possible fault isolation and power grid reconstruction schemes by using a reinforcement learning technology, and carrying out simulation evaluation on each scheme;
S103, sending a formulated intelligent control instruction sequence to related intelligent terminal equipment through a secure encryption channel, gradually executing control operation according to a preset security protocol after the intelligent terminal equipment receives the instruction, collecting state information of each node in real time in the execution process, feeding the information back to a central control system, rapidly processing feedback information by the central control system locally by using an edge computing technology, evaluating the execution effect of a scheme in real time, triggering an emergency response mechanism immediately if abnormal conditions or poor execution effect are detected, dynamically adjusting a control strategy, continuously monitoring the recovery state of the system after fault isolation and power grid reconstruction are completed, generating detailed event reports and optimization suggestions, and providing data support for future fault prevention and system improvement.
In an alternative embodiment of the present invention,
The deep learning model is based on a mixed architecture of a convolutional neural network and a long-short-term memory network, and the rapid identification of fault types and accurate positioning of fault points by analyzing time sequence features and spatial distribution features of the real-time data comprises the following steps:
the method comprises the steps of acquiring voltage, current and power data in real time at a sampling frequency of 100 times per second through intelligent sensors deployed at key nodes of a power distribution network, preprocessing the data, including denoising, standardization and time window segmentation, to form a data format suitable for deep learning model input;
Inputting the preprocessed data into a deep learning model with a multi-channel structure, wherein the multi-channel structure respectively corresponds to voltage, current and power data, the data of each channel firstly passes through a series of convolution layers and pooling layers, the convolution layers use convolution kernels with three sizes of 3×3, 5×5 and 7×7, 64 filters are arranged in each size, and each convolution operation is followed by a maximum pooling layer;
Remodelling the feature images output by the convolution layer and the pooling layer and inputting the remodelling feature images into a long-short-time memory network layer, wherein the long-short-time memory network layer comprises 128 hidden units;
adding a self-attention layer after the long-short-term memory network layer, so that the model can adaptively pay attention to the most relevant time steps and characteristics;
the output of the self-attention layer passes through a full-connection layer, and finally the probability distribution of the fault type is output through a softmax classifier;
training the deep learning model by adopting a method of combining large-scale simulation data and actual fault data, using a cross entropy loss function and an Adam optimizer, setting the learning rate to be 0.001 initially, adopting a learning rate attenuation strategy, and simultaneously applying a dropout technology to prevent overfitting, wherein the discarding rate is set to be 0.5;
And analyzing the input real-time data by using the trained deep learning model, completing fault identification and positioning within millisecond level, and outputting fault type and fault position information, wherein the identification accuracy of the fault type is not lower than 98.5%, and the average error of fault positioning is less than 50 meters.
Illustratively, in order to achieve rapid identification and accurate positioning of power distribution network faults, the present study proposes a hybrid deep learning architecture based on Convolutional Neural Networks (CNNs) and long and short term memory networks (LSTM). The model fully utilizes the advantages of CNN in space feature extraction and the capability of LSTM in time sequence analysis, thereby effectively processing complex fault modes in the power distribution network.
Firstly, voltage, current and power data are collected in real time through intelligent sensors deployed at key nodes of a power distribution network. These data are sampled at a high frequency (e.g., 100 times per second) to capture transient fault signatures. The collected original data is preprocessed, including denoising, standardization and time window segmentation, so that a data format suitable for deep learning model input is formed.
And adopting a multichannel structure at an input layer of the model, wherein the multichannel structure corresponds to voltage, current and power data respectively. The data for each channel is first passed through a series of convolution and pooling layers to extract the spatial distribution features. The size and number of convolution kernels are optimized to maximize capture of the fault signature. Specifically, convolution kernels of 3×3,5×5, and 7×7 are used, and 64 filters are set for each size. Each convolution operation is followed by a maximum pooling layer to reduce feature dimensions and increase translational invariance of the model.
The feature map of the convolution layer output is then reshaped and input to the LSTM layer. The LSTM layer contains 128 hidden units for analysis of time series features. Through gating mechanisms, LSTM is able to effectively capture long-term dependencies, which is critical for identifying certain slowly evolving fault types. The output of the LSTM layer passes through a full connection layer, and finally the probability distribution of the fault type is output through a softmax classifier.
In order to improve the positioning accuracy of the model, attention mechanisms are introduced. After the LSTM layer, a self-attention layer is added to enable the model to adaptively focus on the most relevant time steps and features. The mechanism remarkably improves the positioning capability of the model in complex fault scenes.
