Dynamic reactive power compensation device for distributed power generation and distribution network and control method of dynamic reactive power compensation device

Abstract

The invention provides a distributed power generation distribution network dynamic reactive power compensation device and a control method thereof. The device comprises a data acquisition unit, a data classification unit, a model establishment and training unit, a model fusion unit, a control instruction generation unit and a reactive power compensation execution unit. Based on the neural network model, the real-time characteristic value of the corresponding comprehensive neural network model is firstly collected, the corresponding power fluctuation value is predicted, finally the power fluctuation value in the duration is calculated, the reactive power target value needing to be compensated is calculated, and the reactive power target value is compensated by adjusting the amplitude and phase of the output voltage on the alternating current side of the bridge circuit or directly controlling the current on the alternating current side. Therefore, the reactive current output by the SVG can accurately track the set reactive current target value. A decision tree algorithm is adopted to classify distributed generation (DG) and corresponding parameter data, then artificial neural networks are respectively trained for reactive power fluctuation prediction, a comprehensive neural network model is trained through model nesting, and the accuracy of a predicted value and the precision of reactive power compensation are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power distribution networks, and in particular to a dynamic reactive power compensation device for a distributed power generation and distribution network and a control method thereof. Background Art

[0002] With the rapid development of distributed generation technology, the proportion of distributed power sources such as solar energy and wind energy in the distribution network has been increasing. However, distributed generation has the characteristics of intermittency, volatility and randomness, which makes the voltage fluctuation of the distribution network more severe and the power factor lower, seriously affecting the power quality and stable operation of the distribution network.

[0003] When dealing with changes in reactive power demand in the distribution network after the access of distributed generation, traditional reactive power compensation devices have defects such as slow response speed, low compensation accuracy, and inability to make effective adjustments in real time according to the dynamic working conditions of the distribution network, which leads to problems such as voltage fluctuations and reduced power factor in the distribution network, affecting the power quality and operational stability of the distribution network. For example, fixed capacitor compensation devices cannot flexibly adapt to real-time changes in load and distributed generation output; traditional dynamic reactive power compensation devices based on simple control algorithms are difficult to accurately calculate the required reactive power compensation in a complex distributed generation distribution network environment.

[0004] Therefore, there is an urgent need for a dynamic reactive power compensation device and a corresponding control method that can adapt to the characteristics of distributed power generation and distribution networks to solve the above problems. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a dynamic reactive power compensation device for a distributed power generation distribution network and a control method thereof, so as to solve the problems of slow response speed, low compensation accuracy and inability to make effective adjustments in real time according to the dynamic working conditions of the distribution network when the existing reactive power compensation device copes with the changes in reactive power demand of the distribution network after the distributed power generation is connected.

[0006] To solve the above technical problems, the embodiments of the present invention provide the following technical solutions:

[0007] A dynamic reactive power compensation device for a distributed power generation and distribution network, comprising:

[0008] A data acquisition unit is used to collect parameter data of distributed generation DG, not limited to voltage, current and power;

[0009] A data classification unit is used to classify distributed generation DG and its corresponding parameter data, and to construct a neural network training data set corresponding to each type of distributed generation DG;

[0010] A model building and training unit is used to build a neural network training model corresponding to each type of distributed generation DG, and train artificial neural network 1, artificial neural network 2, ... artificial neural network N based on a neural network training data set for each type of distributed generation DG;

[0011] The model fusion unit integrates and trains a comprehensive neural network model based on artificial neural network 1, artificial neural network 2 ... artificial neural network N, connects the trained comprehensive neural network model to the controller, and calculates the SVG control strategy of the static VAR generator;

[0012] A control instruction generating unit, used for performing linearization processing on the static VAR generator SVG control strategy to form a reactive compensation control amount of the SVG;

[0013] The reactive power compensation execution unit is used to perform reactive power compensation on the distribution network by adjusting the amplitude and phase of the output voltage on the AC side of the bridge circuit or directly controlling the AC side current according to the reactive power compensation control amount of the SVG.

[0014] The present invention also provides a distributed power generation and distribution network dynamic reactive power compensation control method, comprising the following steps:

[0015] The collection of distributed generation DG parameter data is not limited to voltage, current, and power;

[0016] Classify the distributed generation DG and its corresponding parameter data, and construct a neural network training data set corresponding to each type of distributed generation DG;

[0017] Establishing a neural network training model corresponding to each type of distributed generation DG, and training artificial neural network 1, artificial neural network 2, ... artificial neural network N based on each type of distributed generation DG neural network training data set;

[0018] Based on artificial neural network 1, artificial neural network 2 ... artificial neural network N, a comprehensive neural network model is integrated and trained, and the trained comprehensive neural network model is connected to the controller to calculate the SVG control strategy of the static VAR generator;

[0019] Linearizing the static VAR generator SVG control strategy to form a reactive compensation control quantity of the SVG;

[0020] According to the reactive compensation control amount of SVG, reactive compensation is performed on the distribution network by adjusting the amplitude and phase of the output voltage on the AC side of the bridge circuit, or directly controlling the AC side current.

