A method for locating causes of distributed photovoltaic three-phase voltage imbalance

Abstract

The method for locating the cause of three-phase voltage imbalance in distributed photovoltaic systems utilizes PSCAD simulation software to simulate the causes of the three-phase voltage imbalance and export simulation data. The data specifically includes: positive-sequence voltage components, negative-sequence voltage components, positive-sequence current components, negative-sequence current components, three-phase voltage phase difference, three-phase current phase difference, harmonic voltage, and harmonic current. Then, a dataset X is constructed to train the localization model. After the model is trained, the equipment is installed in the distributed photovoltaic power station to predict the cause of the three-phase voltage imbalance. The data is then transmitted to an internet platform using a 4G module, making it particularly suitable for locating the cause of three-phase voltage imbalance in distributed photovoltaic systems.

Description

Technical Field

[0001] This invention relates to a method for locating the cause of three-phase voltage imbalance in distributed photovoltaic systems, belonging to the technical field of photovoltaic three-phase voltage imbalance detection methods. Background Technology

[0002] Power quality is a technical term describing whether power supply is safe, reliable, and continuous. Power quality problems include numerous issues such as harmonics, voltage flicker, and waveform distortion. Among these, three-phase voltage imbalance is one of the most common power quality problems. With the rapid development of technology, distributed photovoltaic (PV) power generation has achieved significant economic and social benefits in recent years. PV power generation is one of the most effective ways to solve the energy crisis, with advantages such as safety and sustainability. PV power generation does not require consideration of other factors; it can generate electricity as long as there is sunlight. However, to achieve the best economic benefits, PV power generation is often set up as a distributed system, flexibly arranged in multiple locations, offering high flexibility and improving energy utilization. Currently, the main application of distributed PV power generation is grid connection to achieve energy integration, but this also brings a series of problems to the distribution network. One of the most significant problems is that residential electricity loads, PV panels, and other components are mostly distributed single-phase power sources. When these single-phase power sources are connected to the grid, they cause three-phase voltage imbalance. If the three-phase voltage imbalance exceeds the specified range, it will cause abnormal load operation, increase motor temperature, affect equipment utilization efficiency, accelerate circuit aging, and may even cause the collapse of some mechanical systems, data loss, or even excessive harmonic current, resulting in fire and causing personal injury and financial loss.

[0003] As is well known, distributed photovoltaic (PV) power generation is prone to three-phase voltage imbalance. The causes of this imbalance can be numerous, including sudden changes in sunlight intensity, rapid temperature fluctuations, dust brought in by strong winds, and snowfall, all of which can affect PV power generation quality and lead to voltage imbalance. Besides inherent issues with distributed PV systems, imbalances can also be caused by asymmetrical faults, such as single-phase grounding short circuits, line breaks, or component burnout. Given these numerous potential causes, manually inspecting and analyzing each problem would be extremely time-consuming and resource-intensive, and could also negatively impact the power supply.

[0004] Current methods for locating three-phase voltage imbalance sources, such as the invention patent application CN 107797023A, while meeting the needs of three-phase voltage imbalance source location to a certain extent, can only determine whether the source is located on the power station side or the load side. This method does narrow the maintenance scope, requiring only maintenance on the power station side or the load side, but it cannot determine the specific cause of the three-phase voltage imbalance, and the problem of not being able to quickly locate the specific cause and perform rapid maintenance remains. Therefore, it is necessary to develop a new method for locating the cause of three-phase voltage imbalance in distributed photovoltaic systems to solve the above-mentioned problems of existing methods. Summary of the Invention

[0005] The purpose of this invention is to provide a method for locating the cause of three-phase voltage imbalance in distributed photovoltaic systems, thereby solving the problem that existing methods for locating three-phase voltage imbalance sources cannot determine the specific cause of the three-phase voltage imbalance.

[0006] The technical solution of this invention is:

[0007] The method for locating the cause of three-phase voltage imbalance in distributed photovoltaic systems includes the following steps:

[0008] (a) Using PSCAD simulation software to simulate the situation causing three-phase voltage imbalance;

[0009] (ii) Export simulation data;

[0010] (III) Constructing dataset X:

[0011] (iv) After constructing the dataset X, it is fed into the localization model for recognition training. The localization model includes principal component analysis, gradient boosting decision tree, and genetic algorithm.