The training of the model adopts a method of combining large-scale simulation data and limited actual fault data. A large number of samples containing various fault types and locations were generated by the power system simulation software, while a historical fault record was collected as a supplement. The training process uses a cross entropy loss function and an Adam optimizer, the learning rate is initially set to 0.001, and a learning rate decay strategy is adopted. To prevent overfitting, a dropout technique was applied, with the discard rate set to 0.5.
Finally, trained models are able to accomplish fault identification and localization within milliseconds. On the test set, the accuracy rate of the model in identifying the fault type reaches 98.5%, and the average error of fault positioning is less than 50 meters. The high-efficiency accurate fault diagnosis provides a reliable basis for subsequent fault isolation and system recovery, so that the reliability and recovery speed of the power distribution network are remarkably improved.
In an alternative embodiment of the present invention,
According to the identified fault type and fault point position, and combining preset power distribution network topological structure information, evaluating the influence range and severity of the fault, and generating a preliminary fault evaluation report comprises the following steps:
Constructing a digital twin model of a power distribution network, wherein the digital twin model of the power distribution network is based on a graph database technology, transformers, switches, lines, loads and power generation equipment in the power distribution network are represented as nodes, and connection relations among the nodes are represented as directed edges, wherein the directed edges represent power flow directions;
setting attribute information for the node, wherein the attribute information comprises equipment identification, rated capacity, current load, voltage level, position coordinates, equipment type, rated current, equipment state, line length, impedance parameters, load type and importance level;
Importing the topological structure and equipment information of the power distribution network into the graph database through a data interface with a power distribution network management system, and establishing a real-time data updating mechanism with a monitoring and data acquisition system to dynamically update the attribute information of the nodes;
Receiving fault node identification and fault type information, and positioning a fault starting point in the digital twin model of the power distribution network based on the fault node identification;
adopting a depth-first search algorithm, and gradually exploring a region affected by the fault along the power flow direction represented by the directed edge by taking the fault starting point as a starting point;
In the searching process of the depth-first searching algorithm, based on attribute information of the current access node and the adjacent node, judging whether the fault can be transmitted to the adjacent node or not by combining the fault type;
Checking whether the non-triggered protection equipment exists between the current node and the adjacent node or not when the fault type is a short circuit fault;
adding the nodes which are judged to be affected by the faults into an affected area list, and recording the father node of each affected node for reconstructing a fault propagation path;
generating a fault impact assessment result comprising an affected node list and a fault propagation order based on the affected region list and a fault propagation path;
Converting the fault influence assessment result into a visual graph by using a graph visualization technology, and intuitively displaying a fault influence range and a propagation path;
Verifying the method based on historical fault data, comparing the predicted influence range with the actual situation, and optimizing propagation judgment logic and parameters in a depth-first search algorithm according to the verification result;
And realizing a self-adaptive learning mechanism, and automatically adjusting and optimizing the judgment parameters and logic of the method according to the feedback information of each actual fault.
Illustratively, the fault influence assessment method based on the digital twin model of the power distribution network and the depth-first search algorithm is an innovative power distribution network fault analysis technology. According to the method, by constructing an accurate digital twin model and combining an efficient searching algorithm, quick and accurate fault influence range evaluation and propagation path tracking are realized. The following is the detailed technical content of the method:
The digital twin model construction of the power distribution network is the basis of the method. Neo4j is selected as a graph database platform, and the complex power distribution network structure is represented by utilizing the efficient graph traversal performance of the graph database platform. In Neo4j, a variety of node types are created, including transformer nodes, switch nodes, line nodes, load nodes, and power generation nodes. Each node type defines a specific set of attributes. For example, the transformer node contains ID, rated capacity, current load, voltage level, position coordinates, etc., the switch node contains ID, type (breaker/disconnector), rated current, status (on/off), etc., the line node contains ID, length, impedance parameters, rated current, etc., and the load node contains ID, type (resident/commercial/industrial), importance level, current power, etc.
The connection relationship between the nodes is represented by directed edges, the direction of which represents the flow of power, the properties of which include the connection type, the direction of current, etc. For example, an edge from a transformer node to a line node represents the flow of power from the transformer to the line. This configuration enables the system to quickly track power flow direction and fault propagation paths.
Data import is a key step in model construction. And extracting complete power distribution network topology and equipment information through a data interface with the existing power distribution network management system. These data are efficiently imported into the graph database using Neo4 j's batch import tool. For example, for a medium-scale distribution network comprising 1000 nodes and 1500 edges, the batch importation may be completed in a matter of minutes.
