Power distribution network intelligent self-recovery method, device, equipment, medium and program

Abstract

Embodiments of the present application provide a power distribution network intelligent self-recovery method, device, equipment, medium and program. It relates to the technical field of power intelligent management. The method comprises: inputting real-time operation data of a power distribution network into a fault prediction model to obtain predicted fault risks of each node in the power distribution network; if the predicted fault risk indicates that a fault exists, inputting the real-time operation data into a fault diagnosis model to obtain fault probabilities of each node in the power distribution network; determining a fault position of the power distribution network according to the predicted fault risks and the fault probabilities of each node in the power distribution network; generating and executing control instructions for switch devices in the power distribution network according to the predicted fault risks and the fault position to isolate a fault area; and generating and executing power supply recovery instructions for a non-fault area based on an adaptability function and constraint conditions of the power distribution network to recover power supply in the non-fault area. The present application is used to reduce the size of a power outage area and improve power supply reliability.

Description

Technical Field

[0001] This application relates to the field of intelligent power management technology, and in particular to an intelligent self-recovery method, device, equipment, medium and program for power distribution networks. Background Technology

[0002] As the scale of power systems gradually increases, it is necessary to conduct inspections of power systems to diagnose power faults, detect problems in a timely manner, and ensure the normal operation of the power system.

[0003] In related technologies, fault diagnosis of power systems is performed through manual inspections. This involves manually checking key nodes within the distribution network one by one and judging whether there are faults in the power system based on human experience. Afterwards, maintenance is carried out on the faulty areas of the distribution network. However, this method has the problem of not being able to accurately predict potential faults in advance, resulting in long power outage times and wide outage areas, seriously affecting the reliability of power supply. Summary of the Invention

[0004] This application provides a method, apparatus, equipment, medium, and program for intelligent self-recovery of power distribution networks, which can reduce the size of power outage areas and improve power supply reliability.

[0005] In a first aspect, embodiments of this application provide a smart self-recovery method for a power distribution network, comprising:

[0006] By inputting real-time operation data of the distribution network into a fault prediction model based on long short-term memory networks, the predicted fault risk of each node in the distribution network can be obtained.

[0007] If the predicted fault risk indicates that a fault exists, the real-time operating data will be input into the fault diagnosis model based on Bayesian network to obtain the fault probability of each node in the distribution network.

[0008] Based on multi-source data fusion technology, the fault location of the distribution network is determined according to the predicted fault risk and the fault probability of each node in the distribution network. The fault location is used to indicate the node where the fault exists.

[0009] Based on the predicted fault risk and fault location, control commands are generated and executed for switchgear in the distribution network to isolate the fault area;

[0010] Based on the fitness function and constraints of the distribution network, a genetic algorithm and fault recovery strategy are used to generate and execute power restoration instructions for non-faulty areas to restore power supply to non-faulty areas. The fitness function is constructed with the goal of maximizing the restored power load and minimizing the restoration path loss of the distribution network. The constraints include constraints on the line capacity, voltage deviation and short-circuit current of the distribution network.

[0011] In one possible implementation, the real-time operating data of the distribution network is obtained in the following way:

[0012] By interacting with acquisition devices deployed on multiple nodes within the distribution network, node status data used to indicate the operating status of the nodes is obtained;

[0013] By interacting with data acquisition devices deployed on multiple power devices within the distribution network, equipment health status data used to indicate the health status of power devices is obtained;

[0014] By integrating node status data and device health status data, real-time operational data is obtained.

[0015] In one possible implementation, after obtaining the real-time runtime data, the process further includes:

[0016] Based on a preset data quality assessment mechanism, the real-time running data is verified to obtain a quality score for the real-time running data.

[0017] If the quality score of the real-time running data is lower than the preset score threshold, the data verification of the real-time running data will be triggered, and the data verification result will be obtained.

[0018] The data verification result indicates that the verification passed;

[0019] If the data verification result indicates that the verification failed, the real-time operation data of the distribution network should be reacquired.

[0020] In one possible implementation, before inputting real-time operational data into a Bayesian network-based fault diagnosis model to obtain the fault probability of each node in the distribution network, the method further includes:

[0021] Acquire real-time meteorological data and equipment aging data;

[0022] Based on real-time meteorological data and equipment aging data, the fault diagnosis model is updated by adjusting the parameters in the fault diagnosis model.

[0023] In one possible implementation, the intelligent self-recovery method for distribution networks further includes:

[0024] Based on the model performance testing strategy, the fault prediction model is periodically tested to obtain the real-time performance of the fault prediction model.

[0025] If the real-time performance is less than the performance threshold, the fault prediction model will be optimized and updated.

[0026] One possible implementation involves optimizing and updating the fault prediction model in the following way:

[0027] The model is trained using the Adam optimizer, with the cross-entropy loss function as the loss function.

[0028] Based on the dynamic learning rate adjustment strategy, the learning rate is exponentially decayed using the formula l_r=l_{r0}\times\gamma^n, where l_{ro} is the initial learning rate, \gamma is the decay factor, n is the total number of training rounds, and r is the current training round.

[0029] Secondly, embodiments of this application provide a smart self-recovery device for a power distribution network, comprising:

[0030] The prediction module is used to input real-time operating data of the distribution network into a fault prediction model based on long short-term memory network to obtain the predicted fault risk of each node in the distribution network.

[0031] The diagnostic module is used to input real-time operating data into a Bayesian network-based fault diagnosis model if the predicted fault risk indicates the presence of a fault, thereby obtaining the fault probability of each node in the distribution network; based on multi-source data fusion technology, it determines the fault location of the distribution network according to the predicted fault risk and the fault probability of each node in the distribution network, and the fault location is used to indicate the node with the fault.

[0032] The processing module generates and executes control commands for switching equipment in the distribution network based on predicted fault risk and fault location to isolate faulty areas. Based on the fitness function and constraints of the distribution network, it uses a genetic algorithm and fault recovery strategy to generate and execute power restoration commands for non-faulty areas to restore power supply to non-faulty areas. The fitness function is constructed with the goal of maximizing the restored power load and minimizing the restoration path loss of the distribution network. The constraints include constraints on the line capacity, voltage deviation, and short-circuit current of the distribution network.

[0033] In one possible implementation, the real-time operating data of the distribution network is obtained in the following way:

[0034] By interacting with acquisition devices deployed on multiple nodes within the distribution network, node status data used to indicate the operating status of the nodes is obtained;

[0035] By interacting with data acquisition devices deployed on multiple power devices within the distribution network, equipment health status data used to indicate the health status of power devices is obtained;

[0036] By integrating node status data and device health status data, real-time operational data is obtained.

[0037] In one possible implementation, after obtaining real-time operational data, the quality inspection module is specifically used for:

[0038] Based on a preset data quality assessment mechanism, the real-time running data is verified to obtain a quality score for the real-time running data.

[0039] If the quality score of the real-time running data is lower than the preset score threshold, the data verification of the real-time running data will be triggered, and the data verification result will be obtained.

