A Voltage Stability Control Method Based on Photovoltaic Power Generation Prediction

Abstract

This invention discloses a voltage stability control method based on photovoltaic power generation prediction, comprising the following steps: predicting photovoltaic power generation, and controlling the voltage stability based on the predicted photovoltaic power generation P. pv (t), combined with the load curve P L (t), Q L (t), adjusting the reactive power Q of the photovoltaic inverter pv This invention achieves "anticipatory automatic control" of the voltage at the grid connection point of distributed power sources, minimizing or limiting the fluctuation of the voltage drop ΔU(t) to a small range, thus ensuring that the grid-connected voltage U2(t) does not exceed the limit. This avoids the voltage fluctuations and instantaneous voltage exceeding limits that often occurred in the early stages of regulation due to the communication delay between the reactive power controller and the inverter's 485 interface. It has significant engineering application value.

Description

Technical Field

[0001] This invention belongs to the field of power system technology, specifically relating to a voltage stability control method based on photovoltaic power generation prediction. Background Technology

[0002] Distributed photovoltaic (PV) output is characterized by strong randomness, volatility, and intermittency. Currently, the situation is that distributed PV is generally controlled by a fixed grid connection point power factor of 1 under local control mode. However, the R / X ratio of the distribution network is generally large. The active power injection of distributed PV will cause the grid connection point voltage to rise. The active power fluctuation may cause large voltage fluctuations, which brings difficulties to the voltage regulation of the distribution network.

[0003] Inverters in distributed photovoltaic systems generally have fast and continuous reactive power regulation capabilities. Incorporating this reactive power and voltage regulation method into the regional reactive power and voltage control of the distribution network will help enhance the reactive power and voltage regulation capabilities of the distribution network.

[0004] Currently, the distributed reactive power control method calculates the required reactive power variable based on the deviation of the grid connection point voltage and then distributes it to the inverter. Due to the communication delay between the reactive power controller and the inverter's 485 communication, the result is often that voltage fluctuations and instantaneous voltage values ​​exceed the standard in the early stage of regulation. Summary of the Invention

[0005] In view of the above-mentioned technical problems, the present invention provides a voltage stability control method based on photovoltaic power generation prediction.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0007] A voltage stability control method based on photovoltaic power generation prediction includes the following steps: predicting photovoltaic power generation, and controlling the voltage stability based on the predicted photovoltaic power generation P. pv (t), combined with the load curve P L (t), Q L (t), adjusting the reactive power Q of the photovoltaic inverter pv (t), so that the voltage drop ΔU(t) fluctuation is minimized or within a small range, ensuring that the grid-connected voltage U2(t) does not exceed the limit.

[0008] Preferably, predicting photovoltaic power generation includes the following steps:

[0009] Data Acquisition: Within one hour prior to the target time, the photovoltaic power generation P is measured every five minutes. pv Perform a sampling to obtain active power data;

[0010] Data Processing: The acquired active power data is preprocessed to obtain raw data samples. Preprocessing includes interpolating and imputing samples with few missing values ​​and discarding samples with many missing values. The preprocessed raw dataset is then normalized to obtain the true sample x of photovoltaic active power output. N ;

[0011] Generating virtual samples: Generative Adversarial Networks (GANs) learn through unsupervised learning. The generative network consists of a generator (G) and a discriminator (D). A set of noise vectors is randomly input into the generator, which then generates virtual samples based on real samples x. N The data distribution in the dataset generates enough virtual samples x';

[0012] Optimize the parameters of the adversarial network: Input both real data samples and virtual samples into the discriminant model, output the results, and then continuously update the parameters of the generator model based on the results. The two play a game and continuously optimize the parameters until Nash equilibrium is reached. At this point, the virtual samples trained by the generator model are basically the same as the real samples.

[0013] Clustering and obtaining predicted values: The K-means clustering algorithm is used to divide the above virtual dataset into several clusters. The cluster with the most data distribution is selected, and its cluster center is used as the predicted photovoltaic power generation.

[0014] The present invention has the following beneficial effects:

[0015] (1) By predicting the active power of photovoltaic power generation, a strategy for adjusting the reactive power of photovoltaic inverters is developed to achieve the goal of stabilizing the voltage fluctuation at the grid connection point of the photovoltaic power station and preventing the grid connection voltage value from exceeding the standard.

