WO2023045278A1 - 数据双驱动的台风灾害下电网故障预测方法、装置和设备 - Google Patents

数据双驱动的台风灾害下电网故障预测方法、装置和设备 Download PDF

Info

Publication number
WO2023045278A1
WO2023045278A1 PCT/CN2022/081314 CN2022081314W WO2023045278A1 WO 2023045278 A1 WO2023045278 A1 WO 2023045278A1 CN 2022081314 W CN2022081314 W CN 2022081314W WO 2023045278 A1 WO2023045278 A1 WO 2023045278A1
Authority
WO
WIPO (PCT)
Prior art keywords
disaster
data
typhoon
samples
power grid
Prior art date
Application number
PCT/CN2022/081314
Other languages
English (en)
French (fr)
Inventor
谢海鹏
汤凌峰
祝昊
别朝红
李更丰
Original Assignee
西安交通大学
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 西安交通大学 filed Critical 西安交通大学
Publication of WO2023045278A1 publication Critical patent/WO2023045278A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/048—Activation functions
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
    • G06Q50/06—Energy or water supply
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T3/00—Geometric image transformations in the plane of the image
    • G06T3/40—Scaling of whole images or parts thereof, e.g. expanding or contracting
    • G06T3/4007—Scaling of whole images or parts thereof, e.g. expanding or contracting based on interpolation, e.g. bilinear interpolation

