Photovoltaic power station grid-connected inverter regulation and control method

Abstract

The invention discloses a photovoltaic power station grid-connected inverter regulation and control method, which comprises the following steps: A, measuring the direct current voltage and direct current of a current photovoltaic power station, and calculating the current output power of the photovoltaic power station; b, changing the MPPT duty ratio, calculating the variation of the output power of the photovoltaic power station, and determining the adjustment direction of the MPPT duty ratio according to the positive and negative change directions of the output power of the photovoltaic power station; and C, adjusting the MPPT duty ratio according to the adjustment direction of the MPPT duty ratio obtained in the step B, and optimizing the operation parameters of the inverter by using a random forest model before each adjustment. The defects in the prior art can be overcome, and the photovoltaic power generation grid-connected control precision is improved.

Description

Technical Field

[0001] The invention relates to the technical field of photovoltaic power generation grid-connected control, and in particular to a photovoltaic power station grid-connected inverter control method. Background Art

[0002] With the gradual increase of photovoltaic power stations in my country, the installed capacity of photovoltaic power generation has also grown rapidly. Since the photovoltaic power generation power is significantly affected by external factors such as light, the photovoltaic system needs to be adjusted when it is connected to the grid. The adjustment is divided into two aspects, one is the power generation power of the photovoltaic system itself, and the other is the matching of the photovoltaic system output with the grid parameters. In the prior art, these two aspects of the adjustment function are usually completed by different controllers respectively, but because there is a mutual correlation between the input and output of the photovoltaic system, this leads to the easy generation of system parameter oscillation when the photovoltaic system is adjusted, affecting the photovoltaic grid-connected power and the stability of the grid. Summary of the invention

[0003] The technical problem to be solved by the present invention is to provide a photovoltaic power station grid-connected inverter control method, which can solve the deficiencies of the prior art and improve the photovoltaic power generation grid-connected control accuracy.

[0004] In order to solve the above technical problems, the technical solutions adopted by the present invention are as follows.

[0005] A photovoltaic power station grid-connected inverter control method comprises the following steps: A. Measure the DC voltage and DC current of the current photovoltaic power station and calculate the current output power of the photovoltaic power station; B. Change the MPPT duty cycle, calculate the output power change of the photovoltaic power station, and determine the adjustment direction of the MPPT duty cycle according to the positive and negative change direction of the photovoltaic power station output power; C. Adjust the MPPT duty cycle according to the adjustment direction of the MPPT duty cycle obtained in step B. Use the random forest model to optimize the inverter operating parameters before each adjustment.

[0006] Preferably, the establishment of the random forest model includes the generation of a training data set and the construction of a decision tree.

[0007] Preferably, the generation of the training data set includes the following steps: Five types of data, namely, inverter DC side current, inverter DC side voltage, inverter AC side current, inverter AC side voltage and inverter reactive power, are used as training data sets. A sliding window with a step size of Δt is used to traverse each type of data to obtain the first-order dynamic reference data of each type of data. According to the linear relationship of the first-order dynamic reference data, the above five types of data are classified and formed into training data subsets. Each training data subset includes five types of data that are linearly related to each other corresponding to the first-order dynamic reference data.

[0008] Preferably, the construction of the decision tree includes the following steps: The number of decision trees is determined according to the number of training data subsets, so that the training data subsets correspond to the decision trees one by one. The maximum depth of recursive splitting of the decision tree is equal to the number of extreme points of the first-order dynamic reference data fitting curve corresponding to the training data subsets. The extreme points of the first-order dynamic reference data fitting curve corresponding to the training data subsets are used as the best segmentation features. The decision tree is recursively split using the training data subsets to obtain the specific structure of the decision tree.

[0009] Preferably, all decision trees are traversed, each decision tree is pruned in turn, and the pruned topological structures are compared. If the same topological structure exists and the topological structure has never appeared in other decision trees, any decision tree with the topological structure pruned is selected for node restoration, and the topological structure is restored to the selected decision tree.

[0010] Preferably, a feedback node is set on each decision tree, the feedback nodes correspond to the input nodes one by one and are directly connected in the topological structure, and the feedback node on each decision tree is connected to a corresponding feedback model.

[0011] Preferably, the feedback model includes a multidimensional frequency domain coordinate system, the dimension of the multidimensional frequency domain coordinate system is the same as the number of feedback nodes on the corresponding decision tree; the data related to the corresponding decision tree is input into the feedback model and mapped to the multidimensional frequency domain coordinate system, a plane wave function is formed in each dimension of the multidimensional frequency domain coordinate system, the standard reference data in the feedback model is mapped to the multidimensional frequency domain coordinate system, a standard reference plane wave function is formed in each dimension of the multidimensional frequency domain coordinate system, and the feedback parameter output by the feedback model to the feedback node is determined according to the average deviation of the spatial distribution of the plane wave function and the standard reference plane wave function.