[0021] Optionally, the classifying of distributed generation DG and its corresponding parameter data specifically includes:

[0022] Select features of the collected data, including energy source features, power generation scale features, and grid access features;

[0023] Data preprocessing: clean the data after feature selection and perform One-Hot encoding on the cleaned data;

[0024] Construct a decision tree, select a splitting criterion, calculate the information gain of each feature, and then select the feature with the largest information gain as the splitting feature of the current node, dividing the data set into different subsets to achieve parameter data classification.

[0025] Optionally, the establishing of a neural network training model corresponding to each type of distributed generation DG is specifically as follows:

[0026] Select different types of neural networks for different distributed generation classification tasks and determine the number of nodes in the input layer, hidden layer and output layer of the neural network;

[0027] Initialize the network parameters of each layer of the neural network, including weights and biases;

[0028] The cross entropy loss function is used as the error detection function between the characteristic value and the predicted value of distributed generation DG, where the characteristic value is the normalized value of various parameters of distributed generation DG that cause its output power fluctuation, and the predicted value is the normalized value of the predicted fluctuation of distributed generation DG output power. At the same time, the optimization algorithm is selected to adjust the weight and bias of the network.

[0029] Optionally, the training of artificial neural network 1, artificial neural network 2, ... artificial neural network N based on each type of distributed generation DG neural network training data set is specifically:

[0030] The corresponding distributed generation DG neural network training data set is divided into a training set and a data set, and the divided training set data is encoded and normalized;

[0031] The normalized training set data is input into the neural network through the input layer nodes in sequence. The input data is propagated and calculated in the neural network in the order from the input layer to the hidden layer and then to the output layer to obtain the output value.

[0032] According to the output results of the output layer and the actual categories in the training set data, the loss value is calculated using the selected loss function;

[0033] Using the calculated loss value, the error signal is transmitted backward from the output layer to the input layer through the back-propagation algorithm to adjust the weights and biases of the network;

[0034] Update the network weights and biases using the selected optimization algorithm based on the weight and bias adjustments obtained through backpropagation;

[0035] Repeat the above steps of forward propagation, loss calculation, back propagation, parameter update, etc. until the predetermined number of training times is reached or the stopping condition is met. The stopping condition is that the loss value reaches a certain minimum value or the number of training times reaches a certain maximum value.

[0036] Use the validation set to evaluate the trained part of the network, input the validation set data into the network, obtain the output result according to the forward propagation step, and then calculate the loss value and accuracy of the validation set based on the output result and the actual category in the validation set data. If the loss value and accuracy meet the conditions, stop training.

[0037] Optionally, the artificial neural network 1, artificial neural network 2 ... artificial neural network N are based on the fusion training of the comprehensive neural network model, and the training of the comprehensive neural network model is specifically realized by nesting of neural network models, wherein the nesting methods include series nesting, parallel nesting or mixed nesting, and the nested combinations are classified according to the principle of proximity of the distribution network line topology relationship, wherein the comprehensive neural network model is one or more.

[0038] Optionally, the calculation of the static VAR generator SVG control strategy is specifically:

[0039] At the time t-Δt, the real-time characteristic values ​​of the corresponding comprehensive neural network model are collected to predict the corresponding power fluctuation value;

[0040] At time t, the voltage and current signals of the power grid are detected, and the voltage and current in the three-phase stationary coordinate system are converted to the synchronous rotating coordinate system using Clarke transformation to obtain the instantaneous reactive current component;

[0041] According to the power fluctuation value and the instantaneous reactive current component, the reactive power target value required to be compensated for the power fluctuation value within the time period Δt is calculated. By adjusting the amplitude and phase of the output voltage on the AC side of the bridge circuit, or directly controlling its AC side current, the reactive current output by the SVG can accurately track the set reactive current target value.

[0042] Optionally, the linearization processing of the static VAR generator SVG control strategy is specifically performed as follows:

[0043] (1) Establish an equivalent model and regard the static VAR generator SVG as a device that can output variable reactive current. Its output current i s With control voltage u c The functional relationship between s =f(u c );

[0044] (2) Taylor series expansion, at a typical working point u c0 Nearby pair function f(uc ) is expanded by Taylor series to obtain:

[0045] f(u c )≈f(u c )≈f(u c0 )+f′(u c0 )(u c -u c0 )

[0046] Among them, f′(u c0 ) is the function f(u c ) in u c0 First derivative of a point;

[0047] (3) Get the linearized model, let i S0 =f(u c0 ), then the linearized model is expressed as:

[0048] i out ≈i S0 +k(u c -u c0 )

[0049] Where k = f′(u c0 ) is the linearized gain function, i out is the current output control quantity after linearization.

