A power distribution network control method considering communication quality

Abstract

The application discloses a power distribution network control method considering communication quality, and relates to the technical field of power electronics. The method comprises the following steps: judging whether a decision instruction of a control center is successfully received by a transformer area; if the decision instruction is successfully received, executing the decision instruction; otherwise, combining transformer area data of all transformer areas at a current time to construct a pre-decision data matrix; searching for a historical transformer area data matrix with the largest similarity to the pre-decision data matrix, and searching for a corresponding decision information matrix; and executing a historical decision instruction in the decision information matrix. The method can generate a local decision instruction when the decision instruction of the control center is failed to be decentralized due to poor communication quality, and the transformer area cannot timely receive the decision instruction, thereby effectively avoiding disordered operation of the transformer area caused by communication delay and instruction transmission failure, and helping to guarantee the real-time performance and accuracy of power distribution network control.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of power electronics, in particular to a power distribution network control method considering communication quality. BACKGROUND

[0002] In recent years, the scale of grid-connected distributed energy (such as photovoltaic and energy storage) is expanding, which aggravates the problem of harmonic pollution and brings challenges to the power supply quality of new power distribution networks. The accumulation of harmonics can cause resonance in the system, which threatens the safe operation of load equipment. The traditional harmonic suppression techniques mainly include two methods: virtual impedance and active filter. The virtual impedance method absorbs the harmonic current in the power grid by adding a virtual impedance to the inverter, thereby suppressing harmonics, but there are still deficiencies in quantifying the contribution of each inverter to the harmonic link of the power distribution network, making it difficult to accurately control the suppression effect of different inverters on harmonics. Active filters can controllably compensate according to the harmonic current on the load side, but the cost is high and the economy is poor, which is not suitable for widespread application. In the new power distribution network, photovoltaic and energy storage are connected through inverters, and inverters, as active elements, can controllably generate compensation currents equal in size and opposite in direction to harmonic currents by using the remaining capacity after active grid connection through advanced control strategies, thereby canceling the harmonic currents emitted by harmonic sources and achieving the purpose of suppressing harmonics. The power distribution network control center issues decision commands based on real-time operation data collected by the intelligent fusion terminal of the transformer area, coordinates the collaborative control of multiple inverters, determines the distorted power and harmonic current that each inverter should generate, and then solves the problem of harmonic overlimiting.

[0003] As can be seen, resource sharing between different subjects in the new power distribution network not only depends on its own control strategy, but also needs reliable communication technology. The existing harmonic suppression and daily control of multi-element resource collaborative operation of the power distribution network are based on the condition that the communication network is reliable and the equipment is stable, ignoring the actual problem of complex and unreliable communication network equipment in the actual power distribution network. In the actual production scene, poor communication quality will also cause the failure of the decision-making instructions of the control center, and the transformer area cannot receive the instructions in time, resulting in disordered operation of the transformer area and system operation risks. SUMMARY

[0004] To solve the above technical problems existing in the prior art, the present application aims to provide a method that can still implement short-term real-time control of the power distribution network when communication quality problems occur.

[0005] Specifically, the present application provides a power distribution network control method considering communication quality, and the specific technical solution includes the following steps:

[0006] Step S1: determining whether the decision-making instructions of the control center are successfully received by the transformer area; if successfully received, executing the decision-making instructions, otherwise continuing;

[0007] Step S2: combine the substation data of all substations at the current time to construct a pre-decision data matrix X A ;

[0008] Step S3: retrieve the historical substation data matrix with the largest similarity from the pre-decision data matrix X A and find the corresponding decision information matrix, and execute the historical decision instructions in the decision information matrix.

[0009] Preferably, the successful reception is specifically:

[0010] Calculate the time delay of the time stamp carried by the decision instruction and the substation receiving time; if the time delay is less than or equal to the time delay threshold, it is determined to be successfully received, otherwise it is determined to be unsuccessfully received.