Definitions

  • the invention belongs to the technical field of power grid failure prediction, and in particular relates to a data double-driven power grid failure prediction method, device and equipment under a typhoon disaster.
  • the research on fault prediction methods of distribution network under typhoon disaster is mainly divided into physical model based on disaster mechanism and data-driven model based on historical data.
  • the research idea of the physical model is to establish the wind load model of the distribution network line and the tower according to the probability distribution of the actual wind speed and the design wind speed of the equipment itself, and combine the geographical environment of the equipment, the service life of the equipment itself and other factors to analyze the model. Correction, so as to obtain the failure probability of lines and towers under typhoon disasters.
  • the research idea of the data-driven model is usually based on historical meteorological information, geographic information and power grid information, constructing a data set including disaster-causing factors and network faults, learning the data set through a machine learning model and establishing a corresponding mapping relationship.
  • the existing research usually uses a synthetic minority Class oversampling technology (SMOTE) generates a balanced data set of minority class samples, or uses cost-sensitive learning methods to assign different penalty coefficients to different classes to improve the model's learning emphasis on minority class samples.
  • SMOTE synthetic minority Class oversampling technology
  • the present invention provides a data double-driven power grid fault prediction method, device and equipment under typhoon disasters, which improves the accuracy and interpretability of the distribution network fault prediction method under typhoon disasters, and enhances the resilience of the distribution network to typhoon disasters .
  • a data double-driven distribution network fault prediction method under a typhoon disaster described in the present invention includes the following steps:
  • Step 1 Collect multivariate impact data of power grid faults under typhoon disasters and the sum of permanent trip times of power grids in the predicted area, and divide them into static data and dynamic data according to the time-domain change attributes of the data, and use static data, dynamic data and prediction area The sum of the number of permanent trips of the power grid is used to construct a disaster-caused data set;
  • Step 2 equalize the disaster data set
  • Step 4 collect the corresponding multivariate impact data of the prediction area under future typhoon disasters, construct a disaster data set, input it into the optimized dual-channel prediction model in step 3, and obtain the power grid fault situation of the research area under future typhoon disasters Predictive value.
  • the static data includes forest coverage, land type, maintenance degree of the power grid, and population density
  • the dynamic data includes the distance between the center of the typhoon and the center of the region, the minimum air pressure at the center of the typhoon, and the maximum wind speed near the center of the typhoon , the moving speed of the typhoon, the moving direction angle of the typhoon, the radius of the seven-level wind circle, the average wind speed in the forecast area, and the precipitation in the forecast area.
  • step 2 use the Borderline-SMOTE1 algorithm to divide the minority class sample set according to the distribution of the disaster-causing data set in the high-dimensional space, and perform sample generation for the minority class samples at the decision boundary after division; then The discriminant model is used to test the difference between the data distribution of the training set and the test set, and the parameters of the Borderline-SMOTE1 algorithm are adjusted according to the size of the difference, and finally the parameter-optimized Borderline-SMOTE1 algorithm is used to balance the disaster data set.
  • step 2 includes the following steps:
  • Step 2.1 use the K nearest neighbor algorithm to calculate the m nearest neighbor samples of each mild fault sample
  • Step 2.2 according to the proportion of mild fault samples in the m nearest neighbor samples of mild fault samples, they are divided into safety samples, dangerous samples and noise samples;
  • Step 2.3 For each dangerous sample x i , select the required number of mild fault samples among its K nearest neighbor samples;
  • Step 2.4 For each selected neighbor sample x′ j , use linear interpolation to generate a new sample x i,j of mild fault class;
  • Step 2.5 Add the generated new samples of mild faults to the original disaster-causing training set to obtain the updated disaster-causing data set;
  • Step 2.6 Check the updated disaster-caused data set. If it meets the requirements, go to step 3. If it does not meet the requirements, adjust the parameters of the Borderline-SMOTE1 algorithm until the disaster-caused data set meets the requirements.
  • step 2.6 includes the following steps:
  • Step 2.6.1 Randomly sample the disaster-causing training set so that the number of samples in the sampled training set and the disaster-causing test set are equal; then set the labels of the training set samples and the test set samples to 0 and 1 respectively, and mix them to form Discriminate the dataset and divide it proportionally into new training and testing sets;
  • Step 2.6.2 based on the new training set and test set, with the cross-entropy function as the loss function, the gradient of each parameter value of the discriminant model is obtained through the error back propagation method, and then all the parameters of the discriminant model are determined by the Adam gradient descent algorithm. The parameters are updated to obtain the discrimination accuracy of the disaster training set and the disaster test set;
  • Step 2.6.3 use the discriminant model to distinguish the ability of the disaster-causing training set and the disaster-causing test set to measure the difference in sample distribution between the two.
  • the discrimination accuracy is higher than the accuracy threshold, the number of nearest neighbor samples for the Borderline-SMOTE1 algorithm, etc. Adjust the parameters; when the discrimination accuracy is lower than the accuracy threshold, go to step 3.
  • step 3 includes the following steps:
  • Step 3.1 extracting static features from static data based on feedforward neural network; extracting dynamic features from dynamic data based on long short-term memory network and multi-head attention mechanism;
  • Step 3.2 Splicing the static features and dynamic features, and mapping them to the predicted probability of each fault situation type of the power grid through the linear layer, taking the maximum probability value corresponding to the disaster type as the predicted fault situation type of the sample, and obtaining the prediction model; using the cross entropy function As a loss function, measure the difference between the predicted value and the actual value; then use the error backpropagation algorithm to obtain the gradient value of the cross entropy function for each parameter in the model; finally combine the learning rate, batch size and the number of neurons in each layer, use The small batch Adam algorithm updates the prediction model parameters;
  • step 3.3 includes the following steps:
  • Step 3.3.1 According to the predicted value obtained after the disaster test set is input to the prediction model, the actual value and the predicted value of whether each sample in the disaster test set belongs to the disaster type are counted, and three binary confusion moments are formed;
  • Step 3.3.2 Obtain a set of true positive TP i , false positive FP i , true negative TN i and false negative FN i corresponding to each confusion matrix according to the matrix elements, and then obtain the corresponding precision rate P i and recall rate R i ;
  • Step 3.3.3 obtain the macro-precision rate macro- P , the macro-recall rate macro-R and the macro-F1 value macro-F1 according to the precision rate P i , the recall rate R i and the F1 measurement;
  • Step 3.3.4 Evaluate the performance of the power grid failure prediction model under typhoon disaster according to the four indicators of macro precision rate, macro recall rate, macro F1 and accuracy rate.
  • a power grid failure prediction device under a typhoon disaster comprising:
  • the collection module is used to collect data, and transmit the collected data to the calculation output module;
  • the data includes the multivariate impact data of the power grid failure caused by the typhoon disaster, the sum of the number of permanent trips of the predicted regional power grid, and real-time typhoon data;
  • the calculation output The module is used to train the prediction model according to the collected data set, and output the power grid fault prediction value according to the prediction model and real-time typhoon data.
  • a computer device comprising an electrically connected memory and a processor, the memory is stored with a calculation program that can run on the processor, and when the processor executes the calculation program, any one of claims 1-8 is realized The steps of the method described in the item.
  • the present invention has at least the following beneficial technical effects:
  • the invention classifies the multivariate influencing factors of distribution network failures under typhoon disasters into static data and dynamic data, and utilizes the feed-forward neural network to extract the characteristics of the static data, and uses the long-short-term memory network enhanced by the multi-head self-attention mechanism to extract The characteristics of dynamic data, and finally use the linear layer to fuse all the extracted features, and establish the mapping relationship between multiple influencing factors and distribution network fault conditions.
  • the dual-channel prediction model constructed by the present invention fully considers the stability of the static data on the disaster situation of the distribution network and the time-varying and cumulative nature of the dynamic data on the fault situation of the distribution network, and constructs a model with higher accuracy and interpretability. Stronger distribution network fault prediction model under typhoon disaster.
  • the Borderline-SMOTE1 algorithm used in the present invention identifies the samples at the decision boundary based on the K nearest neighbor algorithm, and uses random linear interpolation to synthesize new samples, which overcomes the blindness and randomness of the sample generation process in the existing sample imbalance processing method , the subjectivity and cumbersomeness of the way to determine the penalty coefficient, effectively reducing the imbalance degree of the disaster-caused data set, laying a good data foundation for the training of the power grid fault prediction model, and helping to improve the power grid under the typhoon disaster.
  • the accuracy of the power grid fault prediction method can enhance the resilience of the distribution network to typhoon disasters.
  • Figure 1 is a schematic diagram of the disaster data set
  • Figure 2 is a schematic diagram of the classification of mild fault samples of the Borderline-SMOTE1 algorithm
  • Fig. 3 is the schematic diagram of discriminant model inspection sample distribution
  • Figure 4 is a unit structure diagram of LSTM
  • Fig. 5 is a network structure diagram of a dual-channel prediction model
  • Fig. 6 is a schematic diagram of the module structure of the grid fault prediction device provided by the present invention.
  • FIG. 7 is a schematic structural diagram of a computer device provided by the present invention.
  • first and second are used for descriptive purposes only, and cannot be interpreted as indicating or implying relative importance or implicitly specifying the quantity of indicated technical features. Thus, a feature defined as “first” and “second” may explicitly or implicitly include one or more of these features.
  • plural means two or more.
  • installation”, “connection” and “connection” should be understood in a broad sense, for example, it can be a fixed connection or a detachable connection.
  • Connected, or integrally connected it may be mechanically connected or electrically connected; it may be directly connected or indirectly connected through an intermediary, and it may be the internal communication of two components.
  • a distribution network fault prediction method under typhoon disasters based on static and dynamic data dual drives including four parts, respectively constructing disaster-caused data sets, balancing disaster-caused data sets, constructing dual-channel prediction models, and predicting Fault situation of regional distribution network under typhoon disaster in the future.
  • Step 1 From the four perspectives of meteorological information, geographic information, power grid information, and population information, select the multiple factors affecting the distribution network failure under the typhoon disaster, and according to the time domain change attribute of the data (the change range of the data during the typhoon transit) Divide it into static data and dynamic data to construct a disaster data set;
  • Step 2 In view of the unbalanced samples in the disaster data set, the Borderline-SMOTE1 algorithm is used to divide the minority sample set according to the distribution of the disaster data set in the high-dimensional space, and for the minority samples at the decision boundary after division Carry out sample generation; then use the discriminant model to test the difference between the data distribution of the training set and the test set, and adjust the parameters of the Borderline-SMOTE1 algorithm according to the size of the difference, and finally apply the parameter-optimized Borderline-SMOTE1 algorithm to balance the disaster data set;
  • Step 3 Use the feed-forward neural network to extract the characteristics of the static data in the disaster-caused data set, use the long-term short-term memory network (LSTM) and the multi-head self-attention mechanism to extract the sequence features of the dynamic data in the disaster-caused data set, and establish the distribution network under the typhoon disaster
  • LSTM long-term short-term memory network
  • a dual-channel prediction model for faults and based on the disaster-caused data set after sample equalization processing, combined with the cross-entropy loss function and error back propagation method to solve and optimize the model parameters, and finally obtain the optimized dual-channel prediction model, and evaluate its performance. If the performance meets the requirements, go to step 4, otherwise continue to optimize.
  • Step 4 Collect the corresponding data of a research area under future typhoon disasters, construct a disaster-caused data set, and input it into the optimized dual-channel prediction model in Step 3 to obtain the distribution network failure situation of the research area under future typhoon disasters predicted value of .
  • the specific process of each step is as follows:
  • the present invention selects disaster-causing data from four angles of meteorological information, geographic information, power grid information and population information, and divides them into static data and dynamic data according to the time-domain change of the data during the typhoon transit period, and jointly constitutes the disaster-causing data set. Finally, it is combined with the fault type of distribution network under typhoon disaster to form a disaster data set.
  • this type of disaster data is classified as static data, including four types of data, including forest coverage, land type, power grid maintenance degree, and population density.
  • static data including four types of data, including forest coverage, land type, power grid maintenance degree, and population density.
  • dynamic data including the distance between the center of the typhoon and the center of the region, the minimum air pressure at the center of the typhoon, the maximum wind speed near the center of the typhoon, the moving speed of the typhoon, the moving direction angle of the typhoon, and the seven-level wind circle. Radius, the average wind speed of the region, and the precipitation of the region have a total of eight types of data.
  • the static data is composed of data of a single time section
  • dynamic data is composed of continuous 48-hour sequence data.
  • the present invention combines static and dynamic data to form a disaster-caused data set sample, and uses the type of distribution network fault as a data set sample label to jointly form the final disaster-caused data set, and according to the ratio of eight to two The ratio divides it into a hazard training set and a hazard test set.
  • the normal operation class has the largest number of samples
  • the minor fault class has the second largest sample number
  • the severe fault class has the least number of samples.
  • the unbalanced sample of the disaster data set makes the distribution network fault prediction model lack of learning for minority samples during the training process, which eventually leads to low prediction accuracy for minority samples.
  • the present invention reduces the imbalance degree of the disaster-caused data set based on the Borderline-SMOTE1 algorithm, and checks the quality of the generated minority samples through the discriminant model .
  • the SMOTE algorithm commonly used in oversampling technology has greater blindness and randomness when selecting target samples for sample generation, and it is easy to generate new samples that are meaningless or interfere with defining the decision boundary. Therefore, the present invention is based on the Borderline-SMOTE1 algorithm, divides the minority class samples according to the type distribution characteristics around them, and selects the minority class samples that are close to the decision boundary for sample generation to reduce the imbalance of the disaster-causing data set.
  • the process of generating severe fault samples is the same. It should be noted that the sample generation algorithm is only applied to the hazard training set.
  • Step1 Use the K nearest neighbor algorithm to calculate the m nearest neighbor samples for each mild fault sample
  • Step3 For each dangerous sample x i , select the required number of mild fault samples among its K nearest neighbor samples;
  • Step4 For each selected neighbor sample x′ j , use linear interpolation to generate a new sample x i,j of mild fault class, the calculation formula is:
  • ⁇ is a random number between 0 and 1.
  • Step5 Add the generated new samples of mild faults to the original disaster training set.
  • the present invention designs a discriminant model to test the sample distribution of the disaster-causing training set and the disaster-causing test set after adding and generating samples, and adjust the parameter settings of the sample generation method according to the test results.
  • the specific principle is shown in Figure 3 . The following describes the specific process of the discriminant model to test the difference in sample distribution.
  • the discriminant dataset is based on the idea of self-supervised learning, and the sample division of the training set and the test set is used as the label source of the discriminant dataset. Considering that the number of samples in the disaster training set is generally several times that of the disaster test set, random sampling is performed on the disaster training set to ensure that the number of samples in the training set and the disaster test set are equal after sampling. Then set the labels of the training set samples and the test set samples to 0 and 1 respectively, mix them to form a discriminative data set, and divide it into a new training set and a test set according to the ratio of 8:2.
  • the ability of the discriminant model to distinguish the disaster-caused training set from the disaster-caused test set is used to measure the difference in sample distribution between the two.
  • the discrimination accuracy is higher than the accuracy threshold, it means that the distribution of samples in the disaster training set and the disaster test set is quite different, and the disaster training set needs to be reconstructed, that is, the number of nearest neighbor samples of the Borderline-SMOTE1 algorithm
  • the discrimination accuracy is lower than the accuracy threshold, it means that the difference between the two sample distributions is small, and it can be directly used for the training and testing of the prediction model.
  • the accuracy threshold of the discriminant model is generally set at 70%.
  • the feed-forward neural network is composed of an input layer, a hidden layer and an output layer.
  • the neurons in each layer are fully connected, and there is no intra-layer connection structure and cross-layer connection structure, so the information transmission process of the feed-forward neural network is one-way of.
  • the present invention uses a multi-layer feedforward neural network to extract its static features layer by layer for the static data that remains unchanged within 48 hours.
  • long short-term memory (LSTM) networks not only transfer information between layers, but also within the same layer.
  • LSTM has "memory” and “transitive” for data processing.
  • the unit structure of LSTM contains multiple gate structures, which can effectively deal with the gradient disappearance and gradient explosion problems caused by the interlayer connection structure.
  • Each unit of LSTM includes three gate structures: forget gate, input gate and output gate.
  • the unit structure of LSTM is shown in Figure 4. .
  • LSTM calculates the forgetting gate gating signal f t , the input gating gating signal it and the output gating gating signal o t respectively according to the input information x t at the current moment and the short-term memory h t-1 at the previous moment:
  • ⁇ refers to the Sigmoid activation function
  • U f is the connection weight of the current input x t and the forget gate structure
  • U i is the connection weight of the current input x t and the input gate structure
  • U O is the current input x t and the output gate structure
  • W f is the connection weight of the short-term memory h t-1 and the forget gate structure at the previous moment
  • W i is the connection weight of the short-term memory h t-1 and the input gate structure at the previous moment
  • W O is the connection weight of the previous The short-term memory h t-1 and the connection weight of the output gate structure at a moment
  • b f is the bias of the forget gate structure
  • bi is the bias of the input gate structure
  • b o is the bias of the output gate structure.
  • LSTM reprocesses the input information x t at the current moment and the short-term memory h t-1 at the previous moment based on three gating signals, thereby updating the long-term memory c t and short-term memory h t .
  • the specific calculation formula is:
  • connection weight of W c is the short-term memory and the candidate long-term memory
  • connection weight of b c is the candidate long-term memory input bias.
  • the present invention adopts a multi-head attention mechanism, uses multiple mapping subspaces to extract key components in known data in all directions and from multiple angles, and maximizes the use of known data information .
  • the multi-head attention mechanism first maps the data Q to multiple subspaces, and uses the self-attention formula Attention(Q) to calculate the correlation and dependence between the data.
  • the specific calculation formula of the self-attention value head i (Q) corresponding to the i-th head is as follows:
  • the present invention first utilizes the LSTM network to extract the dynamic data features, and then adds the multi-head attention mechanism layer to the LSTM network to further extract the deep dynamic data features in the dynamic data, laying the foundation for the establishment of the final mapping relationship.
  • the present invention uses a feed-forward neural network to process static data, and uses an LSTM network strengthened by a multi-head self-attention mechanism to process dynamic data. Finally, the deep features extracted by the two are spliced, and mapped to each distribution network through a linear layer. For the predicted probability of the type of fault situation, take the maximum probability value corresponding to the type of disaster affected as the sample of the predicted fault situation type.
  • the network structure of the prediction model is shown in Figure 5. Among them, it is necessary to add the corresponding batch normalization layer and nonlinear activation function after the first linear layer of the feedforward neural network to improve the convergence of the prediction model.
  • the present invention uses the cross-entropy function as a loss function on the basis of the disaster training set to measure the difference between the predicted value and the actual value. Then, the gradient value of the cross entropy function to each parameter in the model is obtained through the error back propagation algorithm. Finally, combined with hyperparameters such as learning rate, batch size and the number of neurons in each layer, the small batch Adam algorithm is used to update the prediction model parameters.
  • the prediction of the distribution network fault situation type in the present invention is a three-category problem, and the number of samples in each category of the disaster test set is not equal.
  • the present invention uses the precision rate, recall rate and F1 measurement as the basic index system, and introduces a macro-average mechanism to synthesize Considering the performance of the prediction model in different types of sample sets in the disaster test set, the specific process is described as follows.
  • the actual value and predicted value of whether each sample in the disaster test set belongs to the disaster type is counted, and a total of three binary confusion moments can be formed.
  • a set of true positive TP i , false positive FP i , true negative TN i and false negative FN i corresponding to each confusion matrix is obtained according to the matrix elements, and then the corresponding precision rate P i and recall rate R are obtained. i .
  • three indicators, namely macro-P, macro-recall and macro-F1 are obtained to comprehensively measure the performance of the prediction model.
  • the specific calculation formula is as follows.
  • the present invention selects four indicators of macro precision rate, macro recall rate, macro F1 and accuracy rate to evaluate the performance of the distribution network failure prediction model under typhoon disasters .
  • a power grid fault prediction device under a typhoon disaster includes an acquisition module and a calculation output module;
  • the collection module is used to collect data and transmit the collected data to the calculation output module;
  • the data includes historical dynamic data, static data and real-time typhoon data, and the real-time typhoon data includes dynamic data and static data.
  • the calculation output module is used to train the prediction model according to the historical dynamic data, static data and the sum of the permanent trip times of the predicted regional power grid, and then output the power grid fault prediction value according to the prediction model and real-time dynamic data and static data.
  • a computer device provided by the present invention includes an electrically connected memory and a processor, wherein the memory stores a computing program that can run on the processor, and when the processor executes the computing program , realizing the steps of the above prediction method.
  • the prediction device is realized in the form of a software function unit and sold or used as an independent product, it can be stored in a computer-readable storage medium.
  • the present invention realizes all or part of the processes in the methods of the above embodiments, and can also be completed by instructing related hardware through a computer program.
  • the computer program can be stored in a computer-readable storage medium, and the computer When the program is executed by the processor, the steps in the above-mentioned various method embodiments can be realized.
  • the computer program includes computer program code, and the computer program code may be in the form of source code, object code, executable file or some intermediate form.
  • the computer-readable medium may include: any entity or device capable of carrying the computer program code, a recording medium, a U disk, a removable hard disk, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM, Read-OnlyMemory), Random access memory (RAM, RandomAccessMemory), electric carrier signal, telecommunication signal and software distribution medium, etc. It should be noted that the content contained in the computer-readable medium may be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, computer-readable media Excludes electrical carrier signals and telecommunication signals.
  • the present invention is a general prediction model of distribution network faults under typhoon disasters.
  • Using the method of the present invention for prediction can effectively reduce the inherent unbalanced problem of data sets and improve the quality of generated samples.
  • the prediction method of the present invention takes into account the stability of static data effects and the accumulation of dynamic data effects, further improves the accuracy and interpretability of the prediction model, and provides more accurate predictions for the distribution network to cope with typhoon disasters information.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Health & Medical Sciences (AREA)
  • Business, Economics & Management (AREA)
  • General Health & Medical Sciences (AREA)
  • Software Systems (AREA)
  • Economics (AREA)
  • Evolutionary Computation (AREA)
  • Computational Linguistics (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Biophysics (AREA)
  • Mathematical Physics (AREA)
  • Biomedical Technology (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Data Mining & Analysis (AREA)
  • Human Resources & Organizations (AREA)
  • Strategic Management (AREA)
  • General Business, Economics & Management (AREA)
  • Marketing (AREA)
  • Tourism & Hospitality (AREA)
  • Game Theory and Decision Science (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Operations Research (AREA)
  • Quality & Reliability (AREA)
  • Development Economics (AREA)
  • Public Health (AREA)
  • Water Supply & Treatment (AREA)
  • Primary Health Care (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