[0050] Optionally, the optimization algorithm includes a gradient descent algorithm, an Adagrad algorithm, or an Adam algorithm.

[0051] Optionally, the nested combination classification is nested according to the principle of proximity of the topological relationship of the power distribution network lines, specifically:

[0052] The distribution network is equivalent to several main feeders, and the equivalent main branch calculation is performed;

[0053] Conduct breadth-first search on the equivalent distribution network branches to determine candidate compensation nodes;

[0054] Screen candidate compensation nodes and determine nodes to be compensated;

[0055] The nodes to be compensated are combined according to the principle of proximity in topological relationship, and the artificial neural networks corresponding to the combined distribution network branches are nested.

[0056] The beneficial effects of the above technical solution of the present invention are as follows:

[0057] 1. Due to the complexity and rapidity of reactive power fluctuations in the power grid, in order to get closer to the reactive compensation control target, the control strategy of the present invention first collects the real-time characteristic values ​​of the corresponding comprehensive neural network model based on the neural network model, predicts the corresponding power fluctuation value, and detects the voltage and current signals of the power grid within a controllable time interval Δt, and finally calculates the reactive power target value required for compensation based on the power fluctuation value within the Δt time length. By adjusting the amplitude and phase of the output voltage on the AC side of the bridge circuit, or directly controlling its AC side current, the reactive current output by the SVG can accurately track the set reactive current target value.

[0058] 2. Due to the various types of power sources in the distributed power generation distribution network and the complex grid topology, the present invention adopts a decision tree algorithm to classify the distributed power generation DG and its corresponding parameter data, and then trains artificial neural networks to predict reactive power fluctuations, and trains a comprehensive neural network model through neural network model nesting, wherein the nesting methods include series nesting, parallel nesting or mixed nesting, and the nested combination classification is nested according to the principle of proximity of the distribution network line topology relationship, which further improves the accuracy of the prediction value and the accuracy of reactive power compensation. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 This is a schematic diagram showing the relationship between the modules of the dynamic reactive power compensation device for a distributed power generation and distribution network of the present invention;

[0060] Figure 2 It is a flow chart of the dynamic reactive power compensation method of the distributed power generation and distribution network of the present invention;

[0061] Figure 3 It is a principle block diagram of the reactive power compensation control system of the present invention. DETAILED DESCRIPTION

[0062] In order to make the technical problems, technical solutions and advantages to be solved by the present invention more clear, a detailed description will be given below with reference to the accompanying drawings and specific embodiments.

[0063] like Figure 1 As shown, the present invention provides a dynamic reactive power compensation device for a distributed power generation and distribution network, comprising:

[0064] The data acquisition unit 101 is used to collect parameter data of distributed generation DG, including but not limited to voltage, current and power;

[0065] The data classification unit 102 is used to classify the distributed generation DG and its corresponding parameter data, and construct a neural network training data set corresponding to each type of distributed generation DG;

[0066] A model building and training unit 103 is used to build a neural network training model corresponding to each type of distributed generation DG, and train artificial neural network 1, artificial neural network 2, ... artificial neural network N based on a neural network training data set for each type of distributed generation DG;

[0067] The model fusion unit 104 integrates and trains a comprehensive neural network model based on the artificial neural network 1, the artificial neural network 2 ... the artificial neural network N, connects the trained comprehensive neural network model to the controller, and calculates the control strategy of the static VAR generator SVG;

[0068] The control instruction generating unit 105 is used to perform linearization processing on the static VAR generator SVG control strategy to form the reactive power compensation control quantity of the SVG;

[0069] The reactive power compensation execution unit 106 is used to perform reactive power compensation on the distribution network by adjusting the amplitude and phase of the AC side output voltage of the bridge circuit or directly controlling the AC side current according to the reactive power compensation control amount of the SVG.