[0011] Preferably, the method further comprises a preset time delay threshold, specifically:

[0012] Obtain a historical data set of the response time of the power distribution system, and calculate the average value of each historical data;

[0013] Calculate the difference between each historical data and the average value, and select the top 95% of the historical data with smaller difference, and calculate the average value again as the time delay threshold.

[0014] Preferably, step S3 specifically comprises:

[0015] Step S31: combine the historical substation data matrix at each time and the corresponding decision information matrix to construct a decision augmented matrix X ;

[0016] Step S32: combine the pre-decision data matrix X A Input the first convolutional neural network of the SiamFC network to obtain the first high-dimensional feature f A , the formula is as follows:

[0017] ;

[0018] In the formula, f Ai is the pre-decision data matrix X A The high-dimensional feature of the first i class substation data, i =1,2,..., N , i is the class number of the substation data, N is the number of classes of substation data;

[0019] The decision augmented matrixX The second convolutional neural network input to the SiamFC network yields the second high-dimensional feature. f X The formula is as follows:

[0020] ;

[0021] In the formula, t For time sequence number, T The total number of moments. t =1,2,..., T , f Xt For the first t The third high-dimensional feature of the historical data matrix of each station area. f Ct For the first t The fourth high-dimensional feature of a decision information matrix;

[0022] Step S33: Through M × N window w Traversing the decision augmentation matrix X Calculate the pre-decision data matrix X A With each window w Similarity to the historical data matrix of TV stations S ( w The formula is as follows:

[0023] ;

[0024] Step S34: Select similarity S ( w The largest window w According to the decision augmentation matrix X Locate the corresponding decision information matrix and execute the historical decision instructions contained therein;

[0025] The first and second convolutional neural networks share network parameters.

[0026] Preferably, step S34 specifically includes:

[0027] Similarity S ( w The largest window w Translate to the right N Each unit obtains the corresponding decision information matrix and executes the historical decision instructions contained therein.

[0028] Furthermore, it also includes training the SiamFC network:

[0029] Calculate the first high-dimensional featuref A Second high-dimensional features f X Contrast loss Loss The formula is as follows:

[0030] ;

[0031] ;

[0032] In the formula, For the first i If the labels of the data are similar to those of the station area Then the corresponding historical distribution area data matrix and decision augmentation matrix X The i Similar data for similar districts, if If they are not similar; β This is the boundary threshold; S ( f A， f X ) is the first high-dimensional feature f A Second high-dimensional features f X similarity, f Xti For the first t In the historical data matrix of the first district i High-dimensional characteristics of data similar to those of transformer substations S ( f Ai , f Xti () is the pre-decision data matrix X A and the t In the historical data matrix of the first district i Similarity of data between similar TV stations;

[0033] Among them, if similarity S ( f Ai , f Xti If the value is greater than the preset value, then the pre-decision data matrix is ​​determined. X A and the t In the historical data matrix of the first district i Similar data for similar distribution areas;

[0034] By traversing the decision augmentation matrix X Adjusting the boundary threshold β This results in contrast loss. Loss Minimum;

[0035] find the contrast loss Loss The minimum time corresponds to the historical substation data matrix, and the historical decision instruction in the corresponding decision information matrix is executed.

[0036] Preferably, the substation data includes total harmonic distortion of voltage, fundamental voltage effective value, harmonic voltage effective value, fundamental current effective value, harmonic current effective value and total harmonic distortion of current.

[0037] As can be seen, the technical solution provided by the present application can generate local decision instructions when the decision instructions of the control center fail to be sent due to poor communication quality and the substation cannot timely receive the decision instructions, which can effectively avoid the disordered operation of the substation caused by communication delay and instruction transmission failure, and help to ensure the real-time and accuracy of the power distribution network control. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 It is an application scenario diagram of the power distribution network control method in an embodiment of the present application.