[0012] Preferably, the spatial distribution average deviation is used as the output quantity, and the feedback parameter is calculated through the PID transfer function.

[0013] The beneficial effect brought about by adopting the above technical solution is that the present invention improves the accuracy of synchronous optimization of input and output parameters of photovoltaic inverters and reduces system oscillations by improving the existing random forest model structure and usage method. DETAILED DESCRIPTION

[0014] Due to the particularity of the photovoltaic system, its output power is related to the light intensity and load parameters, so the MPPT algorithm is generally used to track the maximum output power of the photovoltaic system. However, when tracking the maximum output power, the output power, frequency and voltage of the entire photovoltaic system will fluctuate, and these fluctuations will directly affect the stability of the power grid. It is precisely because of the above reasons that the existing photovoltaic inverter control algorithm will cause obvious oscillations in the system during adjustment.

[0015] After research by our R&D team, we found that the main reason for this problem is that the commonly used control algorithms (such as random forest, neural network, PDI, etc.) cannot effectively respond quickly and accurately to the bilateral variables of input and output. Therefore, we optimized the existing random forest algorithm and obtained the following technical solution.

[0016] The specific steps include: A. Measure the DC voltage and DC current of the current photovoltaic power station and calculate the current output power of the photovoltaic power station; B. Change the MPPT duty cycle, calculate the output power change of the photovoltaic power station, and determine the adjustment direction of the MPPT duty cycle according to the positive and negative change direction of the photovoltaic power station output power; C. Adjust the MPPT duty cycle according to the adjustment direction of the MPPT duty cycle obtained in step B. Use the random forest model to optimize the inverter operating parameters before each adjustment.

[0017] The establishment of the random forest model includes the following steps: Five types of data, including the inverter DC side current, inverter DC side voltage, inverter AC side current, inverter AC side voltage, and inverter reactive power, are used as training data sets. A sliding window with a step size of Δt (the value of Δt is determined comprehensively based on the hardware system computing power and control accuracy. The higher the hardware system computing power and the higher the control accuracy requirements, the smaller the value of Δt) is used to traverse each type of data, and the first-order dynamic reference data of each type of data is obtained. The above five types of data are classified according to the linear relationship of the first-order dynamic reference data and form training data subsets. Each training data subset includes five types of data that are linearly related to each other corresponding to the first-order dynamic reference data. The first-order dynamic reference data can well reflect the dynamic change relationship between different types of data, thereby improving the generalization ability of each decision tree.

[0018] The number of decision trees is determined according to the number of training data subsets, so that the training data subsets correspond to the decision trees one by one. The maximum depth of the recursive splitting of the decision tree is equal to the number of extreme points of the first-order dynamic reference data fitting curve corresponding to the training data subset. The extreme points of the first-order dynamic reference data fitting curve corresponding to the training data subset are used as the best segmentation features. The decision tree is recursively split using the training data subset to obtain the specific structure of the decision tree. We abandon the traditional method of generating the best segmentation features. By using the extreme points of the first-order dynamic reference data fitting curve as the best segmentation features, we can optimize the node topology of the decision tree and prevent the decision tree from overfitting.

[0019] Traverse all decision trees, prune each decision tree in turn, compare the pruned topological structures, and if there is an identical topological structure, and the topological structure has never appeared in other decision trees, select any decision tree that has pruned the topological structure to restore the nodes, and restore the topological structure to the selected decision tree. Here, we did not simply adopt the traditional pruning strategy, but made a secondary correction to the decision tree after using the traditional pruning strategy. Because we use the extreme points of the first-order dynamic reference data fitting curve as the best segmentation feature of the decision tree, this leads to extremely high data sensitivity of the decision tree nodes. If a local topological structure is deleted on a large scale, it will cause instability in the entire model. So we solve this problem through secondary correction.

[0020] Feedback nodes are set on each decision tree. The feedback nodes correspond to the input nodes one by one and are directly connected in the topological structure. The feedback nodes on each decision tree are connected to a corresponding feedback model.