[0070] like Figure 2 As shown, the present invention also provides a distributed power generation distribution network dynamic reactive power compensation control method, comprising the following steps:

[0071] S1. Collect parameter data of distributed generation DG, not limited to voltage, current and power;

[0072] S2, classifying the distributed generation DG and its corresponding parameter data, and constructing a neural network training data set corresponding to each type of distributed generation DG;

[0073] S3, establishing a neural network training model corresponding to each type of distributed generation DG, and training artificial neural network 1, artificial neural network 2, ... artificial neural network N based on each type of distributed generation DG neural network training data set;

[0074] S4, based on artificial neural network 1, artificial neural network 2 ... artificial neural network N, a comprehensive neural network model is integrated and trained, the trained comprehensive neural network model is connected to the controller, and the control strategy of the static VAR generator SVG is calculated;

[0075] S5, linearizing the control strategy of the static VAR generator SVG to form a reactive compensation control quantity of the SVG;

[0076] S6. According to the reactive compensation control amount of SVG, reactive compensation is performed on the distribution network by adjusting the amplitude and phase of the output voltage on the AC side of the bridge circuit, or directly controlling the AC side current.

[0077] In this embodiment, the distributed generation DG and its corresponding parameter data are classified, specifically including:

[0078] Select features of the collected data, including energy source features (such as solar energy, wind energy, biomass energy, hydropower, geothermal energy, etc.), power generation scale features (specific capacity values ​​in kilowatts), and grid access methods (independent operation or grid-connected operation);

[0079] Data preprocessing, data cleaning after feature selection, and One-Hot encoding of the cleaned data. Data cleaning includes: checking whether there are erroneous values, missing values, etc. in the collected data. For erroneous values, such as the power generation scale values ​​that are obviously inconsistent with the actual situation, they need to be corrected or deleted. For missing values, there are many methods to deal with them, such as filling with the mean or median (applicable to numerical features), or filling with the mode (applicable to discrete features). One-Hot encoding method converts each category into a binary vector, in which only one element is 1 and the rest are 0. For example, for the five energy sources of solar energy, wind energy, biomass energy, hydropower, and geothermal energy, after One-Hot encoding, the corresponding code for solar energy is [1, 0, 0, 0, 0], the corresponding code for wind energy is [0, 1, 0, 0, 0], and so on.

[0080] To build a decision tree, we first select the splitting criterion. The decision tree construction process is to continuously select appropriate features for splitting to form a tree structure. Commonly used splitting criteria include information gain, information gain ratio, Gini coefficient, etc. Taking information gain as an example, it measures the degree to which the uncertainty of the data set is reduced after splitting using a certain feature. When building a distributed power generation classification decision tree, we will calculate the information gain of each feature, and then select the feature with the largest information gain as the splitting feature of the current node. For example, at the root node, we will calculate the information gain of features such as energy source, power generation scale, and grid access method, and then select the feature with the largest information gain, such as energy source, as the splitting feature of the root node.

[0081] Secondly, the information gain of each feature is calculated, and then the feature with the largest information gain is selected as the split feature of the current node. The information gain of the remaining features (power generation scale, grid access method, etc.) is calculated, and the feature with the largest information gain is selected as the split feature of the next node. The subset division continues, and this process is repeated recursively until the stopping condition is met, thereby dividing the data set into different subsets to achieve parameter data classification.

[0082] In this embodiment, a neural network training model corresponding to each type of distributed generation DG is established, specifically:

[0083] Different types of neural networks are selected for different distributed generation classification tasks. Different types of neural networks can be selected according to the characteristics of distributed generation classification tasks. Common ones include multi-layer perceptron (MLP), convolutional neural network (CNN), recurrent neural network (RNN) and its variants. Determine the number of nodes in the input layer, hidden layer and output layer of the neural network. Among them: the number of nodes in the input layer depends on the number of features selected for classification. When constructing the distributed generation training data set, the features to be used have been determined, such as energy source (there may be multiple nodes after encoding), power generation scale, grid access method, geographical location (latitude and longitude may correspond to different nodes respectively), operating time, maintenance cost, power generation efficiency, local climate conditions, etc. The encoding or value of these features is mapped one by one to the nodes of the input layer. For example, if there are 10 different features for classification, then 10 nodes are set in the input layer.

[0084] The number of hidden layers is generally determined through experiments and experience. For relatively simple distributed generation classification tasks, you can usually try to set 1-3 hidden layers first. Too few hidden layers may not be able to fully learn the complex relationships in the data, while too many hidden layers may lead to overfitting. Number of nodes: The number of nodes in each hidden layer also needs to be determined based on experience and experiments. A common method is to start with a relatively small number of nodes, such as setting each hidden layer to 2-3 times the number of nodes in the input layer, and then gradually adjust it according to the training effect. For example, if the input layer has 10 nodes, then the first hidden layer can try to set 20-30 nodes.

[0085] The number of nodes in the output layer depends on the number of categories to be classified. If the classification is based on energy sources, and there are five types of energy sources, such as solar energy, wind energy, biomass energy, hydropower, and geothermal energy, then the output layer will have five nodes, one for each energy source category. Similarly, if the classification is based on other methods, such as the power generation scale is divided into three categories: small, medium, and large, then the output layer will have three nodes.