[0039] Figure 2 It is a flow diagram of the power distribution network control method in an embodiment of the present application. DETAILED DESCRIPTION

[0040] In the following, the technical solution provided by the present application will be further described in detail in combination with embodiments and drawings.

[0041] Embodiment 1:

[0042] In the new power distribution network, using photovoltaic inverters and energy storage inverters to work together for harmonic compensation is a new emerging harmonic suppression technology, which has certain flexibility and feasibility. The power distribution network control center collects data of the substation intelligent fusion terminal in real time, cooperates with multiple inverters to generate compensation current equal in size and opposite in direction to the harmonic current, and suppresses the propagation of harmonics. Among them, the bidirectional data interaction between the power distribution network control center and each substation intelligent fusion terminal depends on a stable and efficient communication network. In order to solve the problem of high coupling between the power distribution network and the communication network, researchers propose a method based on the collaborative planning model of communication base stations and distributed energy, which optimizes the collaborative work of the communication network and the power distribution network by considering the capacity and geographical location of the base station. Although this scheme has certain potential, in actual application, due to the fixed layout of base stations and power distribution networks, and multiple restrictions such as geographical location and cost, it is very difficult to reconfigure and adjust these layouts.

[0043] In actual application, at the substation level, there are many power distribution network devices and insufficient communication network reliability, limited network bandwidth, and due to the wide distribution of substations, signal interference and data conflict often occur. For example, Figure 1As shown, in the power distribution network control center layer, due to the instability of the communication network, the control command may be delayed or interrupted during transmission, resulting in the command not reaching the target substation in time. In this case, the failure of command transmission may lead to the loss of control of equipment operation, thereby affecting the power supply quality and increasing the operation risk of the system. Therefore, in the scenario of poor communication quality, the intelligent fusion of the substation needs to have certain local control capability, which can make decisions quickly according to the existing data and preset strategy model, avoid excessive dependence on real-time instructions, and prevent the deterioration or loss of control of harmonic problems.

[0044] To overcome the problem of failure of control center command issuance, the embodiment proposes a substation intelligent decision-making model, which uses a twin full convolutional network (SiamFC) to match the most similar data in the historical database and output the most possible compensation decision information, making harmonic suppression decisions quickly and accurately, thereby realizing the rapid execution of harmonic suppression in the scenario of poor communication quality.

[0045] As shown in Figure 2 , when the control center command is failed, the decision-making instruction is generated based on the SiamFC network. Specifically, the following steps are included:

[0046] Step 1: State evaluation of the substation receiving the decision-making instruction.

[0047] When the power distribution network control center issues a decision-making instruction, the instruction data packet will carry accurate timestamp information. After receiving the instruction, the intelligent fusion terminal of each substation will compare the current timestamp with the reception time to estimate the time delay. Specifically, the terminal will calculate the time difference between the issuance of the instruction by the control center and the actual reception of the instruction by the terminal, and compare it with the time delay threshold. The time delay threshold is set according to the network transmission characteristics, historical data and system tolerance, etc., and is usually set based on the 95% distribution of the system response time. If the calculated time delay exceeds the time delay threshold, it means that the substation has failed to successfully receive the instruction within the specified time, and there may be instruction loss or network transmission failure. At this time, the intelligent fusion terminal of the substation will immediately trigger the short-term intelligent decision-making mechanism (i.e. the subsequent steps) to generate a local decision-making instruction, achieving effective control.

[0048] The design method of the time delay threshold includes:

[0049] Step 1: Obtain the historical data set of the system response time of the power distribution network, and calculate the average value of each historical data;

[0050] Step 2: Calculate the difference between each historical data and the average value, and select the top 95% of the historical data with smaller difference, and calculate the average value again;

[0051] Step 3: Take the average value calculated again as the time delay threshold.

[0052] Optionally, step 2 can be repeated until the most recently calculated average is within (1 ± 1) of the previous average. a If the latency is between 0.5% and 0.6%, the latest calculated average value will be used as the latency threshold. Technical personnel can dynamically adjust this threshold according to actual needs. a The size, generally speaking, a Take 0.5.