[0040] The data verification result indicates that the verification passed;

[0041] If the data verification result indicates that the verification failed, the real-time operation data of the distribution network should be reacquired.

[0042] In one possible implementation, the diagnostic module is further configured to:

[0043] Acquire real-time meteorological data and equipment aging data;

[0044] Based on real-time meteorological data and equipment aging data, the fault diagnosis model is updated by adjusting the parameters in the fault diagnosis model.

[0045] In one possible implementation, the prediction module is further configured to:

[0046] Based on the model performance testing strategy, the fault prediction model is periodically tested to obtain the real-time performance of the fault prediction model.

[0047] If the real-time performance is less than the performance threshold, the fault prediction model will be optimized and updated.

[0048] One possible implementation involves optimizing and updating the fault prediction model in the following way:

[0049] The model is trained using the Adam optimizer, with the cross-entropy loss function as the loss function.

[0050] Based on a dynamic learning rate adjustment strategy, using the formula... The learning rate decays exponentially, where, The initial learning rate, is the decay factor, n is the total number of training rounds, and r is the current training round number.

[0051] Thirdly, embodiments of this application provide a smart self-recovery device for a power distribution network, comprising: a memory and a processor;

[0052] The memory stores the instructions that the computer executes;

[0053] The processor executes computer execution instructions stored in memory, causing the processor to perform the first aspect and / or various possible implementations of the first aspect as described above.

[0054] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the first aspect and / or various possible implementations of the first aspect.

[0055] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the first aspect and / or various possible implementations of the first aspect.

[0056] The intelligent self-recovery method, device, equipment, medium, and program for distribution networks provided in this application improve the accuracy and timeliness of distribution network fault prediction, diagnosis, and location by using real-time operation data of the distribution network for fault risk prediction and further obtaining the fault probability of each node after detecting fault risk. Combined with multi-source data fusion technology, the fault location can be determined, thus providing a reliable basis for rapid isolation of fault areas. By generating and executing control commands for switching equipment based on predicted fault risk and fault location, and using a fitness function constructed with the objectives of maximizing the restored power load and minimizing the restoration path loss, combined with constraints such as line capacity, voltage deviation, and short-circuit current, the power restoration commands for non-fault areas are generated and executed using genetic algorithms and fault recovery strategies. This improves the adaptability and reliability of recovery decisions while meeting the operational constraints of the distribution network, thereby achieving efficient power restoration in non-fault areas. Attached Figure Description

[0057] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0058] Figure 1 A schematic diagram illustrating a scenario for the intelligent self-recovery method for power distribution networks provided in this application embodiment;

[0059] Figure 2 A flowchart illustrating the intelligent self-recovery method for power distribution networks provided in this application embodiment;

[0060] Figure 3 This is a schematic diagram of the structure of the intelligent self-recovery device for power distribution networks provided in the embodiments of this application;

[0061] Figure 4 This is a schematic diagram of the structure of the intelligent self-recovery device for power distribution networks provided in the embodiments of this application.

[0062] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0063] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application.

[0064] Distribution network fault handling belongs to the field of power system automation and intelligent operation and maintenance technology. It is mainly applied in urban distribution networks, industrial park distribution networks and medium and low voltage power supply networks with distributed power sources, and is used to realize fault identification, fault isolation and power supply restoration under complex operating conditions.

[0065] In this type of scenario, the operating system typically relies on field sensors, metering devices, communication networks, and scheduling or monitoring platforms to continuously collect data such as voltage, current, power, and switch status, and analyze and control the operating status of the distribution network accordingly.

[0066] In related technologies, power distribution network fault handling schemes typically use rule control, traditional machine learning, or general neural networks to analyze operating data. After detecting an anomaly, they combine status monitoring information to determine the fault type and location, and execute switch opening and closing and load transfer according to preset logic, thereby completing the isolation of the fault area and the restoration of power supply to the non-fault area.

[0067] The aforementioned schemes fail to adequately characterize the continuous temporal changes and long-term correlations in distribution network operation data, making it difficult to effectively reflect the dynamic relationship between load fluctuations, equipment status evolution, and fault formation. Furthermore, the lack of a unified collaborative modeling mechanism among data from different sources results in a weak connection between fault prediction and fault diagnosis results, making it prone to misjudgments, omissions, or location errors under complex weather conditions, equipment aging, or changes in operating modes.

[0068] Furthermore, many restoration decisions in related technologies rely on static rules or a single optimization objective, making it difficult to simultaneously consider factors such as the load to be restored, restoration path losses, and operational constraints like line capacity, voltage deviation, and short-circuit current. This results in insufficient restoration efficiency, suboptimal restoration paths, and even impacts power supply safety and system reliability. Therefore, improving the accuracy of distribution network fault prediction and diagnosis, the timeliness of fault location, and the adaptability of restoration decisions under complex dynamic operating conditions has become an urgent technical problem to be solved.

[0069] This application provides an intelligent self-recovery method for distribution networks. Real-time operational data of the distribution network is input into a fault prediction model based on a long short-term memory network to obtain the predicted fault risk of each node. When a fault is predicted, the real-time operational data is then input into a fault diagnosis model based on a Bayesian network to obtain the fault probability of each node, and the fault location is determined by multi-source data fusion. Control commands for switching equipment are generated and executed based on the predicted fault risk and fault location to isolate the fault area. Furthermore, based on a fitness function constructed with the objectives of maximizing the restored power load and minimizing the restoration path loss, and combined with line capacity, voltage offset, and short-circuit current constraints, a genetic algorithm and fault recovery strategy are used to generate and execute power restoration commands for non-faulty areas. This method is applicable to distribution network monitoring and control architectures composed of distributed sensors, smart meters, communication networks, edge computing nodes, and a central control platform, and can provide a foundation for the development of subsequent technical solutions.

[0070] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0071] Figure 1 This is a schematic diagram illustrating a scenario for the intelligent self-recovery method for power distribution networks provided in an embodiment of this application. For example... Figure 1 As shown, the specific application scenario of this application embodiment includes multiple data acquisition devices 11 and a data center 12 deployed in the power distribution network, wherein:

[0072] The data acquisition device 11 and the data center 12 are communicatively connected. Optionally, the communication connection includes a wired communication connection and a wireless communication connection.

[0073] Data center 12 acquires node status data and equipment health status data collected by acquisition device 11 through a communication connection with acquisition device 11. Optionally, acquisition device 11 actively reports data according to a preset sampling period, or triggers event reporting when an abnormal state is detected, to ensure that data center 12 can keep abreast of the power distribution network operation status.