[0016] (2) The voltage of the distributed power grid connection point can be "automatically controlled according to expectations". This can avoid the situation where voltage fluctuations and instantaneous voltage values ​​exceed the standard in the early stage of regulation due to the delay in communication between the reactive voltage controller and the inverter 485. This has engineering application value. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of a distributed photovoltaic grid-connected system in a voltage stability control method based on photovoltaic power generation prediction according to an embodiment of the present invention;

[0018] Figure 2 This is the basic structure of the GAN network in a voltage stability control method based on photovoltaic power generation prediction according to an embodiment of the present invention;

[0019] Figure 3 This is a flowchart illustrating the photovoltaic power generation prediction process in a voltage stability control method based on photovoltaic power generation prediction, according to an embodiment of the present invention. Detailed Implementation

[0020] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0021] This invention proposes a voltage stability control method based on photovoltaic power generation prediction, comprising the following steps: predicting photovoltaic power generation, and controlling the voltage stability based on the predicted photovoltaic power generation P. pv (t), combined with the load curve P L (t), Q L (t), adjusting the reactive power Q of the photovoltaic inverter pv (t), so that the voltage drop ΔU(t) fluctuation is minimized or within a small range, and the grid-connected voltage U2(t) is guaranteed not to exceed the limit. By combining the predicted photovoltaic active power, the voltage of the distributed power grid connection point can be "automatically controlled according to expectations" to reduce voltage fluctuation and exceedance.

[0022] With attachment Figure 1 Taking the distributed photovoltaic grid-connected system shown as an example, the end power users include the PV power generation system. In the figure, U1 is the voltage at the beginning of the line, U2 is the voltage at the end of the line, P and Q are the active power and reactive power flowing to the end users, respectively, R is the line resistance, and X is the line reactance. The voltage relationship of the line can be obtained as shown in equation (1).

[0023]

[0024] The voltage drop of the line is:

[0025]

[0026] Expressed in terms of instantaneous value, i.e.:

[0027]

[0028] Because the R value of the distribution network is relatively large, P PV Fluctuations in the voltage level significantly affect the fluctuation of ΔU. Furthermore, when photovoltaic power generation is large and cannot be fully absorbed locally, some of the generated power will be fed back into the system through the transmission line. In this case, the direction of the active power P in the above equation will change, meaning PR will become negative. When PR + QX becomes negative, i.e., ΔU is negative, the voltage at the end of the transmission line (the photovoltaic power user end) will be higher than the bus voltage of the sending substation, potentially leading to overvoltage problems.

[0029] Control objectives make Minimum, in predicting Ppv (t) and known (or predicted) P L (t), Q L In the case of (t), the adjustment quantity Q is real time pv The value of (t) is as follows:

[0030]

[0031] The maximum adjustable reactive power of a photovoltaic power station is determined as follows:

[0032]

[0033]

[0034] Where M represents the total number of inverters in the photovoltaic power station. The maximum value of the total reactive power capacity of a photovoltaic power station. S represents the maximum reactive power output capacity of each inverter. Ni The rated capacity of the inverter is the capacity of the photovoltaic system. When the active power output of the photovoltaic system is lower than the rated capacity of the inverter, the remaining capacity can provide reactive power support to the grid. (In actual engineering, the capacity of the inverter is configured to be greater than the rated power of the photovoltaic panel.)

[0035] Reactive power distribution can allocate the total reactive power regulation to each inverter according to its capacity ratio based on the reactive power margin of different inverters, as shown in the following formula.

[0036]

[0037] This allocation method ensures that the reactive power generated by each inverter is within the allowable range.

[0038] In specific application examples, the photovoltaic power prediction process is shown in the appendix. Figure 3 The specific description is as follows:

[0039] Step 1: Data Acquisition. Within the hour preceding the target time, the photovoltaic power generation P is recorded every five minutes. pv Perform a sampling to obtain active power data.

[0040] Step two, data processing. The acquired active power data is preprocessed to obtain raw data samples x. Preprocessing includes interpolating and imputing samples with few missing values, while discarding samples with many missing values. To ensure that each data point has equal impact, the preprocessed raw dataset undergoes Z-score normalization to obtain the true photovoltaic active power output samples x. N :

[0041]

[0042] Where x is the preprocessed original dataset, μ is the sample mean, and S is the standard deviation.

[0043] Step three: Generative Adversarial Networks (GANs) learn through unsupervised learning. (See attached document.) Figure 2 The generative network comprises a generator (G) and a discriminator (D). A set of noise vectors is randomly input into the generator, which then uses real samples x as a basis for its operation. N The data distribution in the dataset generates enough virtual samples x'.