[0053] Step 2: Input data preprocessing.

[0054] If receiving decision instructions from the control center fails, the data from all transformer areas at the current moment will be combined to construct a pre-decision data matrix. X A , X A ∈ R M×N .in, M Number of stations; N This refers to the number of categories of data for distribution substations in the distribution network, which is also a dimension of the feature data. On the other hand, it involves retrieving the total number of data categories for all distribution substations in the database. T Historical data matrix of transformer substations at each moment X B , X B ∈ R M×N For each historical transformer area data matrix X B Each of them has a specially generated decision information matrix. X C , X C ∈ R M ×L .in, L This refers to the number of decision information items in the decision information matrix, i.e., the decision information dimension. This is the data matrix for each historical transformer area. X B With the corresponding decision information matrix X C Combine to generate the corresponding decision augmentation matrix. X , X =( X B | X C )∈ R M×(N+L) Decision augmentation matrix X In the middle, historical station area data matrix X B On the left, the decision information matrix X C On the right.

[0055] In the context of coordinating control of grid-connected light storage inverters to govern harmonics through multi-dimensional data collected by intelligent fusion terminals in each distribution area in the distribution network control center, the required related operating data usually includes total harmonic distortion of voltage, effective value of fundamental voltage, effective value of each harmonic voltage, effective value of fundamental current, effective value of each harmonic current, and total harmonic distortion of current. The decision instruction is usually the distortion power and the amplitude and phase of each harmonic current that each inverter should issue. When used for distribution area reactive power regulation, the required related operating data usually includes pre-treatment voltage, load active power, load reactive power, and photovoltaic active power. The decision instruction is usually the reactive power of the inverter. Therefore, in terms of coordinated governance of power quality in the distribution network, the number of categories of distribution area data N is greater than the dimension of decision information in the decision information matrix L .

[0056] Step three: SiamFC network design.

[0057] The SiamFC (Siamese Fully Convolutional Network) network adopts a Siamese Network-based Fully Convolutional Network (FCN) architecture, which aims to judge and locate data through feature mapping and similarity learning. The SiamFC network structure includes two convolutional neural networks (CNNs) with shared weights. FCN is a special convolutional neural network that only contains convolutional layers and pooling layers, which maps the input data to a high-dimensional space and extracts its spatial features through this architecture. The core goal of this network is to learn a similarity measure to make decisions by calculating the similarity between the distribution area operating data matrices.

[0058] In this process, the pre-decision data matrix X A and the decision augmented matrix X are input into two branch networks (sharing weights with the same convolutional neural network structure). The spatial features of the input matrix are extracted through the convolutional layer to obtain its first high-dimensional feature f A and second high-dimensional feature f X , as follows:

[0059] ;

[0060] ;

[0061] wherein,i This refers to the category number of the data for the transformer substation. N The number of categories in the data for the transformer area. i =1,2,..., N , f Ai Pre-decision data matrix X A No. i High-dimensional characteristics of data similar to those of transformer substations; t For time sequence number, T The total number of moments. t =1,2,..., T , f Xt For the first t The third high-dimensional feature of the historical data matrix of each station area. f Ct For the first t The fourth high-dimensional feature of a decision information matrix.

[0062] The loss function chosen is contrastive loss, used to measure the similarity between two input data points. Specifically, it determines the correlation between the pre-decision monitoring area data and historical data by calculating the similarity between two feature representations. Cosine similarity is used for measurement, as shown in the following formula:

[0063] ;

[0064] ;

[0065] In the formula, Loss The first high-dimensional feature f A Second high-dimensional features f X The comparative loss, For the first i If the labels of the data are similar to those of the station area Then the corresponding historical distribution area data matrix and decision augmentation matrix X The i Similar data for similar districts, if If they are not similar; β This is the boundary threshold, used to adjust the distance for similarity. S ( f A， f X ) is the first high-dimensional feature f A Second high-dimensional features f X similarity, f Xti For the firstt In the historical data matrix of the first district i High-dimensional characteristics of data similar to those of transformer substations S ( f Ai , f Xti () is the pre-decision data matrix X A and the t In the historical data matrix of the first district i Similarity of data in similar districts.

