A method for predicting the operation volume of power dispatching fault processing
By quantifying the historical data of power grid faults and balancing samples, building a deep confidence network model, the problem of inadequate prediction of power scheduling fault handling operations in the existing technology is solved, and accurate prediction of operation volume in the next 7 days is achieved. Reasonable arrangement of on-duty personnel has improved the risk control of power dispatching and operation.
Patent Information
- Application Number
- CN202111520565.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-12-13
- Publication Date
- 2024-11-05
- Estimated Expiration
- 2041-12-13
AI Technical Summary
The prior art cannot predict the operation amount of power scheduling fault processing in advance, resulting in problems such as long fault processing time, overwork, and blockage of workload due to insufficient scheduling duty during the period of frequent failures.
By obtaining the historical data of power grid faults, performing quantization and labeling, using the SMOTE algorithm for sample balancing, building a DBN model, and using the DBN model to predict the operation amount of power scheduling fault processing.
It has achieved effective prediction of the operation volume of power dispatching fault handling in the next 7 days, helping to arrange on-duty personnel in advance, avoiding problems caused by insufficient personnel, and improving the risk control of power dispatching operation.
Smart Images

Figure CN114219146B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of power grid dispatching control, and in particular to a method for predicting an operation amount of power dispatching fault processing. Background Art
[0002] The number of people on duty should be reasonably arranged for power dispatching duty to meet the requirements of duty workload. At present, the dispatching duty mode of Guangxi Power Grid is mainly arranged according to the planned workload. During the period of frequent faults, the long fault handling time, overwork of personnel, and workload congestion caused by insufficient dispatching duty personnel have long existed. Reliable prediction of the power dispatching fault handling operation volume is a prerequisite for realizing reasonable dispatching tidal duty. If the dispatching fault handling operation volume cannot be predicted in advance, the duty personnel will not be able to be well scheduled in advance. Therefore, it is necessary to improve the existing technology. Summary of the invention
[0003] The purpose of the present invention is to provide a method for predicting the amount of power dispatching fault handling operations, which can solve the problems of the prior art that cannot predict the amount of dispatching fault handling operations in advance, resulting in long fault handling time, overwork of personnel, workload congestion, etc. due to insufficient dispatching on-duty personnel during periods of frequent faults.
[0004] The objective of the present invention is achieved through the following technical solutions:
[0005] The present invention provides a method for predicting an operation amount of a power dispatching fault processing, comprising the following steps:
[0006] Step S1, obtaining power grid fault historical data;
[0007] Step S2: quantify the historical data of power grid faults to form power dispatch fault processing operation data, and label the meteorological data and power load;
[0008] Step S3, performing sample balancing on the quantized data using the SMOTE algorithm;
[0009] Step S4: Building a linear fitting model based on the balanced data, and selecting key training data types according to data weights;
[0010] Step S5: determine and build a DBN model;
[0011] Step S6: adjusting network parameters, selecting a suitable training algorithm, and using historical grid fault data to train the DBN model;
[0012] Step S7: Use the DBN model to predict the amount of power dispatching fault handling operations in the future.
[0013] Furthermore, the quantification processing of the power grid fault history data includes:
[0014] Step S201: Use a certain time unit as an interval division, assign a score to each power grid fault type within the certain time unit, and the same type of power grid fault type will have the same score. The power grid fault will be converted into a score that directly reflects the amount of dispatch fault processing operations;
[0015] Step S202: Count the number of times the same type of power grid fault type occurs within the certain time unit, calculate the total score of the same type of power grid fault type, and form quantified power grid fault operation data.
[0016] Furthermore, the time unit is in hours or days.
[0017] Furthermore, the total score of the same type of power grid fault type=the single score of each power grid fault type×the number of times the power grid fault type occurs.
[0018] Furthermore, the key training data types include temperature, rainfall, cloud cover, air pressure, wind speed, and power load.
[0019] Furthermore, the power grid fault history data includes meteorological data and power load at the time of the fault.
[0020] The present invention proposes a prediction algorithm for the amount of power grid fault handling operations applied to tidal scheduling, which can effectively predict the amount of power dispatch fault handling operations in the next 7 days. As the basis for the power dispatch duty plan, it is possible to arrange enhanced duty in advance when it is known in advance that the amount of fault handling operations is large. The number of on-duty personnel for a specific period of time can be reasonably arranged in advance, effectively avoiding a series of problems caused by insufficient dispatch personnel during the period of frequent faults. It will be possible to predict the amount of power dispatch fault handling operations in the short term in the future, which is of great significance for the advance risk control of power dispatch operation, and changes the existing dilemma of post-emergency handling. BRIEF DESCRIPTION OF THE DRAWINGS
[0021] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.