In order to maintain the real-time performance of the model, a real-time data updating mechanism of the SCADA system is designed. Every 5 seconds, the system automatically acquires the latest equipment state data, such as the current load of the transformer, the on-off state of the switch and the like, from the SCADA, and updates the attribute of the corresponding node. This ensures that the digital twin model accurately reflects the dynamic state of the actual distribution network.
Depth First Search (DFS) algorithms are central to fault impact assessment. When the fault node ID and fault type information are received, the algorithm first locates the fault starting point in the digital twin model. DFS is then implemented using the stack data structure, progressively exploring the affected area along the power flow direction, starting from the point of failure.
In the DFS process, it is critical to determine whether a fault will propagate to neighboring nodes. This decision logic varies depending on the type of fault and node characteristics. For example, for a short circuit fault, the algorithm checks if there is an un-triggered protection device between the current node and the neighboring node. If there is an un-triggered protection device, the fault will propagate to the neighboring node. For open circuit faults, the algorithm determines whether the neighboring node depends on the current node for power. If a dependency exists, the neighboring nodes may be affected.
The algorithm uses the aggregate data structure to record the accessed nodes, avoiding repeated accesses. Meanwhile, parent node information of each affected node is stored using a dictionary structure for subsequent reconstruction of the fault propagation path. This approach both ensures that no affected nodes are missed and efficiently tracks the exact path of fault propagation.
In order to improve the processing efficiency of the large-scale network, a parallel DFS algorithm is realized. The power distribution network is divided into a plurality of subgraphs, and each subgraph can independently perform DFS. For example, for a large power distribution network comprising 10000 nodes, the power distribution network can be divided into 10 subgraphs, and 10 parallel threads are used for simultaneous processing, so that the calculation time is remarkably reduced.
The outcome output and visualization is to facilitate the decision maker's understanding and use of the assessment results. The system generates an output in JSON format containing the affected node list and fault propagation paths. For example, a typical output may contain 50 affected nodes and 35 propagation relationships between them. The NetworkX library is used for converting the data into a visual graph, and the fault influence range and the propagation path are visually displayed. The nodes may be represented in different colors and sizes to indicate their type and importance, and the thickness of the edges may indicate the current magnitude.
Model verification and optimization are key to ensuring method accuracy and adaptability. And verifying the model by using the historical fault data, and comparing the predicted influence range of the model with the actual situation. For example, a backtracking test is performed on 100 failure events occurring in the past year, and the accuracy of model prediction is calculated. If the accuracy is below 90%, the propagation decision logic and weight parameters in the DFS algorithm need to be adjusted.
And an adaptive learning mechanism is realized, and model parameters are automatically optimized according to feedback of actual faults each time. For example, if the model underestimates the scope of influence of a certain type of fault multiple times, the system will automatically increase the weight of the corresponding propagation determination logic. The mechanism enables the model to be continuously learned and improved, and is suitable for dynamic changes of the power distribution network.
Through the detailed technical steps, the method can complete influence evaluation within a few seconds after the fault occurs, and provides accurate decision basis for fault isolation and system recovery. For a typical urban distribution network, which comprises 5000 nodes and 7500 edges, the method can complete one-time complete fault influence evaluation within 3 seconds, and the accuracy rate is more than 95%. The high-efficiency and accurate evaluation capability has important significance for quickly responding to faults of the power distribution network, minimizing the influence range of the faults and improving the reliability and toughness of the power distribution network.