[0066] when Ideally, S ( f A， f X )≥ β At this time, max(0, β - S ( f A， f X When ))=0, the loss is 0. Ideally, S ( f A， f X )≤ β At this time, max(0, S ( f A， f X )- β =0, loss is 0. By adjusting network parameters through backpropagation, the loss function is minimized, maximizing the similarity of similar sample pairs (approaching 1) and minimizing the similarity of dissimilar sample pairs (approaching 0). β This establishes a clear decision boundary for the threshold. β The mathematical essence of the difference between the similarity score and the expected boundary is to construct an interval through a boundary threshold β, forcing similar sample pairs to have a similarity score higher than the expected boundary threshold. β The similarity between dissimilar sample pairs is lower than that of other samples. β In the formula, the max operation (max{0, β - S ( f A， f X )} and max{0, S ( f A， f X ) -β}) that embodies this interval constraint.

[0067] Step four: output decision information.

[0068] During the training process, the historical data in the transformer database is retrieved using a sliding window. Assuming that the size of the historical data augmentation matrix is KM × ( N + L ), the sliding window size is set to be consistent with the dimension of the pre-decision transformer data X A , that is, the dimension of the window is M × N . By traversing the historical transformer data matrix X B part of the decision augmentation matrix using the sliding window, the similarity w of each window S position is calculated w , and the formula is as follows:

[0069] ;

[0070] In the formula, f X ( w ) represents the feature representation corresponding to the window w position. The historical transformer data matrix X B with the highest similarity score is selected as the most similar historical data to the pre-decision transformer, which is used for subsequent decision information generation.

[0071] During the entire sliding window traversal process, after multiple iterations of training and backpropagation, the network parameters are optimized, and the boundary threshold β is updated to minimize the loss function value, obtaining the positioning X B of the most similar historical transformer data matrix w best . After obtaining N , it is shifted to the right by one unit, and the corresponding decision information matrix is generated. The calculation process of the position box information is represented as:

[0072] ;

[0073] In the formula, represents the position of the decision information matrix.

[0074] Finally, the decision matrix information matrix is taken as the output P , and the formula is as follows:

[0075] .

[0076] It can be seen from the embodiment that the method provided by the application can be applied to any scene requiring control decision decentralization in a power distribution network environment, and is not limited to the control aspect of harmonic suppression.

[0077] It can be seen from the above embodiment and the accompanying drawings that the technical solution provided by the application can generate local decision instructions when the communication quality is poor, the decision instructions of the control center fail to be decentralized, and the transformer area cannot receive the decision instructions in time, which can effectively avoid the disordered operation of the transformer area caused by communication delay and instruction transmission failure, and is helpful to guarantee the real-time performance and accuracy of power distribution network control.

[0078] Further, the preset time delay threshold can be accurately set in combination with historical data to accurately determine whether the decision instructions are successfully received, which is helpful to avoid the disordered operation of the transformer area; the SiamFC network is trained, which is helpful to call the control method most suitable for the power distribution network at present, and is helpful to the accurate control of the power distribution network; and the accurate setting of the preset time delay threshold and the training of the SiamFC network are both helpful to the real-time performance and accuracy of the power distribution network control.