[0012] (v) Check the accuracy of the positioning model

[0013] After training the localization model, check if its accuracy reaches 95%. If not, repeat the above steps to strengthen the training of the localization model. If it reaches the target, proceed to the next step.

[0014] (vi) Installation

[0015] Installed on the power grid, it reads data in real time.

[0016] (vii) Use

[0017] Once the localization model is trained, it can begin normal operation.

[0018] The system reads and monitors various data from the three-phase line in real time, and then uses these data to solve the characteristic relationship in the positioning model. The results are then compared with those obtained from training the positioning model, and the closest result is used as the output. The output shows the cause of the voltage imbalance in the three lines, and the data is uploaded to the cloud via a 4G module.

[0019] (ix) Troubleshooting three-phase voltage imbalance

[0020] When the voltage approaches the dangerous threshold of three-phase voltage imbalance, an alarm is sent to maintenance personnel for timely repair.

[0021] The beneficial effects of this invention are:

[0022] (1) The present invention adopts the principal component analysis algorithm, which can quickly achieve the effect of dimensionality reduction and quickly and efficiently select the most effective feature vector that can predict three-phase voltage imbalance.

[0023] (2) In predicting the three-phase voltage imbalance, the gradient boosting decision tree algorithm is adopted. It is one of the most advanced ensemble learning algorithms at present. It has fast calculation speed in the prediction stage, can be parallelized between trees, has good robustness, and has good generalization and expressive ability.

[0024] (3) This invention addresses the problem of three-phase voltage imbalance in distributed photovoltaic systems. It can also be applied to other power generation methods such as wind power and hydropower, and has advanced significance in solving the overall problem of three-phase voltage imbalance.

[0025] (4) Compared with manual maintenance, the present invention can predict and check the cause of three-phase voltage imbalance faster and more accurately, with low cost and accurate results.

[0026] (5) It solves a major drawback of distributed photovoltaic power generation, is beneficial to the development of clean energy photovoltaic power generation, and has a favorable factor in solving the energy crisis.

[0027] (6) The relationship between each eigenvector and other eigenvectors was analyzed, the data analysis was more thorough, the results were more convincing, and each cause of three-phase voltage imbalance could be accurately identified, resulting in more accurate output results.

[0028] (7) This system includes a 4G transmission module, which can upload information to the cloud in real time and send it to the maintenance personnel, making it more interactive. Attached Figure Description

[0029] Figure 1 This is a diagram of the three-phase voltage structure.

[0030] Figure 2 A system for locating the cause of three-phase voltage imbalance;

[0031] Figure 3This is a flowchart of the method for locating the three causes of voltage imbalance in this invention;

[0032] Figure 4 This is the schematic diagram of a standard GBDT;

[0033] Figure 5 This is a PSCAD simulation diagram of a three-phase imbalance caused by a sudden change in illumination. Detailed Implementation

[0034] The method for locating the cause of three-phase voltage imbalance in distributed photovoltaic systems includes the following steps:

[0035] (a) Using PSCAD simulation software to simulate the situation causing three-phase voltage imbalance;

[0036] The causes of three-phase voltage imbalance in distributed photovoltaic (PV) power generation include sudden changes in sunlight intensity, rapid temperature changes, dust brought in by strong winds, snowfall, electrical malfunctions, and short circuits. All of these factors can affect PV power generation quality and cause three-phase voltage imbalance. PSCAD simulation software is used to build relevant models and simulate these situations. The specific simulation method is as follows:

[0037] Several scenarios that could cause three-phase voltage imbalance were simulated using PSCAD software. The resulting data included the positive-sequence voltage components, negative-sequence voltage components, positive-sequence current components, negative-sequence current components, three-phase voltage phase difference, three-phase current phase difference, harmonic voltage, and harmonic current for phases A, B, and C. This data was used to construct dataset X. (For PSCAD simulation diagrams of three-phase imbalance caused by sudden changes in illumination, please refer to the instruction manual appendix.) Figure 5 A photovoltaic single-phase grid-connected model was built and connected to phase A of the power grid. Phases B and C were connected to ordinary loads. A sudden change in illumination was performed for simulation. The simulation results show that before the illumination change, there was no obvious three-phase voltage imbalance. After the sudden change in illumination, a significant three-phase imbalance appeared (see the instruction manual appendix). Figure 5 The corresponding data at this time is the same as the aforementioned data, including the positive-sequence voltage component, negative-sequence voltage component, positive-sequence current component, negative-sequence current component, three-phase voltage phase difference, three-phase current phase difference, harmonic voltage, and harmonic current of the three-phase electricity A, B, and C. This data is derived as the dataset X for the three-phase voltage imbalance phenomenon in this case and used to train the localization model. The localization model includes principal component analysis, gradient boosting decision tree, and genetic algorithm (see the appendix of the instruction manual). Figure 3 );

[0038] (ii) Export simulation data;

[0039] After simulating the three-phase voltage imbalance using PSCAD simulation software, the data of the positive-sequence voltage component, negative-sequence voltage component, positive-sequence current component, negative-sequence current component, three-phase voltage phase difference, three-phase current phase difference, harmonic voltage and harmonic current of the three-phase electricity A, B and C are exported.

[0040] (III) The process and principles for constructing dataset X are as follows:

[0041] After exporting the simulation data from step (II), the next step is to construct dataset X (the structure of the three-phase electricity is as follows). Figure 1 As shown, the exported simulation data includes: positive-sequence voltage components, negative-sequence voltage components, positive-sequence current components, negative-sequence current components, three-phase voltage phase difference, three-phase current phase difference, harmonic voltage, and harmonic current of the three-phase electricity (A, B, and C).

[0042] Construct dataset X using the simulation data exported in step (II);

[0043]

[0044] This represents the sampling point of the positive sequence component of the voltage of the i-th phase line; This represents the sampling point of the negative sequence component of the voltage of the i-th phase line; This represents the sampling point of the positive sequence component of the current in the i-th phase line; This represents the sampling point of the negative sequence component of the current in the i-th phase line; This represents the voltage phase difference between phase i and phase u. Let M represent the current phase difference between phase i and phase u, M represent the harmonic voltage characteristic set, and H represent the harmonic current characteristic set.

[0045] Specifically, N represents the number of sampled values, totaling N sampled values; A, B, and C represent the three phases of the three-phase voltage (see the instruction manual appendix). Figure 1 ).

[0046]

[0047]

[0048]

[0049] In the above formula, This represents the s-th sampling point of the b-th harmonic voltage.

[0050]

[0051] In the above formula, This represents the s-th sampling point of the b-th harmonic current.

[0052] This completes the construction of dataset X;

[0053] (iv) After constructing the dataset X, input it into the localization model for recognition training. The localization model includes principal component analysis, gradient boosting decision tree, and genetic algorithm (see the instruction manual appendix). Figure 3 )

[0054] (1) The principal component analysis structure is as follows

[0055] Principal component analysis refers to the ability to reduce an m-dimensional matrix to a smaller one. l Since the set X has a high dimension, it is a high-dimensional matrix, which is time-consuming and laborious to process. In order to improve the running efficiency and detection efficiency and shorten the running time, we first perform principal component analysis on the dataset X.

[0056] Assuming an m-dimensional dataset We need to reduce it to k dimensions. The specific process is as follows:

[0057] ① First, preprocessing is performed, including mean reduction and normalization.

[0058] Demeaning:

[0059]

[0060] In layman's terms, it means subtracting the average value of each column of data in X from the average value of that column to obtain a new matrix. ;

[0061] Then normalize:

[0062]

[0063] In layman's terms, it involves dividing each data point by the maximum value in its column, thus reducing all data in the matrix to the range of 0-1, resulting in the matrix... ;

[0064] ② Calculate the matrix covariance matrix The specific calculation formula is as follows:

[0065]

[0066] ③ For the covariance matrix mentioned above, the "eigenvalue decomposition method" can be used to further solve for the covariance matrix. eigenvalues ​​and eigenvectors;

[0067] The method for finding the eigenvalues ​​and eigenvectors of a matrix is ​​a conventional method and will not be elaborated here.