In an alternative embodiment of the present invention,
Based on the generated preliminary fault assessment report, a multi-objective optimization algorithm is started, wherein the algorithm comprehensively considers factors of power supply reliability, economy and user importance, and the factors comprise:
receiving a preliminary fault assessment report, wherein the preliminary fault assessment report comprises an affected area, a fault type, a fault position and an estimated influence degree;
Starting a multi-objective optimization algorithm based on the preliminary fault evaluation report, wherein the multi-objective optimization algorithm adopts an improved non-dominant ranking genetic algorithm II so as to balance three objectives of power supply reliability, economy and user importance;
For the power supply reliability target, calculating a system average interrupt frequency index and a system average interrupt duration index, and minimizing the two indexes; for the economic goal, consider the equipment replacement cost, cost of labor and cost of electric energy consumption, and minimize the total cost;
setting the population scale as 100, the evolution algebra as 500, adopting single-point cross operation, the cross probability as 0.8, adopting uniform mutation operation and the mutation probability as 0.1;
Introducing a self-adaptive crossover and mutation probability adjustment mechanism, calculating an average fitness value of a population every 10 generations, and increasing crossover probability by 0.05 and mutation probability by 0.01 when the average fitness value calculated for three continuous times changes by less than 1 percent until the average fitness value reaches a preset upper limit;
Setting network topology constraint, voltage constraint and current constraint, wherein the voltage constraint requires that the voltage deviation of each node is not more than +/-7% of rated voltage, the current constraint requires that the line current is not more than 80% of rated capacity, punishing individuals violating constraint by adopting a penalty function method, layering the individuals by adopting a rapid non-dominant sorting method in each generation of evolution process, and using crowding degree distance calculation to maintain population diversity, and outputting a group of non-dominant solutions as a pareto optimal solution set for a decision maker to select the most suitable fault recovery scheme.
Illustratively, a multi-objective optimization algorithm is initiated based on the generated preliminary fault assessment report that comprehensively considers factors of power supply reliability, economy, and user importance to provide key inputs to the multi-objective optimization algorithm, including affected area, fault type, fault location, and estimated extent of impact. Based on this information, a multi-objective optimization algorithm is designed to generate an optimal failure recovery scheme. The algorithm employs an improved non-dominant ranking genetic algorithm II (NSGA-II) to balance the three conflicting goals of power supply reliability, economy, and user importance.
First, for a power supply reliability target, the algorithm considers a System Average Interrupt Frequency Index (SAIFI) and a System Average Interrupt Duration Index (SAIDI). The SAIFI calculation method is that the annual user power outage times are divided by the total number of users, and SAIDI is that the annual user power outage time sum is divided by the total number of users. For example, if 10 power failure accidents occur within a certain regional year, the total number of users is 1000, the cumulative power failure time is 500 hours, the SAIFI is 0.01 times per user/year, and the SAIDI is 0.5 hours per user/year. The algorithm aims at minimizing the two indexes and improving the overall power supply reliability.
Second, economic goals primarily consider cost factors in the fault recovery process. This includes equipment replacement costs, labor costs, power consumption costs, etc. For example, the cost of replacing a 10kV/400V transformer is about 10 ten thousand yuan, and the labor cost of one emergency repair is about 5000 yuan/hour. The algorithm achieves the economic goal by minimizing the total cost.
Third, the user importance goal reflects the difference in the needs of different users for power supply continuity. The algorithm adopts five-level importance classification, namely particularly important users (such as hospitals and data centers), important users (such as large industrial enterprises), important users (such as small and medium-sized enterprises), general users (such as residential areas) and non-important users (such as temporary electricity utilization). Each class of users is given a different weight, e.g. a particularly important user weight of 1.0, whereas the non-important user weight is 0.2. The algorithm preferably ensures the restoration of power to the high weight user.
In the implementation process, the chromosome coding of the NSGA-II algorithm adopts an integer coding mode, and each gene represents the state of a controllable switch (0 represents open and 1 represents closed). The population size was set at 100 and the algebra of evolution was 500. The crossover operation adopts single-point crossover, the crossover probability is 0.8, the mutation operation adopts uniform mutation, and the mutation probability is 0.1.
In order to speed up convergence and improve the quality of the solution, adaptive crossover and mutation probability adjustment mechanisms are introduced. When the population diversity is reduced, the crossover and mutation probability is increased, otherwise, the crossover and mutation probability is reduced. Specifically, the average fitness value of the population is calculated every 10 generations, if the average fitness value of three continuous calculation changes by less than 1%, the crossover probability is increased by 0.05, and the mutation probability is increased by 0.01 until the preset upper limit is reached.
Constraints of the algorithm include network topology constraints, voltage constraints, and current constraints. The network topology constraint ensures that the recovered network structure is radial, the voltage constraint requires that the voltage deviation of each node is not more than +/-7% of rated voltage, and the current constraint ensures that the line current is not more than 80% of rated capacity. Individuals who violate the constraint punish through a penalty function method, and the selected probability of the individuals is reduced.
During each generation of evolution, individuals are stratified by a rapid non-dominant ranking method, and crowd diversity is maintained by using crowding distance calculation. Finally, the algorithm outputs a set of non-dominant solutions, i.e., pareto optimal solution sets. The decision maker may choose the most suitable failure recovery scheme from the set of solutions according to the actual situation.