[0086] Initialize the network parameters of each layer of the neural network, including weights and biases. In a neural network, the connections between neurons have corresponding weights, and each neuron also has a bias term. The initial values ​​of these weights and biases will affect the training effect of the neural network. Common initialization methods include random initialization, Xavier initialization, He initialization, etc. For the distributed power generation neural network model, the Xavier initialization method is used. It determines the initial value of the weight based on the number of input and output neurons, which can avoid the problem of gradient disappearance or gradient explosion in the early stage of training to a certain extent.

[0087] The present invention uses the cross entropy loss function as the error detection function between the characteristic value and the predicted value of the distributed generation DG, wherein the characteristic value is the normalized value of various parameters of the distributed generation DG that cause its output power fluctuation, and the predicted value is the normalized value of the predicted fluctuation of the distributed generation DG output power, and the optimization algorithm is selected to adjust the weight and bias of the network. The cross entropy loss function measures the degree of difference between the predicted category and the actual category. When the predicted category completely matches the actual category, the cross entropy loss is 0, and when the predicted category is significantly different from the actual category, the cross entropy loss increases. Among them, the output power fluctuation of distributed generation (DG) may be affected by the following parameters: natural factor related parameters, such as light intensity, wind speed, temperature; equipment related parameters, such as equipment aging, equipment failure; energy storage system related parameters: energy storage capacity, charging and discharging efficiency, energy storage system response time. These influencing factors are input into the neural network as characteristic values ​​to predict the output power fluctuation of distributed generation (DG).

[0088] In this embodiment, artificial neural network 1, artificial neural network 2, ... artificial neural network N are trained based on each type of distributed generation DG neural network training data set, specifically:

[0089] The corresponding distributed generation DG neural network training data set is divided into a training set and a data set, and the divided training set data is encoded and normalized. Through normalization, the data with different features are mapped to a relatively small and uniform value range (such as [0, 1] or [-1, 1], etc.), so that when calculating the gradient, the gradient update step size in different feature directions is relatively balanced, which enhances the generalization ability of the neural network and improves the numerical stability of the neural network.

[0090] The normalized training set data is input into the neural network through the input layer nodes in sequence. The input data is propagated and calculated in the neural network in the order from the input layer to the hidden layer and then to the output layer to obtain the output value.

[0091] The selected loss function is used to calculate the loss value based on the output result of the output layer and the actual category in the training set data. For example, if the output result of the output layer is to predict that a distributed power generation system belongs to the solar energy category, and the actual category is the wind energy category, then the difference between the two is calculated based on the cross entropy loss function, that is, the loss value.

[0092] Using the calculated loss value, the error signal is transmitted backward from the output layer to the input layer through the back propagation algorithm to adjust the weights and biases of the network. For example, starting from the neurons in the output layer, the amount of adjustment of their own weights and biases is calculated based on the error signal and their own input data, and then the error signal is transmitted to the neurons in the hidden layer. The neurons in the hidden layer also calculate the amount of adjustment of their own weights and biases based on the error signal and their own input data, and this is repeated until the neurons in the input layer are reached.

[0093] According to the weight and bias adjustment amount obtained by back propagation, the weight and bias of the network are updated using the selected optimization algorithm; for example, the Adam optimization algorithm updates the weight and bias of the network according to the adjustment amount and its own algorithm rules, so that the performance of the network is improved. Among them, the optimization algorithms include the gradient descent algorithm, the Adagrad algorithm, and the Adam algorithm.

[0094] Repeat the above steps of forward propagation, loss calculation, back propagation, parameter update, etc. until the predetermined number of training times is reached or the stopping condition is met. The stopping condition is that the loss value reaches a certain set minimum value, or the number of training times reaches a certain set maximum value.

[0095] Use the validation set to evaluate the trained part of the network. Input the validation set data into the network and get the output result according to the forward propagation step. Then calculate the loss value and accuracy of the validation set based on the output result and the actual category in the validation set data. If the loss value and accuracy meet the conditions, stop training.

[0096] In this embodiment, based on artificial neural network 1, artificial neural network 2 ... artificial neural network N, a comprehensive neural network model is integrated and trained, and the comprehensive neural network model is trained specifically by neural network model nesting, wherein the nesting method includes series nesting, parallel nesting or mixed nesting, and the nested combination classification is nested according to the principle of proximity of the topological relationship of the distribution network line, wherein the comprehensive neural network model is one or more. For example, in series nesting, a recurrent neural network (RNN) is first constructed, such as a long short-term memory network (LSTM), which is used to process the time series of historical power generation data and related meteorological data (such as wind speed, light intensity, etc.). LSTM can capture the changing rules of these data in the time series and extract the time series features related to power generation. Then, the features extracted by LSTM are passed as input to a multi-layer perceptron (MLP). MLP further processes these features, and finally outputs the predicted power generation value through the calculation of its hidden layer and output layer. In this way, through the series nesting of RNN and MLP, the processing ability of RNN for sequence data is first utilized, and then the general classification and computing ability of MLP are used to realize the prediction of distributed power generation.