[0068] ④ The eigenvalues ​​obtained in step ③ above need to be sorted from largest to smallest, and the largest one should be selected. l1; then take its corresponding l The eigenvectors are used as column vectors in sequence to form the eigenvector matrix P;

[0069] ⑤ Convert the data to l In the new space constructed by the eigenvectors, that is:

[0070]

[0071] After principal component analysis, a new feature vector set Y is obtained. Next, this feature vector set Y is used for gradient boosting decision tree algorithm analysis.

[0072] (2) The gradient boosting decision tree structure is as follows:

[0073] Using PSCAD simulation software, each scenario causing three-phase voltage imbalance is simulated sequentially. Principal component analysis is then used to obtain the dataset Y for each cause (see Formula 14). Gradient boosting decision tree algorithm is then used to analyze and solve the dataset Y. Finally, a specific column is analyzed sequentially. With the remainder Relationships between dimensional matrices, using the remainder dimensional vector derivation This involves identifying the relationships between these data vectors; for each different cause, identifying the distinct feature relationships between these data vectors; and then, after training the localization model, reverse deduction is performed to deduce, based on the detected data, which cause the three-phase voltage imbalance; the specific gradient boosting decision tree algorithm is as follows:

[0074] After principal component analysis, the dataset Y was obtained.

[0075]

[0076] Sequentially from 1 to l The column is used as output data, and the remaining data is used as input data:

[0077] With the first l List the output data, 1 to l-1 Let's take the input data as an example to illustrate, and let... For input data

[0078] To output the value, a functional relationship needs to be constructed:

[0079] Predict output values ​​based on features with minimal error.

[0080] The English name for Gradient Boosting Decision Tree is Gradient Boosting Decision Tree, so its abbreviation is GBDT (see the instruction manual). Figure 4 (This is a schematic diagram of a standard GBDT). This gradient boosting decision tree can contain several classification and regression trees. The English name for this classification and regression tree is CART, hence the abbreviation. Here, we can assume that the k-th classification and regression tree is denoted as […]. Then, the predicted values ​​of the first k classification and regression trees are the predicted values ​​of the gradient boosting decision tree, which can be expressed as:

[0081]

[0082] Gradient Boosting Decision Tree (GBDT) is essentially an additive model, where the final prediction of the gradient boosting decision tree is obtained by summing the predictions of all the classification and regression trees (CART). Alternatively, it can be represented recursively, making it simpler; that is, representing k classification and regression trees (CART) in a recursive manner. It can also be understood as using a forward, stepwise calculation and incremental approach to optimize the overall gradient boosting decision tree (GBDT) model, as shown below.

[0083]

[0084] The objective function is also set for the CART (Classification and Regression Tree) system, specifically the objective function for the k-th CART tree to be trained. This can be expressed as follows:

[0085]

[0086] Gradient boosting decision trees, as the name suggests, use gradient descent to minimize the objective function as quickly as possible. The objective function is... The gradient is: That is, the new function obtained along this direction can make the corresponding loss function smaller. The knowledge points involved are generalized functions and gradient descent flow of generalized functions.

[0087] Based on the gradient descent flow concept of generalized functions, the function can be further optimized. That is, it can be expressed in the following form

[0088]

[0089] From the formula and formula We can obtain:

[0090]

[0091] As shown in Formula 20, the objective function of the k-th classification and regression tree CART is the negative gradient of the objective function (the output values ​​of the previous (k-1) CART classification and regression trees). This can be easily understood based on the generalized function gradient descent flow. Another key factor is the learning rate. In order to obtain the fastest objective function descent rate, we set the learning rate to 1.

[0092] The objective function can logically be expressed as the sum of squared residuals, as shown in the following formula:

[0093]

[0094] Furthermore, it can be deduced that:

[0095]

[0096] In short, the goal, or task, of each classification and regression tree CART in the Gradient Boosting Decision Tree (GBDT) is to fit the residual remaining after the sum of all previous classification and regression tree CARTs.