[0074] Data Center 12 integrates node status data and device health status data into real-time operational data. The integration process includes data time synchronization, format unification, quality verification, and multi-source fusion, forming a structured and standardized real-time operational dataset for the distribution network. Data Center 12 can also execute intelligent self-recovery methods for the distribution network based on real-time operational data to isolate faulty areas and restore power to non-faulty areas within the distribution network. Specifically, Data Center 12 first uses real-time operational data for fault detection and location, identifying the location and type of fault. Then, based on the fault location results, it automatically generates a fault isolation strategy, operating relevant switching equipment via remote commands to quickly isolate the faulty area from the distribution network. Finally, after fault isolation is completed, normal power supply to non-faulty areas is restored through network reconfiguration and backup power switching, minimizing the scope and duration of power outages and improving the reliability and intelligent operation and maintenance level of the distribution network.

[0075] Figure 2 This is a flowchart illustrating the intelligent self-recovery method for power distribution networks provided in an embodiment of this application. Figure 2 As shown, the method includes:

[0076] S201. Input the real-time operation data of the distribution network into the fault prediction model based on the long short-term memory network to obtain the predicted fault risk of each node in the distribution network.

[0077] The fault prediction model based on long short-term memory network is used to model the time-series operation data of the distribution network that is continuously collected, and output the fault risk value of each node at the current time or within the preset prediction window.

[0078] In practice, the acquired real-time operational data can be preprocessed using conventional methods and then input into the fault prediction model to form time series samples for the corresponding nodes. Specifically, if Ri represents the predicted fault risk of the i-th node, then the closer Ri is to 1, the higher the probability that the node will fail within the preset time window.

[0079] By inputting real-time operational data into a long short-term memory network in a time-series manner, it is possible to identify trend changes related to the formation of a fault before the fault is fully manifested. This provides a priori risk basis for subsequent diagnosis and isolation steps, transforming fault handling from a reactive response to a proactive response with predictive features.

[0080] Optionally, the fault prediction model based on long short-term memory networks is trained in the following way:

[0081] First, the preprocessed data is divided into training and testing sets according to time series to avoid interference with model evaluation due to future information leakage. Based on this, a network structure is constructed consisting of an input layer, a Long Short-Term Memory (LSTM) layer, a Dropout layer, a fully connected layer, and an output layer.

[0082] The input layer receives sequence data consisting of multiple time steps and multi-dimensional features. The LSTM layer, as the core of the model, uses a forget gate to filter out historical information that needs to be discarded. The forget gate can be represented as:

[0083]

[0084] Among them, the Gate of Oblivion Decide which information from the cell state Ct−1 at the previous moment should be retained and which should be discarded; This indicates that the hidden state from the previous moment will be restored. Input at the current moment splicing; Represented as the weight matrix of the forget gate; The bias term representing the forget gate; The Sigmoid activation function is used to perform the forget gate. The output value is compressed to the (0,1) range.

[0085] like A value close to 1 indicates that historical information for that dimension is retained; if... When the value is close to 0, it indicates that the information has been forgotten. In fault prediction scenarios, the forget gate can automatically identify and filter out early normal operating condition data that is unrelated to equipment degradation.

[0086] LSTM layers control the writing of new information through input gates and candidate cell states. The input gate... It can be represented as:

[0087]

[0088] Input gate Decide on the information to input at the current moment What percentage of the data will be written into the cell state? This represents the weight matrix of the input gate; This represents the bias term of the input gate. The sigmoid activation function is used to gate the input. The output value is compressed to the (0,1) range. When new abnormal fluctuations occur in the monitored data, the input gate will increase accordingly, allowing more new features to enter the memory unit.

[0089] Candidate cell state It can be represented as:

[0090]

[0091] Candidate cell state This represents the candidate value for new information to be written at the current moment. The weight matrix representing the candidate states; The bias term represents the candidate state; simultaneously, through the activation function... A nonlinear transformation is applied to the input to generate new candidate memory content. Candidate cell states. Will be related to the input gate After multiplication, the mixture is selectively superimposed onto the cellular state.

[0092] LSTM layers regulate the output of cell states through output gates, which include:

[0093]

[0094] Output gate Control the current cell state How much information will be output as a hidden state? ; This represents the weight matrix of the output gate; This represents the bias term of the output gate. The design of the output gate allows LSTM to selectively expose its internal memory: in fault prediction, when the model determines that the current time step contains key degradation features, the output gate will increase, passing important information to subsequent layers; conversely, it will suppress the propagation of irrelevant information.

[0095] Meanwhile, the LSTM layer completes the long-term transmission of information based on the cell state update formula and the hidden state formula. The cell state update formula can be expressed as:

[0096]

[0097] The cell state update formula is the core mechanism by which LSTM achieves long-term memory. It consists of two linearly superimposed parts: the first part For the "forgetting path", through the forgetting gate Selective preservation of historical cell states; Part Two For "input path", through the input gate New information about candidates Selective absorption is achieved. Because the update process corresponding to the cell state update formula is additive, the gradient can be stably propagated along the time step, effectively alleviating the gradient vanishing problem in traditional recurrent neural networks (RNNs). In fault prediction, the cell state can continuously accumulate information about the gradual changes in the device from normal to degraded.

[0098] The hidden state formula can be expressed as:

[0099]

[0100] Hidden state It is the final output of the LSTM layer, and also serves as the input for the next time step in the loop.

[0101] Hidden state The calculation is divided into two steps: first, through The function will determine the cell state. The values ​​are mapped to the interval (-1, 1) to achieve information normalization and activation. Then, the output gate is used. For activation value Weighted filtering is performed to obtain the hidden representation currently exposed to the outside world. Hidden state. It carries both filtered long-term memory information and has a numerical range suitable for input into fully connected layers, ultimately used for fault prediction and decision-making.

[0102] The above gating mechanisms work together to enable LSTM to effectively capture long-term dependencies in time series.

[0103] A Dropout layer is added after the LSTM layer to randomly drop some neurons with a probability of 0.2, suppressing the model's overfitting to the training data. The fully connected layer maps the high-dimensional temporal features output by the LSTM layer to the target output dimension. The output layer provides the fault prediction result in a binary classification form, where 0 indicates no fault and 1 indicates a fault.

[0104] The model optimizes the loss function on the training set using the backpropagation algorithm, and verifies its generalization ability on the test set, ultimately obtaining an LSTM fault prediction model that can be used for real-time prediction of equipment faults.

[0105] Optionally, an optimizer can be used to accelerate training during the fault prediction model training process. The choice of optimizer directly affects the model's convergence speed and final performance. Commonly used optimizers include stochastic gradient descent, adaptive gradient algorithm, root mean square propagation algorithm, and adaptive moment estimation algorithm. Among them, the Adam optimizer combines the advantages of momentum method and adaptive learning rate. By calculating the first and second moment estimates of the gradient, it dynamically adjusts the learning rate of each parameter, enabling rapid convergence in the early stages of training and fine-tuning of parameters in the later stages. It is suitable for training most fault prediction models.

[0106] For example, in the training process of the fault prediction model, the Adam optimizer is used to train the model, and the loss function is the cross-entropy loss function. A dynamic learning rate adjustment strategy is introduced, initially using a larger learning rate to accelerate convergence, and in the r-th training round, the learning rate used in the r-th round is determined by the learning rate decay formula. :

[0107]

[0108] in, The initial learning rate, Attenuation factor ( The range of values ​​is usually 100. ), where n is the current training epoch. As the training epoch n increases, the learning rate... By gradually decreasing the parameters exponentially, the model can quickly approach the optimal solution in the early stages of training. In the later stages of training, the parameters are finely adjusted to avoid oscillations around the optimal solution, thereby improving the convergence stability and prediction accuracy of the fault prediction model.

[0109] S202. If the predicted fault risk indicates the presence of a fault, the real-time operating data will be input into the fault diagnosis model based on Bayesian networks to obtain the fault probability of each node in the distribution network.

[0110] A fault diagnosis model based on Bayesian networks is used to perform probabilistic inference on observation information related to node operation after a fault risk is triggered, and to form the posterior probability of each node's fault.

[0111] Specifically, the first step is to determine whether the predicted failure risk indicates the existence of a failure.

[0112] Optionally, the system can determine whether the predicted fault risk of each node exceeds a preset risk threshold. When the predicted fault risk of at least one node exceeds the preset risk threshold, the system determines that a fault exists and inputs the real-time running data into the Bayesian network-based fault diagnosis model.

[0113] Optionally, after inputting real-time running data into a fault diagnosis model based on Bayesian networks, the failure probability of each node is calculated based on Bayesian inference.

[0114] Optionally, corresponding fault diagnosis models can be established based on Bayesian networks for different ranges, that is, the fault probability can be determined by different fault diagnosis models in different nodes of the distribution network.

[0115] Alternatively, a unified fault diagnosis model can be used to centrally infer the areas where risks are triggered in all nodes of the distribution network.

[0116] Based on the above analysis, it is clear that this step does not involve inputting real-time operational data into the Bayesian network-based fault diagnosis model at all times. Instead, it is initiated only after the fault prediction results meet the triggering conditions. Therefore, while controlling the computational load, probabilistic inference can be used to compensate for the shortcomings of Long Short-Term Memory (LSM) networks in expressing causal relationships. The predicted fault risk obtained through the LSM-based fault prediction model reflects the temporal evolution trend; the fault probability obtained through the Bayesian network-based fault diagnosis model reflects the diagnostic reliability under current observational evidence. The combination of predicted fault risk and fault probability can reduce misjudgments and omissions under complex operating conditions.

[0117] S203. Based on multi-source data fusion technology, the fault location of the distribution network is determined according to the predicted fault risk and the fault probability of each node in the distribution network. The fault location is used to indicate the node where the fault exists.

[0118] Multi-source data fusion technology is used to uniformly and collaboratively process predicted fault risks and fault probabilities from different sources and with different mechanisms. Among them, the predicted fault risk reflects time-series prediction information, and the fault probability reflects evidence-based diagnostic information. After fusion, a node-level comprehensive judgment result is formed, and the fault location is completed accordingly.

[0119] Specifically, multi-source data fusion technology uses nodes as the object to perform fusion analysis on predicted fault risks and fault probabilities to obtain a comprehensive fault judgment for each node. Then, the fault location is determined based on the comprehensive fault judgment.

[0120] For example, if the overall fault assessment points to a single node, then that node is identified as the fault location.

[0121] For example, if the comprehensive fault diagnosis points to multiple adjacent nodes, the corresponding area is identified as the fault area for subsequent isolation control.

[0122] By unifying and integrating the two types of results—fault risk and fault probability—a connection mechanism is established between risk prediction and fault location. This enables the fault location to no longer rely solely on the output of a single algorithm, and to obtain more stable location results under complex operating conditions.

[0123] Optionally, for N nodes in the distribution network, a weighted analysis is performed by predicting fault risk and fault probability to obtain the comprehensive fault judgment for each node.

[0124] For example, comprehensive fault judgment can be expressed as the comprehensive fault probability of each node. Wherein, the comprehensive fault probability... It can be represented as:

[0125]

[0126] in, Indicates the predicted failure risk. Indicates the probability of failure; This represents the weighting coefficient.

[0127] Subsequently, a dual mechanism of threshold discrimination and association verification is used to determine the fault location. When the maximum value of the overall fault probability of a node is greater than or equal to the node's fault threshold, the node is determined to be the fault location.

[0128] Furthermore, electrical correlation verification, temporal correlation verification, and topological correlation verification are performed on nodes identified as fault locations. Electrical correlation verification requires consistency with upstream protection action information; temporal correlation verification requires that the anomaly detection time of each data source conforms to the fault propagation direction; and topological correlation verification requires that the nodes be located within the intersection range of the fault indicator action segments.

[0129] S204. Based on the predicted fault risk and fault location, generate and execute control commands for switchgear in the distribution network to isolate the fault area.

[0130] Control commands are used to drive switching equipment in the distribution network to perform corresponding operations, thereby eliminating the impact of faults at the network topology level.

[0131] Specifically, based on the node or region corresponding to the fault location, the switching equipment related to fault isolation is determined in the distribution network topology model, and corresponding control commands are generated.

[0132] Control commands are sent to field execution units or switch controllers via the power distribution automation communication link, and the corresponding execution feedback is returned after field execution.

[0133] Centered on the fault location, the faulty area is cut off from the operating network through executable and feedback-enabled switching control actions, forming a non-faulty power supply area and a region to be restored, providing an isolated network structure basis for subsequent power restoration calculations.

[0134] S205. Based on the fitness function and constraints of the distribution network, a genetic algorithm and fault recovery strategy are used to generate and execute power restoration instructions for non-faulty areas to restore power supply to non-faulty areas. The fitness function is constructed with the goal of maximizing the restored power load and minimizing the restoration path loss of the distribution network. The constraints include constraints on the line capacity, voltage deviation and short-circuit current of the distribution network.

[0135] Genetic algorithms are used to search for feasible recovery schemes in the network state after fault isolation. Fault recovery strategies are used to specify the path selection for power restoration, the sequence of switching operations, and the order of load restoration.

[0136] First, multiple candidate recovery scheme codes are generated based on the isolated real-time topology. Each candidate recovery scheme code corresponds to a set of switch open / closed states and a corresponding power supply path. Optionally, the candidate recovery scheme code should at least include information such as the tie switch state, the key section switch state, and the connection relationship of the load section to be restored. Then, a fitness function is constructed with the objectives of maximizing the restored power supply load and minimizing the recovery path loss. The restored power supply load is calculated based on the total load of the nodes after restoration, and the recovery path loss is calculated based on the line resistance, current distribution, and path length on the candidate paths.

[0137] For example, the fitness function F can be expressed as F = λL - μP, where L represents the restored power load, P represents the restoration path loss, and λ and μ are target weight coefficients. If a unified dimension is required, L and P should be standardized before being substituted into the calculation.