本发明公开了一种数据双驱动的台风灾害下电网故障预测方法、装置和设备，所述方法包括：构建致灾数据集、均衡化致灾数据集、构建双通道预测模型以及用双通道预测模型进行预测，将台风灾害下配电网受灾情况的多元影响因素归类为静态数据和动态数据，并利用前馈神经网络提取静态数据的特征，利用多头自注意力机制强化后的长短期记忆网络提取动态数据的特征，最终采用线性层对提取的所有特征进行融合，建立多元影响因素与配电网受灾情况的映射关系。充分考虑了静态数据对配电网受灾情况作用的稳定性和动态数据对配电网受灾情况作用的时变性与累积性，构建了准确率更高、可解释性更强的台风灾害下配电网故障预测模型。

Description

数据双驱动的台风灾害下电网故障预测方法、装置和设备 技术领域
本发明属于电网故障预测技术领域，具体涉及一种数据双驱动的台风灾害下电网故障预测方法、装置和设备。
背景技术
台风灾害影响范围大、持续时间长，近年来随着全球气候的变化，台风及以上强度的热带气旋占比在不断上升，对沿海地区输配电网络的正常运作造成了巨大的威胁。相比输电网，配电网的设备数量更多，设备的老化问题严重，更易受台风等自然灾害的影响。因此，需要针对台风的破坏性和配电网的脆弱性，研究台风灾害下配电网故障预测方法，为配电网的弹性增强策略提供可靠的先验信息。
对台风灾害下配电网故障预测方法的研究主要分为基于致灾机理的物理模型和基于历史数据的数据驱动模型。物理模型的研究思路是根据实际风速的概率分布和设备自身的设计风速，建立配电网线路和杆塔的风荷载模型，并结合设备所处的地理环境、设备自身的使用年限等因素对模型进行修正，从而得到台风灾害下线路和杆塔的故障概率。数据驱动模型的研究思路通常是基于历史的气象信息、地理信息和电网信息，构建包含致灾因素和网络故障的数据集，通过机器学习模型对数据集进行学习并建立相应的映射关系。同时，考虑到台风灾害下配电网的故障数据中包含大量故障数量为零的样本，使得机器学习模型在故障数量不为零的样本中产生较大的预测偏差，现有研究通常采用合成少数类过采样技术(SMOTE)生成少数类样本均衡数据集，或者采用代价敏感学习方法给不同的类别赋予不同的惩罚系数，提高模型对少数类样本的学习侧重度。
物理模型在建模复杂度的限制下，难以对配电网设备故障的影响因素进行全面化、精细化的建模，从而损失一定的预测精度。随着电力部门和气象部门的数据收集与管理系统日益完善，当前的研究多聚焦于通过数据驱动模型预测台风灾害下的配电网故障情况。但现有的数据驱动模型只考虑了每个时间截面中各影响因素与配电网故障之间的关系，没有考虑到部分因素对配电网故障作用的累积性。同时目前研究采用的SMOTE算法在样本合成对象的选择过程中具有一定的盲目性和随机性，生成的少数类样本质量较差，而代价敏感学习方法在确定各类别的惩罚系数时，需要根据模型性能对参数进行反复调整，且调整的方向较为主观。
技术问题
本发明提供了一种数据双驱动的台风灾害下电网故障预测方法、装置和设备，提高台风灾害下配电网故障预测方法的准确率和可解释性，增强配电网应对台风灾害的抵御能力。
技术解决方案
为达到上述目的，本发明所述一种数据双驱动的台风灾害下配电网故障预测方法，包括以下步骤：
步骤1，采集台风灾害下电网故障的多元影响数据以及被预测区域电网永久跳闸次数总和，并根据数据的时域变化属性将其划分为静态数据和动态数据，利用静态数据、动态数据和预测区域电网永久跳闸次数总和构建致灾数据集；
步骤2，对致灾数据集进行均衡化处理；
步骤3，利用前馈神经网络提取致灾数据集中静态数据的特征，利用长短期记忆网络和多头自注意力机制提取致灾数据集中动态数据的序列特征，建立台风灾害下电网故障的双通道预测模型，并基于样本均衡处理后的致灾数据集，进行模型参数的求解和调优，最终得到优化后的双通道预测模型；并对其性能进行评估；若性能符合要求则进行步骤4，否则继续进行优化；
步骤4，收集未来台风灾害下预测区域相应的多元影响数据，并构建致灾数据集，将其输入至步骤3中优化后的双通道预测模型，得到未来台风灾害下该研究区域电网故障情况的预测值。
进一步的，步骤1中，静态数据包括森林覆盖率、土地类型、电网的维护程度和人口密度，所述动态数据包括台风中心与区域中心的距离、台风的中心最低气压、台风的近中心最大风速、台风的移动速度、台风的移动方向角、七级风圈半径、预测区域的平均风速和预测区域的降水量。
进一步的，步骤2的过程为：用Borderline-SMOTE1算法，根据高维空间中致灾数据集的分布对少数类样本集合进行划分，并针对划分后决策边界处的少数类样本进行样本生成；接着通过判别模型检验训练集和测试集数据分布的差异，并根据差异大小对Borderline-SMOTE1算法进行参数调优，最终应用参数优化后的Borderline-SMOTE1算法均衡致灾数据集。
进一步的，步骤2包括以下步骤：
步骤2.1、使用K近邻算法计算每一个轻度故障类样本的m个最近邻样本；
步骤2.2、根据轻度故障类样本的m个最近邻样本中轻度故障样本的占比，将其分安全类样本、危险类样本和噪声类样本；
步骤2.3、针对每一个危险类样本x i，在其K个最近邻样本中选择所需数量的轻度故障类样本；
步骤2.4、对于每一个被选择的近邻样本x′ j，使用线性插值生成轻度故障类新样本x i,j；
步骤2.5、将生成的轻度故障类新样本添加至原致灾训练集中，得到更新后的致灾数据集；
步骤2.6、对更新后的的致灾数据集进行检验，若符合要求则进行步骤3，若不符合要求对Borderline-SMOTE1算法进行调参，直至致灾数据集符合要求。
进一步的，步骤2.6包括以下步骤：
步骤2.6.1、针对致灾训练集进行随机采样，使采样后的训练集与致灾测试集的样本数量相等；接着将训练集样本和测试集样本的标签分别置为0和1，混合形成判别数据集，并按比例将其划分为新的训练集和测试集；
步骤2.6.2、以新的训练集和测试集为基础，以交叉熵函数为损失函数，通过误差反向传播法得到判别模型每个参数值的梯度，进而通过Adam梯度下降算法对判别模型所有参数进行更新，得到致灾训练集与致灾测试集的判别准确率；
步骤2.6.3、使用判别模型区分致灾训练集与致灾测试集的能力衡量两者的样本分布差异，当判别准确率高于准确率阈值时，对Borderline-SMOTE1算法的最近邻样本数等参数进行调整；当判别准确率低于准确率阈值时，执行步骤3。
进一步的，步骤3包括以下步骤：
步骤3.1、基于前馈神经网络从静态数据中提取静态特征；基于长短期记忆网络与多头注意力机制从动态数据中提取动态特征；
步骤3.2、将静态特征和动态特征进行拼接，并通过线性层映射为电网各故障情况类型的预测概率，取最大概率值对应受灾类型为样本的预测故障情况类型，得到预测模型；使用交叉熵函数作为损失函数，衡量预测值与实际值的差异程度；接着通过误差反向传播算法得到交叉熵函数对模型中每一个参数的梯度值；最后结合学习率、批大小以及各层神经元数量，使用小批量Adam算法对预测模型参数进行更新；
3.3、以查准率和查全率为基本指标体系，并引入宏平均机制综合考虑预测模型在致灾测试集中不同类型样本集合中的表现，对预测模型进行评估。
进一步的，步骤3.3包括以下步骤：
步骤3.3.1、根据致灾测试集输入至预测模型后得到的预测值，统计致灾测试集中每个样本是否属于该受灾类型的实际值与预测值，共形成三个二分类混淆矩；
步骤3.3.2、根据矩阵元素得到每一个混淆矩阵对应的一组真阳性TP i、假阳性FP i、真阴性TN i和假阴性FN i，进而得到相应的查准率P i和查全率R i；
步骤3.3.3、根据查准率P i、查全率R i和F1度量得到宏查准率macro-P、宏查全率macro-R和宏F1值macro-F1；
步骤3.3.4、根据宏查准率、宏查全率、宏F1和准确率共四个指标对台风灾害下电网故障情况预测模型的性能进行评估。
一种台风灾害下电网故障预测装置，包括：
采集模块，用于采集数据，并将采集的数据传递至计算输出模块；所述数据包括致台风灾害下电网故障的多元影响数据以及被预测区域电网永久跳闸次数总和，以及实时台风数据；计算输出模块，用于根据采集的数据集训练预测模型，并根据预测模型和实时台风数据输出电网故障预测值。
一种计算机设备，包括电连接的存储器和处理器，所述存储器上存储有可在处理器上运行的计算程序，所述处理器执行所述计算程序时，实现权利要求1-8中任意一项所述的方法的步骤。
有益效果
与现有技术相比，本发明至少具有以下有益的技术效果：
本发明将台风灾害下配电网故障情况的多元影响因素归类为静态数据和动态数据，并利用前馈神经网络提取静态数据的特征，利用多头自注意力机制强化后的长短期记忆网络提取动态数据的特征，最终采用线性层对提取的所有特征进行融合，建立多元影响因素与配电网故障情况的映射关系。本发明构建的双通道预测模型充分考虑了静态数据对配电网受灾情况作用的稳定性和动态数据对配电网故障情况作用的时变性与累积性，构建了准确率更高、可解释性更强的台风灾害下配电网故障预测模型。
