Power distribution network abnormal working condition detection method based on quantile regression gradient boosting tree

Abstract

The invention discloses a power distribution network abnormal working condition detection method based on a quantile regression gradient boosting tree, and the method comprises the following steps: solving a current quantile value based on a quantile theory; a quantile regression gradient boosting tree model is introduced to describe the current distribution situation, and the current value is accurately estimated; reading a current value at the time t, and calculating a corresponding instantaneous abnormal current threshold IAT and a timing abnormal current threshold PAT; and performing logic judgment based on the current value, the instantaneous abnormal current threshold IAT and the timing abnormal current threshold PAT, and judging whether abnormity occurs or not. According to the method, the requirements for sampling information and sampling frequency are not high, the good detection and recognition rate is achieved for abnormal working conditions of various current fluctuation degrees, especially the abnormal condition of small abnormal current, and the abnormal working conditions can be detected timely and accurately.

Description

Technical Field

[0001] The present invention belongs to the technical field of abnormal operating condition detection of low-voltage distribution networks, and specifically relates to a distribution network abnormal operating condition detection method based on a quantile regression gradient boosting tree. Background Art

[0002] With the progress and development of society, the power industry is also constantly improving and improving in the torrent of time. The network scale is growing, and the types of power energy are becoming more and more diverse. With the research and development of green renewable energy and the update and improvement of smart grids, more and more new energy sources are being put into daily power production, forming a more complex power network. In the "generation, transmission, transformation, distribution and use" architecture of the power system, the distribution network directly faces the vast number of users and is an important guarantee for the reliability of power supply. With the progress of society, economic development, and the improvement of people's living standards, higher requirements are put forward for the power industry to provide more flexible, more stable and more reliable energy supply. In various survey data, it can be found that the probability of abnormal conditions in the distribution network area is higher than that of transmission lines. The distribution network itself covers a large area, with complex line topology, diverse operating conditions, and a wide variety of equipment. Under different operating conditions, abnormal conditions of different degrees may occur in different locations. In the operation of the power grid, if the monitoring is not timely, the fault will occur and it is difficult to eliminate it automatically. The voltage fluctuation caused by it will cause greater harm to the equipment, and there may be regional power outages, and even threaten personal safety. If we can determine as quickly and accurately as possible after an accident occurs, whether an accident has occurred, the type of accident, the location of the accident, etc., or even provide early warning of possible accidents, the losses caused by the accident will be greatly reduced. Therefore, it is very necessary to detect abnormal working conditions.

[0003] Quantile, also called quantile point, is a numerical point that divides the probability distribution range of a random variable into several equal parts. It has extremely wide applications in statistics. It can help us understand the central trend of data, and can reliably screen and identify outliers and outliers in the data. It plays an important role in data analysis.

[0004] At present, there are many studies on abnormal condition detection of distribution networks. According to the data characteristics of the methods used, the diagnosis and positioning methods can be roughly divided into the following categories: methods based on steady-state signals, such as the voltage, current, power, etc. of the system for abnormal condition detection; methods based on transient signals, such as the fluctuation of voltage and current caused by disturbances for abnormal condition detection; or by injecting characteristic signals into the system through signal injection equipment, and comparing the differences in the injected signals in different areas for abnormal condition detection. However, low-voltage distribution networks have variable operating modes, various equipment transformation situations, and messy lines. Many methods used in medium-voltage distribution networks are not applicable to low-voltage distribution networks.

[0005] The present invention combines quantile theory with abnormal operating condition detection of low-voltage distribution network, and realizes relatively sensitive and accurate detection of abnormal operating condition of distribution network by dynamically changing the set current threshold, so as to timely understand the operation status of power grid, provide basis and support for subsequent maintenance, avoid casualties as much as possible, and reduce the losses caused by them. Summary of the invention

[0006] The present invention is proposed to solve the problems existing in the prior art, and its purpose is to provide a distribution network abnormal condition detection method based on quantile regression gradient boosting tree.