[0022] Figure 1 It is a flow chart of the method for predicting the operation amount of power dispatching fault processing of the present invention;
[0023] Figure 2 This is a schematic diagram of the SMOTE algorithm idea;
[0024] Figure 3 This is a schematic diagram of the process of selecting key impact data based on the LIME algorithm;
[0025] Figure 4 Schematic diagram of the process of predicting fault handling volume using DBN model;
[0026] Figure 5 This is a flowchart for training a DBN model;
[0027] Figure 6 This is a schematic diagram of the deep belief network model structure;
[0028] Figure 7 A comparison chart of the power dispatching fault handling operation volume predicted by the method of the present invention and the actual value in the Beihai power grid area;
[0029] Figure 8 A comparison chart of the predicted operation amount of power dispatching fault processing in Guilin power grid area and the actual value using the method of the present invention;
[0030] Fig. 9 A comparison chart of the power dispatching fault handling operation volume and the actual value predicted by the method of the present invention in Yulin network area;
[0031] Fig.10 A comparison chart of the power dispatching fault handling operation volume predicted by the method of the present invention and the actual value in Liuzhou power grid area;
[0032] Fig.11 This is a statistical chart of the accuracy of power dispatch fault volume prediction. DETAILED DESCRIPTION
[0033] The embodiments of the present disclosure are described in detail below with reference to the accompanying drawings.
[0034] The following describes the embodiments of the present disclosure through specific examples, and those skilled in the art can easily understand other advantages and effects of the present disclosure from the contents disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present disclosure, rather than all of the embodiments. The present disclosure can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present disclosure. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict. Based on the embodiments in the present disclosure, all other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present disclosure.
[0035] The method for predicting the operation amount of power dispatching fault processing of the present invention comprises the following steps:
[0036] Step S1, obtaining power grid fault historical data.
[0037] The historical data of power grid faults include meteorological data, power load and other forecast data at the time of the fault.
[0038] Step S2: quantify the historical data of power grid faults to form power dispatch fault processing operation data, and label the meteorological data and power load.
[0039] Furthermore, the quantification of the historical data of power grid faults includes:
[0040] Step S201: Use a certain time unit as the interval division. The time unit is generally in hours or days. Each power grid fault type within the certain time unit is assigned a score. The same type of power grid fault type has the same score. The power grid fault will be converted into a score that directly reflects the amount of dispatch fault processing operations, as shown in Table 1.
[0041]
[0042]
[0043] Table 1
[0044] The type of power grid fault is set according to actual conditions and is not limited to the types given in Table 1. Table 1 is only an example to illustrate the specific process of the present invention.
[0045] Step S202: Count the number of times the same type of power grid fault type occurs within the certain time unit, calculate the total score of the same type of power grid fault type, and form quantified power grid fault operation data, as shown in Table 2.
[0046] The total score of the same type of power grid fault type = the single score of each power grid fault type × the number of times the power grid fault type occurs.
[0047]
[0048]
[0049] Table 2
[0050] Step S3: Perform sample balancing on the quantized data using the SMOTE algorithm.
[0051] Due to the uneven distribution of fault handling operations in power dispatching, the historical sample data is unbalanced, which will make it difficult for the model to converge. Therefore, the SMOTE (synthetic minority oversampling technique) algorithm is used to balance the number of samples according to the magnitude of fault handling operations. Taking the Beihai network area in Guangxi as an example, the number of fault-free days accounts for 34.6% of the total. The imbalance of samples will make it difficult for model training to converge. Therefore, it is necessary to oversample the collected samples to balance the samples. This will be achieved through the SMOTE algorithm. The core idea is to use KNN (nearest neighbor classifier) to generate multiple adjacent samples in the feature space for minority class samples. The algorithm idea is shown in Figure 2 .
[0052] Step S4: Build a linear fitting model based on the balanced data and select key training data types according to data weights (LIME algorithm).
[0053] Selection of training data types: The coupling degree between prediction information and prediction target fundamentally determines the upper limit of prediction accuracy. In order to accurately evaluate the influencing factors of power dispatch fault handling operation volume, a linear fitting model will be established through the local interpretable model-agnostic explanation (LIME) algorithm to interpret multiple samples. Finally, data such as temperature, rainfall, cloud cover, air pressure, wind speed, and power load will be selected as key influencing information of power dispatch fault handling operation volume. The LIME algorithm sample selection process can be found in Figure 3 ,The model data sample types are shown in Table 3.