[0021] The feedback model includes a multidimensional frequency domain coordinate system, the dimension of which is the same as the number of feedback nodes on the corresponding decision tree; the data related to the corresponding decision tree is input into the feedback model and mapped to the multidimensional frequency domain coordinate system, a plane wave function is formed in each dimension of the multidimensional frequency domain coordinate system, the standard reference data in the feedback model is mapped to the multidimensional frequency domain coordinate system, a standard reference plane wave function is formed in each dimension of the multidimensional frequency domain coordinate system, and the feedback parameter output by the feedback model to the feedback node is determined according to the average deviation of the spatial distribution of the plane wave function and the standard reference plane wave function. The average deviation of the spatial distribution is used as the output, and the feedback parameter is calculated through the PID transfer function.

[0022] If the existing decision tree is connected to feedback data, the robustness of the decision tree against disturbances will be reduced, thereby increasing its instability. Since we have optimized the structure of the decision tree, the nodes of the decision tree can perform feature gains of different amplitudes on the original data and feedback data respectively, thereby solving the technical problem of the existing decision tree connecting to feedback data. The feedback model processes data in the multi-dimensional frequency domain, which can decouple different series of data, reduce nonlinear problems, and enable the decision tree to respond quickly and accurately to feedback data. By introducing feedback control in the decision tree, the matching degree between the output of the photovoltaic inverter and the grid demand is effectively improved.

[0023] The above shows and describes the basic principles and main features of the present invention and the advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments. The above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A photovoltaic power station grid-connected inverter control method, characterized in that The following steps are involved: A. Measure the DC voltage and DC current of the current photovoltaic power station and calculate the current output power of the photovoltaic power station; B. Change the MPPT duty cycle, calculate the output power change of the photovoltaic power station, and determine the adjustment direction of the MPPT duty cycle according to the positive and negative change direction of the photovoltaic power station output power; C. Adjust the MPPT duty cycle according to the adjustment direction of the MPPT duty cycle obtained in step B. Use the random forest model to optimize the inverter operating parameters before each adjustment.

2. The photovoltaic power station grid-connected inverter control method according to claim 1, characterized in that: The establishment of the random forest model includes the generation of training data sets and the construction of decision trees.

3. The photovoltaic power station grid-connected inverter control method according to claim 2, characterized in that: The generation of training dataset includes the following steps: Five types of data, namely, inverter DC side current, inverter DC side voltage, inverter AC side current, inverter AC side voltage and inverter reactive power, are used as training data sets. A sliding window with a step size of Δt is used to traverse each type of data to obtain the first-order dynamic reference data of each type of data. According to the linear relationship of the first-order dynamic reference data, the above five types of data are classified and formed into training data subsets. Each training data subset includes five types of data that are linearly related to each other corresponding to the first-order dynamic reference data.

4. The photovoltaic power station grid-connected inverter control method according to claim 3, characterized in that: Construction of decision tree The following steps are included: The number of decision trees is determined according to the number of training data subsets, so that the training data subsets correspond to the decision trees one by one. The maximum depth of recursive splitting of the decision tree is equal to the number of extreme points of the first-order dynamic reference data fitting curve corresponding to the training data subsets. The extreme points of the first-order dynamic reference data fitting curve corresponding to the training data subsets are used as the best segmentation features. The decision tree is recursively split using the training data subsets to obtain the specific structure of the decision tree.

5. The photovoltaic power station grid-connected inverter control method according to claim 4, characterized in that: Traverse all decision trees, prune each decision tree in turn, compare the pruned topological structures, and if there is an identical topological structure that has never appeared in other decision trees, select any decision tree that has pruned the topological structure to restore the nodes, and restore the topological structure to the selected decision tree.

6. The photovoltaic power station grid-connected inverter control method according to claim 5, characterized in that: Feedback nodes are set on each decision tree. The feedback nodes correspond to the input nodes one by one and are directly connected in the topological structure. The feedback nodes on each decision tree are connected to a corresponding feedback model.

7. The photovoltaic power station grid-connected inverter control method according to claim 6, characterized in that: The feedback model includes a multidimensional frequency domain coordinate system, and the dimension of the multidimensional frequency domain coordinate system is the same as the number of feedback nodes on the corresponding decision tree; the data related to the corresponding decision tree is input into the feedback model and mapped to the multidimensional frequency domain coordinate system, and a plane wave function is formed in each dimension of the multidimensional frequency domain coordinate system; the standard reference data in the feedback model is mapped to the multidimensional frequency domain coordinate system, and a standard reference plane wave function is formed in each dimension of the multidimensional frequency domain coordinate system; the feedback parameter output by the feedback model to the feedback node is determined according to the average deviation of the spatial distribution of the plane wave function and the standard reference plane wave function.

8. The photovoltaic power station grid-connected inverter control method according to claim 7, characterized in that: The spatial distribution average deviation is taken as the output and the feedback parameter is calculated through the PID transfer function.