本发明使用的Borderline-SMOTE1算法基于K近邻算法对决策边界处的样本进行识别，并使用随机线性插值进行新样本合成，克服了现有样本不均衡处理方式中样本生成过程的盲目性 和随机性、惩罚系数确定方式的主观性和繁琐性等缺点，有效降低了致灾数据集的不均衡度，为电网故障情况预测模型的训练奠定了较好的数据基础，有助于提高台风灾害下配电网故障预测方法的准确性，进而增强配电网应对台风灾害的抵御能力。
附图说明
图1为致灾数据集的示意图；
图2为Borderline-SMOTE1算法的轻度故障类样本分类示意图；
图3为判别模型检验样本分布的原理图；
图4为LSTM的单元结构图；
图5为双通道预测模型的网络结构图；
图6为本发明提供的电网故障预测装置的模块结构示意图；
图7为本发明提供的计算机设备的结构示意图。
本发明的实施方式
为了使本发明的目的和技术方案更加清晰和便于理解。以下结合附图和实施例，对本发明进行进一步的详细说明，此处所描述的具体实施例仅用于解释本发明，并非用于限定本发明。在本发明的描述中，需要理解的是，术语“中心”、“纵向”、“横向”、“上”、“下”、“前”、“后”、“左”、“右”、“竖直”、“水平”、“顶”、“底”、“内”、“外”等指示的方位或位置关系为基于附图所示的方位或位置关系，仅是为了便于描述本发明和简化描述，而不是指示或暗示所指的装置或元件必须具有特定的方位、以特定的方位构造和操作，因此不能理解为对本发明的限制。此外，术语“第一”、“第二”仅用于描述目的，而不能理解为指示或暗示相对重要性或者隐含指明所指示的技术特征的数量。由此，限定有“第一”、“第二”的特征可以明示或者隐含地包括一个或者更多个该特征。在本发明的描述中，除非另有说明，“多个”的含义是两个或两个以上。在本发明的描述中，需要说明的是，除非另有明确的规定和限定，术语“安装”、“相连”、“连接”应做广义理解，例如，可以是固定连接，也可以是可拆卸连接，或一体地连接；可以是机械连接，也可以是电连接；可以是直接相连，也可以通过中间媒介间接相连，可以是两个元件内部的连通。对于本领域的普通技术人员而言，可以具体情况理解上述术语在本发明中的具体含义。
实施例1
参照图1，一种基于静动态数据双驱动的台风灾害下配电网故障预测方法，包括四大部分，分别为构建致灾数据集、均衡化致灾数据集、构建双通道预测模型和预测未来台风灾害下区域配电网的故障情况。
步骤1，从气象信息、地理信息、电网信息、人口信息四个角度出发，选取台风灾害下配电网故障的多元影响因素，并根据数据的时域变化属性(台风过境期间数据的变化幅度)将其划分为静态数据和动态数据，构建致灾数据集；
步骤2，针对致灾数据集中的样本不均衡现象，使用Borderline-SMOTE1算法，根据高维空间中致灾数据集的分布对少数类样本集合进行划分，并针对划分后决策边界处的少数类样本进行样本生成；接着通过判别模型检验训练集和测试集数据分布的差异，并根据差异大小对Borderline-SMOTE1算法进行参数调优，最终应用参数优化后的Borderline-SMOTE1算法均衡致灾数据集；
步骤3，利用前馈神经网络提取致灾数据集中静态数据的特征，利用长短期记忆网络(LSTM) 和多头自注意力机制提取致灾数据集中动态数据的序列特征，建立台风灾害下配电网故障的双通道预测模型，并基于样本均衡处理后的致灾数据集，结合交叉熵损失函数和误差反向传播法等进行模型参数的求解和调优，最终得到优化后的双通道预测模型，并对其性能进行评估。若性能符合要求则进行步骤4，否则继续进行优化。
步骤4，收集未来台风灾害下某研究区域相应的数据，并构建致灾数据集，将其输入至步骤3中优化后的双通道预测模型，得到未来台风灾害下该研究区域配电网故障情况的预测值。各步骤的具体过程如下：
1、构建致灾数据集
本发明从气象信息、地理信息、电网信息和人口信息四个角度选取致灾数据，并根据台风过境期间数据的时域变化情况将其分为静态数据和动态数据，共同构成致灾数据集的样本，最后与台风灾害下配电网的故障情况类型结合形成致灾数据集。
在台风过境的时间尺度内，部分致灾数据基本不发生变化，对配电网故障情况的影响具有稳定性。故将此类致灾数据归类为静态数据，包括森林覆盖率、土地类型、电网的维护程度和人口密度共四种数据。而部分致灾数据随时间变化较大，对配电网故障情况的影响具有时变性和累积性。故将此类致灾数据归类为动态数据，包括台风中心与区域中心的距离、台风的中心最低气压、台风的近中心最大风速、台风的移动速度、台风的移动方向角、七级风圈半径、区域的平均风速、区域的降水量共八种数据。需要注意的是，静态数据由单个时间截面的数据构成，动态数据由连续48小时的序列数据构成。
考虑到台风灾害伴随有强风和强降雨，对配电网的架空线路、地下电缆和杆塔等电力元件均会造成一定程度的破坏，因此本发明对台风灾害下配电网连续24个小时的永久跳闸次数进行求和，跳闸次数总和为0时认为该区域配电网正常运行，跳闸次数总和为1至9次时认为该区域配电网轻度故障，跳闸次数总和大于9次时认为该区域配电网重度故障，以配电网故障情况的这三种受灾情况类型作为致灾数据集的标签。
2、综上所述，本发明结合静、动态数据形成致灾数据集样本，并以配电网故障情况类型作为数据集样本标签，共同构成最终的致灾数据集，并按照八比二的比例将其划分为致灾训练集和致灾测试集。样本及样本标签的示意图如图1所示，其中f 1、f 2、f 3和f 4为为静态数据，f ,i，j为动态数据，i＝5,6，……12；j＝1,2，……48；第i项第j个小时的动态数据，n LO,k为第k个小时的跳闸次数，k＝1,2，……24；均衡化致灾数据集。
台风作为极端自然灾害的一种，发生的概率较低，覆盖的区域范围也比较有限。因此，致灾数据集中正常运行类的样本数量最多、轻度故障类的样本数量其次、重度故障类的样本数量最少，即轻度故障类样本和重度故障类样本均为少数类样本。
致灾数据集的样本不均衡现象，使得配电网故障情况预测模型在训练过程中缺乏针对少数类样本的学习，最终导致其对少数类样本的预测准确率偏低。考虑到代价敏感学习的方法具有一定的主观性，且参数调整过程比较繁琐，因此，本发明基于Borderline-SMOTE1算法降低致灾数据集的不均衡度，并通过判别模型检验生成少数类样本的质量。
1)Borderline-SMOTE1样本生成算法
过采样技术中常用的SMOTE算法在选择目标样本进行样本生成时具有较大的盲目性与随机性，容易生成对界定决策边界无意义或有干扰的新样本。所以本发明基于Borderline-SMOTE1 算法，根据少数类样本周围的类型分布特点对其进行划分，并选择距离决策边界近的少数类样本进行样本生成，降低致灾数据集的不均衡度。以生成轻度故障类样本为例对Borderline-SMOTE1的算法步骤进行阐述，重度故障类样本的生成过程同理。需要注意的是，样本生成算法仅应用于致灾训练集。
Step1：使用K近邻算法计算每一个轻度故障类样本的m个最近邻样本；
Step2：根据轻度故障类样本的m个最近邻样本中轻度故障样本的占比，将其分为以下三类，分类示意图如图2所示。
(1)安全类样本：最近邻样本中一半以上的样本均为轻度故障样本，如图2中的A样本；
(2)危险类样本：最近邻样本中一半以下的样本为轻度故障样本，如图2中的B样本；
(3)噪声类样本：最近邻样本无轻度故障样本，如图2中的C样本；
Step3：针对每一个危险类样本x i，在其K个最近邻样本中选择所需数量的轻度故障类样本；
Step4：对于每一个被选择的近邻样本x′ j，使用线性插值生成轻度故障类新样本x i,j，计算公式为：
x i,j＝x i+γ(x′ j-x i)     (1)
式中，γ为介于0到1之间的随机数。
Step5：将生成的轻度故障类新样本添加至原致灾训练集中。
2)对致灾数据集进行检验
考虑到生成少数类样本的添加人为改变了致灾训练集的数据分布，当生成样本的质量较低时，致灾训练集和致灾测试集的样本分布差异会增大，进而导致预测模型在致灾测试集上的泛化能力降低。故本发明设计了判别模型，对添加生成样本后的致灾训练集和致灾测试集进行样本分布检验，并根据检验结果对样本生成方法的参数设置进行调整，具体的原理如图3所示。下面介绍判别模型检验样本分布差异的具体过程。
(1)判别数据集的构建：判别数据集基于自监督学习的思想，利用训练集和测试集的样本划分情况作为判别数据集的标签来源。考虑到致灾训练集的样本数量一般是致灾测试集样本数量的数倍，故针对致灾训练集进行随机采样，保证采样后的训练集与致灾测试集的样本数量相等。接着将训练集样本和测试集样本的标签分别置为0和1，混合形成判别数据集，并按8:2的比例将其划分为新的训练集和测试集.
(2)判别模型的训练过程：以新的训练集和测试集为基础，以交叉熵函数为损失函数，通过误差反向传播法得到判别模型每个参数值的梯度，进而通过Adam梯度下降算法对判别模型所有参数进行更新，得到致灾训练集与致灾测试集的判别准确率。
(3)判别模型的测试结果分析：使用判别模型区分致灾训练集与致灾测试集的能力衡量两者的样本分布差异。当判别准确率高于准确率阈值时，说明致灾训练集和致灾测试集的样本分布差异较大，需要对致灾训练集进行重构处理，即对Borderline-SMOTE1算法的最近邻样本数等参数进行调整；当判别准确率低于准确率阈值时，说明两者的样本分布差异小，可以直接用于预测模型的训练和测试。其中，判别模型的准确率阈值一般定为70％。