[0097] The nested combination classification is nested according to the principle of proximity of the topological relationship of the distribution network line. Specifically, the distribution network is equivalent to several main feeders, and the equivalent main branch calculation is performed; the equivalent distribution network branches are subjected to breadth-first search to determine the candidate compensation nodes; the candidate compensation nodes are screened to determine the nodes to be compensated; the nodes to be compensated are combined according to the principle of proximity of the topological relationship, and the artificial neural networks corresponding to the combined distribution network branches are nested. For distributed power lines, this method can be used to first calculate the reactive compensation position, install reactive compensation devices at reasonable locations to reduce the active power loss of the line, and maintain the voltage of nearby lines within a reasonable range during photovoltaic power generation. At the same time, the artificial neural networks corresponding to the combined distribution network branches are nested according to the principle of proximity of the topological relationship to obtain a nested prediction model. The reason for adopting the principle of proximity of the topological relationship is to maximize the prediction accuracy of the nested neural network, because the reactive power change-influencing characteristic values ​​of the power lines with close topological relationships are often similar, so the results of the various prediction models within the range of close topological relationships are more correlated, and the prediction results of each prediction model are further nested for prediction, and the accuracy is naturally better.

[0098] In this embodiment, the static VAR generator SVG control strategy is calculated as follows:

[0099] At the time t-Δt, the real-time characteristic values ​​of the corresponding comprehensive neural network model are collected to predict the corresponding power fluctuation value;

[0100] At time t, the voltage and current signals of the power grid are detected, and the voltage and current in the three-phase stationary coordinate system are converted to the synchronous rotating coordinate system using Clarke transformation to obtain the instantaneous reactive current component;

[0101] According to the power fluctuation value and the instantaneous reactive current component, the reactive power target value required to be compensated for the power fluctuation value within the time period Δt is calculated. By adjusting the amplitude and phase of the output voltage on the AC side of the bridge circuit, or directly controlling its AC side current, the reactive current output by the SVG can accurately track the set reactive current target value.

[0102] In this embodiment, the control strategy of the static VAR generator SVG is linearized, specifically:

[0103] (1) Establish an equivalent model and regard the static VAR generator SVG as a device that can output variable reactive current. Its output current i s With control voltage u c The functional relationship between s =f(u c );

[0104] (2) Taylor series expansion, at a typical working point uc0 Nearby pair function f(u c ) is expanded by Taylor series to obtain:

[0105] f(u c )≈f(u c )≈f(u c0 )+f′(u c0 )(u c -u c0 )

[0106] Among them, f′(u c0 ) is the function f(u c ) in u c0 First derivative of a point;

[0107] (3) Get the linearized model, let i S0 =f(u c0 ), then the linearized model is expressed as:

[0108] i out ≈i S0 +k(u c -u c0 )

[0109] Where k = f′(u c0 ) is the linearized gain function, i out is the current output control quantity after linearization.

[0110] When controlling reactive power compensation according to the changes in reactive power demand of the power grid, it is difficult to accurately achieve the expected reactive power output value due to the nonlinear relationship. After linearization, a more accurate control algorithm can be designed based on the linearized model, which can more accurately adjust reactive power according to demand and improve control accuracy. At the same time, the response speed of the nonlinear system may be affected when facing grid voltage fluctuations, sudden changes in reactive power demand, etc. For example, when the reactive power demand of the power grid suddenly increases, the nonlinear control may not be able to quickly make appropriate switching adjustments to meet the demand. After linearization, the control strategy designed using the linearized model can make the static VAR generator SVG respond to external changes more quickly and adjust the reactive power output in time to adapt to the dynamic changes of the power grid.

[0111] After the reactive power compensation control quantity of SVG is formed by linear processing in the above manner, reactive power compensation is performed on the distribution network by adjusting the amplitude and phase of the output voltage on the AC side of the bridge circuit or directly controlling the current on the AC side according to the reactive power compensation control quantity of SVG. For example, when SVG is used for compensation, its working principle is to calculate the reactive power Q of the power grid branch according to the reactive power Q of the power grid branch. s The reactive power Q generated by the SVG device is continuously controlled and changed according to the changes in svc, making it equal to the reactive power Q of the distributed generation d The sum is a fixed value or limited to a certain range to ensure the stability of reactive power on the system side. Ideally, there is Q s =Q svc +Q d .

