A method for identifying sensitive load types based on steady-state power quality monitoring data

Abstract

The present invention relates to a sensitive load type identification method based on steady-state power quality monitoring data. Considering the diversity and unknown nature of sensitive loads contained in users, a non-invasive monitoring method is adopted to combine transient power quality indicators with multi-indicator steady-state power quality monitoring data to accurately identify voltage sag sensitive loads contained in important users, which is beneficial for power suppliers to provide differentiated services and for power users to implement customized governance.

Description

Technical Field

[0001] The invention belongs to the field of load type identification and relates to a sensitive load type identification method based on steady-state power quality monitoring data. Background Art

[0002] Voltage sag is an unavoidable power quality problem in power grid operation, which can interrupt the normal production of industrial users, causing huge economic losses and safety hazards. Complaints caused by voltage sag account for more than 80% of complaints about power quality problems. The high harm and high proportion of voltage sag events make it an important power problem for power users and power supply companies.

[0003] In recent years, the number of voltage sag-sensitive loads used in modern industry has increased year by year. The high sensitivity of such loads to voltage disturbances and the contradiction between the inevitable voltage sag events have become increasingly prominent. In order to reduce user losses and mitigate the hazards of voltage sags, power supply companies have taken many measures, but due to insufficient understanding of user-side loads, it is difficult to achieve accurate optimization of power supply services. Therefore, in the process of taking measures, it is still necessary to solve the problem of accurate identification of sensitive loads contained in important users, so as to help power suppliers achieve differentiated services that reduce costs, increase efficiency, and improve accuracy, and for electricity users to implement targeted, effective, and timely customized governance.

[0004] However, due to the large differences in the types of sensitive loads contained by users in different industries, it is difficult for most users to make professional judgments on the nature of the loads. Currently, most sensitive load identification is based on load tolerance characteristics evaluated by repeated experiments or capacity calculations under known load type conditions, without fully considering the difficulty of experimenting with already commissioned equipment and the unknown nature of loads contained in actual industrial enterprises. In the case of unknown load types, non-invasive methods should be considered to achieve type identification of sensitive loads.

[0005] As power users pay more attention to power quality issues, most sensitive users have completed the installation of power quality monitoring terminals in accordance with the Technical Specifications for Power Quality Monitoring Systems. The monitoring data of the devices contains multi-feature, large-volume long-term power quality monitoring data, as well as information on voltage sag events, which provides a solid and sufficient data foundation for further adopting non-invasive methods to identify sensitive loads based on steady-state power quality monitoring data. Therefore, this paper proposes a method for identifying voltage sag sensitive loads based on steady-state power quality monitoring data.

[0006] At present, the nature of user sensitive loads is mostly determined by conducting a large number of repeated experiments on voltage sag tolerance characteristics for a single load, and fitting the voltage tolerance characteristics of a single load based on the load action conditions obtained from events with different sag severity, thereby achieving load sensitivity assessment. This method is only applicable to independent sensitive loads that have not yet been put into operation on the user side, and it is difficult to identify the overall sensitive load on the user side. Or, under the premise of knowing the type of user sensitive load, the sensitive load identification problem is converted into a load capacity ratio problem, but in practice, it is difficult to accurately obtain load process parameters, and it is difficult to define the nature of the user's load in advance. Summary of the invention

[0007] The purpose of the present invention is to provide a sensitive load type identification method based on steady-state power quality monitoring data, which is conducive to the power supplier to achieve differentiated services of reducing costs, increasing efficiency and improving accuracy, and the power user to implement targeted, effective and timely customized management.

[0008] To achieve the above object, the technical solution of the present invention is: a method for identifying sensitive load types based on steady-state power quality monitoring data, comprising:

[0009] (1) A time segmentation method based on active power monitoring data is proposed, which can identify the start and end times of the periods before and after the sag event;

[0010] (2) A dynamic clustering method based on multi-indicator steady-state power quality monitoring data is proposed, which can accurately identify abnormal load operation caused by temporary sag events of different severity.

[0011] (3) A method for mapping clusters to typical sensitive load action areas is proposed. Combined with boundary fitting, it can realize the type identification of voltage sag sensitive loads contained in users.

[0012] (4) A method for identifying voltage sag sensitive loads based on steady-state power quality monitoring data is proposed.

[0013] Compared with the prior art, the present invention has the following beneficial effects:

[0014] The present invention obtains the start and end events of the preceding and following sections of each event through the user transient monitoring value and the sag event index; by analyzing the changes in multi-index steady-state power quality monitoring data caused by sag events of different severity levels, it is proposed to use a dynamic clustering method to classify clusters, and to map clusters to action areas to achieve identification of sensitive loads.

[0015] advantage:

[0016] Different from the method of performing a large number of independent repeated experiments on a single load and fitting the load voltage tolerance characteristic curve, the method of the present invention is conducive to the power supplier to achieve differentiated services such as reducing costs, increasing efficiency and improving accuracy, and the power user to implement targeted, effective and timely customized management. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a schematic diagram of the process of the present invention;

[0018] Figure 2 This is a schematic diagram of voltage sag sensitive load identification in the present invention. DETAILED DESCRIPTION

[0019] The technical solution of the present invention is described in detail below in conjunction with the accompanying drawings.

[0020] like Figure 1 , 2 As shown, the present invention provides a sensitive load type identification method based on steady-state power quality monitoring data, comprising:

[0021] Step S1, acquisition of original power quality monitoring data. To realize the segmentation of sag events, user active power real-time monitoring transient data is selected as the data source. To construct the steady-state data set to be identified, multi-indicator steady-state power quality monitoring data is selected as the data source.

[0022] Step S2: Segmentation of the sag event based on transient data: After preprocessing the active power monitoring value by using the maximum value normalization and discrete wavelet transformation, the sliding mean event segmentation method is used to obtain the front and rear time of the sag event.

[0023] Step S3, construction and dynamic clustering of the steady-state data set to be identified. Select basic electrical parameters such as power, current, and voltage, as well as multi-feature steady-state power quality data that reflect the power quality emission characteristics such as current imbalance and voltage distortion rate, and calculate the average difference of multi-index power quality data before and after the sag event as the sample set to be processed. Use dynamic kmeans to cluster samples, and use the silhouette coefficient as the evaluation index to obtain the optimal clustering cluster.

[0024] Step S4, sensitive load type identification based on steady-state data: Fit the boundaries and inflection points of each cluster of the above optimal clustering class, and compare them with the preset VTC curve. After correctness verification with confusion matrix indicators, the sensitive loads contained in the user are identified.

[0025] Furthermore, the sag event segmentation based on transient data described in step S2 is specifically the following steps:

[0026] Step S201: First, the active power value is normalized by using the maximum value normalization method, and the active power monitoring data of a certain user is recorded as P = {p 1 ,p 2 ,......,p n}, the maximum value normalization method is used for processing, the formula is as follows:

[0027]

[0028] where p i represents the i-th active power monitoring point of the user, and the per-unit power value is recorded as P'={p' 1 ,p' 2 ,...,p' i ,...,p' n}.

[0029] Step S202: For the per-unit value of active power, discrete wavelet transform (DWT) is used to separate the data into high-frequency part and low-frequency part. The high-frequency and low-frequency parts of the x-th layer of active power P' are expressed as follows:

[0030]

[0031]

[0032] Among them, P′ x,L [n], P′ x,H [n] represents the low-frequency and high-frequency components of the nth layer of the active power per unit value; K represents the weight coefficient length; k represents the length of the wavelet right shift; l[k], h[k] represent low-pass and high-pass filters.

[0033] The noise components are mainly concentrated in the high-frequency detail components, so the default threshold is used to denoise the active power wavelet. The per-unit value of the active power after DWT denoising is P″={p″ 1 ,p″ 2 ,...,p″ i ,...,p″ n}.

[0034] Step S203: Use the sliding mean method to segment the active power trajectory and identify the starting point of each event segment. Select the length of the sliding window as w 1 , the active power data set in the jth sliding window is Calculate the range change rate ΔP within each sliding window j " and standard deviation σP j ”, the formula is as follows:

[0035]

[0036]

[0037]

[0038] The range change rate data set calculated by the sliding window is recorded as The standard deviation data set is denoted as Then, the quartile probability model is used to obtain the mean of the third quartile of the range change rate and standard deviation data set, which is recorded as Gradually increase the sliding window length w and calculate the above mean until

[0039] Step S204: w j is the sliding window length, and the power data is segmented and recorded as Calculate the range change rate and standard deviation for each small segment of data. Calculate according to the following formula:

[0040]

[0041]

[0042] where x∈[1+aw j ,1+(1+a)w j ];

[0043] The first and last jump point data in the calculation result are recorded as the start and end time of the power change segment, that is, the load state area stabilization time after the start and end of the sag event. So far, the end time t of the front section of the sag event is completed based on the active power sampling value. 1 , the start time of the latter segment t 4 Get.

[0044] Furthermore, the construction and dynamic clustering of the steady-state data set to be identified in step S3 are specifically the following steps:

[0045] Step S301: Based on the division of the front and back segments of the temporary drop event, obtain the front segment of the event [t 1- ,t 1 ], the latter part [t 4 ,t 4+ ] of multidimensional steady-state data, such as for feature X m In the ith sag event, the data are the data sets before and after the event, respectively {x i1- ,x i2- ,...,x iN-}, {x i1+ ,x i2+,...,x iN+}, calculate the average difference Δx before and after the event im , the average difference of all features is recorded as x i , the user b temporary drop events are combined into a data source to be processed X = {x 1 ,x 2 ,...,x i ...,x b}.

[0046]

[0047] Step S302: For the above unlabeled data sources, this paper adopts a dynamic Kmeans method to change the number of initial clustering clusters, and uses the silhouette coefficient to evaluate the clustering effect.

[0048] Randomly select c cluster centers from the unlabeled samples to be clustered X, denoted as {c 1 ,c 2 ,......,c c}.

[0049] Calculate the Euclidean distance from each element in the sample to the initial cluster center, and update the c clusters according to the distance formula, which is as follows:

[0050]

[0051] The cluster centers of the k clusters that have been clustered are recalculated, and the calculation formula is as follows:

[0052]

[0053] Calculate the silhouette coefficient under the initial number of clusters. The calculation formula is as follows:

[0054]

[0055]

[0056]

[0057] in, are the intra-cluster cohesion and inter-cluster separation, respectively. is the silhouette coefficient.

[0058] Update the initial cluster center to make c = c + 1, repeat steps 2-4, when and When , the iteration stops and c at this time is selected as the final number of clusters.

[0059] Step S303: The user b temporary downtime events are combined into a data source to be processed X = {x 1 ,x2 ,...,x i ...,x b}Perform dynamic clustering to obtain various clustering clusters.

[0060] The inflection point fitting process is performed on each cluster data obtained by clustering to obtain the user's VTC fitting curve. The fitting curve is compared with the preset sensitive load VTC curve to realize the sensitive load identification of the unknown load user.

[0061] Step S304: Use confusion matrix to quantify the reliability of the results, which is used to represent the various sensitive loads and combinations identified by the model relative to the true values. The accuracy of the confusion matrix is ​​S PPV , recall rate S TRP , specificity S TNR , S F1 The value is calculated as follows:

[0062]

[0063]

[0064]

[0065]

[0066] Accuracy PPV Represents the accuracy evaluation of the model on load classification, and the recall rate S TRP and specificity S TNR Recognition accuracy and completeness of the reaction model, S F1 The value reflects the comprehensive characteristics of the clustering method. Among them, the accuracy S PPV and S F1 The higher the value, the more accurate the clustering is.

[0067] The above are preferred embodiments of the present invention. Any changes made according to the technical solution of the present invention, as long as the resulting functions do not exceed the scope of the technical solution of the present invention, belong to the protection scope of the present invention.

Claims

1. A sensitive load type identification method based on steady-state power quality monitoring data, It is characterized in that The steps include: S1. Acquisition of original power quality monitoring data: In order to realize the segmentation of temporary sag events, the user's active power real-time monitoring transient data is selected as the data source; In order to construct the steady-state data set to be identified, multi-index steady-state power quality monitoring data are selected as the data source; S2. Segmentation of sag events based on transient data: After preprocessing the real-time monitoring transient data of active power of users by using the maximum value normalization and discrete wavelet transformation, the sliding mean event segmentation method is used to obtain the front and rear time of the sag event; S3. Construction and dynamic clustering of the steady-state data set to be identified: Select basic electrical parameters including power, current, and voltage, as well as multi-index steady-state power quality data that reflect power quality emission characteristics including current imbalance and voltage distortion rate, and calculate the average difference of multi-index steady-state power quality monitoring data before and after the sag event as the sample set to be processed; use the dynamic kmeans method to cluster samples, and use the silhouette coefficient as the evaluation index to obtain the optimal clustering cluster; S4. Sensitive load type identification based on steady-state data: Fit the boundaries and inflection points of each cluster of the optimal clustering cluster, and compare them with the preset VTC curve. After correctness verification with confusion matrix indicators, the sensitive loads contained in the user are identified; S2 is as follows: S201, using the maximum value normalization method to perform per-unit processing on the active power value, record the real-time monitoring transient data of active power of a certain user as P = {p 1 ,p 2 ,......,p n }, the maximum value normalization method is used for processing, the formula is as follows: p i is the monitoring value of the i-th active power monitoring point of the user, and the per-unit active power value is recorded as P'={p' 1 ,p' 2 ,...,p' i ,...,p' n }; S202, for P', using discrete wavelet transform DWT to separate the data into a high-frequency part and a low-frequency part; S3 is as follows S301, based on the division of the front and back segments of the temporary drop event, obtain the front segment of the event [t 1- ,t 1 ], the latter part [t 4 ,t 4+ ], for the multidimensional steady-state data of feature X m In the ith sag event, the pre-event data set is {x i1- ,x i2- ,...,x iN- }, the post-event data set is {x i1+ ,x i2+ ,...,x iN+ }, calculate the average difference Δx before and after the event im , the average difference of all features is recorded as x i , the user b temporary drop events are combined into a data source to be processed X = {x 1 ,x 2 ,...,x i ...,x b }; S302: For unlabeled data sources, a dynamic Kmeans method is used to change the number of initial clusters, and a silhouette coefficient is used to evaluate the clustering effect; Randomly select c cluster centers from the unlabeled data source X to be processed, denoted as {c 1 ,c 2 ,......,c c }; Calculate the Euclidean distance from each element in the sample to the initial cluster center, and update the c clusters according to the distance formula, which is as follows: The cluster centers of the k clusters that have been clustered are recalculated using the following formula: Calculate the silhouette coefficient under the initial number of clusters. The formula is as follows: are the intra-cluster cohesion and inter-cluster separation, respectively. is the silhouette coefficient; Update the initial cluster center to make c = c + 1, repeat steps S2-S4, when and When , the iteration stops and c at this time is selected as the final number of clusters.

2. According to claim 1, a sensitive load type identification method based on steady-state power quality monitoring data, It is characterized in that The x-th layer low-frequency part and high-frequency part of the active power per unit value P' in S202 are expressed as follows: P′ x,L [n] represents the nth low frequency per unit value of active power, P′ x,H [n] represents the nth layer high-frequency component of the active power per unit value; K represents the weight coefficient length; k represents the length of the wavelet right shift; l[k] represents a low-pass filter, and h[k] represents a high-pass filter; The noise components are concentrated in the high-frequency detail components, so the default threshold denoising is adopted to achieve wavelet denoising of active power; the per-unit value of active power after DWT denoising is P″={p″ 1 ,p″ 2 ,...,p″ i ,...,p″ n }; After executing S202, the following steps are further performed: S203, using the sliding mean method, the active power trajectory is segmented and processed to identify the starting point of each event segment; the length of the sliding window is selected as w 1 , the active power data set in the jth sliding window is Calculate the range change rate ΔP″ within each sliding window j and standard deviation σP″ j , the formula is as follows: The range change rate data set calculated by the sliding window is recorded as The standard deviation data set is denoted as Then, the quartile probability model is used to obtain the mean of the third quartile of the range change rate and standard deviation data set, which is recorded as Gradually increase the sliding window length w and calculate the mean until S204, w j is the sliding window length, and the power data is segmented and recorded as Calculate the range change rate and standard deviation for each small segment of data. Calculate according to the following formula: Among them, x∈[1+aw j ,1+(1+a)w j ]; The first and last jump point data in the calculation results are recorded as the start and end time of the power change segment, that is, the load state area stabilization time after the sag event starts and ends; At this point, the end time t of the temporary sag event is realized based on the active power sampling value. 1 , the start time of the latter segment t 4 Get.

3. According to claim 1, a sensitive load type identification method based on steady-state power quality monitoring data, It is characterized in that After executing S302, the following steps are further performed: S303, the user b temporary drop events are combined into a data source to be processed X = {x 1 ,x 2 ,...,x i ...,x b } Perform dynamic clustering to obtain clusters; perform inflection point fitting processing on each cluster data obtained by clustering to obtain the user's VTC fitting curve; compare the fitting curve with the preset sensitive load VTC curve to realize sensitive load identification of unknown load users; S304. Use confusion matrix to quantify the reliability of the results, which is used to represent the various sensitive loads and combinations identified by the model relative to the true values. The accuracy of the confusion matrix S PPV , recall rate S TRP , specificity S TNR , S F1 The value is calculated as follows: Accuracy PPV Represents the accuracy evaluation of the model on load classification, and the recall rate S TRP and specificity S TNR Recognition accuracy and completeness of the reaction model, S F1 The comprehensive characteristics of the value-reflecting clustering method; the accuracy S PPV and S F1 The higher the value, the more accurate the clustering is.

Images