[0138] The constraints are verified simultaneously during the coding and evaluation of each candidate recovery scheme. Among them, the line capacity constraint requires that the power flow of each branch does not exceed its rated current carrying capacity, the voltage offset constraint requires that the voltage of each node after recovery be within the allowable range, and the short-circuit current constraint requires that the short-circuit current level of the network after recovery does not exceed the breaking capacity and thermal stability limits of the switching equipment and lines.

[0139] For example, the requirement that the power flow of each branch should not exceed its rated current carrying capacity can be expressed as:

[0140]

[0141] in, Indicates the tidal current of a branch line; Indicates branch current Rated current carrying capacity.

[0142] The initial population of the genetic algorithm consists of multiple candidate recovery scheme codes that satisfy the basic topological connectivity condition. Then, fitness calculation, individual selection, crossover, and mutation operations are sequentially performed on the candidate recovery scheme codes. When the population of candidate recovery scheme codes reaches a preset number of generations or the optimal fitness no longer improves over several consecutive generations, the recovery scheme is output.

[0143] In this process, individual selection can employ roulette wheel or tournament methods to retain high-fitness individuals. Crossover is used to swap path combinations in the encoding of different candidate recovery schemes. Mutation is used to randomly change the on / off states within the encoding of candidate recovery schemes to expand the search space. Simultaneously, the genetic algorithm also retains the current best individual for the next generation to maintain search convergence stability.

[0144] The fault recovery strategy is embedded as a constraint rule in the genetic algorithm, including restoring important load nodes first and then general load nodes, prioritizing connection paths with sufficient upstream backup capacity and voltage support capability, and limiting the number of switches that operate simultaneously during the recovery process.

[0145] If a candidate solution violates any constraint in the fault recovery strategy, it is determined to be an infeasible individual and a penalty is imposed or it is directly eliminated.

[0146] Afterwards, power restoration instructions are generated according to the restoration plan, and closing or adjustment instructions are issued to the corresponding tie switches and sectionalizing switches in the order specified in the restoration plan. After each operation, the line current, node voltage and network topology status are checked by measurement feedback. The next operation is executed only after the constraints are confirmed to be met.

[0147] By incorporating the maximization of restored power load and the minimization of restoration path losses into a unified objective, and simultaneously satisfying line capacity, voltage offset, and short-circuit current constraints during the search process, the power restoration command is not a static rule output, but rather an executable solution calculated based on the current isolated network state. Furthermore, by iteratively screening various switch combinations using a genetic algorithm, restoration paths and sequences adapted to the current fault scenario can be formed under complex topologies and multiple constraints, thereby completing the power restoration of non-faulty areas.

[0148] This application provides an intelligent self-recovery method for distribution networks. By connecting fault prediction, fault diagnosis, fault location, fault isolation, and recovery reconstruction through a unified data link and control link, a closed-loop self-recovery process is formed for complex dynamic operating conditions. Specifically, a Long Short-Term Memory (LSTM) network is used to extract temporal evolution information from operational data; a Bayesian network is used for probabilistic diagnosis; multi-source data fusion is used to output more consistent fault locations; isolation control is used to construct clear recovery boundaries; and a genetic algorithm and fault recovery strategy are used to search for recovery schemes under operational constraints. This enables integrated processing of fault identification, location, isolation, and recovery of non-faulty areas under conditions of load fluctuations, weather changes, and equipment status changes, ensuring that the fault processing results remain consistent with the current actual network state.

[0149] In one possible implementation, the real-time operation data of the distribution network is obtained by: interacting with acquisition devices deployed on multiple nodes within the distribution network to obtain node status data indicating the node's operating status; interacting with acquisition devices deployed on multiple power devices within the distribution network to obtain device health status data indicating the power device's health status; and integrating the node status data and device health status data to obtain real-time operation data.

[0150] Node status data may specifically include at least one of the following: node voltage, line current, switch open / closed status, power flow, and node load change information. Equipment health status data may specifically include at least one of the following: temperature, vibration, partial discharge level, insulation status, and operational alarm information.

[0151] The data acquisition equipment can be installed in locations such as ring main units, distribution transformers, pole-mounted switches, circuit breakers, or smart meters within the power distribution network. Optionally, data can be exchanged with the acquisition equipment via wired or wireless communication to obtain node status data and equipment health status data.

[0152] Optionally, the data acquisition device can be a terminal device that meets the communication protocol for power distribution automation to achieve continuous uploading of field data. In practical applications, other models of the data acquisition device can also be selected, and this embodiment does not limit this.

[0153] The data acquisition device collects node status data and device health status data according to a preset sampling period. The preset sampling period is a value set based on actual operating conditions. Then, it interacts with the acquisition device via wired or wireless communication to obtain the node status data and device health status data. Further, the node status data and device health status data collected by different acquisition devices are then preprocessed. Optionally, the data preprocessing includes at least one of the following: timestamp alignment, format standardization, quality verification, missing value completion, and noise reduction.

[0154] Optionally, if the data preprocessing involves denoising, methods such as time-domain filtering, frequency-domain filtering, or wavelet thresholding can be used. Time-domain filtering smooths the original data over time, suppressing random noise interference. Frequency-domain filtering transforms the data to the frequency domain, filters out high-frequency noise components, and then transforms it back to the time domain. Wavelet thresholding utilizes the multi-resolution analysis characteristics of wavelet transform to separate signal and noise at different scales, and performs thresholding on the noisy wavelet coefficients to achieve denoising.

[0155] For example, methods such as mean filtering can be used to eliminate abnormal fluctuations in node status data and device health status data. Mean filtering calculates the arithmetic mean of data within a fixed-length window in the data sequence as the filter output of the current point, which can effectively smooth out spikes and peaks caused by random errors of sensors or instantaneous interference in communication.

[0156] Optionally, if data preprocessing involves normalization, methods such as Min-Max normalization, Z-Score standardization, or decimal scaling normalization can be used. Min-Max normalization linearly maps data to a specified interval (usually [0,1]), suitable for scenarios where the data distribution range is clear and stable. Z-Score standardization converts data into a standard normal distribution with a mean of 0 and a standard deviation of 1, suitable for scenarios where the data contains outliers or the distribution is approximately normal. Decimal scaling normalization normalizes data by shifting the decimal point, suitable for scenarios where the data volume varies significantly.

[0157] For example, the node status data and equipment health status data are normalized using a normalization formula to eliminate the differences between features of different dimensions and orders of magnitude, so that each feature has equal weight contribution in subsequent model training and fault analysis, and avoids the model from being biased towards features with larger values ​​due to differences in dimensions.

[0158] For example, the normalization formula is:

[0159]

[0160] in, The original data values, It is the minimum value in the dataset for this feature dimension. The maximum value in the dataset for this feature dimension. This represents the normalized data value. This formula linearly transforms the original data to the interval [0,1].

[0161] when hour, ;when hour, When X is between its minimum and maximum values, is proportionally mapped to the (0,1) interval. Through this normalization process, node state data and equipment health state data with different dimensions are unified into the same dimensionless value range, which facilitates subsequent multi-source data fusion analysis and model input.