[0054]
[0055] Table 3
[0056] Step S5: Determine and build a DBN model (determine the depth and breadth).
[0057] The change in the prediction of the operation volume of power dispatching fault handling can be regarded as a time series model, but due to its high data discreteness, stronger randomness, and extremely uneven sample distribution, taking the Beihai network area in Guangxi as an example, the number of fault-free days accounts for 34.6% of the total number, so it is necessary to perform SMOTE oversampling on the samples to balance the samples, and the oversampling of samples will destroy the sequence, so it is not suitable for recurrent neural networks such as LSTM and GRU. In summary, the model type uses DBN modeling that supports pre-training, and the topological structure of the model is shown in Figure 4 In this model, DBN will receive pre-processed multi-dimensional information such as rainfall, wind power, and load, and calculate the predicted value of the output power dispatch fault handling operation volume.
[0058] Step S6: Adjust network parameters, select a suitable training algorithm, and use historical power grid fault data to train the DBN model.
[0059] For ease of understanding, the training principle of the deep belief network is as follows:
[0060] DBN is a multi-layer neural network composed of multiple restricted Boltzmann machines (RBM) stacked end to end. Its architecture is as follows: Figure 6 Compared with the traditional multi-layer perceptron, the advantage of DBN is that it can support pre-training of each RBM layer by layer, which greatly reduces the difficulty of training deep networks. In addition, the pre-trained DBN will be easier to converge in formal training and will not be limited to suboptimal solutions. When DBN is connected to the Softmax layer for training, DBN can be applied to pattern recognition.
[0061] Different from the traditional BP neural network, DBN is a model defined based on energy, and its training process is energy-driven.
[0062] The energy function of DBN can be expressed as:
[0063]
[0064] In the formula, w ij is the connection weight, b i is the bias of the visible neuron, c j is the bias of the hidden neuron, and θ is the parameter set of all connection weights and biases. i and hidden neurons h j The product of represents the degree of correlation between the two, which is related to the connection weight w ij When they are consistent, the maximum likelihood estimator of θ can be obtained.
[0065] And the activation of neurons in DBN is based on probability distribution, and the joint probability distribution of (v,h) states can be expressed as:
[0066]
[0067] Where Z(θ) is a normalization function, the purpose of which is to make the sum of all probability distributions equal to 1, which can be expressed as:
[0068] Z(θ)=∑ v,h e -E(v,h丨θ) (3)
[0069] The activation probability of each visible neuron and hidden neuron in DBN can be expressed as:
[0070] P(v i= 1|h,θ) = σ(b i + ∑ j h j w ij )(4)
[0071] P(h j = 1丨v,θ) = σ(c j + ∑ i v i w ij )(5)
[0072] where σ(·) is the activation function, and in this paper, the sigmoid function is taken
[0073] The training process of the DBN model can be divided into the following two parts:
[0074] Step S601: Use the CD algorithm to perform unsupervised layer-by-layer initialization training on each RBM layer to obtain appropriate weights and biases within the RBM layer. This is the pre-training process of the DBN, which is reflected as follows:
[0075] Step S6011: Initialize the weights, biases, and other network parameters;
[0076] Step S6012: Let the training data (input data) be v i , calculate the probability of the corresponding hidden layer nodes according to Equation (5), and obtain the activation vector data h of the hidden layer nodes from this probability distribution j ;
[0077] Step S6013: Calculate the outer product of v i and h j , which is the "positive gradient";
[0078] Step S6014: According to Equation (4), return from h j and obtain the activation vector data v' of a reconstructed visible layer node i ; then, according to Equation (31), obtain the activation vector data h' of the hidden layer nodes again from v' i (If the process of steps 1 - 3 loops K times, it is the CD-K algorithm); j
[0079] Step S6015: Calculate the outer product of v' i and h' j , which is the "negative gradient";
[0080] Step S6016: Calculate the KL divergence according to the positive and negative gradients, and correct the weights and biases with a certain learning rate ε, as
[0081] Step S6017, determine whether the energy function reaches the training target value, if so, the training ends; otherwise, repeat step S6016.
[0082] Step S602: stack the RBMs end to end to form a DBN, and use the back propagation algorithm to perform supervised training on the DBN. The specific implementation steps are:
[0083] Step S6021, adding training data (input data and output data) by batch;
[0084] Step S6022, activating the neuron according to the input data and the activation function σ(·) to obtain the corresponding DBN output value;
[0085] Step S6023: Calculate the loss function (cross entropy) of the DBN output and the added output data, that is,
[0086] Step S6024: According to the gradient of the cross entropy, etc., the weights and biases are corrected, such as (Gradient descent method);
[0087] Step S6025, determine whether the loss function L reaches the set target accuracy. If so, the training ends; otherwise, repeat steps S6021 to S6024.