[0007] The technical solution of the present invention is: a method for detecting abnormal operating conditions of a distribution network based on a quantile regression gradient boosting tree, comprising the following steps:

[0008] A. Based on the quantile theory, the current quantile value is obtained;

[0009] B. The quantile regression gradient boosting tree model is introduced to characterize the current distribution situation and make a more accurate estimate of the current value;

[0010] C. Read the current value at time t, and calculate the corresponding instantaneous abnormal current threshold IAT and timing abnormal current threshold PAT;

[0011] D. Perform logical judgment based on the current value, instantaneous abnormal current threshold IAT, and timing abnormal current threshold PAT to determine whether an abnormality occurs.

[0012] Furthermore, step A is based on the quantile theory to obtain the current quantile value, and the specific process is as follows:

[0013] First of all, quantiles are points that divide the probability distribution range of a random variable into several equal parts. Common quantiles include quantiles, quartiles, and percentiles.

[0014] Then, according to the corresponding theoretical concepts, the corresponding quantile expression is obtained:

[0015]

[0016] Among them, I is a random variable, indicating the corresponding current value; X is an independent variable used to predict I; F is the cumulative distribution function of the current I; Q represents the conditional quantile of I given X; t represents the corresponding measurement time point; inf is a function, indicating the minimum value of the subsequent function. In the above formula, the set it represents corresponds to the minimum value, which is the maximum boundary of all elements that meet the conditions; τ is the quantile probability level, indicating the corresponding probability value P r ,Right now That is, given X t In the case of I t Less than or equal to its corresponding quantile Q t probability.

[0017] Furthermore, the quantile regression gradient boosting tree model in step B is based on the quantile regression tree, and its quantile regression tree is as follows:

[0018] First, the quantile constructs a binary tree through binary division, and the two branches are divided into two again, and the iteration is repeated until the stopping condition is met. Finally, the whole tree is divided into multiple different leaf nodes. By judging which leaf node the input test variable belongs to, the output value on the corresponding leaf node is obtained as the quantile prediction value;

[0019] Then, the kth explanatory variable and its corresponding value v are selected as the splitting variable and splitting point, and the corresponding training set is divided into two sub-areas S1 and S2;

[0020] S1(k,v)={x|x (k) ≤v},S2(k,v)={x|x (k) >v}(2)

[0021] Among them, S represents the sub-region, k is the sequence number corresponding to the selected explanatory variable, v is the value of the selected segmentation variable, and x is the independent variable;

[0022] Then, the corresponding output value of the sub-region is:

[0023]

[0024] in, is the output value of the corresponding area, m is the number of the area, and n is the number of elements in the corresponding area;

[0025] Then, in order to obtain the final partition node with better prediction effect, the partition node must satisfy the following formula:

[0026]

[0027] The above division process is repeated iteratively until the termination condition is met. At this time, the training set is divided into L sub-regions or leaf nodes S1, S2, …, S L , the response variable on its leaf node is:

[0028] y l ={y|x i ∈S l},l=1,2,...,L(5)

[0029] Finally, the quantile regression tree, at the corresponding τ quantile point, the corresponding output value on its leaf node is:

[0030]

[0031] Among them, inf is the lower bound; F(·) is the empirical conditional distribution function, and its formula is as follows:

[0032]

[0033] Furthermore, the quantile regression gradient boosting tree in step B performs ensemble learning on the basis of a single quantile regression tree to improve the prediction accuracy of the model. The model uses the steepest descent method to fit and generate a quantile regression tree with the negative gradient of the loss function as the approximate value of the residual. Through continuous iteration, multiple quantile regression trees are finally generated, and these trees are linearly added to obtain the final model.

[0034] Furthermore, in step C, the instantaneous abnormal current threshold IAT will immediately sound an alarm when the collected current value exceeds the set value, indicating an abnormality has occurred. The calculation formula is as follows:

[0035]

[0036] in, is the predicted value corresponding to the above quantile regression gradient boosting tree model, and γ is a parameter set by human experience, usually 1.05-1.3.

[0037] Furthermore, in step C, the timing abnormal current threshold PAT is set. When the collected current value exceeds the set value for a certain period of time, an alarm is issued, indicating that an abnormality has occurred. The calculation formula is as follows:

[0038]

[0039] in, is the predicted value corresponding to the above quantile regression gradient boosting tree model.

[0040] Furthermore, step D performs a logical judgment based on the current value, the instantaneous abnormal current threshold IAT, and the timing abnormal current threshold PAT to determine whether an abnormality occurs. The specific process is as follows:

[0041] d1. Compare the current value with the instantaneous abnormal current threshold IAT

[0042] If the current value is greater than the instantaneous abnormal current threshold IAT t , then an exception occurs;

[0043] If the current value is not greater than the instantaneous abnormal current threshold IAT t , compare the current value with the timing abnormal current threshold PAT t Make comparisons;

[0044] d2. If the current value is less than the timing abnormal current threshold PAT t , then there is no abnormality and it is in normal operation;

[0045] If the current value is greater than the timing abnormal current threshold PAT t , there may be an abnormality, read the current value at the next time point and calculate the corresponding current threshold;

[0046] d3. At the next time point, if the current value is greater than the instantaneous abnormal current threshold at the next time point, an abnormality occurs; if the current value is less than the instantaneous abnormal current threshold at the next time point, the current value is compared with the timing abnormal current threshold at the next time point. If the current value is less than the timing abnormal current threshold, the cycle ends and there is no abnormality; if the current value is greater than the timing abnormal current threshold, the current value at the next time point is read and the corresponding threshold is calculated;

[0047] d4. Repeat step d3 until n-1 cycles are completed and the current value I is obtained. t+n-1 , n represents the timing time period of the timing abnormal current threshold.

[0048] Furthermore, if the current value I t+n-1 If the current is greater than the timed abnormal current threshold or the instantaneous abnormal current threshold of the corresponding time, an alarm will be triggered immediately, indicating that an abnormality has occurred.

[0049] The beneficial effects of the present invention are as follows:

[0050] The present invention applies quantile correlation theory to the detection of abnormal operating conditions in low-voltage distribution networks. Based on the quantile regression gradient boosting tree model, the current change trend is predicted and characterized, the abnormal current threshold is dynamically changed, and the abnormal operating conditions in the distribution network are detected through a comparison algorithm between the threshold and the current measurement value.

[0051] The present invention has low requirements on sampling information and sampling frequency, has a good detection and recognition rate for abnormal working conditions with various degrees of current fluctuation, especially abnormal conditions with small abnormal currents, and can detect abnormal working conditions in a timely and accurate manner. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 It is a flowchart of overall anomaly detection based on quantile regression gradient boosting tree of the present invention;

[0053] Figure 2 is a flow chart of the abnormality detection link of the present invention;

[0054] Figure 3 It is a diagram for adjusting the number parameters of the gradient boosting tree model tree of the timing abnormal current in the present invention. DETAILED DESCRIPTION

[0055] Hereinafter, the present invention will be described in detail with reference to the accompanying drawings and embodiments:

[0056] like Figures 1 to 3 As shown, the distribution network abnormal condition detection method based on quantile regression gradient boosting tree includes the following steps:

[0057] A. Based on the quantile theory, the current quantile value is obtained.

[0058] Quantile, also called quantile, is a numerical point that divides the probability distribution range of a random variable into several equal parts. Common quantiles include quantiles, quartiles, percentiles, etc.

[0059] According to the corresponding theoretical concepts, the corresponding quantile expression can be obtained:

[0060]

[0061] Among them, I is a random variable, indicating the corresponding current value; X is an independent variable used to predict I; F is the cumulative distribution function of the current I; Q represents the conditional quantile of I given X; t represents the corresponding measurement time point; inf is a function, indicating the minimum value of the subsequent function. In this formula, the set it represents corresponds to the minimum value, which is the maximum boundary of all elements that meet the conditions; τ is the quantile probability level, indicating the corresponding probability value P r ,Right now That is, given X t In the case of I t Less than or equal to its corresponding quantile Q t probability.

[0062] Specifically, the collected current data is used to predict the current threshold based on the quantile. Determining the quantile probability level is equivalent to limiting the probability distribution of the corresponding value. When the quantile probability level is high, the predicted current threshold will include all current values ​​within the corresponding probability range. Then, when a value outside the threshold range appears, it can be determined that an abnormality has occurred.

[0063] B. Introduce the quantile regression gradient boosting tree model to characterize the current distribution trend and make a more accurate estimate of the current value

[0064] b1. Quantile regression tree

[0065] Regression tree is a machine learning algorithm that constructs a binary tree through binary search. On this basis, the two branches are divided into two again, and the iteration is repeated until the stopping condition is met. Finally, the whole is divided into multiple different leaf nodes. By judging which leaf node the input test variable belongs to, the output value on the corresponding leaf node is obtained as the quantile prediction value.

[0066] Specifically, the kth explanatory variable and its corresponding value v are selected as the splitting variable and splitting point, and the corresponding training set is divided into two sub-areas S1 and S2.

[0067] S1(k,v)={x|x (k) ≤v},S2(k,v)={x|x (k) >v}(2)

[0068] Among them, S represents the sub-region, k is the sequential number corresponding to the selected explanatory variable, v is the value of the selected segmentation variable, and x is the independent variable.

[0069] The corresponding output value of the sub-region is:

[0070]

[0071] in, is the output value of the corresponding area, m is the number of the area, and n is the number of elements in the corresponding area.

[0072] In order to obtain a partition node with better final prediction effect, the partition node must satisfy the following formula:

[0073]

[0074] The above division process is repeated repeatedly until the termination condition is met. At this point, the training set is divided into L sub-regions or leaf nodes S1, S2, …, S L , the response variable on its leaf node is:

[0075] y l ={y|x i ∈S l},l=1,2,...,L(5)

[0076] Quantile regression tree, at the corresponding τ quantile point, the corresponding output value on its leaf node is:

[0077]

[0078] Among them, inf is the lower bound; F(·) is the empirical conditional distribution function, and its formula is as follows:

[0079]

[0080] b2. Quantile Regression Gradient Boosting Tree

[0081] Quantile regression gradient boosting tree is an ensemble learning based on a single quantile regression tree, which effectively improves the prediction accuracy of the model. This model uses the steepest descent method and the negative gradient of the loss function as the approximate value of the residual to fit and generate a quantile regression tree. Through continuous iteration, multiple quantile regression trees are finally generated, and these trees are linearly added to obtain the final model. The specific steps are as follows:

[0082] b21. Initialize g [0] (x i )=0,i=1,2,…,n;

[0083] b22. Calculate the negative gradient and use it as an approximation of the residual;

[0084] b23. Fit the residuals to the independent variables and construct a quantile regression tree in

[0085] is the parameter of the quantile regression tree, and its expression is:

[0086]

[0087] b24. Update the boosted tree model:

[0088]

[0089] b25. Repeat the iteration until the difference in the residual sum of squares is less than 10 -6 , and get the quantile regression gradient

[0090] Tree-raising model:

[0091]

[0092] When the corresponding test data set z is input, the corresponding prediction value is

[0093]

[0094] Specifically, in the actual model training process, it is necessary to adjust the model parameters. The quantile regression gradient boosting tree model is trained using the training data set. The selection of key parameters directly affects the performance of the model and the accuracy of prediction.

[0095] Specifically, in the present invention, it is necessary to determine the parameters of the tree depth, the number of regression trees, and the number of variables. Increasing the number of regression trees usually improves the complexity and predictive ability of the model, but correspondingly, the problem of overfitting will also occur. Controlling the depth of each tree can prevent the learned model from being too complex. For these two parameters, the k-fold cross-validation method is used to find the optimal parameter values, and the mean absolute error is used as the evaluation indicator, which can reflect the overall prediction deviation.

[0096] Specifically, k-fold cross validation is to randomly divide the data set into k non-overlapping subsets of the same size, use k-1 subsets as training sets to train the model, and use the remaining subset as a test set for testing, calculate the evaluation index, repeat the above steps k times, so that all data sets become test sets, obtain k evaluation index values, and take the average of the k evaluation indexes as the final result. The formula for its evaluation index is as follows:

[0097]

[0098] C. Read the current value at time t and calculate the corresponding instantaneous abnormal current threshold IAT and timing abnormal current threshold PAT

[0099] On the basis of the above quantile solution, the instantaneous abnormal current threshold and the timing abnormal current threshold are obtained.

[0100] The instantaneous abnormal current threshold IAT will immediately sound an alarm when the collected current value exceeds the set value, indicating an abnormality. The calculation formula is as follows:

[0101]

[0102] The timed abnormal current threshold PAT, when the collected current value exceeds the set value for a certain period of time, an alarm is issued, and an abnormality occurs. The calculation formula is as follows:

[0103]

[0104] in, is the prediction value corresponding to the above quantile regression gradient boosting tree model, γ is a parameter set by human experience, usually 1.05-1.3, and the corresponding τ in IAT and PAT is 0.99 and 0.97 based on experience.

[0105] Specifically, a certain number of data sets are used to train the model, and data is collected in real time and used as input to input the model for threshold prediction. Each time data is collected, the threshold is recalculated and updated to generate a dynamically changing threshold, which is convenient for discovering abnormal situations with smaller disturbances.

[0106] D. Make logical judgment based on current value, instantaneous abnormal current threshold IAT, and timing abnormal current threshold PAT to determine whether an abnormality occurs

[0107] d1. Compare the current value with the instantaneous abnormal current threshold. If the current value is greater than IAT t , an abnormality occurs; otherwise, the current value is compared with the timing abnormal current threshold;

[0108] d2. If the current value is less than PAT t , then there is no abnormality now and it is in normal operation; otherwise, there may be an abnormality, read the current value at the next time point and calculate the corresponding current threshold;

[0109] d3. At the next time point, if the current value is greater than the instantaneous abnormal current threshold at the next time point, an abnormality occurs; if the current value is less than the instantaneous abnormal current threshold at the next time point, the current value is compared with the timing abnormal current threshold at the next time point. If the current value is less than the timing abnormal current threshold, the cycle ends and there is no abnormality; if the current value is greater than the timing abnormal current threshold, the current value at the next time point is read and the corresponding threshold is calculated;

[0110] d4. Repeat step d3 until n-1 cycles are completed and the current value I is obtained. t+n-1 , n represents the timing time period of the timing current threshold;

[0111] d5. If I t+n-1 If the current is greater than the timing current threshold or instantaneous current threshold of the corresponding time, an alarm will be triggered immediately, indicating that an abnormality has occurred.

[0112] Embodiment 1

[0113] Using the actual data collected by Tianjin Keppel as load operation data, a 44-node 0.4kV low-voltage distribution network circuit model was built, and simulation software was used for operation simulation. The current data was acquired with a sampling interval of one minute. In practice, the fixed fault current threshold is usually set to 1.1 times the maximum operating current. For the occurrence of large current abnormal conditions, the fixed current threshold can also play a corresponding abnormal detection effect. Therefore, in terms of experimental settings, this experiment will focus on the abnormalities or disturbances of small current fluctuations. In order to verify the detection effect of different abnormal current levels, the abnormal current is based on the daily normal current size and is divided into 1.3 times, 1.4 times, and 1.5 times for experiments. These three related abnormalities were designed at different random 200 time points for experiments. In the experiment, the number of cycles n was taken as 3, and the experimental results were recorded.

[0114] In this embodiment, 8 independent variables are used as the overall independent variable input, which can better grasp the change trend of the overall data without data redundancy. One day's data is used as training data, and test data is selected from another day's running data. In terms of adjusting model parameters, k-fold cross validation is used to select appropriate parameter values. Taking the number parameter adjustment of the gradient boosting tree model tree of timing abnormal current as an example, Figure 3 The parameter changes from 60 to 160, and the evaluation index value graph obtained by using k-fold cross validation is shown in Figure 3 When the number of trees is less than 120, the downward trend is obvious. When it is greater than 130, the change tends to be stable. In this paper, 130 is selected as the number of trees. For the instantaneous abnormal current threshold, this experiment selects the number of regression trees as 180 and the depth of the tree as 40; for the timing abnormal current threshold, this experiment selects the number of regression trees as 130 and the depth of the tree as 16. This experiment uses the traditional fixed current threshold and Gaussian distribution model to bring into the corresponding process for comparison. The experimental results are shown in Table 1.

[0115] Table 1 Experimental results data

[0116]

[0117] According to the data shown in Table 1, the anomaly recognition rate is obtained by dividing the number of points that identify anomalies by the total number of set anomaly points, which indicates the detection effect for abnormal situations. It can be seen from the data in the table that after using dynamic thresholds, these three methods can detect small disturbance anomalies very sensitively, and relatively speaking, the larger the abnormal current generated by the disturbance, the greater the probability of its detection. The method proposed in this paper has better detection effect than the Gaussian distribution model and the linear quantile model. At the level with relatively large disturbance, the detection accuracy of the method proposed in this paper can reach 97%, which is greater than the accuracy of the other two models, and in terms of relatively small disturbances, the effect of the method proposed in this paper is more prominent, with an accuracy of more than 90%, which can detect abnormal situations more accurately.

[0118] The present invention dynamically changes the set current abnormality threshold through quantile theory, and timely and accurately identifies and recognizes various abnormal working conditions, especially small disturbances and small current fluctuation abnormalities, to avoid further losses.

[0119] Although the present invention has been described above in conjunction with the accompanying drawings, the present invention is not limited to the above-mentioned specific embodiments, which are merely illustrative rather than restrictive. Under the guidance of the present invention, ordinary technicians in this field can make many modifications without departing from the purpose of the present invention, which all fall within the scope of protection of the present invention.

Claims

1. A distribution network abnormal condition detection method based on quantile regression gradient boosting tree, characterized by: The following steps are involved: A. Based on the quantile theory, the current quantile value is obtained; B. The quantile regression gradient boosting tree model is introduced to characterize the current distribution situation and make a more accurate estimate of the current value; C. Read the current value at time t, and calculate the corresponding instantaneous abnormal current threshold IAT and timing abnormal current threshold PAT; D. Perform logical judgment based on the current value, instantaneous abnormal current threshold IAT, and timing abnormal current threshold PAT to determine whether an abnormality occurs.

2. The method for detecting abnormal operating conditions of a distribution network based on quantile regression gradient boosting tree according to claim 1 is characterized in that: Step A is based on the quantile theory to obtain the current quantile value. The specific process is as follows: First of all, quantiles are points that divide the probability distribution range of a random variable into several equal parts. Common quantiles include quantiles, quartiles, and percentiles. Then, according to the corresponding theoretical concepts, the corresponding quantile expression is obtained: Among them, I is a random variable, indicating the corresponding current value; X is an independent variable used to predict I; F is the cumulative distribution function of the current I; Q represents the conditional quantile of I given X; t represents the corresponding measurement time point; inf is a function, indicating the minimum value of the subsequent function. In the above formula, the set it represents corresponds to the minimum value, which is the maximum boundary of all elements that meet the conditions; τ is the quantile probability level, indicating the corresponding probability value P r ,Right now That is, given X t In the case of I t Less than or equal to its corresponding quantile Q t probability.