In order to verify the effectiveness of the algorithm, simulation tests are performed on the urban distribution network. The distribution network comprises 200 distribution transformers, 300 lines and 500 customer nodes. Medium-scale faults involving 30 user nodes together were simulated. The algorithm runs on an Intel Core i7 processor, a computer with 16GB memory, and takes about 2 minutes to get the optimal solution set. The final selected recovery scheme reduces SAIDI by 15%, SAIDI by 20%, while controlling recovery costs to within 90% of the budget and ensuring that all particularly important users and 95% of the important users resume power within 30 minutes.
Through the multi-objective optimization method, not only can the fault recovery scheme for balancing factors in all aspects be rapidly made, but also a plurality of alternatives are provided for a decision maker, and the flexibility of decision making is enhanced. With continuous optimization of the algorithm and application of more practical cases, the method is expected to become a powerful tool for fault recovery decision of the power distribution network, and the overall reliability and toughness of the power distribution network are improved.
In an alternative embodiment of the present invention,
Simulating a plurality of possible fault isolation and power grid reconstruction schemes by using reinforcement learning technology, and carrying out simulation evaluation on each scheme, wherein the selection of the optimal fault isolation and power grid reconstruction scheme through iterative optimization comprises the following steps:
The method comprises the steps of constructing a power distribution network environment model, defining an environment state, setting an action space, defining a specific fault isolation and network reconstruction scheme for each action, designing a reward function, wherein the environment state is represented by a high-dimensional vector and comprises voltage, current and power factor parameters of each node and the on-off state of a switch;
constructing a deep Q network, wherein the deep Q network comprises a plurality of layers of convolutional neural networks and a full-connection layer, an input layer receives the environment state, and an output layer corresponds to the Q value of each possible action;
Setting an experience playback mechanism, storing a transfer sample by using an experience pool with fixed capacity, setting a target network updating mechanism, updating target network parameters once every fixed training step, adopting an epsilon-greedy strategy to perform action selection, setting the initial epsilon value of the epsilon-greedy strategy to be 1, attenuating fixed proportion in each round until reaching the minimum value, and executing a training process, wherein the method comprises the following substeps:
Resetting an environment model of the power distribution network to obtain an initial state, selecting actions according to an epsilon-greedy strategy, executing the selected actions to obtain rewards and a next state, storing a transfer sample into an experience pool, randomly extracting a batch of samples from the experience pool for training, updating a depth Q network parameter by using a mean square error loss function and an Adam optimizer, and periodically updating a target network;
Performing test evaluation on the trained deep Q network, wherein the test evaluation comprises single fault, multiple faults and cascading failure scenes; and selecting an optimal fault isolation and power grid reconstruction scheme based on the test evaluation result.
Illustratively, a power distribution network environment model is first constructed. The model contains key information such as network topology, load distribution, generator position, switch state and the like. Taking a medium-scale distribution network as an example, the network comprises 200 nodes, 300 lines, 50 controllable switches and 10 distributed generators. The network topology is represented by a graph structure, and each node stores its connection relationship, type (load, generator or transformer) and parameters. The load distribution is generated according to the historical data and the load prediction model, and the daily load curve and the seasonal variation are considered. The generator position and capacity are input according to actual configuration, including renewable energy sources such as photovoltaic, wind power and the like. The switch state is initialized to a normal operating state.
The environmental state is defined next. The environmental state is represented by a high-dimensional vector, which contains the voltage, current, power factor, etc. parameters of each node, and the open and close states of all switches. For the 200 node distribution network described above, the state vector dimension is 800 (3 parameters per node plus 50 switch states). The voltage is expressed in per unit value, typically ranging between 0.95 and 1.05. The current is also normalized by the rated capacity using per unit value. The power factor range is typically between 0.8 and 1. The switch states are indicated by 0 and 1 for closed and open.
Setting the action space is the next step. The action space is defined as the combination of the operations of all controllable switches. For 50 controllable switches, there are theoretically 2≡50 combinations. However, considering practical operational constraints and network security constraints, invalid actions may be reduced by preprocessing, such as disabling simultaneous operation of adjacent switches. A discrete motion space is finally obtained comprising about 10000 effective motions. Each action corresponds to a particular fault isolation and network reconfiguration scheme.