[0088] The layer-by-layer pre-training process of DBN improves the ability to extract key features in complex models, and reduces the difficulty of training deep networks to a considerable extent. After training, DBN has the characteristics of training data and can identify and output appropriate categories based on the input prediction information.
[0089] Step S7: Use the DBN model to predict the amount of power dispatching fault handling operations in the future.
[0090] A daily prediction model for the amount of power dispatching fault handling operations is established based on the deep belief network (DBN). This part designs deep learning theory: after modeling, the DBN model receives input forecast meteorological data such as temperature, rainfall, cloud cover, air pressure, wind speed, and power load, and can predict the output of the amount of power dispatching fault handling operations in the short term in the future.
[0091] In order to illustrate the beneficial effects of the present invention, the following is described in conjunction with specific experimental data:
[0092] At present, the effectiveness of the method of the present invention has been verified in practice in Beihai, Yulin, Guilin, Liuzhou, Qinzhou and other power grid areas. The following is a prediction display. The modeling data time range is from January 1, 2017 to December 31, 2020. The model predicts the power dispatching fault handling operation volume for Beihai, Guilin, Yulin, Liuzhou and other power grid areas. The prediction time range is February-May 2021. The prediction effect is shown in Figure 7 to Fig.10 The prediction accuracy is shown in Table 4 and Fig.11 .
[0093]
[0094]
[0095] Table 4
[0096] As can be seen from the above figures and tables, the predicted value of the power dispatching fault handling operation volume has a high degree of fit with the actual value, and the average prediction accuracy of each network area is about 80%. The DBN model has good prediction capabilities. In particular, during severe weather, it can have a good early warning effect on abnormal power dispatching fault handling operation volume. It can serve as an early warning basis for emergency duty.
[0097] The above is only for explaining the implementation mode of the present invention and is not intended to limit the present invention. For those skilled in the art, any modification, equivalent substitution, improvement, etc. made within the spirit and principle of the present invention without creative work should be included in the protection scope of the present invention.
Claims
1. A method for predicting the operation amount of power dispatching fault processing, characterized in that: The following steps are involved: Step S1, obtaining power grid fault history data, wherein the power grid fault history data includes meteorological data and power load at the time of the fault; Step S2: quantify the historical data of power grid faults to form the operation volume data of power dispatch fault processing, and label the meteorological data and power load; The quantification processing of the power grid fault historical data includes: Step S201: Use a certain time unit as an interval division, assign a score to each power grid fault type within the certain time unit, and the same type of power grid fault type will have the same score. The power grid fault will be converted into a score that directly reflects the amount of dispatch fault processing operations; Step S202: Count the number of times the same type of power grid fault type occurs within the certain time unit, calculate the total score of the same type of power grid fault type, and form quantified power grid fault operation data, wherein the total score of the same type of power grid fault type = the single score of each power grid fault type × the number of times the power grid fault type occurs; Step S3, performing sample balancing on the quantized data using the SMOTE algorithm; Step S4: Building a linear fitting model based on the balanced data, and selecting key training data types according to data weights; Step S5: determine and build a DBN model; Step S6: adjusting network parameters, using CD algorithm and back propagation algorithm, and using power grid fault history data to train the DBN model; Step S7: Use the DBN model to predict the amount of power dispatching fault handling operations in the future.
2. The method for predicting the operation amount of power dispatching fault processing according to claim 1, characterized in that: The time unit is in hours or days.
3. The method for predicting the operation amount of power dispatching fault processing according to claim 1, characterized in that: The total score of the same type of power grid fault type=the single score of each power grid fault type×the number of times the power grid fault type occurs.
4. The method for predicting the operation amount of power dispatching fault processing according to claim 1, characterized in that: The key training data types include temperature, rainfall, cloud cover, air pressure, wind speed, and power load.
Citation Information
Patent Citations
Emergency material demand prediction method and system for power grid meteorological disasters
CN111639817A
Power distribution network fault risk early warning method and system based on deep learning
CN112749904A
Multi-class imbalance fault classification method based on reinforcement learning and knowledge distillation
CN113222035A
Power system future time period simulation method and system based on real-time and prediction data
CN113361108A
Power grid event diagnosis autonomous learning method and system based on artificial intelligence
CN113361663A