3. The method for detecting abnormal operating conditions of a distribution network based on quantile regression gradient boosting tree according to claim 1 is characterized in that: The quantile regression gradient boosting tree model in step B is based on the quantile regression tree, and its quantile regression tree is as follows: First, the quantile constructs a binary tree through binary division, and the two branches are divided into two again, and the iteration is repeated until the stopping condition is met. Finally, the whole tree is divided into multiple different leaf nodes. By judging which leaf node the input test variable belongs to, the output value on the corresponding leaf node is obtained as the quantile prediction value; Then, the kth explanatory variable and its corresponding value v are selected as the splitting variable and splitting point, and the corresponding training set is divided into two sub-areas S1 and S2; S1(k,v)={x|x (k) ≤v},S2(k,v)={x|x (k) >v} (2) Among them, S represents the sub-region, k is the sequence number corresponding to the selected explanatory variable, v is the value of the selected segmentation variable, and x is the independent variable; Then, the corresponding output value of the sub-region is: in, is the output value of the corresponding area, m is the number of the area, and n is the number of elements in the corresponding area; Then, in order to obtain the final partition node with better prediction effect, the partition node must satisfy the following formula: The above division process is repeated iteratively until the termination condition is met. At this time, the training set is divided into L sub-regions or leaf nodes S1, S2, …, S L , the response variable on its leaf node is: y l ={y|x i ∈S l },l=1,2,...,L (5) Finally, the quantile regression tree, at the corresponding τ quantile point, the corresponding output value on its leaf node is: Where inf is the lower bound; F(·) is the empirical conditional distribution function, and its formula is as follows:

4. The method for detecting abnormal operating conditions of a distribution network based on quantile regression gradient boosting tree according to claim 3 is characterized in that: The quantile regression gradient boosting tree in step B performs ensemble learning on the basis of a single quantile regression tree to improve the prediction accuracy of the model. The model uses the steepest descent method and takes the negative gradient of the loss function as the approximate value of the residual to fit and generate a quantile regression tree. Through continuous iteration, multiple quantile regression trees are finally generated, and these trees are linearly added to obtain the final model.

5. The method for detecting abnormal operating conditions of a distribution network based on quantile regression gradient boosting tree according to claim 1 is characterized in that: In step C, the instantaneous abnormal current threshold IAT, if the collected current value exceeds the set value, an alarm will be issued immediately, and an abnormality will occur. The calculation formula is as follows: in, is the predicted value corresponding to the above quantile regression gradient boosting tree model, and γ is a parameter set by human experience, usually 1.05-1.

3.

6. The method for detecting abnormal operating conditions of a distribution network based on quantile regression gradient boosting tree according to claim 1 is characterized in that: In step C, if the collected current value exceeds the set value for a certain period of time, an alarm is issued and an abnormality occurs. The calculation formula is as follows: in, is the predicted value corresponding to the above quantile regression gradient boosting tree model.

7. The method for detecting abnormal operating conditions of a distribution network based on quantile regression gradient boosting tree according to claim 1 is characterized in that: Step D performs logical judgment based on the current value, the instantaneous abnormal current threshold IAT, and the timing abnormal current threshold PAT to determine whether an abnormality occurs. The specific process is as follows: d1. Compare the current value with the instantaneous abnormal current threshold IAT If the current value is greater than the instantaneous abnormal current threshold IAT t , then an exception occurs; If the current value is not greater than the instantaneous abnormal current threshold IAT t , compare the current value with the timing abnormal current threshold PAT t Make comparisons; d2. If the current value is less than the timing abnormal current threshold PAT t , then there is no abnormality and it is in normal operation; If the current value is greater than the timing abnormal current threshold PAT t , there may be an abnormality, read the current value at the next time point and calculate the corresponding current threshold; d3. At the next time point, if the current value is greater than the instantaneous abnormal current threshold at the next time point, an abnormality occurs; if the current value is less than the instantaneous abnormal current threshold at the next time point, the current value is compared with the timing abnormal current threshold at the next time point. If the current value is less than the timing abnormal current threshold, the cycle ends and there is no abnormality; if the current value is greater than the timing abnormal current threshold, the current value at the next time point is read and the corresponding threshold is calculated; d4. Repeat step d3 until n-1 cycles are completed and the current value I is obtained. t+n-1 , n represents the timing time period of the timing abnormal current threshold.

8. The method for detecting abnormal operating conditions of a distribution network based on quantile regression gradient boosting tree according to claim 7 is characterized in that: If the current value I t+n-1 If the current is greater than the timed abnormal current threshold or the instantaneous abnormal current threshold of the corresponding time, an alarm will be triggered immediately, indicating that an abnormality has occurred.

Images