[0097] Gradient boosting decision trees have a serial structure, where the first classification and regression tree (CART) can be represented as follows: The remaining residuals of this classification and regression tree CART can be represented as :

[0098]

[0099] Next, our gradient boosting decision tree (GBDT) model can be represented as:

[0100]

[0101] The objective of the second classification and regression tree (CART) is to obtain and The relationship between them can be represented in the following form:

[0102]

[0103] In the above formula That is, our Gradient Boosting Decision Tree (GBDT) model... The residuals remaining after prediction can be expressed in the following form:

[0104]

[0105] Furthermore, better and more precise results were obtained. It can be represented as

[0106]

[0107] Furthermore, this process continues...; if a total of k classification and regression trees (CART) are needed, then k iterations of this process are required until our gradient boosting decision tree (GBDT) is trained. After training k classification and regression trees (CART), a very accurate... The output predicted value is the sum of k CART classification and regression trees. The formula can be expressed as follows:

[0108]

[0109] (3) Genetic Algorithm

[0110] After the Gradient Boosting Decision Tree (GBDT) algorithm, a range of values ​​for parameters such as the number of iterations, area under the operational feature curve, and depth needs to be obtained. Then, a genetic algorithm can be used to obtain the optimal solution. Here, the genetic algorithm (GA) refers to the method of searching for the optimal solution proposed by John Holland in the United States. The genetic algorithm searches for the optimal solution for the number of iterations, area under the operational feature curve, and depth after the GBDT algorithm within the range of values.

[0111] (v) Check the accuracy of the positioning model

[0112] After training the localization model, check if its accuracy reaches 95%. If not, repeat the above steps to strengthen the training of the localization model. If it reaches the target, proceed to the next step.

[0113] (vi) Installation

[0114] Installed on the power grid, it reads data in real time.

[0115] (vii) Use

[0116] Once the localization model is trained, it can begin normal operation.

[0117] The system reads and monitors various data from the three-phase line in real time, and then uses these data to solve the characteristic relationship in the positioning model. The results are then compared with those obtained from training the positioning model, and the closest result is used as the output. The output shows the cause of the voltage imbalance in the three lines, and the data is uploaded to the cloud via a 4G module.

[0118] (ix) Troubleshooting three-phase voltage imbalance

[0119] When the voltage approaches the dangerous threshold of three-phase voltage imbalance, an alarm is sent to maintenance personnel for timely repair.

[0120] The invention patent with publication number CN 107797023A only uses detection equipment to read the voltage and current values ​​at the point of common coupling, and then calculates the negative sequence voltage value, negative sequence current value, and imbalance ratio to determine whether the three-phase voltage imbalance source originates from the power station side or the load side. However, it suffers from limited data types and volume, lacks model training, and cannot pinpoint the specific cause of the three-phase imbalance. This invention utilizes PSCAD to simulate the causes of three-phase voltage imbalance, exports data, constructs a dataset X, and analyzes it through model training to identify the corresponding characteristic quantities and relationships of dataset X for each cause. The system is then installed on the power grid, and the model analyzes the characteristic quantities and relationships of the power grid dataset in real time. Once a three-phase imbalance problem occurs, the characteristic quantities and relationships of the power grid dataset X are immediately compared with those of the simulation dataset X to quickly identify the cause of the three-phase imbalance. The system then notifies maintenance personnel via a 4G module for rapid repair.

[0121] The method for locating the cause of three-phase voltage imbalance in distributed photovoltaic systems utilizes PSCAD simulation software to simulate the causes of the three-phase voltage imbalance and export simulation data (specifically including: positive-sequence voltage component, negative-sequence voltage component, positive-sequence current component, negative-sequence current component, three-phase voltage phase difference, three-phase current phase difference, harmonic voltage, and harmonic current). Then, a dataset X is constructed (as shown in Equation 1) to train the model. After the model is trained, the equipment is installed in the distributed photovoltaic power station to predict the cause of the three-phase voltage imbalance. The data is then transmitted to an internet platform using a 4G module, making it particularly suitable for locating the cause of three-phase voltage imbalance in distributed photovoltaic systems.

Claims

1. A method for locating the cause of distributed photovoltaic three-phase voltage imbalance, characterized in that: The distributed photovoltaic three-phase voltage imbalance reason positioning method comprises the following steps: (1) simulate the situation of causing three-phase voltage imbalance by using PSCAD simulation software; (2) export simulation data; After simulating three-phase voltage imbalance by using PSCAD simulation software, export the voltage positive sequence component, voltage negative sequence component, current positive sequence component, current negative sequence component, three-phase voltage phase difference, three-phase current phase difference, harmonic voltage and harmonic current data of A, B and C three-phase power at this time; (3) the process and principle of constructing data set X are as follows: After exporting the simulation data in step (2), the next step is to construct data set X using the simulation data exported in step (2); ； a positive sequence component sample point representing the jth voltage of the ith phase line; a negative sequence component sample point representing the jth voltage of the ith phase line; a positive sequence component sample point representing the jth current of the ith phase line; a negative sequence component sample point representing the jth current of the ith phase line; a voltage phase difference between the ith phase and the u phase, a current phase difference between the ith phase and the u phase, M represents a harmonic voltage feature set, and H represents a harmonic current feature set. Specifically, N represents the number of sampling values, a total of N sampling values; A, B and C represent three-phase voltage of three-phase power respectively; ； ； In the above formula, denotes the s-th sample point of the b-th harmonic voltage; ； In the above formula, denotes the s-th sample point of the b-th harmonic current; In this way, the construction of data set X is completed; (4) after constructing data set X, it is brought into the positioning model for recognition training, and the positioning model comprises principal component analysis, gradient boosting decision tree and genetic algorithm; (1) the structure of principal component analysis is as follows Principal component analysis refers to reducing an m-dimensional matrix to l a lower-dimensional matrix by first performing principal component analysis on the data set X; Suppose we have an m-dimensional data set To reduce to k dimensions, the procedure is as follows: ① first, pre-processing, and de-meaning and normalization are performed; De-meaning: ； Subtract the average of each column of data from X to get a new matrix ; Then, normalization processing is performed: ； Divide each data by the maximum value of the column where it is located, and let all data in the matrix vary to the range of 0-1, to obtain the matrix ; ii. Compute the matrix covariance matrix of , where the specific formula is as follows ； ③ For the covariance matrix described above, the eigenvalues and eigenvectors of the covariance matrix are further solved by using "eigenvalue decomposition method" . ④ The characteristic values solved in step ③ above need to be sorted from large to small, and the largest one is selected l ; then the corresponding l eigenvector is sequentially taken as a column vector to form an eigenvector matrix P; (5) converting the data into a new space constructed by l a number of feature vectors, i.e.: ； After principal component analysis, a new feature vector set Y is obtained, and then gradient boosting decision tree algorithm analysis is performed using the vector set Y; (2) the structure of gradient boosting decision tree is as follows: After simulating each cause of three-phase voltage imbalance by using PSCAD simulation software, the data set Y of each cause is obtained by using principal component analysis algorithm; the data set Y is analyzed and solved by using gradient boosting decision tree algorithm; a certain column is analyzed in sequence with the remaining with the remaining dimensional vector derivation , find the relationship between them; for each reason that causes three-phase voltage imbalance, find the characteristic relationship between these data vectors and their corresponding features; then, after training the positioning model, deduce which reason caused the three-phase voltage imbalance through the detected data. The specific gradient boosting decision tree algorithm is as follows: After principal component analysis, the data set Y is obtained ； Output data in the first column and input data in the second column l Output data in the first column and input data in the second column l Output data in the first column and input data in the second column l- 1 Output data in the first column and input data in the second column Output data in the first column and input data in the second column Output data in the first column and input data in the second column Output data in the first column and input data in the second column The English name of the gradient boosting decision tree is Gradient Boosting Decision Tree, so it is called GBDT. The gradient boosting decision tree contains a number of classification and regression trees, and the English name of the classification and regression tree is classification and regression tree, so it is called CART. Suppose the kth classification and regression tree is denoted as The prediction value of the first k classification and regression trees is the prediction value of the gradient boosting decision tree, which is denoted as ； Gradient boosting decision tree, namely GBDT, is essentially an additive model, that is, the final prediction value of gradient boosting decision tree is obtained by adding the prediction values of all classification and regression trees CART; that is, in the form of recursion; the k classification and regression trees CART are expressed in the form of recursion; the forward step-by-step calculation