[0162] Optionally, if the data preprocessing is missing value imputation, methods such as statistical interpolation, machine learning prediction, or spatio-temporal correlation inference can be adopted. Statistical interpolation includes linear interpolation, spline interpolation, Lagrange interpolation, etc., which estimates missing values based on the time-series variation rule of the data itself. The machine learning prediction method uses historical data to train a regression model or a neural network model, and predicts missing values based on adjacent moments or related features. The spatio-temporal correlation inference method utilizes the data correlation of adjacent nodes or similar equipment in the distribution network to infer missing data through spatial neighborhood information.

[0163] As an example, the linear interpolation method is used to interpolate missing values to achieve missing value imputation. The linear interpolation method assumes that the change between adjacent valid data points presents a linear relationship, and proportionally calculates the estimated value of the missing point according to the values and relative positions of the known data points before and after the missing point. It has the characteristics of simple calculation and good real-time performance, and is suitable for short-term missing scenarios where data changes are relatively gentle.

[0164] Optionally, the formula adopted by the linear interpolation method may be:

[0165]

[0166] wherein, is the missing value to be imputed, is the value of the valid data point immediately before the missing value, is the value of the valid data point immediately after the missing value, j is the time index of the previous valid data point, k is the time index of the subsequent valid data point, i is the time index of the position where the missing value is located, and satisfies j<i<k. In the formula, represents the ratio of the time distance of the missing point relative to the interval between the front and rear valid points, is the numerical variation between the front and rear valid points, and the product of the two gives the estimated increment from to the missing point, then adding obtains the linear estimation result of the missing value. This formula is based on the equal ratio assumption, which considers that the data changes uniformly between adjacent valid points, and realizes reasonable inference of missing values through time weight distribution, which can effectively maintain the continuity and trend consistency of the data sequence while ensuring calculation efficiency.

[0167] Next, the processed node status data and device health status data are mapped to a unified data structure for integration, forming real-time operational data that can be used for subsequent fault prediction and fault diagnosis. The data structure can employ a unified field table, a time-series vector, or a relational table to maintain the correspondence between node status and device health status.

[0168] Optionally, the integration of node status data and device health status data can be completed within the edge computing node or the central control platform.

[0169] By integrating node-collected information and equipment health information into real-time operational data, the distribution network's operational status reflects not only changes in electrical quantities at the node level but also the health evolution of the equipment. This provides consistent data input for subsequent fault risk assessment, fault probability calculation, and fault location determination. Because the real-time operational data undergoes synchronous integration and standardization, its temporal relevance and comparability are maintained, thus supporting subsequent joint analysis of the distribution network's dynamic operational status by the model.

[0170] Optionally, the following security measures and communication guarantee mechanisms are adopted in the process of obtaining node status data by interacting with acquisition devices deployed on multiple nodes in the distribution network, and in the process of obtaining equipment health status data by interacting with acquisition devices deployed on multiple power devices in the distribution network.

[0171] Security measures include: using AES encryption algorithm to encrypt transmitted and stored data to ensure the confidentiality and integrity of data during transmission and storage; deploying intrusion detection systems and firewalls to monitor and defend against network attacks in real time, preventing unauthorized intrusion and data tampering; and establishing a strict user access permission management mechanism to allocate differentiated operation permissions based on user roles and responsibilities, thereby achieving hierarchical protection of data and access with the least privilege.

[0172] The communication guarantee mechanism includes: building a redundant communication network of fiber optic and wireless (4G / 5G, LoRa, etc.) links; when the main communication link fails, the system automatically switches to the backup link to ensure the continuity and reliability of data transmission; and designing an offline self-healing strategy to perform limited fault handling based on the data stored locally on the edge device and preset rules in the event of a communication interruption, maintain basic monitoring and protection functions, and synchronize data and restore normal operation mode after communication is restored.

[0173] In one possible implementation, after obtaining the real-time operating data, the method further includes: performing quality verification on the real-time operating data based on a preset data quality assessment mechanism to obtain a quality score for the real-time operating data; if the quality score of the real-time operating data is lower than a preset score threshold, then triggering data verification of the real-time operating data to obtain a data verification result; determining that the data verification result indicates that the verification passed; if the data verification result indicates that the verification failed, then re-acquiring the real-time operating data of the distribution network.

[0174] A data quality assessment mechanism is used to quantitatively evaluate real-time operational data from data acquisition devices at distribution network nodes and power equipment. A quality score characterizes the degree to which the real-time operational data meets the input requirements for subsequent fault prediction and diagnosis. Data validation is used to further verify abnormal data when the score is below a threshold.

[0175] The quality score can be calculated by combining three indicators: completeness, accuracy, and consistency. Completeness is used to characterize the proportion of missing values, accuracy is used to characterize the degree of deviation of outliers, and consistency is used to characterize the degree of matching between data collected from multiple sources at the same time.

[0176] Based on a pre-defined data quality assessment mechanism, when verifying the quality of real-time operational data, the system analyzes the timestamps, sampling periods, field integrity, and numerical ranges of the data, and calculates scores for each indicator according to pre-defined weights. The scores for each indicator are then aggregated to form a quality score.

[0177] If the quality score is lower than the preset score threshold, the data verification module is activated to verify outliers, missing values ​​and duplicate values ​​in the real-time running data, and cross-compares them with redundant sampled data from different nodes to output the data verification results.

[0178] When the data verification result indicates that the verification is successful, the real-time running data is sent to the subsequent fault prediction or fault diagnosis stage.

[0179] When the data verification result indicates that the verification fails, a new data acquisition request is sent to the data acquisition devices corresponding to multiple nodes and multiple power devices in the distribution network to reacquire real-time operating data.

[0180] By employing the above methods, the data entering subsequent models maintains a stable quality level, ensuring the availability of real-time operational data in fault prediction, fault diagnosis, and recovery decision-making, and reducing input deviations caused by abnormal acquisition, communication jitter, or local distortion, thereby improving the reliability of the results of intelligent self-recovery processing of the distribution network.

[0181] In one possible implementation, before inputting real-time operational data into a Bayesian network-based fault diagnosis model to obtain the fault probability of each node in the distribution network, the method further includes: acquiring real-time meteorological data and equipment aging data; and updating the fault diagnosis model by adjusting the parameters in the fault diagnosis model based on the real-time meteorological data and equipment aging data.

[0182] The model parameters of the fault diagnosis model can characterize the prior distribution and conditional probability relationship of faults under different meteorological conditions and different equipment health states.

[0183] Real-time meteorological data can be obtained from meteorological monitoring terminals, including rainfall, temperature, humidity, wind speed, and lightning warning information, and converted into environmental feature vectors corresponding to the diagnostic model.

[0184] The specific data on equipment aging can be determined by comprehensively considering the equipment's service life, maintenance records, partial discharge characteristics, abnormal temperature rise, and historical fault records, and then used to form an aging coefficient or health score input model.