Designing the bonus function is critical to reinforcement learning. The bonus function takes into account four aspects of power restoration, network loss, voltage quality, and number of operations. The power supply restoration degree is represented by a load ratio of restoration power supply, and ranges from 0 to 1. Network loss is expressed as the rate of increase of loss relative to normal operating conditions, and should ideally be close to 1. The voltage quality is expressed as the maximum value of the voltage deviation of all nodes, and the smaller the deviation is, the better the deviation is. The number of operations is expressed in terms of the ratio of the number of switching actions to the maximum number of allowed operations, encouraging the reconstruction to be completed with a minimum of operations. The four indexes are weighted and summed to obtain a final rewarding value, and the weights are adjusted according to actual requirements and can be set to be 0.4, 0.3, 0.2 and 0.1.
Constructing a deep Q network is the core for implementing decisions. The network structure comprises three layers of convolutional neural networks and two layers of full-connection layers. The input layer receives the 800-dimensional state vector, and the output layer corresponds to 10000 possible Q values of actions. The method comprises the following specific structures that a first convolution layer uses 32 3x3 convolution kernels with a step length of 1, a second convolution layer uses 64 3x3 convolution kernels with a step length of 1, a third convolution layer uses 64 3x3 convolution kernels with a step length of 1, a first full connection layer comprises 256 neurons, a second full connection layer comprises 128 neurons, and an output layer comprises 10000 neurons. All hidden layers use a ReLU activation function, and the output layer uses a linear activation function.
To improve learning efficiency and stability, an experience playback mechanism is provided. A pool of experience with a capacity of 100000 is used to store transition samples, each sample containing current state, actions performed, rewards earned and next state. When the experience pool is full, the old samples are replaced with a first-in first-out strategy.
The target network update mechanism is used to stabilize the training process. And updating the target network parameters once every 1000 training steps, and directly copying the parameters of the main network to the target network. And adopting epsilon-greedy strategy to select the action. The initial epsilon value was set to 1, decaying 0.995 per training round, and the minimum value was set to 0.01. Thus, more exploration is performed in the initial stage of training, and learned knowledge is gradually utilized as learning progresses.
Performing the training process includes a number of sub-steps. First, the distribution network environment model is reset, and an initial fault state is randomly generated. Then selecting the action according to epsilon-greedy strategy, randomly selecting epsilon probability, and selecting the action with the largest Q value according to 1-epsilon probability. And executing the selected action, obtaining a new network state through load flow calculation, and calculating a reward value. This transfer sample is stored in an experience pool. 256 samples were randomly drawn from the experience pool to form a batch for training. Using the mean square error as a loss function, the Adam optimizer is used to update the depth Q network parameters with a learning rate set to 0.001. The target network is updated every 1000 steps. Repeating the above steps until reaching the preset termination condition, such as training round number reaching 10000 or average rewarding value becoming stable.
And testing and evaluating the trained deep Q network. Multiple fault scenarios are set, including single faults (e.g., single line tripping), multiple faults (e.g., multiple lines tripping simultaneously), and cascading faults (one fault triggering a cascading reaction). The test is repeated 100 times for each scene, and indexes such as average recovery time, recovery rate, network loss, voltage quality and the like are recorded.
And finally, selecting an optimal fault isolation and power grid reconstruction scheme based on the test evaluation result. A scheme which shows balance on various indexes is generally selected, if fault isolation and reconstruction can be completed within 10 seconds, more than 95% of load is recovered, network loss is controlled within 105% of the original level, and voltage deviation of all nodes is not more than +/-5% of rated voltage. The scheme can quickly restore power supply and ensure the economical efficiency and the safety of the reconstructed network.
In an alternative embodiment of the present invention,
The method comprises the steps of sending a formulated intelligent control instruction sequence to related intelligent terminal equipment through a secure encryption channel, gradually executing control operation according to a preset security protocol after the intelligent terminal equipment receives the instruction, collecting state information of each node in real time in the execution process, feeding the information back to a central control system, rapidly processing feedback information by the central control system locally by using an edge computing technology, evaluating an executing effect of a scheme in real time, and immediately triggering an emergency response mechanism if abnormal conditions or poor executing effect are detected, wherein the dynamic adjustment control strategy comprises the following steps:
performing test evaluation on the trained deep Q network, wherein the test evaluation comprises single fault, multiple faults and cascading failure scenes;
Based on the test evaluation result, selecting an optimal fault isolation and power grid reconstruction scheme;
Converting the selected optimal scheme into an intelligent control instruction sequence, and transmitting the intelligent control instruction sequence to related intelligent terminal equipment through a secure encryption channel based on elliptic curve cryptography;
After receiving the instruction, the intelligent terminal device gradually executes control operation according to a preset safety protocol, wherein the safety protocol comprises environment inspection before operation, operation interval time control and operation sequence forced execution;
in the execution process, the state information of each node is collected in real time through a high-precision sensor network, and the information is fed back to a central control system through a multi-level data convergence structure;
the central control system utilizes a layered edge computing architecture to rapidly process feedback information locally and evaluate the scheme execution effect in real time;
if abnormal conditions or poor execution effects are detected, an emergency response mechanism is immediately triggered, and a multi-objective optimization algorithm is adopted to dynamically adjust a control strategy.