[0112] like Figure 3 As shown, it is a block diagram of the reactive power compensation control system of the present invention. PT1 and CT1 are the voltage sensor and current sensor of the power grid branch, respectively. PT2 and CT2 are the voltage sensor and current sensor of the distributed generation DG side, respectively. In the figure, Qref is the reactive power setting value. The actual reactive power of the power system fluctuates. This fluctuating power is collected by A / D acquisition and data processing of the power grid branch, and then compared with the reactive setting value Qref and accumulated to form a deviation sequence. The PI regulator quickly responds to voltage changes by proportionally and integrally adjusting the voltage deviation. The A / D acquisition and data processing on the DG side collect the reactive power Q of the distributed power source. d , and the reactive power Q of the grid branch s The control strategy of the static VAR generator SVG is obtained after the calculation. Due to the complexity and rapidity of the reactive power fluctuation of the power grid, in order to get closer to the reactive compensation control target, the control strategy of the controller of the present invention first collects the real-time characteristic value of the corresponding comprehensive neural network model based on the neural network model, predicts the corresponding power fluctuation value, and detects the voltage and current signals of the power grid within a controllable time interval Δt, and uses Clarke transformation to convert the voltage and current in the three-phase stationary coordinate system to the synchronous rotating coordinate system to obtain the instantaneous reactive current component. According to the power fluctuation value and the instantaneous reactive current component, the reactive power target value required to be compensated for the power fluctuation value within the Δt time length is calculated, and the reactive current output by the SVG can accurately track the set reactive current target value by adjusting the amplitude and phase of the output voltage on the AC side of the bridge circuit, or directly controlling the current on the AC side.

[0113] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A dynamic reactive power compensation device for a distributed power generation and distribution network, characterized in that: include: A data acquisition unit is used to collect parameter data of distributed generation DG, not limited to voltage, current and power; A data classification unit is used to classify distributed generation DG and its corresponding parameter data, and to construct a neural network training data set corresponding to each type of distributed generation DG; A model building and training unit is used to build a neural network training model corresponding to each type of distributed generation DG, and train artificial neural network 1, artificial neural network 2, ... artificial neural network N based on a neural network training data set for each type of distributed generation DG; The model fusion unit integrates and trains a comprehensive neural network model based on artificial neural network 1, artificial neural network 2 ... artificial neural network N, connects the trained comprehensive neural network model to the controller, and calculates the SVG control strategy of the static VAR generator; A control instruction generating unit, used for performing linearization processing on the static VAR generator SVG control strategy to form a reactive compensation control amount of the SVG; The reactive power compensation execution unit is used to perform reactive power compensation on the distribution network by adjusting the amplitude and phase of the output voltage on the AC side of the bridge circuit or directly controlling the AC side current according to the reactive power compensation control amount of the SVG.

2. A dynamic reactive power compensation control method for a distributed power generation distribution network, characterized in that: The following steps are involved: The collection of distributed generation DG parameter data is not limited to voltage, current, and power; Classify the distributed generation DG and its corresponding parameter data, and construct a neural network training data set corresponding to each type of distributed generation DG; Establishing a neural network training model corresponding to each type of distributed generation DG, and training artificial neural network 1, artificial neural network 2, ... artificial neural network N based on each type of distributed generation DG neural network training data set; Based on artificial neural network 1, artificial neural network 2 ... artificial neural network N, a comprehensive neural network model is integrated and trained, and the trained comprehensive neural network model is connected to the controller to calculate the SVG control strategy of the static VAR generator; Linearizing the static VAR generator SVG control strategy to form a reactive compensation control quantity of the SVG; According to the reactive compensation control amount of SVG, reactive compensation is performed on the distribution network by adjusting the amplitude and phase of the output voltage on the AC side of the bridge circuit, or directly controlling the AC side current.

3. The method for dynamic reactive power compensation control of a distributed power generation distribution network according to claim 2, characterized in that: The classification of distributed generation DG and its corresponding parameter data specifically includes: Select features of the collected data, including energy source features, power generation scale features, and grid access features; Data preprocessing: clean the data after feature selection and perform One-Hot encoding on the cleaned data; Construct a decision tree, select a splitting criterion, calculate the information gain of each feature, and then select the feature with the largest information gain as the splitting feature of the current node, dividing the data set into different subsets to achieve parameter data classification.

4. The method for dynamic reactive power compensation control of a distributed power generation distribution network according to claim 2, characterized in that: The establishment of a neural network training model corresponding to each type of distributed generation DG is specifically as follows: Select different types of neural networks for different distributed generation classification tasks and determine the number of nodes in the input layer, hidden layer and output layer of the neural network; Initialize the network parameters of each layer of the neural network, including weights and biases; The cross entropy loss function is used as the error detection function between the characteristic value and the predicted value of distributed generation DG, where the characteristic value is the normalized value of various parameters of distributed generation DG that cause its output power fluctuation, and the predicted value is the normalized value of the predicted fluctuation of distributed generation DG output power. At the same time, the optimization algorithm is selected to adjust the weight and bias of the network.