[0044] Step 4, convert the real data sample x N Both the virtual sample x' and the discriminator are input into the discriminator, and the discriminator outputs the judgment result.

[0045] Step 5: Calculate the error based on the objective function. The generative model uses a convolutional neural network (CNN). The addition of convolutional layers allows the generator G and discriminator D to accurately capture the dataset. The objective function of the GAN model is:

[0046]

[0047] Where, x N Given real data (z) and random data (z), the generator G(z) generates virtual samples x'. The value of D(*) is determined by the discriminator and ranges from 0 to 1. If D(*) = 0, it means the discriminator considers the input data to be fake; if D(*) = 0, it means the discriminator considers the input data to be fake. * ) = 1 indicates that the discriminator considers the input data to be real data.

[0048] In the objective functions of the GANs mentioned above, the training objectives for G and D are different.

[0049] The optimization loss function for generator G can be written as:

[0050] Loss G =E z～N(z|0，1) log(1-D(G(z))) (10)

[0051] Of course, the optimized loss function of the discriminator D can be written as:

[0052]

[0053] p g For the data distribution generated by the generator, during training, the generator G aims to make the discriminator D(G(z)) approach 1 as much as possible for the generated virtual sample x', that is, to minimize the objective function V(D,G), thus minimizing the difference between the virtual sample generated by the generator and the real sample; while the discriminator D aims to input the real sample data x'. NWhen the input is a real sample x', it is best to have D(x) = 1; while when the input is a virtual sample x', it is best to have D(G(z)) = 0, that is, to maximize the objective function V(D, G), which is to maximize the probability that the discriminator judges the real sample as "true" and the virtual sample as "false".

[0054] Step Six, Update Parameters: During discriminator training, keep the generator parameters unchanged, update the discriminator parameters using the characteristics of the sample data, then the discriminator outputs the results, and continuously update the generator parameters based on the results. That is, the generator parameter updates come from the backpropagation of the discriminator, and p... g The updates continue, and the two sides iterate alternately, constantly optimizing their own parameters, creating a competitive situation until p(z) = p(x). N When both reach Nash equilibrium, the objective functions of D and G are both optimal. At this point, the output value of D is 0.5. The virtual samples trained by the generator are basically the same as the real samples, and then proceed to the next step. Otherwise, return to step three and continue training.

[0055] Step seven: This patent employs the commonly used Euclidean distance clustering algorithm—the K-means algorithm. The advantages of the K-means clustering algorithm include: relatively simple data processing, low time complexity, and fast computation speed.

[0056] K-means clustering divides a given dataset of n data objects into k clusters (n and k are positive integers), each cluster being a separate cluster. Its characteristics include: each data object can only belong to one cluster; data within the same cluster have high similarity; and the similarity between clusters is low. This process is repeated until the clustering is complete. The algorithm steps are as follows: 1) Determine cluster centers: Randomly select k data objects from the real data sample as initial cluster centers; 2) Data partitioning: Calculate the distance from all data points to the cluster centers and partition the data into the corresponding clusters based on the principle of minimum distance; 3) Update cluster centers: Based on the virtual sample x', recalculate the average value of the data in each cluster, iteratively update the cluster centers, and repeat steps 2) and 3) until a termination condition is met (e.g., minimum error, number of iterations, etc.). The cluster with the most data distribution is selected, and its cluster center is used as the predicted power of photovoltaic power generation.

[0057] The adjustable reactive power of the photovoltaic inverter is calculated based on the rated value of the power generation unit and the predicted active power output, to limit the reactive power output of all power generation units in the station, so as to maximize the utilization of the reactive power regulation capability of the power generation unit.

[0058] Furthermore, based on the reactive power required for control and in accordance with the aforementioned limitations, it is proportionally allocated to each inverter to jointly achieve reactive power regulation.

[0059] The embodiments of the present invention use "expected automatic control" to solve the problem of voltage fluctuation and rise at the user end of distributed photovoltaic access, which has engineering application value.

[0060] It should be understood that the exemplary embodiments described herein are illustrative and not restrictive. Although one or more embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art will understand that various changes in form and detail may be made without departing from the spirit and scope of the invention as defined by the appended claims.