Illustratively, to ensure secure transmission of control instructions, an encryption scheme based on elliptic curve cryptography is employed. First, the central control system exchanges a public key with each intelligent terminal apparatus in advance. Upon sending the instruction, the central system encrypts the instruction content using the AES-256 algorithm, the key being generated by the ECDH key exchange protocol. The encrypted content is digitally signed using the ECDSA algorithm to ensure instruction integrity and non-repudiation.
The transport channel employs TLS 1.3 protocol to provide end-to-end encryption protection. In order to prevent man-in-the-middle attacks, a certificate fixing technology is implemented, and a public key certificate of a legal server is pre-arranged in a terminal device. In addition, a bidirectional authentication mechanism is introduced, and the terminal equipment is required to provide certificates, so that the security is further improved.
In view of the complexity of the distribution network, a distributed instruction issuing strategy is employed. The central system first breaks the global instruction into a plurality of local instructions and then sends the local instructions to the relevant terminals through multiplexing and parallel transmission. Each instruction contains a unique sequence number and time stamp for subsequent synchronous execution and audit trails.
After receiving the instruction, the intelligent terminal equipment gradually executes control operation according to a preset safety protocol;
After receiving the encryption instruction, the intelligent terminal equipment firstly performs decryption and signature verification operations. After verification is passed, the device will check the timeliness and sequence integrity of the instruction, preventing replay attacks and instruction tampering. For the distributed instruction, the device waits for all relevant instructions to arrive and then processes the instructions, so that the consistency of the operation is ensured.
When the control operation is executed, the terminal device strictly follows a preset security protocol. The protocol comprises the mechanisms of environment checking before operation, operation interval time control, operation sequence enforcement and the like. For example, before performing a switching operation, the device checks whether the line current is within a safe range, the operation interval of adjacent switches is not less than 100 ms to prevent grid oscillations, and for a multi-step operation, it is necessary to perform in a predetermined order, no skip or repetition is allowed.
The terminal device also implements limited local decision-making capability in order to cope with possible communication interruption situations. When an interruption in communication with the central system is detected to exceed a preset threshold (e.g., 5 seconds), the device may autonomously perform a portion of the critical operations according to a locally stored emergency plan to maintain the system substantially stable.
In the execution process, the state information of each node is collected in real time, and the information is fed back to a central control system;
In order to realize real-time monitoring of the state of the power distribution network, a high-precision sensor network is deployed at key nodes. The sensors comprise a voltage transformer, a current transformer, a power factor meter and the like, and the sampling frequency can reach 60 times per second, so that transient changes are ensured to be captured. Meanwhile, the intelligent switch equipment also has a state monitoring function, and can provide information such as switch positions, operation times, operation time and the like.
The collected data is transmitted to a central control system through a multi-stage data convergence structure after being locally preprocessed. And an edge computing technology is adopted, an edge server is deployed at a position close to a data source, primary analysis and compression are carried out on the original data, and the transmission bandwidth requirement is reduced. The data transmission adopts a publish-subscribe mode, and the MQTT protocol is used to ensure the real-time performance and reliability of the data.
To cope with large-scale data flows, a distributed flow processing framework, such as APACHE FLINK, is adopted at the central control system end to realize millisecond-level data processing delay. The system also integrates a data quality control module, and ensures the accuracy of feedback information through methods such as data consistency check, abnormal value detection and the like.
The central control system utilizes an edge computing technology to rapidly process feedback information locally and evaluate the scheme execution effect in real time;
The central control system adopts a hierarchical architecture to distribute data processing tasks to computing nodes of different levels. At the edge layer, a lightweight neural network model, such as a compressed version of a convolutional neural network, is deployed for quickly identifying abnormal patterns of grid parameters. The models can adapt to the power grid characteristics of different areas through optimization of the transfer learning technology.