5. The distributed power generation and distribution network dynamic reactive power compensation control method according to claim 4, characterized in that: The training of artificial neural network 1, artificial neural network 2, ... artificial neural network N based on each type of distributed generation DG neural network training data set is specifically: The corresponding distributed generation DG neural network training data set is divided into a training set and a data set, and the divided training set data is encoded and normalized; The normalized training set data is input into the neural network through the input layer nodes in sequence. The input data is propagated and calculated in the neural network in the order from the input layer to the hidden layer and then to the output layer to obtain the output value. According to the output results of the output layer and the actual categories in the training set data, the loss value is calculated using the selected loss function; Using the calculated loss value, the error signal is transmitted backward from the output layer to the input layer through the back propagation algorithm to adjust the weights and biases of the network; Update the network weights and biases using the selected optimization algorithm based on the weight and bias adjustments obtained through backpropagation; Repeat the above steps of forward propagation, loss calculation, back propagation, parameter update, etc. until the predetermined number of training times is reached or the stopping condition is met. The stopping condition is that the loss value reaches a certain minimum value or the number of training times reaches a certain maximum value. Use the validation set to evaluate the trained part of the network, input the validation set data into the network, obtain the output result according to the forward propagation step, and then calculate the loss value and accuracy of the validation set based on the output result and the actual category in the validation set data. If the loss value and accuracy meet the conditions, stop training.

6. The method for dynamic reactive power compensation control of a distributed power generation distribution network according to claim 2, characterized in that: The artificial neural network 1, artificial neural network 2 ... artificial neural network N are based on the fusion training of the comprehensive neural network model, and the training of the comprehensive neural network model is specifically realized by nesting the neural network models, wherein the nesting methods include series nesting, parallel nesting or mixed nesting, and the nested combinations are classified according to the principle of proximity of the distribution network line topology relationship, wherein the comprehensive neural network model is one or more.

7. The method for dynamic reactive power compensation control of a distributed power generation and distribution network according to claim 2, characterized in that: The calculation of the static VAR generator SVG control strategy is specifically as follows: At the time t-Δt, the real-time characteristic values ​​of the corresponding comprehensive neural network model are collected to predict the corresponding power fluctuation value; At time t, the voltage and current signals of the power grid are detected, and the voltage and current in the three-phase stationary coordinate system are converted to the synchronous rotating coordinate system using Clarke transformation to obtain the instantaneous reactive current component; According to the power fluctuation value and the instantaneous reactive current component, the reactive power target value required to be compensated for the power fluctuation value within the time period Δt is calculated. By adjusting the amplitude and phase of the output voltage on the AC side of the bridge circuit, or directly controlling its AC side current, the reactive current output by the SVG can accurately track the set reactive current target value.

8. The method for dynamic reactive power compensation control of a distributed power generation distribution network according to claim 2, characterized in that: The linearization process of the SVG control strategy is specifically as follows: (1) Establish an equivalent model and regard the static VAR generator SVG as a device that can output variable reactive current. Its output current i s With control voltage u c The functional relationship between s =f(u c ); (2) Taylor series expansion, at a typical working point u c0 Nearby pair function f(u c ) is expanded by Taylor series to obtain: f(in c )≈f(u c )≈f(u c0 )+f′(u c0 )(in c -in c0 ) Among them, f′(u c0 ) is the function f(u c ) in u c0 First derivative of a point; (3) Get the linearized model, let i S0 =f(u c0 ), then the linearized model is expressed as: and out and S0 +k(u c -in c0 ) Where k = f′(u c0 ) is the linearized gain function, i out is the current output control quantity after linearization.

9. The method for dynamic reactive power compensation control of a distributed power generation and distribution network according to claim 5, characterized in that: The optimization algorithms include gradient descent algorithm, Adagrad algorithm and Adam algorithm.

10. The method for dynamic reactive power compensation control of a distributed power generation and distribution network according to claim 6, characterized in that: The nested combination classification is nested according to the principle of proximity of the distribution network line topology relationship, specifically: The distribution network is equivalent to several main feeders, and the equivalent main branch calculation is performed; Conduct breadth-first search on the equivalent distribution network branches to determine candidate compensation nodes; Screen candidate compensation nodes and determine nodes to be compensated; The nodes to be compensated are combined according to the principle of proximity in topological relationship, and the artificial neural networks corresponding to the combined distribution network branches are nested.

Images