Claims

1. A power distribution network control method considering communication quality, characterized in that, Includes the following steps: Step S1: Determine whether the control area has successfully received the decision instruction from the control center; if successfully received, execute the decision instruction; otherwise, continue. Step S2: Combine the data from all transformer areas at the current time to construct a pre-decision data matrix X. A ; Step S3: Retrieve and pre-decision data matrix X A Find the historical data matrix of the most similar transformer stations, locate the corresponding decision information matrix, and execute the historical decision instructions in the decision information matrix; Step S3 specifically includes: Step S31: Combine the historical station area data matrix at each moment with the corresponding decision information matrix to construct the decision augmentation matrix X; Step S32: Calculate the pre-decision data matrix X A The first convolutional neural network input to the SiamFC network yields the first high-dimensional feature f. A The formula is as follows: ； In the formula, f Ai For the pre-decision data matrix X A The high-dimensional features of the i-th type of transformer area data, i=1,2,...,N, where i is the category number of the transformer area data and N is the number of categories of the transformer area data; The decision augmentation matrix X is input into the second convolutional neural network of the SiamFC network to obtain the second high-dimensional feature f. X The formula is as follows: ； In the formula, t is the time sequence number, T is the total number of times, t=1,2,...,T, f Xt f is the third high-dimensional feature of the data matrix of the t-th historical station area. Ct Let be the fourth high-dimensional feature of the t-th decision information matrix; Step S33: Traverse the decision augmentation matrix X through an M×N window w to calculate the pre-decision data matrix X. A The similarity S(w) between each window w and the historical station data matrix is ​​calculated using the following formula: ； Step S34: Select the window w with the largest similarity S(w), find the corresponding decision information matrix according to the decision augmentation matrix X, and execute the historical decision instructions in it; The first and second convolutional neural networks share network parameters.

2. The power distribution network control method considering communication quality as described in claim 1, characterized in that, The successful reception specifically refers to: Calculate the delay between the timestamp carried in the decision command and the reception time of the station area; if the delay is less than or equal to the delay threshold, the reception is considered successful; otherwise, the reception is considered unsuccessful.

3. The power distribution network control method considering communication quality as described in claim 2, characterized in that, It also includes a method for presetting the delay threshold, specifically: Obtain historical datasets of the power distribution network system response time and calculate the average value of each historical dataset. Calculate the difference between each historical data point and the average value, and select the top 95% of historical data points with the smallest differences. Then, calculate the average value again as the time delay threshold.

4. The power distribution network control method considering communication quality as described in claim 1, characterized in that, Step S34 specifically includes: Shift the window w with the highest similarity S(w) to the right by N units to obtain the corresponding decision information matrix, and execute the historical decision instructions in it.

5. A power distribution network control method considering communication quality as described in claim 1, characterized in that, This also includes training the SiamFC network: Calculate the first high-dimensional feature f A Second high-dimensional feature f X The contrast loss is calculated using the following formula: ； ； In the formula, Let i be the label for the i-th type of transformer area data, if If the corresponding historical transformer area data matrix and the decision augmentation matrix X are similar to the i-th type of transformer area data, then... If they are dissimilar, then they are not; β is the boundary threshold; S(f A f X ) is the first high-dimensional feature f A Second high-dimensional feature f X similarity, f Xti Let S(f) be a high-dimensional feature of the i-th type of transformer data in the t-th historical transformer data matrix. Ai f Xti X is the pre-decision data matrix. A The similarity between the data of the i-th type of station in the t-th historical station data matrix; Where, if the similarity S(f) Ai f Xti If the value is greater than the preset value, then the pre-decision data matrix X is determined. A Similar to the data of the i-th type of station in the t-th historical station data matrix; By traversing the decision augmentation matrix X, the boundary threshold β is adjusted to minimize the contrastive loss. Find the historical data matrix of the transformer area where the loss is minimized, and execute the historical decision instructions in the corresponding decision information matrix.

6. A power distribution network control method considering communication quality as described in any one of claims 1 to 5, characterized in that, The transformer area data includes total harmonic distortion (THD) of voltage, effective value of fundamental voltage, effective values ​​of each harmonic voltage, effective value of fundamental current, effective values ​​of each harmonic current, and total harmonic distortion (THD) of current.

Images