[0185] After acquiring real-time meteorological data and equipment aging data, the parameter weights in the fault diagnosis model are reallocated based on the real-time meteorological data and equipment aging data, and the parameter items corresponding to meteorological sensitivity and equipment degradation sensitivity are corrected.

[0186] Optionally, the process of redistributing parameter weights in the fault diagnosis model can be implemented based on Bayesian parameter correction, that is, using real-time meteorological data to adjust the prior probabilities of different fault types, using equipment aging data to correct the node fault conditional probability table, or using a sliding time window to incrementally train recent samples to obtain an updated fault diagnosis model.

[0187] Optionally, to adapt to different distribution network scales, the parameter weights in the fault diagnosis model can be reallocated within the edge computing node or the central control platform.

[0188] Through the above update method, the fault diagnosis model can synchronously correct its internal parameters by combining the current environmental state and the equipment degradation state, so that the node failure probability output when subsequent real-time operating data is input is more in line with the actual operating conditions, and improves the diagnostic consistency and positioning reliability under complex weather and aging equipment conditions.

[0189] In one possible implementation, the intelligent self-recovery method for distribution networks further includes: periodically performing performance testing on the fault prediction model based on a model performance testing strategy to obtain the real-time performance of the fault prediction model; if the real-time performance is less than the performance threshold, optimizing and updating the fault prediction model.

[0190] Optionally, the model performance detection strategy can be invoked by the monitoring and scheduling module at preset time intervals, or it can be triggered after the fault prediction model completes a preset number of inference tasks.

[0191] When periodically testing the performance of a fault prediction model based on a model performance testing strategy, recently collected operational data and corresponding real fault results or confirmed labeled results are input into the evaluation unit. Evaluation metrics are calculated based on the consistency between the predicted results and the real results. Optionally, the evaluation metrics include at least one of the following: accuracy, recall, F1 score, or area under the curve (AUC).

[0192] The calculated evaluation index is then used as the real-time performance of the fault prediction model.

[0193] Furthermore, the performance threshold can be pre-set as a single indicator threshold or a comprehensive indicator threshold based on historical operating data, initial model training results, and distribution network operation scenario requirements.

[0194] When real-time performance falls below the performance threshold, the fault prediction model can be retrained or its parameters fine-tuned by calling the incremental training dataset. Optimization and updates can be completed by combining learning rate reset, weight decay adjustment, and sample reweighting.

[0195] When real-time performance falls below a performance threshold, the fault prediction model is updated, enabling its parameters to adjust synchronously with changes in distribution network topology, load fluctuations, and fault mode evolution, thus maintaining the effectiveness of fault risk prediction. Because a clear triggering relationship is established between real-time performance and the performance threshold, the fault prediction model can be optimized and updated before performance degrades to the point of affecting self-recovery decisions, thereby maintaining the continuity and stability of subsequent fault location and power restoration processes.

[0196] One possible implementation involves optimizing and updating the fault prediction model using the following method: training the model with the Adam optimizer and the cross-entropy loss function; and adjusting the learning rate dynamically using the formula... The learning rate decays exponentially, where, The initial learning rate, is the decay factor, n is the total number of training rounds, and r is the current training round number.

[0197] The fault prediction model can be constructed as a classification prediction network based on a Long Short-Term Memory (LSTM) network. The output of the fault prediction model generates the fault risk probability for each node. The Adam optimizer iteratively updates the parameters of the LSTM network and uses first and second moments to estimate the adaptive adjustment of the parameter update magnitude. The cross-entropy loss function measures the difference between the predicted probability and the fault label. The fault labels used to optimize and update the fault prediction model can be obtained from historical fault records, alarm records, and manual review results.

[0198] A dynamic learning rate adjustment strategy is used to control the step size of parameter updates during training. A preset initial learning rate is used in the early stages of training to accelerate model convergence. As the number of training epochs r increases, the learning rate decreases according to the exponential decay formula. Learning rate for the r-th round Update. Attenuation factor. Used to limit the rate at which the learning rate decreases, allowing the model to continue approaching the optimal parameters with smaller steps in subsequent training phases. Initial learning rate Attenuation factor The number of training epochs, r, can be pre-written into the training configuration file, and a new learning rate can be calculated in real time based on the current epoch number after each training epoch.

[0199] During the optimization and update of the fault prediction model, when the training error obtained based on cross-entropy loss drops below the preset threshold, the training process continues to perform parameter fine-tuning according to the decayed learning rate, thereby completing the optimization and update of the fault prediction model.

[0200] By employing the Adam optimizer and cross-entropy loss function in conjunction with an exponentially decaying learning rate, the model can maintain a relatively fast parameter correction speed in the early stages of training and maintain a stable convergence state in the later stages, making the fault risk probability output more consistent with the real fault distribution.

[0201] Figure 3 This is a schematic diagram of the structure of the intelligent self-recovery device for power distribution networks provided in the embodiments of this application, as shown below. Figure 3 As shown, the intelligent self-recovery device 30 for power distribution networks provided in this embodiment includes:

[0202] The prediction module 301 is used to input the real-time operation data of the distribution network into the fault prediction model based on the long short-term memory network to obtain the predicted fault risk of each node in the distribution network.

[0203] The diagnostic module 302 is used to input real-time operating data into a fault diagnosis model based on a Bayesian network if the predicted fault risk indicates that a fault exists, to obtain the fault probability of each node in the distribution network; based on multi-source data fusion technology, it determines the fault location of the distribution network according to the predicted fault risk and the fault probability of each node in the distribution network, and the fault location is used to indicate the node with the fault.

[0204] The processing module 303 is used to generate and execute control commands for switching equipment in the distribution network based on the predicted fault risk and fault location to isolate the fault area; based on the fitness function and constraints of the distribution network, it uses a genetic algorithm and fault recovery strategy to generate and execute power supply recovery commands for non-fault areas to restore power supply to non-fault areas. The fitness function is constructed with the goal of maximizing the restored power supply load and minimizing the restoration path loss of the distribution network. The constraints include constraints on the line capacity, voltage deviation and short-circuit current of the distribution network.

[0205] In one possible implementation, the real-time operating data of the distribution network is obtained in the following way:

[0206] By interacting with acquisition devices deployed on multiple nodes within the distribution network, node status data used to indicate the operating status of the nodes is obtained;

[0207] By interacting with data acquisition devices deployed on multiple power devices within the distribution network, equipment health status data used to indicate the health status of power devices is obtained;

[0208] By integrating node status data and device health status data, real-time operational data is obtained.

[0209] In one possible implementation, after obtaining the real-time operating data, the quality detection module 304 is specifically used for:

[0210] Based on a preset data quality assessment mechanism, the real-time running data is verified to obtain a quality score for the real-time running data.

[0211] If the quality score of the real-time running data is lower than the preset score threshold, the data verification of the real-time running data will be triggered, and the data verification result will be obtained.

[0212] The data verification result indicates that the verification passed;

[0213] If the data verification result indicates that the verification failed, the real-time operation data of the distribution network should be reacquired.

[0214] In one possible implementation, the diagnostic module 302 is further configured to:

[0215] Acquire real-time meteorological data and equipment aging data;

[0216] Based on real-time meteorological data and equipment aging data, the fault diagnosis model is updated by adjusting the parameters in the fault diagnosis model.

[0217] In one possible implementation, the prediction module 301 is further configured to:

[0218] Based on the model performance testing strategy, the fault prediction model is periodically tested to obtain the real-time performance of the fault prediction model.

[0219] If the real-time performance is less than the performance threshold, the fault prediction model will be optimized and updated.

[0220] One possible implementation involves optimizing and updating the fault prediction model in the following way:

[0221] The model is trained using the Adam optimizer, with the cross-entropy loss function as the loss function.

[0222] Based on a dynamic learning rate adjustment strategy, using the formula... The learning rate decays exponentially, where, The initial learning rate, is the decay factor, n is the total number of training rounds, and r is the current training round number.

[0223] The intelligent self-recovery device for power distribution networks provided in this embodiment can execute the methods provided in the above-described method embodiments. Its implementation principle and technical effects are similar, and will not be described in detail here.

[0224] Figure 4 This is a schematic diagram of the structure of the intelligent self-recovery device for power distribution networks provided in an embodiment of this application. Figure 4 As shown, the intelligent self-recovery device 40 for power distribution networks provided in this embodiment includes at least one processor 401 and a memory 402. Optionally, the intelligent self-recovery device 40 for power distribution networks also includes a communication component 403. The processor 401, memory 402, and communication component 403 are connected via a bus 404.

[0225] In a specific implementation, at least one processor 401 executes computer execution instructions stored in memory 402, causing at least one processor 401 to perform the above-described method.

[0226] The specific implementation process of processor 401 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0227] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0228] The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.

[0229] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0230] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0231] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the above-described method.

[0232] The aforementioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random-Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0233] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an application-specific integrated circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in the device.

[0234] The division of units is merely a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0235] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0236] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0237] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0238] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0239] Finally, it should be noted that other embodiments of this application will readily conceive of by those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein, and is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes may be made without departing from its scope.

Claims

1. A method for intelligent self-recovery of a power distribution network, characterized in that, include: The real-time operation data of the distribution network is input into a fault prediction model based on a long short-term memory network to obtain the predicted fault risk of each node in the distribution network. If the predicted fault risk indication indicates a fault, the real-time operating data is input into a fault diagnosis model based on a Bayesian network to obtain the fault probability of each node in the distribution network. Based on multi-source data fusion technology, the fault location of the distribution network is determined according to the predicted fault risk and the fault probability of each node in the distribution network. The fault location is used to indicate the node where the fault exists. Based on the predicted fault risk and the fault location, control commands are generated and executed for the switching equipment in the distribution network to isolate the fault area; Based on the fitness function and constraints of the distribution network, a genetic algorithm and fault recovery strategy are used to generate and execute power restoration instructions for non-faulty areas to restore power supply to non-faulty areas. The fitness function is constructed with the goal of maximizing the restored power load and minimizing the restoration path loss of the distribution network. The constraints include constraints on the line capacity, voltage deviation and short-circuit current of the distribution network.

2. The intelligent self-recovery method for distribution networks according to claim 1, characterized in that, The real-time operation data of the power distribution network is obtained through the following methods: By interacting with the acquisition devices deployed on each node in the power distribution network, node status data used to indicate the operating status of the nodes is obtained. By interacting with data acquisition devices deployed on multiple power devices within the power distribution network, equipment health status data used to indicate the health status of the power devices is obtained; The node status data and the device health status data are integrated to obtain the real-time operating data.

3. The intelligent self-recovery method for distribution networks according to claim 2, characterized in that, After obtaining the real-time operational data, the process further includes: Based on a preset data quality assessment mechanism, the real-time running data is quality verified to obtain a quality score for the real-time running data. If the quality score of the real-time running data is lower than the preset score threshold, the data verification of the real-time running data is triggered, and the data verification result is obtained. The data verification result indicates that the verification passed; If the data verification result indicates that the verification failed, the real-time operation data of the power distribution network will be reacquired.

4. The intelligent self-recovery method for distribution networks according to any one of claims 1 to 3, characterized in that, Before inputting the real-time operating data into the Bayesian network-based fault diagnosis model to obtain the fault probability of each node in the distribution network, the method further includes: Acquire real-time meteorological data and equipment aging data; Based on the real-time meteorological data and the equipment aging data, the fault diagnosis model is updated by adjusting the parameters in the fault diagnosis model.

5. The intelligent self-recovery method for distribution networks according to any one of claims 1 to 3, characterized in that, The intelligent self-recovery method for power distribution networks also includes: Based on the model performance testing strategy, the fault prediction model is periodically tested to obtain the real-time performance of the fault prediction model. If the real-time performance is less than the performance threshold, the fault prediction model will be optimized and updated.

6. The intelligent self-recovery method for power distribution networks according to claim 5, characterized in that, The fault prediction model is optimized and updated based on the following methods: The model is trained using the Adam optimizer, with the cross-entropy loss function as the loss function. Based on a dynamic learning rate adjustment strategy, using the formula... The learning rate decays exponentially, where, The initial learning rate, is the decay factor, n is the total number of training rounds, and r is the current training round number.

7. A smart self-recovery device for power distribution networks, characterized in that, include: The prediction module is used to input the real-time operation data of the distribution network into a fault prediction model based on a long short-term memory network to obtain the predicted fault risk of each node in the distribution network. The diagnostic module is used to input the real-time operating data into a Bayesian network-based fault diagnosis model if the predicted fault risk indication indicates a fault, in order to obtain the fault probability of each node in the distribution network. Based on multi-source data fusion technology, the fault location of the distribution network is determined according to the predicted fault risk and the fault probability of each node in the distribution network. The fault location is used to indicate the node where the fault exists. The processing module is used to generate and execute control commands for switching equipment in the distribution network based on the predicted fault risk and the fault location to isolate the fault area; and to generate and execute power restoration commands for non-fault areas based on the fitness function and constraints of the distribution network, using a genetic algorithm and fault recovery strategy to restore power supply to the non-fault areas. The fitness function is constructed with the goal of maximizing the restored power load and minimizing the restoration path loss of the distribution network, and the constraints include constraints on the line capacity, voltage deviation, and short-circuit current of the distribution network.

8. A smart self-recovery device for power distribution networks, characterized in that, include: Memory, processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory, causing the processor to perform the distribution network intelligent self-recovery method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed, are used to implement the distribution network intelligent self-recovery method as described in any one of claims 1 to 6.

10. A computer program product, characterized in that, It includes a computer program that, when executed, implements the intelligent self-recovery method for power distribution networks as described in any one of claims 1 to 6.

Images