and incremental form is adopted to realize the overall optimization of the gradient boosting decision tree GBDT model, which can be expressed as follows ； The setting of the objective function is also set for the classification and regression tree CART, that is, the objective function of the kth classification and regression tree CART to be trained, which is expressed as ； Gradient boosting decision trees employ gradient descent to minimize the objective function as quickly as possible. The objective function is... The gradient is: That is, the new function obtained along this direction makes the corresponding loss function smaller; According to the gradient descent flow idea of the universal function, the function is further optimized i.e. expressed in the form ； From the equation and the equation it follows that ； As shown in formula (20), the objective function of the kth classification and regression tree CART is to obtain the negative gradient of the objective function, and the learning rate is set to 1 in order to obtain the fastest objective function descent rate; For the objective function, it is expressed in the form of residual sum of squares, as shown in the following formula: ； Further, it is derived that: ; The objective of each classification and regression tree CART in the gradient boosting decision tree GBDT, or the task, is to fit the residual left by the sum of all previous classification and regression trees CART; Gradient boosting decision trees are in a cascade structure, where the first classification and regression tree, CART, is represented as The residual left by this classification and regression tree, CART, is represented as : ； Next, the gradient boosting decision tree GBDT model is expressed as: ； For the second classification regression tree CART, the goal is to obtain the relationship between and is expressed in the form of ； In the above formula is the residual error left by the GBDT model when predicting, and is specifically expressed in the following form: ； Further, a better, more accurate is represented as ； Further, always analogize…; if a total of k classification and regression trees CART are needed, then k times of analogization is needed until the gradient boosting decision tree GBDT is trained completely; If k classification and regression trees (CART) are trained, then a very accurate classification and regression tree can be obtained. The output predicted value is the sum of k classification and regression trees (CART); the formula is expressed in the following form: ； (3) genetic algorithm After the gradient boosting decision tree GBDT algorithm, a series of value ranges of the iteration number, the area under the operating characteristic curve, and the depth parameter need to be obtained. Next, the genetic algorithm is used to obtain the optimal solution. The genetic algorithm is used to search for the optimal solution of the iteration number, the area under the operating characteristic curve, and the depth parameter after the GBDT algorithm in the value range. (Five) Check the accuracy of the positioning model After training the positioning model, check whether the accuracy of the positioning model reaches 95%. If it does not reach, repeat the above steps to strengthen the training of the positioning model. If it reaches, go to the next step. (Six) Installation Install on the power grid and read data in real time. (Seven) Use After the positioning model is trained, it can start normal work: Real-time detection of monitoring data of three-phase lines, which is also brought into the positioning model to solve the characteristic relationship; Then compare with the results obtained by the positioning model training. The closest result is taken as the output. The output causes the three-phase voltage imbalance. Through the 4G module, it is uploaded to the cloud; (Eight) Three-phase voltage imbalance problem repair: When the three-phase voltage imbalance degree is close to the danger threshold, send an alarm to the maintenance personnel for timely repair. The specific steps of simulating the three-phase voltage imbalance situation by using PSCAD simulation software in step (one) are as follows: For the reasons causing the three-phase voltage imbalance of distributed photovoltaic power generation, including sudden change of light intensity, too fast temperature change, dust brought by strong wind, snowfall, electrical fault, and line short circuit. Related models are built by using PSCAD simulation software to simulate the corresponding situation. The specific simulation method is as follows: Several situations that can cause three-phase voltage imbalance are simulated by PSCAD software to obtain corresponding data including A, B, and C three-phase voltage positive sequence component, voltage negative sequence component, current positive sequence component, current negative sequence component, three-phase voltage phase difference, three-phase current phase difference, harmonic voltage, and harmonic current, so as to construct data set X. The photovoltaic single-phase grid-connected model is connected to the power grid A phase, B phase and C phase are connected to the ordinary load. The light is suddenly changed to simulate. From the simulation results, before the light is changed, there is no obvious three-phase voltage imbalance. After the light is suddenly changed, there is obvious three-phase imbalance. At this time, the corresponding data is exported, including A, B, and C three-phase voltage positive sequence component, voltage negative sequence component, current positive sequence component, current negative sequence component, three-phase voltage phase difference, three-phase current phase difference, harmonic voltage, and harmonic current. Export as data set X for training the positioning model under the condition of three-phase voltage imbalance. The positioning model includes principal component analysis, gradient boosting decision tree, and genetic algorithm.

Images