3、构建双通道预测模型
为了考虑静态数据作用的稳定性和动态数据作用的时变性与累积性，本发明提出了一种具备可解释性的神经网络架构，分别对静态数据和动态数据进行特征提取，进而建立其与台风灾害下配电网故障情况类型的映射关系。下面分别对静动态数据的特征提取过程以及双通道预测模型的训练方法进行详细阐述。
3.1基于前馈神经网络的静态特征提取
前馈神经网络由输入层、隐藏层和输出层构成，各层神经元之间是全连接的，不存在层内连接结构与跨层连接结构，故前馈神经网络的信息传递过程是单向的。考虑到台风灾害的静态数据对配电网故障情况作用的稳定性，故本发明针对48小时内保持不变的静态数据，采用多层的前馈神经网络逐层提取其静态特征。
3.2基于长短期记忆网络与多头注意力机制的动态特征提取
与前馈神经网络不同，长短期记忆(LSTM)网络不仅在层与层之间传递信息，而且在同一层内传递信息。通过这种层内连接结构的加入，LSTM对数据的处理具有“记忆性”和“传递性”。同时，LSTM的单元结构中含有多个门结构，可以有效处理层间连接结构带来的梯度消失与梯度爆炸问题。
LSTM的每一个单元均包含遗忘门、输入门和输出门三种门结构，在同一层之间传递信息时同时传递门结构处理后的长期记忆和短期记忆，LSTM的单元结构如图4所示。LSTM根据当前时刻的输入信息x t和上一时刻的短期记忆h t-1，分别计算遗忘门门控信号f t、输入门门控信号i t和输出门门控信号o t：
f t＝σ(U fx t+W fh t-1+b f)    (2)
i t＝σ(U ix t+W ih t-1+b i)    (3)
o t＝σ(U ox t+W oh t-1+b o)   (4)
式中,σ指Sigmoid激活函数，U f为当前输入x t与遗忘门结构的连接权重，U i为当前输入x t与输入门结构的连接权重，U O为当前输入x t与输出门结构的连接权重，W f为上一时刻的短期记忆h t-1与遗忘门结构的连接权重，W i为上一时刻的短期记忆h t-1与输入门结构的连接权重，W O为上一时刻的短期记忆h t-1与输出门结构的连接权重，b f为遗忘门结构的偏置，b i为输入门结构的偏置，b o为输出门结构的偏置。
LSTM基于三种门控信号对当前时刻的输入信息x t和上一时刻的短期记忆h t-1进行再处理，从而更新长期记忆c t和短期记忆h t，具体的计算式为：
Figure PCTCN2022081314-appb-000001
Figure PCTCN2022081314-appb-000002
Figure PCTCN2022081314-appb-000003
式中，
Figure PCTCN2022081314-appb-000004
为候选长期记忆，U c为输入信息与候选长期记忆
Figure PCTCN2022081314-appb-000005
的连接权重，W c为短期记忆与候选长期记忆
Figure PCTCN2022081314-appb-000006
的连接权重，b c为候选长期记忆
Figure PCTCN2022081314-appb-000007
的输入偏置。
为了进一步增强网络对动态数据的特征提取能力，本发明采用了多头注意力机制，利用多个映射子空间全方位、多角度地提取已知数据中的关键成分，最大化利用已知的数据信息。多头注意力机制首先将数据Q映射至多个子空间，并利用自注意力公式Attention(Q)计算数据间的关联性与依赖性。第i个头对应的自注意力值head i(Q)的具体计算式如下：
Figure PCTCN2022081314-appb-000008
Figure PCTCN2022081314-appb-000009
式中，d Q为输入数据Q的维度，i＝1,2,...,h为注意力机制的头数，
Figure PCTCN2022081314-appb-000010
分别是第i个头对应的子空间变换矩阵。
接着将所有头的输出进行拼接，通过线性层映射为最终的注意力加权后的值，即MultiHead(Q)：
MultiHead(Q)＝Concat(head 1,...,head h)W O     (10)
式中，Concat为拼接操作，W o为输出映射矩阵。
本发明首先利用LSTM网络提取动态数据特征，接着将多头注意力机制层添加至LSTM网络后，进一步提取动态数据中的深层动态数据特征，为最终映射关系的建立奠定基础。
3.3双通道预测模型的网络结构与训练方法
本发明采用前馈神经网络对静态数据进行处理，采用多头自注意力机制强化的LSTM网络对动态数据进行处理,最终将两者提取的深层特征进行拼接，并通过线性层映射为配电网各故障情况类型的预测概率，取最大概率值对应受灾类型为样本的预测故障情况类型，预测模型的网络结构如图5所示。其中，需要在前馈神经网络的第一个线性层后添加相应的批标准化层和非线性激活函数，提高预测模型的收敛性。
由于台风灾害下配电网故障情况类型的预测属于分类问题，故本发明在致灾训练集的基础上，使用交叉熵函数作为损失函数，衡量预测值与实际值的差异程度。接着通过误差反向传播算法得到交叉熵函数对模型中每一个参数的梯度值。最后结合学习率、批大小以及各层神经元数量等超参数，使用小批量Adam算法对预测模型参数进行更新。
3.4双通道预测模型的评估方法
由于本发明中配电网故障情况类型的预测为三分类问题，且致灾测试集各类别的样本数量不相等。为了缓解多数类样本评估结果对预测准确率的主导作用、全面考虑预测模型在各类别的性能表现，本发明以查准率、查全率和F1度量为基本指标体系，并引入宏平均机制综合考虑预测模型在致灾测试集中不同类型样本集合中的表现，具体过程阐述如下。
首先根据致灾测试集输入至预测模型后得到的预测值，统计致灾测试集中每个样本是否属于该受灾类型的实际值与预测值，共可形成三个二分类混淆矩。矩阵形成后，根据矩阵元素得到每一个混淆矩阵对应的一组真阳性TP i、假阳性FP i、真阴性TN i和假阴性FN i，进而得到相应的查准率P i和查全率R i。最后根据宏平均机制得到宏查准率macro-P、宏查全率macro-R和宏F1值macro-F1共三个指标，综合衡量预测模型的性能，具体的计算公式如下。
Figure PCTCN2022081314-appb-000011
Figure PCTCN2022081314-appb-000012
Figure PCTCN2022081314-appb-000013
考虑到预测准确率能直观突出模型的性能，故本发明选取宏查准率、宏查全率、宏F1和准确率共四个指标对台风灾害下配电网故障情况预测模型的性能进行评估。
4、用双通道预测模型进行预测
采集台风过境前气象部门发布的气象预测数据和各研究区域的地理数据、人口数据和电网数据，构建相应的致灾数据集，并将其输入至参数优化后的双通道预测模型中，得到未来台风灾害下各研究区域配电网故障情况类型的预测值。
实施例2
如图6所示，本发明提供的一种一种台风灾害下电网故障预测装置，包括采集模块和计算输出模块；
其中，采集模块，用于采集数据，并将采集的数据传递至计算输出模块；所述数据包括历史动态数据、静态数据和实时台风数据，实时台风数据包括动态数据和静态数据。
计算输出模块，用于根据历史动态数据、静态数据和被预测区域电网永久跳闸次数总和训练预测模型，然后根据预测模型和实时动态数据、静态数据输出电网故障预测值。
实施例3
如图7所示，本发明提供的一种计算机设备，包括电连接的存储器和处理器，其中，存储器上存储有可在处理器上运行的计算程序，所述处理器执行所述计算程序时，实现上述的预测方法的步骤。
实施例4
所述预测装置如果以软件功能单元的形式实现并作为独立的产品销售或使用时，可以存储在一个计算机可读取存储介质中。基于这样的理解，本发明实现上述实施例方法中的全部或部分流程，也可以通过计算机程序来指令相关的硬件来完成，所述的计算机程序可存储于一计算机可读存储介质中，该计算机程序在被处理器执行时，可实现上述各个方法实施例的步骤。其中，所述计算机程序包括计算机程序代码，所述计算机程序代码可以为源代码形式、对象代码形式、可执行文件或某些中间形式等。所述计算机可读介质可以包括：能够携带所述计算机程序代码的任何实体或装置、记录介质、U盘、移动硬盘、磁碟、光盘、计算机存储器、只读存储器(ROM，Read-OnlyMemory)、随机存取存储器(RAM，RandomAccessMemory)、电载波信号、电信信号以及软件分发介质等。需要说明的是，所述计算机可读介质包含的内容可以根据司法管辖区内立法和专利实践的要求进行适当的增减，例如在某些司法管辖区，根据立法和专利实践，计算机可读介质不包括电载波信号和电信信号。
本发明是一种通用的台风灾害下配电网故障情况预测模型，利用本发明的方法进行预测，可以有效降低固有的数据集不均衡问题，提高生成样本的质量。同时，本发明的预测方法考虑了静态数据作用的稳定性和动态数据作用的累计性，进一步提高了预测模型的准确性和可解释性，为配电网应对台风灾害提供了更为准确的预测信息。
以上内容仅为说明本发明的技术思想，不能以此限定本发明的保护范围，凡是按照本发明提出的技术思想，在技术方案基础上所做的任何改动，均落入本发明权利要求书的保护范围之内。