Claims

1. A voltage stability control method based on photovoltaic power generation prediction, characterized in that, Includes the following steps: Predicting photovoltaic power generation capacity, based on the predicted photovoltaic power generation capacity P pv (t), combined with the load curve P L (t), Q L (t), adjusting the reactive power Q of the photovoltaic inverter pv (t), so that the fluctuation of voltage drop ΔU(t) is minimized or within a small range, ensuring that the grid-connected voltage U2(t) does not exceed the limit. Predicting photovoltaic power generation includes the following steps: Data Acquisition: Within one hour prior to the target time, the photovoltaic power generation P is measured every five minutes. pv Perform a sampling to obtain active power data; Data Processing: The acquired active power data is preprocessed to obtain raw data samples. Preprocessing includes interpolating and imputing samples with few missing values ​​and discarding samples with many missing values. The preprocessed raw dataset is then normalized to obtain the true sample x of photovoltaic active power output. N ; Generating Virtual Samples: The learning method of Generative Adversarial Networks (GANs) is unsupervised learning. The GAN consists of a generator G and a discriminator D. A set of noise vectors is randomly input into the generator, which then generates virtual samples based on real samples x. N The data distribution in the model generates enough virtual samples x'. The objective function of the GAN model is: Where, x N Given real data and random data, the generator G(z) generates virtual samples x'. The value of D(*) is determined by the discriminator and ranges from 0 to 1. If D(*) = 0, it means that the discriminator considers the input data to be fake data; if D(*) = 1, it means that the discriminator considers the input data to be real data. The optimization loss function for generator G is written as: LoSS G =E z～N(z|0，1) log(1-D(G(z))) The optimized loss function of discriminator D is written as: p g For the data distribution generated by the generator, during training, the goal of the generator G is: for the generated virtual sample x', to make the discriminator D(G(z)) approach 1 as much as possible, that is, to minimize the objective function V(D,G), and to minimize the difference between the virtual sample generated by the generator and the real sample; while the goal of the discriminator D is: input real sample data x N When x is the input, it is best to have D(x) = 1; while when the input is the generated virtual sample x', it is best to have D(G(z)) = 0, that is, to maximize the objective function V(D, G). Optimize the parameters of the adversarial network: Input both real data samples and virtual samples into the discriminant model, output the results, and then continuously update the parameters of the generator model based on the results. The two play a game and continuously optimize the parameters until Nash equilibrium is reached. At this point, the virtual samples trained by the generator model are basically the same as the real samples. Clustering and obtaining predicted values: The virtual dataset is divided into several clusters using the K-means clustering algorithm. The cluster with the most data distribution is selected, and its cluster center is used as the predicted photovoltaic power generation. The adjustable reactive power of a photovoltaic inverter is based on the rated value of the power generation unit and the predicted active power output. The limit of reactive power output of all power generation units in the station is calculated to maximize the utilization of the reactive power regulation capability of the power generation unit. In a distributed photovoltaic grid-connected system, the end-user electricity includes the PV power generation system. U1 is the voltage at the beginning of the line, U2 is the voltage at the end of the line, P and Q are the active power and reactive power flowing to the end-user, respectively, R is the line resistance, and X is the line reactance. The voltage relationship of the line is shown in equation (1): The voltage drop of the line is: Expressed in terms of instantaneous value, i.e.: Because the R value of the distribution network is relatively large, P PV The fluctuation of the value has a significant impact on the fluctuation of ΔU. When the photovoltaic power generation is large and cannot be fully consumed locally, part of the power generated will be fed back to the system through the line. At this time, the direction of the active power P in the above formula will change, that is, PR will become negative. When PR+QX becomes negative, that is, ΔU is negative, it will cause the voltage at the end of the line to be higher than the bus voltage of the sending substation, and even cause overvoltage problems. Control objectives make Minimum, in predicting P pv (t) and known (or predicted) P L (t), Q L In the case of (t), the adjustment quantity Q is real time pv The value of (t) is as follows: The maximum adjustable reactive power of a photovoltaic power station is determined as follows: Where M represents the total number of inverters in the photovoltaic power station. The maximum value of the total reactive power capacity of a photovoltaic power station. S represents the maximum reactive power output capacity of each inverter. Ni The rated capacity of the inverter is the capacity of the photovoltaic system. When the active power output of the photovoltaic system is lower than the rated capacity of the inverter, the remaining capacity can provide reactive power support to the grid. Reactive power allocation is based on the reactive power margin of different inverters, distributing the total reactive power regulation to each inverter according to its capacity ratio, as shown in the following formula: This allocation method ensures that the reactive power generated by each inverter is within the allowable range.

Images