The middle layer adopts fog calculation nodes and is responsible for more complex analysis tasks such as short-term load prediction and tide calculation. The integrated learning method is used, and the prediction accuracy is improved by combining the output of a plurality of machine learning models (such as random forests, gradient lifting trees and the like). And comparing the predicted result with actual feedback data for evaluating the executing effect of the scheme.
The core layer implements global optimization and decision making. And adopting reinforcement learning technology, such as a depth deterministic strategy gradient algorithm (DDPG), and dynamically adjusting parameters of a control strategy according to historical data and current states. The system also integrates a digital twin model, can perform quick simulation, and predicts potential influences of different control strategies.
If abnormal conditions or poor execution effects are detected, an emergency response mechanism is immediately triggered, a control strategy is dynamically adjusted, and a multi-dimensional and multi-scale method is adopted for abnormality detection. In the time dimension, using a long-short-term memory network (LSTM) to detect anomalies in the time-series data, and in the space dimension, using a graph neural network to analyze anomaly correlations between nodes. The system also integrates rule-based expert systems, which encode the empirical knowledge of the power expert, for identifying complex failure modes.
When an anomaly is detected, the system triggers different levels of emergency response based on the severity and scope of the anomaly. For local minor anomalies, the system may make small adjustments, such as fine tuning the timing of operation of certain switches, under the current control strategy framework. For extensive or severe anomalies, the system may initiate a predefined emergency plan, such as a fast switch to backup power or performing load shedding.
When the control strategy is dynamically adjusted, the system can comprehensively consider a plurality of targets, including power supply reliability, power quality, economy and the like. A group of pareto optimal adjustment schemes are rapidly generated by adopting a multi-objective optimization algorithm, such as NSGA-III. And then selecting the most suitable scheme to execute by combining the weight of each target under the current condition through a fuzzy comprehensive evaluation method.
The whole process forms a closed loop control system. The system can continuously monitor the effect of the regulated control strategy and further optimize the regulated control strategy according to new feedback information, so that the power distribution network can keep a stable and efficient running state when facing various abnormal conditions.
Fig. 2 is a schematic structural diagram of an intelligent control system for fault location and rapid reliability recovery of a power distribution network according to an embodiment of the present invention, as shown in fig. 2, where the system includes:
The first unit is used for collecting real-time data of each node of the power distribution network, including voltage, current and power data, and inputting the real-time data into a pre-trained deep learning model; the deep learning model is based on a mixed architecture of a convolutional neural network and a long-short-time memory network, and rapidly identifies fault types and accurately positions fault points by analyzing time sequence features and space distribution features of the real-time data;
The second unit is used for starting a multi-objective optimization algorithm based on the generated preliminary fault evaluation report, wherein the algorithm comprehensively considers factors such as power supply reliability, economy, user importance and the like, simulating a plurality of possible fault isolation and power grid reconstruction schemes by using a reinforcement learning technology, and carrying out simulation evaluation on each scheme;
The third unit is used for sending the formulated intelligent control instruction sequence to the related intelligent terminal equipment through the safe encryption channel, the intelligent terminal equipment gradually executes control operation according to a preset safe protocol after receiving the instruction, collects the state information of each node in real time in the execution process and feeds the information back to the central control system, the central control system rapidly processes the feedback information locally by utilizing the edge computing technology, evaluates the execution effect of the scheme in real time, immediately triggers an emergency response mechanism if abnormal conditions or poor execution effect are detected, dynamically adjusts the control strategy, continuously monitors the recovery state of the system after fault isolation and power grid reconstruction are completed, and generates detailed event reports and optimization suggestions to provide data support for future fault prevention and system improvement.
In a third aspect of an embodiment of the present invention,
There is provided an electronic device including:
A processor;
a memory for storing processor-executable instructions;
wherein the processor is configured to invoke the instructions stored in the memory to perform the method described previously.
In a fourth aspect of an embodiment of the present invention,
There is provided a computer readable storage medium having stored thereon computer program instructions which, when executed by a processor, implement the method as described above.
The present invention may be a method, apparatus, system, and/or computer program product. The computer program product may include a computer readable storage medium having computer readable program instructions embodied thereon for performing various aspects of the present invention.
It should be noted that the above embodiments are merely for illustrating the technical solution of the present invention and not for limiting the same, and although the present invention has been described in detail with reference to the above embodiments, it should be understood by those skilled in the art that the technical solution described in the above embodiments may be modified or some or all of the technical features may be equivalently replaced, and these modifications or substitutions do not make the essence of the corresponding technical solution deviate from the scope of the technical solution of the embodiments of the present invention.