Claims (9)

  1. 数据双驱动的台风灾害下电网故障预测方法，其特征在于，包括以下步骤：
    步骤1，采集台风灾害下电网故障的多元影响数据以及被预测区域电网永久跳闸次数总和，并根据数据的时域变化属性将其划分为静态数据和动态数据，利用静态数据、动态数据和预测区域电网永久跳闸次数总和构建致灾数据集；
    步骤2，对致灾数据集进行均衡化处理；
    步骤3，利用前馈神经网络提取致灾数据集中静态数据的特征，利用长短期记忆网络和多头自注意力机制提取致灾数据集中动态数据的序列特征，建立台风灾害下电网故障的双通道预测模型，并基于样本均衡处理后的致灾数据集，进行模型参数的求解和调优，最终得到优化后的双通道预测模型；并对其性能进行评估；若性能符合要求则进行步骤4，否则继续进行优化；
    步骤4，收集未来台风灾害下预测区域相应的多元影响数据，并构建致灾数据集，将其输入至步骤3中优化后的双通道预测模型，得到未来台风灾害下该研究区域电网故障情况的预测值。
  2. 根据权利要求1所述的数据双驱动的台风灾害下电网故障预测方法，其特征在于，所述步骤1中，静态数据包括森林覆盖率、土地类型、电网的维护程度和人口密度，所述动态数据包括台风中心与区域中心的距离、台风的中心最低气压、台风的近中心最大风速、台风的移动速度、台风的移动方向角、七级风圈半径、预测区域的平均风速 和预测区域的降水量。
  3. 根据权利要求1所述的数据双驱动的台风灾害下电网故障预测方法，其特征在于，所述步骤2的过程为：用Borderline-SMOTE1算法，根据高维空间中致灾数据集的分布对少数类样本集合进行划分，并针对划分后决策边界处的少数类样本进行样本生成；接着通过判别模型检验训练集和测试集数据分布的差异，并根据差异大小对Borderline-SMOTE1算法进行参数调优，最终应用参数优化后的Borderline-SMOTE1算法均衡致灾数据集。
  4. 根据权利要求1所述的数据双驱动的台风灾害下电网故障预测方法，其特征在于，所述步骤2包括以下步骤：
    步骤2.1、使用K近邻算法计算每一个轻度故障类样本的m个最近邻样本；
    步骤2.2、根据轻度故障类样本的m个最近邻样本中轻度故障样本的占比，将其分安全类样本、危险类样本和噪声类样本；
    步骤2.3、针对每一个危险类样本x i，在其K个最近邻样本中选择所需数量的轻度故障类样本；
    步骤2.4、对于每一个被选择的近邻样本x′ j，使用线性插值生成轻度故障类新样本x i,j；
    步骤2.5、将生成的轻度故障类新样本添加至原致灾训练集中，得到更新后的致灾数据集；
    步骤2.6、对更新后的的致灾数据集进行检验，若符合要求则进行步骤3，若不符合要求对Borderline-SMOTE1算法进行调参，直至致灾 数据集符合要求。
  5. 根据权利要求4所述的数据双驱动的台风灾害下电网故障预测方法，其特征在于，所述步骤2.6包括以下步骤：
    步骤2.6.1、针对致灾训练集进行随机采样，使采样后的训练集与致灾测试集的样本数量相等；接着将训练集样本和测试集样本的标签分别置为0和1，混合形成判别数据集，并按比例将其划分为新的训练集和测试集；
    步骤2.6.2、以新的训练集和测试集为基础，以交叉熵函数为损失函数，通过误差反向传播法得到判别模型每个参数值的梯度，进而通过Adam梯度下降算法对判别模型所有参数进行更新，得到致灾训练集与致灾测试集的判别准确率；
    步骤2.6.3、使用判别模型区分致灾训练集与致灾测试集的能力衡量两者的样本分布差异，当判别准确率高于准确率阈值时，对Borderline-SMOTE1算法的最近邻样本数等参数进行调整；当判别准确率低于准确率阈值时，执行步骤3。
  6. 根据权利要求1所述的数据双驱动的台风灾害下电网故障预测方法，其特征在于，所述步骤3包括以下步骤：
    步骤3.1、基于前馈神经网络从静态数据中提取静态特征；基于长短期记忆网络与多头注意力机制从动态数据中提取动态特征；
    步骤3.2、将静态特征和动态特征进行拼接，并通过线性层映射为电网各故障情况类型的预测概率，取最大概率值对应受灾类型为样本的预测故障情况类型，得到预测模型；使用交叉熵函数作为损失函数， 衡量预测值与实际值的差异程度；接着通过误差反向传播算法得到交叉熵函数对模型中每一个参数的梯度值；最后结合学习率、批大小以及各层神经元数量，使用小批量Adam算法对预测模型参数进行更新；
    3.3、以查准率和查全率为基本指标体系，并引入宏平均机制综合考虑预测模型在致灾测试集中不同类型样本集合中的表现，对预测模型进行评估。
  7. 根据权利要求1所述的数据双驱动的台风灾害下电网故障预测方法，其特征在于，所述步骤3.3包括以下步骤：
    步骤3.3.1、根据致灾测试集输入至预测模型后得到的预测值，统计致灾测试集中每个样本是否属于该受灾类型的实际值与预测值，共形成三个二分类混淆矩；
    步骤3.3.2、根据矩阵元素得到每一个混淆矩阵对应的一组真阳性TP i、假阳性FP i、真阴性TN i和假阴性FN i，进而得到相应的查准率P i和查全率R i；
    步骤3.3.3、根据查准率P i、查全率R i和F1度量得到宏查准率macro-P、宏查全率macro-R和宏F1值macro-F1；
    步骤3.3.4、根据宏查准率、宏查全率、宏F1和准确率共四个指标对台风灾害下电网故障情况预测模型的性能进行评估。
  8. 一种台风灾害下电网故障预测装置，其特征在于，包括：
    采集模块，用于采集数据，并将采集的数据传递至计算输出模块；所述数据包括致台风灾害下电网故障的多元影响数据以及被预测区域电网永久跳闸次数总和，以及实时台风数据；
    计算输出模块，用于根据采集的数据集训练预测模型，并根据预测模型和实时台风数据输出电网故障预测值。
  9. 一种计算机设备，其特征在于，包括：电连接的存储器和处理器，所述存储器上存储有可在处理器上运行的计算程序，所述处理器执行所述计算程序时，实现权利要求1-8中任意一项所述的方法的步骤。
PCT/CN2022/081314 2021-09-27 2022-03-17 数据双驱动的台风灾害下电网故障预测方法、装置和设备 WO2023045278A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202111139000.6 2021-09-27
CN202111139000.6A CN113837477B (zh) 2021-09-27 2021-09-27 数据双驱动的台风灾害下电网故障预测方法、装置和设备

Publications (1)

Publication Number Publication Date
WO2023045278A1 true WO2023045278A1 (zh) 2023-03-30

Family

ID=78970684

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/081314 WO2023045278A1 (zh) 2021-09-27 2022-03-17 数据双驱动的台风灾害下电网故障预测方法、装置和设备

Country Status (2)

Country Link
CN (1) CN113837477B (zh)
WO (1) WO2023045278A1 (zh)

Cited By (25)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113537552A (zh) * 2021-04-28 2021-10-22 桂林电子科技大学 基于混合深度学习及注意力机制的学生表现预测方法
CN116522764A (zh) * 2023-04-17 2023-08-01 华中科技大学 一种考虑气候变化影响下的热浪-洪水复合灾害评估方法
CN116629465A (zh) * 2023-07-26 2023-08-22 成都源轮讯恒科技有限公司 智能电网视频监控与风险预测响应系统
CN116910493A (zh) * 2023-09-12 2023-10-20 山东能源数智云科技有限公司 基于多源特征提取的设备故障诊断模型的构建方法及装置
CN116933981A (zh) * 2023-09-15 2023-10-24 安徽方能电气技术有限公司 基于停复电事件数据监测的区域电力稳定性分析评价方法
CN116956197A (zh) * 2023-09-14 2023-10-27 山东理工昊明新能源有限公司 基于深度学习的能源设施故障预测方法、装置及电子设备
CN117034755A (zh) * 2023-08-07 2023-11-10 兰州理工大学 一种融合多头注意力机制的冷轧钢力学性能预测方法
CN117056402A (zh) * 2023-10-12 2023-11-14 国网浙江省电力有限公司余姚市供电公司 一种基于多源信号的电机诊断方法、装置及存储介质
CN117056734A (zh) * 2023-10-12 2023-11-14 山东能源数智云科技有限公司 基于数据驱动的设备故障诊断模型的构建方法及装置
CN117150243A (zh) * 2023-10-27 2023-12-01 湘江实验室 一种基于故障影响解耦网络的故障隔离与估计方法
CN117216485A (zh) * 2023-11-09 2023-12-12 国网山东省电力公司电力科学研究院 一种基于客观赋权的输电录波鸟害故障判别方法及系统
CN117347803A (zh) * 2023-10-25 2024-01-05 爱科特科技(海南)有限公司 一种局部放电检测方法、系统、设备及介质
CN117390948A (zh) * 2023-08-10 2024-01-12 苏州黑盾环境股份有限公司 基于多头注意力长短时记忆神经网络冷水机组监测方法
CN117422438A (zh) * 2023-10-07 2024-01-19 国家电网有限公司华东分部 一种输电线路加固方案的确定方法及装置
CN117495147A (zh) * 2023-12-22 2024-02-02 中国石油大学(华东) 一种考虑段间压裂工艺差异的裂缝网络扩展智能预测方法
CN117522177A (zh) * 2024-01-08 2024-02-06 国网江苏省电力有限公司信息通信分公司 一种智能电网稳定性预测方法
CN117742240A (zh) * 2023-12-28 2024-03-22 广州和兴机电科技有限公司 数控机床的远程监控方法及系统
CN117828490A (zh) * 2024-03-06 2024-04-05 南京信息工程大学 一种基于集成学习的对台风灾害预报方法及系统
CN117909719A (zh) * 2024-03-19 2024-04-19 勤源(江苏)科技有限公司 基于自主学习算法的终端物联设备监控预警方法及系统
CN117973440A (zh) * 2024-04-02 2024-05-03 长江三峡集团实业发展(北京)有限公司 基于LSTM-Transformer模型的区域电离层延迟预测方法
CN117992861A (zh) * 2024-04-04 2024-05-07 国网湖北省电力有限公司 一种电力数据精准度稽查方法及系统
CN118195361A (zh) * 2024-05-17 2024-06-14 国网吉林省电力有限公司经济技术研究院 基于大数据的能源管理方法及系统
CN118211053A (zh) * 2024-04-11 2024-06-18 浙江工业大学 一种基于多模态双提示扩散模型的台风预测方法
CN118364865A (zh) * 2024-06-14 2024-07-19 云南省交通规划设计研究院股份有限公司 公路交通气象环境能见度分钟级预测方法及预警系统
CN118428408A (zh) * 2024-07-04 2024-08-02 浙江省水利河口研究院(浙江省海洋规划设计研究院) 一种基于多ai模型融合及参数优化的山洪预报方法

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113837477B (zh) * 2021-09-27 2023-06-27 西安交通大学 数据双驱动的台风灾害下电网故障预测方法、装置和设备
CN114692615B (zh) * 2022-03-09 2024-08-06 大连理工大学 一种针对小语种的小样本意图识别方法
CN116016219A (zh) * 2022-12-20 2023-04-25 缀初网络技术(上海)有限公司 边缘云服务器流失预测方法及装置

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20130132045A1 (en) * 2011-11-21 2013-05-23 International Business Machines Corporation Natural Disaster Forecasting
CN111191832A (zh) * 2019-12-25 2020-05-22 国电南瑞科技股份有限公司 一种台风灾害配电网杆塔故障预测方法及系统
CN113191585A (zh) * 2021-03-23 2021-07-30 广东电网有限责任公司东莞供电局 一种输电线路的台风致灾风险评估方法
CN113837477A (zh) * 2021-09-27 2021-12-24 西安交通大学 数据双驱动的台风灾害下电网故障预测方法、装置和设备

Family Cites Families (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101354757B (zh) * 2008-09-08 2010-08-18 中国科学院地理科学与资源研究所 一种精细尺度下的动态风险及易损性预测方法
US10050990B2 (en) * 2014-12-29 2018-08-14 Guidewire Software, Inc. Disaster scenario based inferential analysis using feedback for extracting and combining cyber risk information
JP2020042705A (ja) * 2018-09-13 2020-03-19 いすゞ自動車株式会社 故障予測装置、故障予測方法及びプログラム
CN109814527B (zh) * 2019-01-11 2020-11-13 清华大学 基于lstm循环神经网络工业设备故障预测方法及装置
CN109933881A (zh) * 2019-03-06 2019-06-25 武汉大学 一种基于优化深度信念网络的电力电子电路故障诊断方法

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20130132045A1 (en) * 2011-11-21 2013-05-23 International Business Machines Corporation Natural Disaster Forecasting
CN111191832A (zh) * 2019-12-25 2020-05-22 国电南瑞科技股份有限公司 一种台风灾害配电网杆塔故障预测方法及系统
CN113191585A (zh) * 2021-03-23 2021-07-30 广东电网有限责任公司东莞供电局 一种输电线路的台风致灾风险评估方法
CN113837477A (zh) * 2021-09-27 2021-12-24 西安交通大学 数据双驱动的台风灾害下电网故障预测方法、装置和设备

Cited By (39)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113537552A (zh) * 2021-04-28 2021-10-22 桂林电子科技大学 基于混合深度学习及注意力机制的学生表现预测方法
CN116522764A (zh) * 2023-04-17 2023-08-01 华中科技大学 一种考虑气候变化影响下的热浪-洪水复合灾害评估方法
CN116522764B (zh) * 2023-04-17 2023-12-19 华中科技大学 一种考虑气候变化影响下的热浪-洪水复合灾害评估方法
CN116629465A (zh) * 2023-07-26 2023-08-22 成都源轮讯恒科技有限公司 智能电网视频监控与风险预测响应系统
CN116629465B (zh) * 2023-07-26 2024-01-12 李波 智能电网视频监控与风险预测响应系统
CN117034755A (zh) * 2023-08-07 2023-11-10 兰州理工大学 一种融合多头注意力机制的冷轧钢力学性能预测方法
CN117390948A (zh) * 2023-08-10 2024-01-12 苏州黑盾环境股份有限公司 基于多头注意力长短时记忆神经网络冷水机组监测方法
CN116910493B (zh) * 2023-09-12 2023-12-08 山东能源数智云科技有限公司 基于多源特征提取的设备故障诊断模型的构建方法及装置
CN116910493A (zh) * 2023-09-12 2023-10-20 山东能源数智云科技有限公司 基于多源特征提取的设备故障诊断模型的构建方法及装置
CN116956197A (zh) * 2023-09-14 2023-10-27 山东理工昊明新能源有限公司 基于深度学习的能源设施故障预测方法、装置及电子设备
CN116956197B (zh) * 2023-09-14 2024-01-19 山东理工昊明新能源有限公司 基于深度学习的能源设施故障预测方法、装置及电子设备
CN116933981A (zh) * 2023-09-15 2023-10-24 安徽方能电气技术有限公司 基于停复电事件数据监测的区域电力稳定性分析评价方法
CN116933981B (zh) * 2023-09-15 2023-12-08 安徽方能电气技术有限公司 基于停复电事件数据监测的区域电力稳定性分析评价方法
CN117422438B (zh) * 2023-10-07 2024-03-29 国家电网有限公司华东分部 一种输电线路加固方案的确定方法及装置
CN117422438A (zh) * 2023-10-07 2024-01-19 国家电网有限公司华东分部 一种输电线路加固方案的确定方法及装置
CN117056734A (zh) * 2023-10-12 2023-11-14 山东能源数智云科技有限公司 基于数据驱动的设备故障诊断模型的构建方法及装置
CN117056402B (zh) * 2023-10-12 2024-04-02 国网浙江省电力有限公司余姚市供电公司 一种基于多源信号的电机诊断方法、装置及存储介质
CN117056734B (zh) * 2023-10-12 2024-02-06 山东能源数智云科技有限公司 基于数据驱动的设备故障诊断模型的构建方法及装置
CN117056402A (zh) * 2023-10-12 2023-11-14 国网浙江省电力有限公司余姚市供电公司 一种基于多源信号的电机诊断方法、装置及存储介质
CN117347803A (zh) * 2023-10-25 2024-01-05 爱科特科技(海南)有限公司 一种局部放电检测方法、系统、设备及介质
CN117150243A (zh) * 2023-10-27 2023-12-01 湘江实验室 一种基于故障影响解耦网络的故障隔离与估计方法
CN117150243B (zh) * 2023-10-27 2024-01-30 湘江实验室 一种基于故障影响解耦网络的故障隔离与估计方法
CN117216485A (zh) * 2023-11-09 2023-12-12 国网山东省电力公司电力科学研究院 一种基于客观赋权的输电录波鸟害故障判别方法及系统
CN117216485B (zh) * 2023-11-09 2024-01-30 国网山东省电力公司电力科学研究院 一种基于客观赋权的输电录波鸟害故障判别方法及系统
CN117495147A (zh) * 2023-12-22 2024-02-02 中国石油大学(华东) 一种考虑段间压裂工艺差异的裂缝网络扩展智能预测方法
CN117495147B (zh) * 2023-12-22 2024-03-22 中国石油大学(华东) 一种考虑段间压裂工艺差异的裂缝网络扩展智能预测方法
CN117742240A (zh) * 2023-12-28 2024-03-22 广州和兴机电科技有限公司 数控机床的远程监控方法及系统
CN117522177B (zh) * 2024-01-08 2024-03-12 国网江苏省电力有限公司信息通信分公司 一种智能电网稳定性预测方法
CN117522177A (zh) * 2024-01-08 2024-02-06 国网江苏省电力有限公司信息通信分公司 一种智能电网稳定性预测方法
CN117828490A (zh) * 2024-03-06 2024-04-05 南京信息工程大学 一种基于集成学习的对台风灾害预报方法及系统
CN117828490B (zh) * 2024-03-06 2024-05-17 南京信息工程大学 一种基于集成学习的对台风灾害预报方法及系统
CN117909719A (zh) * 2024-03-19 2024-04-19 勤源(江苏)科技有限公司 基于自主学习算法的终端物联设备监控预警方法及系统
CN117909719B (zh) * 2024-03-19 2024-05-24 勤源(江苏)科技有限公司 基于自主学习算法的终端物联设备监控预警方法及系统
CN117973440A (zh) * 2024-04-02 2024-05-03 长江三峡集团实业发展(北京)有限公司 基于LSTM-Transformer模型的区域电离层延迟预测方法
CN117992861A (zh) * 2024-04-04 2024-05-07 国网湖北省电力有限公司 一种电力数据精准度稽查方法及系统
CN118211053A (zh) * 2024-04-11 2024-06-18 浙江工业大学 一种基于多模态双提示扩散模型的台风预测方法
CN118195361A (zh) * 2024-05-17 2024-06-14 国网吉林省电力有限公司经济技术研究院 基于大数据的能源管理方法及系统
CN118364865A (zh) * 2024-06-14 2024-07-19 云南省交通规划设计研究院股份有限公司 公路交通气象环境能见度分钟级预测方法及预警系统
CN118428408A (zh) * 2024-07-04 2024-08-02 浙江省水利河口研究院(浙江省海洋规划设计研究院) 一种基于多ai模型融合及参数优化的山洪预报方法

Also Published As

Publication number Publication date
CN113837477B (zh) 2023-06-27
CN113837477A (zh) 2021-12-24

Similar Documents

Publication Publication Date Title
WO2023045278A1 (zh) 数据双驱动的台风灾害下电网故障预测方法、装置和设备
CN115270965B (zh) 一种配电网线路故障预测方法和装置
CN110766212B (zh) 用于历史数据缺失电场的超短期光伏功率预测方法
CN111680875B (zh) 基于概率基线模型的无人机状态风险模糊综合评价方法
Wu et al. A hybrid support vector regression approach for rainfall forecasting using particle swarm optimization and projection pursuit technology
CN108133225A (zh) 一种基于支持向量机的覆冰闪络故障预警方法
CN115542431B (zh) 一种基于地基云雷达和卫星数据的对流初生监测方法
CN114186641B (zh) 一种基于深度学习的滑坡易发性评价方法
CN106656357A (zh) 一种工频通信信道状态评估系统和方法
Manninen et al. Health index prediction of overhead transmission lines: a machine learning approach
CN113689053B (zh) 一种基于随机森林的强对流天气架空线停电预测方法
CN115603446A (zh) 基于卷积神经网络和云边协同作用的配电台区运行监测系统
CN112434887B (zh) 一种结合网络核密度估计和svm的供水管网风险预测方法
CN113884807A (zh) 基于随机森林和多层架构聚类的配电网故障预测方法
Mei et al. A data‐driven approach to state assessment of the converter valve based on oversampling and Shapley additive explanations
CN117556312A (zh) 一种极端天气下的电网输电线路故障预测方法及系统
CN116415724A (zh) 一种光伏电站运维检修时间的预测方法及装置
CN116151799A (zh) 一种基于bp神经网络的配电线路多工况故障率快速评估方法
CN115630320A (zh) 一种云边协同下基于深度迁移网络的配电台区异常工况诊断方法
Liu et al. Wind Speed Forecasting Method Based on Deep Learning Strategy Using Long Short Term Memory Neural Network And Transformer Model
CN112232592A (zh) 基于Adaboost的输电线路山火灾害风险预警方法
Zhao et al. Power outage classification prediction method based on bagging ensemble learning
Wang et al. Research on House Price Forecast Based on Hyper Parameter Optimization Gradient Boosting Regression Model
CN117131947B (zh) 架空输电线路故障预测方法、装置、设备及存储介质
Li et al. A Fault and Capacity Loss Prediction Method of Wind Power Station under Extreme Weather

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 22871343

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE