A power electronic distribution network source-grid-load optimization operation control method and system

Abstract

The present invention discloses a power electronic distribution network source-grid-load optimization operation control method and system. A power electronic distribution network source-grid-load optimization operation control method of the present invention comprises the following steps: first, a fully connected neural network model is constructed to predict the day-ahead output of the photovoltaic power generation system; second, the historical data of the distribution network is used as input to establish a power electronic distribution network coordination optimization model, and the particle swarm optimization algorithm is used to solve the optimal control strategy of the intelligent device as a training set; then, a fully connected neural network and a convolutional neural network are constructed to train the control strategies of the continuous and discrete devices in the training set; finally, the load demand daily ratio is used to calculate the load data required for the day-ahead scheduling, and the trained network is used to obtain the day-ahead time-division coordination optimization strategy of the distribution network; the above method can reduce network losses and avoid node voltage exceeding the limit, thereby ensuring the safe and stable operation of the network.

Description

Technical Field

[0001] The present invention relates to the field of distribution network optimization control, and in particular to a power electronic distribution network source-grid-load optimization operation control method and system. Background Art

[0002] With the introduction of the "dual carbon" policy, my country's distributed photovoltaics have developed rapidly in recent years. Many provinces have formed a development trend of "many points and wide areas, and local high-density grid connection". The high-penetration distributed power sources have changed the flow and voltage distribution of medium and low voltage distribution networks. In some areas, power reverse transmission, voltage exceeding the limit and increased network losses have occurred.

[0003] By effectively coordinating the operation of various intelligent devices within the system, it is possible to reduce network losses and ensure that the voltage of system nodes does not exceed the limit. Therefore, studying a coordinated optimization operation strategy of "source, grid and load" in power electronic distribution network is of great significance to increasing the degree of new energy consumption in distribution network and promoting the transformation of power system under the background of "dual carbon".

[0004] Coordinated control and optimization of distribution networks is an important means to reduce network losses and ensure system voltage quality. Heuristic artificial intelligence algorithms such as genetic algorithms can coordinate a variety of intelligent control devices when dealing with complex problems such as distribution network optimization. They have been used more frequently in previous studies, but since heuristic algorithm optimization requires repeated flow calculations, the time cost is relatively high. At the same time, the degree of power electronics in distribution networks is increasing. How to coordinate power generation forecasting, day-ahead load data acquisition, and day-ahead optimization control to establish a power electronics distribution network source-grid-load optimization operation control method is also an important issue. Summary of the invention

[0005] In view of the defects of the above-mentioned prior art, the purpose of the present invention is to provide a method for constructing a fully connected neural network model to predict the photovoltaic output information on the day before; according to the coordinated optimization control strategy output by the distribution network coordinated optimization control model, a regression model of the fully connected neural network and a classification model of the convolutional network are constructed to train the continuous adjustable devices and discrete adjustable devices in the network respectively, so as to form a control model of the corresponding equipment, thereby accelerating the acquisition of the dispatching decision of the coordinated optimization model on the day before; the load distribution of the distribution network can be obtained more quickly and conveniently on the day before through the daily ratio of user load demand; thereby establishing a complete method from data prediction and acquisition to dispatch acceleration and then to the acquisition of the day before optimization scheme, realizing the effective control of the network loss and node voltage quality of the power electronic distribution network containing distributed power sources, and effectively improving the calculation efficiency, the scheme is simple, and the power electronic distribution network source-grid-load optimization operation control method and system are provided.

[0006] To achieve the above object, the first technical solution adopted by the present invention is:

[0007] A power electronic distribution network source-grid-load optimization operation control method comprises the following steps:

[0008] S1: Build a fully connected neural network model to predict the output data of the photovoltaic power generation system required for day-ahead coordination optimization, and obtain photovoltaic output prediction data;

[0009] S2: Build a distribution network coordinated optimization control model with the minimum active network loss and voltage deviation as the objective function;

[0010] S3: Obtain historical distribution network load data and photovoltaic historical output data, and use the particle swarm algorithm to solve the distribution network coordinated optimization control model in S2 to obtain the corresponding coordinated optimization control strategy;

[0011] S4: According to the coordinated optimization control strategy in S3, a regression model of a fully connected neural network and a classification model of a convolutional network are respectively constructed to control the output of the adjustable device;

[0012] The regression model uses the continuous variables in the coordinated optimization control strategy in S3 as label values ​​to control the continuously adjustable equipment;

[0013] The classification model uses the discrete variables in the coordinated optimization control strategy in S3 as label values ​​to control discrete adjustable devices;

[0014] S5: The load of each node in a day is calculated by the daily ratio of user load demand to obtain the load data required for coordinated optimization control, and combined with the photovoltaic output forecast data obtained in S1, it is input into the distribution network coordinated optimization control model in S2 as the feature quantity of the neural network to obtain the coordinated optimization control strategy for the time period before the day. Combined with the regression model and classification model in S4, the continuous adjustable equipment and discrete adjustable equipment in the distribution network are optimized. The control strategy can be controlled to minimize the system network loss as much as possible before the day and ensure that the node voltage is within an acceptable range, ensuring the safe and stable operation of the system.

[0015] After continuous exploration and experimentation, the present invention constructs a fully connected neural network model to predict the photovoltaic output information on the day before; according to the coordinated optimization control strategy output by the distribution network coordinated optimization control model, a regression model of the fully connected neural network and a classification model of the convolutional network are constructed to respectively train the continuous adjustable devices and the discrete adjustable devices in the network to form control models of the corresponding devices, thereby accelerating the acquisition of the dispatching decision of the coordinated optimization model on the day before; the load distribution of the distribution network can be obtained more quickly and conveniently on the day before through the daily ratio of user load demand; thereby establishing a complete method from data prediction and acquisition to dispatch acceleration and then to the acquisition of the day before optimization solution, realizing the effective control of the network loss and node voltage quality of the power electronic distribution network containing distributed power sources, and effectively improving the calculation efficiency, and the solution is simple and practical.

[0016] As the preferred technical measures:

[0017] The adjustable equipment includes a static VAR generator, a switching capacitor bank and a transformer tap.

[0018] The static VAR generator is a continuously adjustable device, and the switching capacitor bank and the transformer tap are discretely adjustable devices.

[0019] As a preferred technical measure: the photovoltaic power generation system output data prediction steps in S1 are as follows:

[0020] S11: Calculate the maximum information coefficient MIC value between each characteristic value in the historical weather information in the historical photovoltaic data set and the output value of the photovoltaic power generation system at the corresponding time to find the strength of the correlation between each characteristic value and the photovoltaic output;

[0021] S12: According to the maximum information coefficient MIC value between each characteristic variable and the photovoltaic output value calculated in S1, the characteristic variable with a higher maximum information coefficient MIC is selected as the characteristic value required for the neural network input, and the characteristic value is standardized to eliminate the dimension, and the training set required for the photovoltaic output forecast on the day before is obtained in combination with the historical photovoltaic output data;

[0022] S13: Build a fully connected neural network including 1 input layer, 2 fully connected layers, and 1 output layer to train the training set obtained in step S12. Use the gradient descent optimization algorithm Adam to solve the weights and biases in the fully connected neural network. For the fully connected neural network model, the mean absolute error mae is used to measure the prediction error of the model. The specific parameters of the fully connected neural network model are:

[0023] Input layer: There are N units, each unit is used to receive data of one feature value;

[0024] Hidden layer: includes two fully connected layers, each with 64 units and the activation function is ReLU function;

[0025] Output layer: The output layer contains one unit and does not need to set an activation function;

[0026] S14: Divide the data predicted by the fully connected neural network model in S13 by hour, and take the average value to obtain the photovoltaic output forecast data for each hour of the day as the photovoltaic output data set required for day-ahead optimization.

[0027] As the preferred technical measures:

[0028] The distribution network coordinated optimization control model in S2 includes constructing an objective function, setting power flow constraints, constructing control variable constraints, and establishing system safety constraints;

[0029] The objective of the objective function is to minimize the active network loss and node voltage deviation. The specific construction method is as follows:

[0030] minf=f ploss +Zf u ;

[0031] Z is the weight coefficient of node voltage deviation;

[0032] The calculation formula of active network loss is as follows:

[0033]

[0034] Where n is the number of network nodes; G ij is the real part of the node admittance matrix; B ij is the imaginary part of the node admittance matrix; δ ij is the voltage phase angle difference between node i and node j;

[0035] The calculation formula of node voltage deviation is as follows:

[0036]

[0037] Where n is the number of network nodes; Δu i is the difference between node i and node voltage boundary value; u max and u min is the node voltage boundary value;

[0038] The calculation formula of the power flow constraint condition is as follows:

[0039]

[0040] Where N is the number of nodes; P Gi and Q Gi The active and reactive power injected by the power supply of node i; U i and U j is the node voltage of node i and node j; G ij and B ij are the real and imaginary parts of the node admittance matrix; δ ij is the voltage phase angle difference between node i and node j; P Di and Q Di is the active and reactive power of the load at node i; P DGi and Q DGi Inject active and reactive power into node i distributed generation; Q ci is the total reactive compensation amount of node i;

[0041] The calculation formula of the control variable constraint condition is as follows:

[0042]

[0043] In the formula and It is the upper and lower limits of the reactive power compensation of the static VAR compensator. When it is negative, it means reactive power is absorbed, and when it is positive, it means reactive power is emitted. K tmin and K tmax is the gear limit of the transformer tap; Q C1 is the capacity of a single group of switched capacitors; Q Cmax is the maximum capacity of the switched capacitor; t and t max The current gear and the highest gear of the switched capacitor;

[0044] The calculation formula of system safety constraint is as follows:

[0045]

[0046] In the formula and The upper and lower limits allowed for the voltage at node i.

[0047] As the preferred technical measures:

[0048] The particle swarm algorithm in S3 includes the following contents:

[0049] S31: Obtain the historical load data of the distribution network and the historical output information of photovoltaic power generation, perform power flow calculation to obtain the value of the objective function as the particle fitness calculation function; the distributed photovoltaic system is connected to the grid through a current control inverter and is regarded as a PI node in the power flow calculation. The active power is obtained from the historical photovoltaic power generation data, and the reactive power is calculated by the following formula:

[0050]

[0051] e k +jf k =U k

[0052] Where Q k+1 is the reactive power value of the distributed generation in the k+1th iteration; e k 、f k are the real and imaginary parts of the voltage obtained at the kth iteration, j is the imaginary number code, U k is the node voltage at the kth iteration; I is the constant current vector amplitude of the distributed power source; P is the known constant active power;

[0053] Before each iteration, the reactive output power of the photovoltaic power generation system is calculated, and the PI node is converted into a PQ node for power flow calculation;

[0054] S32: List all combinations of discrete variables in the optimization variables by permutation and combination, and use the particle swarm algorithm to iteratively solve the continuous variables in each case;

[0055] S33: Initialize the parameters of the particle swarm algorithm and the speed and position of the particles. The position of the particles is the output power of the static VAR compensator. Set the constraints of the particle swarm algorithm, including the limit output power of the static VAR generator, and calculate the initial fitness of each particle in the population.

[0056] S34: Set individual optimal and group optimal. Individual optimal is the position where the individual obtains the best fitness in the calculation. Group optimal is the position where all particles in the population search for the best fitness. The fitness is calculated once a particle is updated, and the group optimal and individual optimal positions are updated.

[0057] S35: In each iteration, particles pass through individual extreme values ​​and group extreme values ​​and combine inertia weights to update their own speed and position. Since smaller inertia weights are conducive to local search and larger inertia weights are conducive to global search, in order to balance the local and global search capabilities of the algorithm, an empirical formula of linear inertia weight is used, which is as follows:

[0058]

[0059] Where w star is the initial inertia weight; w end is the inertia weight when the iteration reaches the maximum number; k is the current iteration number; T max is the maximum number of iterations;

[0060] S36: until the iteration is finished, the compensation results of the static VAR compensator under various discrete variable combinations are output, each result is compared and the result with the best fitness is selected as the optimal control strategy.

[0061] As the preferred technical measures:

[0062] In S4, the regression model is constructed as follows:

[0063] S41: The historical load data of the distribution network and the output data of the photovoltaic system are standardized and dimensionless as the eigenvalue input, and the optimization result of a single static VAR compensator is used as the output of the model;

[0064] S42: Build a fully connected neural network including 1 input layer, 2 fully connected layers, and 1 output layer. Use the gradient descent optimization algorithm Adam algorithm to solve the weights and biases in the fully connected neural network. Add a Dropout layer and use L1 regularization to reduce overfitting.

[0065] The parameters of the regression model are as follows:

[0066] Input layer: There are N units, each unit is used to receive data of one feature value;

[0067] The first fully connected layer: has 64 units and the activation function is the ReLU function;

[0068] The second fully connected layer: has 64 units and the activation function is the ReLU function;

[0069] Output layer: The output layer contains one unit that outputs the result of a single SVC.

[0070] As the preferred technical measures:

[0071] The classification model is constructed as follows:

[0072] Step 41: Output the dispatch result of a single discrete device as a label, form matrix data with the historical load data of the distribution network and the photovoltaic output data, and input the matrix data into the convolution layer for convolution operation to extract features. In addition, since the size of the input matrix data is much smaller than that of general image data, there is no need to connect the pooling layer. The flattening layer is directly used to convert the input data into a one-dimensional vector for subsequent network classification tasks;

[0073] Step 42: Construct a convolutional neural network model including 1 input layer, 1 convolution layer, 1 flat layer and 2 fully connected layers, and use the RMSprop gradient descent algorithm to iteratively solve the network parameters;

[0074] The parameters of the classification model are as follows:

[0075] Input layer: represents load and PV output data as input data and specifies the data size;

[0076] Convolution layer: The number of convolution kernels is 32, ReLU is used as the activation function, and the convolution kernel size is 3*3;

[0077] Flattening layer: converts the matrix output by the convolutional layer into a one-dimensional vector for subsequent classification tasks in the fully connected layer;

[0078] The first fully connected layer: has 64 units and the activation function is ReLU;

[0079] The second fully connected layer: The number of units is the number of all possible values ​​of a single discrete variable, and the activation function uses the Softmax function for multi-classification tasks.

[0080] As the preferred technical measures:

[0081] In S5, the method for calculating the load data required for day-ahead scheduling using the user load demand daily ratio is as follows:

[0082] The daily ratio of user load demand is used to describe the user load demand. The daily ratio of load demand is the ratio of various types of user loads to the total load of the node within 24 hours a day. The load of each node in the distribution network is divided into three categories: commercial load, residential load and industrial load. In the absence of interference from special factors, it can be considered that the daily ratio of power load of each type of user is relatively fixed. For a given distribution network structure, the load calculation formula at node i at time point t is as follows:

[0083]

[0084] Where: K represents the type of load, 1 represents commercial load, 2 represents residential load, and 3 represents industrial load; p i0 represents the rated load power at node i; r K，i represents the load ratio coefficient of the Kth load to the node i, which varies with different nodes; m K，j is the daily ratio coefficient of the load demand of the Kth load at the jth moment, which changes with time.

[0085] To achieve the above object, the second technical solution adopted by the present invention is:

[0086] A power electronic distribution network source-grid-load optimization operation control system, applying the above-mentioned power electronic distribution network source-grid-load optimization operation control method;

[0087] It includes a photovoltaic power generation system output prediction module, a power electronic distribution network coordinated optimization control strategy acquisition module, a day-ahead equipment coordinated optimization control model training module, and a distribution network coordinated optimization control strategy output module.

[0088] The present invention predicts the photovoltaic output information of the day-ahead through the photovoltaic power generation system output prediction module; uses the day-ahead equipment coordination optimization control model training module to train the continuous adjustable equipment and the discrete adjustable equipment in the network to form the control model of the corresponding equipment, thereby accelerating the acquisition of the dispatching decision of the day-ahead coordination optimization model; the load day ratio algorithm can obtain the load distribution of the distribution network more quickly and conveniently on the day-ahead; and effectively controls the network loss and node voltage quality of the power electronic distribution network containing distributed power sources by establishing a complete method and system from data prediction and acquisition to dispatch acceleration and then to day-ahead optimization solution acquisition.

[0089] As the preferred technical measures:

[0090] Photovoltaic power generation system output prediction module: by calculating the maximum information coefficient MIC value between the photovoltaic historical data characteristics and output information, the strongest feature value is found as input, and a fully connected neural network prediction model is built to obtain the photovoltaic output forecast value for the 24 hours before the day;

[0091] Power electronic distribution network coordinated optimization control strategy acquisition module: build a distribution network coordinated optimization control model based on distributed power sources and various intelligent control devices and solve it by particle swarm algorithm to obtain the optimal control strategy of each device for subsequent training of each device control model;

[0092] Day-ahead equipment coordination optimization control model training module: The corresponding control model is trained based on the optimal control strategy of each intelligent device based on historical load and photovoltaic data. The continuous adjustable device uses the regression model of the fully connected neural network, and the discrete adjustable device uses the classification model of the convolutional neural network.

[0093] Distribution network coordinated optimization control strategy output module: The 24-hour load distribution is calculated the day before by the load day ratio method. The control model of each device responds quickly to the obtained load and predicted photovoltaic output to obtain the coordinated optimization of the power electronic distribution network in different time periods.

[0094] Compared with the prior art, the present invention has the following gain effects:

[0095] After continuous exploration and experimentation, the present invention constructs a fully connected neural network model to predict the photovoltaic output information on the day before; according to the coordinated optimization control strategy output by the distribution network coordinated optimization control model, a regression model of the fully connected neural network and a classification model of the convolutional network are constructed to respectively train the continuous adjustable devices and the discrete adjustable devices in the network to form control models of the corresponding devices, thereby accelerating the acquisition of the dispatching decision of the coordinated optimization model on the day before; the load distribution of the distribution network can be obtained more quickly and conveniently on the day before through the daily ratio of user load demand; thereby establishing a complete method from data prediction and acquisition to dispatch acceleration and then to the acquisition of the day before optimization solution, realizing the effective control of the network loss and node voltage quality of the power electronic distribution network containing distributed power sources, and effectively improving the calculation efficiency, and the solution is simple and practical. BRIEF DESCRIPTION OF THE DRAWINGS

[0096] Figure 1 This is a flow chart of the coordinated optimization control method of the power electronic distribution network "source, grid and load" proposed by the present invention;

[0097] Figure 2 The original IEEE33 node structure diagram provided for the application example of the present invention;

[0098] Figure 3 The results of power generation prediction of the photovoltaic power generation system provided by the application example of the present invention;

[0099] Figure 4 A diagram of the load demand daily ratio and the load proportion of each node provided for the application example of the present invention;

[0100] Figure 5 A diagram showing changes in network loss before and after optimization provided for an application example of the present invention;

[0101] Figure 6 This is a node voltage distribution diagram after optimization implementation provided in an application example of the present invention. DETAILED DESCRIPTION

[0102] In order to more clearly illustrate the technical solutions and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0103] On the contrary, the present invention covers any substitution, modification, equivalent method and scheme made on the essence and scope of the present invention as defined by the claims. Further, in order to make the public have a better understanding of the present invention, some specific details are described in detail in the detailed description of the present invention below. Those skilled in the art can fully understand the present invention without the description of these details.

[0104] Example 1

[0105] like Figure 1 As shown, a power electronic distribution network source-grid-load optimization operation control method comprises the following steps:

[0106] S1: Build a fully connected neural network model to predict the output data of the photovoltaic power generation system required for day-ahead coordination and optimization.

[0107] The steps for photovoltaic power generation data prediction are as follows:

[0108] S11: Calculate the maximum information coefficient MIC (maximum information coefficient) value between each characteristic value in the historical weather information in the historical photovoltaic data set and the output value of the photovoltaic power generation system at the corresponding time to find the strength of the correlation between each characteristic value and the photovoltaic output.

[0109] When the amount of data is large enough, the calculation formula of the maximum information coefficient MIC is as follows:

[0110]

[0111]

[0112] Where x is each weather characteristic value variable, y is the photovoltaic output information at the corresponding time, X, Y are equal divisions after quantizing the variables x, y into scatter plots, p(X, Y) is the joint probability density function, and B is the 0.55th power of the total amount of data.

[0113] S12: According to the maximum information coefficient MIC value between each characteristic variable and the photovoltaic output value calculated by S1, the characteristic variable with a higher maximum information coefficient MIC is selected as the characteristic value required for the neural network input, and the characteristic value is standardized to eliminate the dimension, and the training set required for the day-ahead photovoltaic output prediction is obtained in combination with the historical photovoltaic output data.

[0114] The calculation formula for standardization is:

[0115] X=(X-mean) / std

[0116] Where X is the data of each eigenvalue, mean is the average value, and std is the variance. It should be emphasized that when processing the training set and the test set, the mean and variance calculated in the training set can be directly used when standardizing the test set data.

[0117] S13: Build a fully connected neural network including 1 input layer, 2 fully connected layers, and 1 output layer to train the data obtained in step S12. Use the gradient descent optimization algorithm Adam algorithm to solve the weights and biases in the fully connected neural network. For the obtained model, use the mean absolute error mae to measure the prediction error of the model. The specific parameters of the model are:

[0118] Input layer: There are N units, each unit is used to accept data of one feature value.

[0119] Hidden layer: includes two fully connected layers, each with 64 units and the activation function is ReLU function.

[0120] Output layer: The output layer contains one unit and does not require an activation function.

[0121] S14: Divide the predicted data by hour, and take the average value to obtain the photovoltaic output prediction data for each hour in a day as the photovoltaic output data set required for day-ahead optimization.

[0122] S2: Build a distribution network coordination optimization model with minimizing active network loss and voltage deviation as the objective function.

[0123] The construction method of the distribution network coordinated optimization model is as follows:

[0124] S21: The objective function step is to minimize the active network loss and node voltage deviation:

[0125] minf=f ploss +Zf u ; Z is the weight coefficient of node voltage deviation

[0126] The calculation formula of active network loss is as follows:

[0127]

[0128] Where n is the number of network nodes; G ij is the real part of the node admittance matrix; B ij is the imaginary part of the node admittance matrix; δ ij is the voltage phase angle difference between node i and node j.

[0129] The calculation formula of node voltage deviation is as follows:

[0130]

[0131] Where n is the number of network nodes; Δu i is the difference between node i and node voltage boundary value; u max and u min is the node voltage boundary value.

[0132] S22: The steps of power flow constraint conditions are:

[0133]

[0134] Where N is the number of nodes; P Gi and Q Gi The active and reactive power injected by the power supply of node i; U i and U j is the node voltage of node i and node j; G ij and B ij are the real and imaginary parts of the node admittance matrix; δ ij is the voltage phase angle difference between node i and node j; P Di and Q Di is the active and reactive power of the load at node i; P DGi and Q DGi Inject active and reactive power into node i distributed generation; Q ci is the total reactive compensation amount of node i.

[0135] S23: The control variable constraint condition step is:

[0136]

[0137] In the formula and It is the upper and lower limits of the reactive power compensation of the static VAR compensator. When it is negative, it means reactive power is absorbed, and when it is positive, it means reactive power is emitted. Kt min and K t max is the gear limit of the transformer tap; Q C1 is the capacity of a single group of switched capacitors; Q Cmax is the maximum capacity of the switched capacitor; t and t max It is the current gear and the highest gear of the switched capacitor.

[0138] S24: System security constraint steps are:

[0139]

[0140] In the formula and The upper and lower limits allowed for the voltage at node i.

[0141] S3: The particle swarm algorithm is used to solve the model of S2, and the historical data of the distribution network is used as input to obtain the corresponding coordinated optimization control strategy.

[0142] The particle swarm algorithm includes the following:

[0143] S31: Obtain the historical load data of the distribution network and the historical output information of photovoltaic power generation, perform power flow calculation to obtain the value of the objective function in step S21 as the particle fitness calculation function. It should be noted that when performing power flow calculation, the distributed photovoltaic system is regarded as connected to the grid through a current control inverter, and is regarded as a PI node in the power flow calculation. The active power is obtained from the historical photovoltaic power generation data, and the reactive power is calculated by the following formula:

[0144]

[0145] e k +jf k =U k

[0146] Where Q k+1 is the reactive power value of the distributed generation in the k+1th iteration; e k 、f k are the real and imaginary parts of the voltage obtained at the kth iteration, j is the imaginary number code, U k is the node voltage at the kth iteration; I is the constant current vector amplitude of the distributed power source; P is the known constant active power.

[0147] Through the above method, the reactive injection size of the photovoltaic power generation system can be calculated before each iteration, so that the PI node can be converted into a PQ node for power flow calculation.

[0148] S32: List all combinations of discrete variables in the optimization variables by permutation and combination, and perform iterative solution of the continuous variables using the particle swarm algorithm in each case.

[0149] S33: Initialize the parameters of the particle swarm algorithm and the speed and position of the particles. The position of the particles is the output power of the static VAR compensator. Set the constraints of the particle swarm algorithm, including the limit output power of the static VAR generator, and calculate the initial fitness of each particle in the population.

[0150] The particle velocity is initialized to 0, and the particle position is a random value taken within the upper and lower limits of reactive power compensation as the initial position of the particle.

[0151] S34: Set individual optimal and group optimal. Individual optimal is the position where the individual obtains the best fitness in the calculation. Group optimal is the position with the best fitness searched by all particles in the population. The fitness is calculated once a particle is updated, and the group optimal and individual optimal positions are updated.

[0152] S35: In each iteration, particles update their own speed and position through individual extreme values ​​and group extreme values ​​combined with parameters such as inertia weight. Since a smaller inertia weight is conducive to local search and a larger inertia weight is conducive to global search, in order to balance the local and global search capabilities of the algorithm, an empirical formula of linear inertia weight is used here, which is as follows:

[0153]

[0154] Where w star is the initial inertia weight; w end is the inertia weight when the iteration reaches the maximum number; k is the current iteration number, T max is the maximum number of iterations.

[0155] The particle updates its velocity and position according to the following formula:

[0156] v t+1 =wv t +c1r1(P t -x t )+c2r2(G t -x t )

[0157] x t+1 =x t +v t+1

[0158] Where x represents the position of the particle; v represents the velocity of the particle; w is the inertia weight; c1 and c2 are acceleration weight constants; r1 and r2 are random numbers between [0, 1]; P t is the optimal position that the particle has searched so far; G t is the optimal position searched by the entire particle swarm so far.

[0159] In addition, when the speed and position of the particle exceed the boundary, it is processed according to the boundary value.

[0160] S36: until the iteration is finished, the compensation results of the static VAR compensator under various discrete variable combinations are output, each result is compared and the result with the best fitness is selected as the optimal control strategy.

[0161] S4: The historical distribution network load data and photovoltaic power generation system output data are normalized and used as the feature input of the neural network. The continuous variables in the reactive power optimization strategy in S3 are used as label values ​​to build a regression model of the fully connected neural network. The discrete variables in the reactive power optimization strategy in S3 are used as label values ​​to build a classification model of the convolutional network.

[0162] Model training for continuous and discrete variables includes the following steps:

[0163] S41: The historical load data of the distribution network and the output data of the photovoltaic system are standardized and dimensionless and used as the eigenvalue input, and the optimization result of a single static VAR compensator is used as the output of the model.

[0164] S42: Build a fully connected neural network including 1 input layer, 2 fully connected layers, and 1 output layer. Use the gradient descent optimization algorithm Adam algorithm to solve the weights and biases in the fully connected neural network, add a Dropout layer and use L1 regularization to reduce overfitting. The model parameters are as follows:

[0165] Input layer: There are N units, each unit is used to accept data of one feature value.

[0166] The first fully connected layer: has 64 units and the activation function is the ReLU function;

[0167] The second fully connected layer: has 64 units and the activation function is the ReLU function;

[0168] Output layer: The output layer contains only one unit and outputs the result of a single SVC;

[0169] S43: The dispatch result of a single discrete device is used as a label output, the historical load data of the distribution network and the photovoltaic output data are used to form matrix data, and the matrix data is input into the convolution layer for convolution operation to obtain features. In addition, since the size of the input matrix data is much smaller than that of general image data, there is no need to connect the pooling layer. The flattening layer is used directly to convert the input data into a one-dimensional vector for subsequent network classification tasks.

[0170] S44: Construct a convolutional neural network model consisting of 1 input layer, 1 convolutional layer, 1 flat layer and 2 fully connected layers. Use the RMSprop gradient descent algorithm to iteratively solve the network parameters. The model parameters are as follows:

[0171] Input layer: represents load and PV output data as input data and specifies the data size;

[0172] Convolution layer: The number of convolution kernels is 32, ReLU is used as the activation function, and the convolution kernel size is 3*3;

[0173] Flattening layer: converts the matrix output by the convolutional layer into a one-dimensional vector for subsequent classification tasks in the fully connected layer;

[0174] The first fully connected layer: has 64 units and the activation function is ReLU;

[0175] The second fully connected layer: The number of units is the number of all possible values ​​of a single discrete variable, and the activation function uses the Softmax function for multi-classification tasks.

[0176] S5: The load of each node in a day is calculated by the daily ratio of user load demand to obtain the load data required for day-ahead scheduling. Combining the photovoltaic processing data obtained in S1 and the model in S4 can accelerate the acquisition of a 24-hour day-ahead time-segment coordinated optimization control strategy.

[0177] The calculation method of using the daily ratio of user load demand to calculate the load data required for day-ahead optimization control is as follows:

[0178] The present invention describes the user load demand by the user load demand daily ratio. The load demand daily ratio is the ratio of each type of user load to the total load of the node within 24 hours a day. The load of each node in the distribution network can generally be divided into three categories: commercial load, residential load and industrial load. In the absence of interference from special factors, it can be considered that the daily ratio of the power load of each type of user is relatively fixed. For a given distribution network structure, the load at node i at time point t can be expressed as:

[0179]

[0180] Where: K represents the type of load, 1 represents commercial load, 2 represents residential load, and 3 represents industrial load; p i0 represents the rated load power at node i; r K，i represents the load ratio coefficient of the Kth load to the node i, which varies with different nodes; m K，j is the daily ratio coefficient of the load demand of the Kth load at the jth moment, which changes with time.

[0181] Real-time Example 2

[0182] The present embodiment provides a power electronic distribution network source-grid-load optimized operation control system, which includes a photovoltaic power generation system output prediction module, a power electronic distribution network coordinated optimization control strategy acquisition module, a day-ahead equipment coordinated optimization control model training module, and a distribution network coordinated optimization control strategy output module.

[0183] Photovoltaic power generation system output prediction module: By calculating the maximum information coefficient MIC value between the photovoltaic historical data characteristics and the output information, the characteristic value with the strongest relationship is found as input, and a fully connected neural network prediction model is built to obtain the photovoltaic output forecast value for the 24 hours before the day.

[0184] Power electronic distribution network coordinated optimization control strategy acquisition module: Build a distribution network coordinated optimization model containing distributed power sources and various intelligent control devices and solve it using a particle swarm algorithm to obtain the optimal control strategy for each device, which will be used for the subsequent training of each device control model.

[0185] Day-ahead equipment coordinated optimization control model training module: The corresponding control model is trained based on the optimal control strategy of each intelligent device based on historical load and photovoltaic data. The continuously adjustable equipment uses the regression model of the fully connected neural network, and the discrete adjustable equipment uses the classification model of the convolutional neural network.

[0186] Distribution network coordinated optimization control strategy output module: The 24-hour load distribution is calculated the day before by the load day ratio method. The control model of each device responds quickly to the obtained load and predicted photovoltaic output to obtain the coordinated optimization of the distribution network containing distributed power sources in the day before and in different time periods.

[0187] For details not described in this embodiment, please refer to Embodiment 1.

[0188] Application Example In this application example, a power electronic distribution network source-grid-load optimization operation control method proposed by the present invention is adopted. The distribution network structure is an improved IEEE33 node graph. The IEEE33 node original graph is as follows: Figure 2 As shown, the improvement method is as follows: Distributed photovoltaics are installed on node 18, and the configuration of smart devices in the distribution network is shown in Table 1:

[0189] Table 1

[0190] Device Name Installation Node Number Capacity or gear Static VAR Compensator 5，15，31 -100-1000kvar Switching capacitor bank 6 100kvar per group, 6 groups in total transformer 1 Adjustable gear is 1±2*2.5%

[0191] In order to verify the effectiveness of the power generation prediction method, the data set used for power generation prediction is the photovoltaic power generation data of the Australian Solar Energy Research and Development Center, which records the photovoltaic output data for five minutes every day in most months over three years and influencing factor data including solar radiation, solar scattered radiation, weather temperature, weather relative humidity, wind speed, wind direction, and rainfall information; since some data are missing but this part accounts for a small proportion, it is filled in according to the principle of filling in adjacent data.

[0192] When predicting the day-ahead power generation data, the maximum information coefficient MIC value between the calculated data is shown in Table 2:

[0193] Table 2

[0194] feature Maximum Information Coefficient MIC Solar radiation 0.985 Diffuse solar radiation 0.447 Weather Temperature 0.287 Weather relative humidity 0.277 Wind speed 0.218 wind direction 0.142 Rainfall 0.111

[0195] The first six features with higher maximum information coefficients MIC are selected as the feature values ​​of the fully connected network input.

[0196] The prediction results of photovoltaic power generation system are as follows Figure 3 As shown, the forecast result is a value every five minutes, with a total of 288 forecast data per day.

[0197] The parameter settings of the particle swarm algorithm are shown in Table 3:

[0198] Table 3

[0199] Parameter name Numeric Dimension d 3 <![CDATA[Maximum number of iterations T max > 50 Number of particles N 50 <![CDATA[Initial inertia weight w star > 0.9 <![CDATA[End of inertia weight w end > 0.4 <![CDATA[Learning factors c1, c2]]> 2 <![CDATA[Maximum particle velocity v max > 2

[0200] Calculate the daily ratio of load demand in each period and the proportion of load at each node required for day-ahead coordination optimization. Figure 4 , where a is the ratio diagram of the three types of loads at 33 nodes, and b is the daily ratio diagram of the three types of load demands in 24 hours.

[0201] By using the method described in the present invention, the 24-hour network loss changes of the system before and after optimization are as follows: Figure 5 As shown, it can be seen that the network loss has decreased, indicating that the method of the present invention is effective in reducing network loss.

[0202] By using the method of the present invention, the voltage of each node of the optimized system within 24 hours is as follows: Figure 6 As shown, it can be seen that all node voltages are between 0.95-1.05, indicating that the method of the present invention is effective in ensuring voltage quality.

[0203] It can be seen from the simulation results that the present invention can solve the problems of high network loss and voltage exceeding the limit in the power electronic distribution network, and ensure its safe and stable operation.

[0204] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0205] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0206] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0207] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0208] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A power electronic distribution network source-grid-load optimization operation control method, characterized in that: The following steps are involved: S1: Build a fully connected neural network model to predict the output data of the photovoltaic power generation system required for day-ahead coordination optimization, and obtain photovoltaic output prediction data; S2: Build a distribution network coordinated optimization control model with the minimum active network loss and voltage deviation as the objective function; S3: Obtain historical distribution network load data and photovoltaic historical output data, and use the particle swarm algorithm to solve the distribution network coordinated optimization control model in S2 to obtain the corresponding coordinated optimization control strategy; S4: According to the coordinated optimization control strategy in S3, a regression model of a fully connected neural network and a classification model of a convolutional network are respectively constructed to control the output of the adjustable device; The regression model uses the continuous variables in the coordinated optimization control strategy in S3 as label values ​​to control the continuously adjustable equipment; The classification model uses the discrete variables in the coordinated optimization control strategy in S3 as label values ​​to control discrete adjustable devices; S5: The load of each node in a day is calculated through the daily ratio of user load demand to obtain the load data required for coordinated optimization control, and combined with the photovoltaic output forecast data obtained in S1, it is input into the distribution network coordinated optimization control model in S2 as the feature quantity of the neural network to obtain the coordinated optimization control strategy for the day before and in different time periods; then combined with the regression model and classification model in S4, the continuous adjustable equipment and discrete adjustable equipment in the distribution network are optimized.

2. A power electronic distribution network source-grid-load optimization operation control method as claimed in claim 1, characterized in that: The adjustable equipment includes a static VAR generator, a switching capacitor bank and a transformer tap. The static VAR generator is a continuously adjustable device, and the switching capacitor bank and the transformer tap are discretely adjustable devices.

3. A power electronic distribution network source-grid-load optimization operation control method as claimed in claim 1, characterized in that: The steps for predicting the output data of the photovoltaic power generation system in S1 are as follows: S11: Calculate the maximum information coefficient MIC value between each characteristic value in the historical weather information in the historical photovoltaic data set and the output value of the photovoltaic power generation system at the corresponding time to find the strength of the correlation between each characteristic value and the photovoltaic output; S12: According to the maximum information coefficient MIC value between each characteristic variable and the photovoltaic output value calculated in S1, the characteristic variable with a higher maximum information coefficient MIC is selected as the characteristic value required for the neural network input, and the characteristic value is standardized to eliminate the dimension, and the training set required for the photovoltaic output forecast on the day before is obtained in combination with the historical photovoltaic output data; S13: Build a fully connected neural network including 1 input layer, 2 fully connected layers, and 1 output layer to train the training set obtained in step S12. Use the gradient descent optimization algorithm Adam to solve the weights and biases in the fully connected neural network. For the fully connected neural network model, the mean absolute error mae is used to measure the prediction error of the model. The specific parameters of the fully connected neural network model are: Input layer: There are N units, each unit is used to receive data of one feature value; Hidden layer: includes two fully connected layers, each with 64 units and the activation function is ReLU function; Output layer: The output layer contains one unit and does not require an activation function. S14: Divide the data predicted by the fully connected neural network model in S13 by hour, and take the average value to obtain the photovoltaic output forecast data for each hour of the day as the photovoltaic output data set required for day-ahead optimization.

4. A power electronic distribution network source-grid-load optimization operation control method as claimed in claim 1, characterized in that: The distribution network coordinated optimization control model in S2 includes constructing an objective function, setting power flow constraints, constructing control variable constraints, and establishing system safety constraints; The objective of the objective function is to minimize the active network loss and node voltage deviation. The specific construction method is as follows: min f=f ploss +Zf u ; Z is the weight coefficient of node voltage deviation; The calculation formula of active network loss is as follows: Where n is the number of network nodes; G ij is the real part of the node admittance matrix; B ij is the imaginary part of the node admittance matrix; δ ij is the voltage phase angle difference between node i and node j; The calculation formula of node voltage deviation is as follows: Where n is the number of network nodes; Δu i is the difference between node i and node voltage boundary value; u max and u min is the node voltage boundary value; The calculation formula of the power flow constraint condition is as follows: Where N is the number of nodes; PG i and QG i Active and reactive power injected into the power supply of node i; U i and U j is the node voltage of node i and node j; G ij and B ij are the real and imaginary parts of the node admittance matrix; δ ij is the voltage phase angle difference between node i and node j; P Di and Q Di is the active and reactive power of the load at node i; P DGi and Q DGi Inject active and reactive power into the distributed generation at node i; Q ci is the total reactive compensation amount of node i; The calculation formula of the control variable constraint condition is as follows: In the formula and It is the upper and lower limits of the reactive power compensation of the static VAR compensator. When it is negative, it means reactive power is absorbed, and when it is positive, it means reactive power is emitted. K tmin and K tmax is the gear limit of the transformer tap; Q C1 is the capacity of a single group of switched capacitors; Q Cmax is the maximum capacity of the switched capacitor; t and t max The current gear and the highest gear of the switched capacitor; The calculation formula of system safety constraint is as follows: i = 1, 2, ... N; In the formula and The upper and lower limits allowed for the voltage at node i.

5. The power electronic distribution network source-grid-load optimization operation control method according to claim 1, characterized in that: The particle swarm algorithm in S3 includes the following contents: S31: Obtain the historical load data of the distribution network and the historical output information of photovoltaic power generation, perform power flow calculation to obtain the value of the objective function as the particle fitness calculation function; the distributed photovoltaic system is connected to the grid through a current control inverter and is regarded as a PI node in the power flow calculation. The active power is obtained from the historical photovoltaic power generation data, and the reactive power is calculated by the following formula: yes k +jf k =U k Where Q k+1 is the reactive power value of the distributed generation in the k+1th iteration; e k 、f k are the real and imaginary parts of the voltage obtained at the kth iteration, j is the imaginary number sign, U k is the node voltage at the kth iteration; I is the constant current vector amplitude of the distributed power source; P is the known constant active power; Before each iteration, the reactive output power of the photovoltaic power generation system is calculated, and the PI node is converted into a PQ node for power flow calculation; S32: List all combinations of discrete variables in the optimization variables by permutation and combination, and use the particle swarm algorithm to iteratively solve the continuous variables in each case; S33: Initialize the parameters of the particle swarm algorithm and the speed and position of the particles. The position of the particles is the output power of the static VAR compensator. Set the constraints of the particle swarm algorithm, including the limit output power of the static VAR generator, and calculate the initial fitness of each particle in the population. S34: Set individual optimal and group optimal. Individual optimal is the position where the individual obtains the best fitness in the calculation. Group optimal is the position where all particles in the population search for the best fitness. The fitness is calculated once a particle is updated, and the group optimal and individual optimal positions are updated. S35: In each iteration, the particle updates its own speed and position through individual extreme values ​​and group extreme values ​​combined with inertia weights, using an empirical formula of linear inertia weights, as shown below: Where w star is the initial inertia weight; w end is the inertia weight when the iteration reaches the maximum number; k is the current iteration number; T max is the maximum number of iterations; S36: until the iteration is finished, the compensation results of the static VAR compensator under various discrete variable combinations are output, each result is compared and the result with the best fitness is selected as the optimal control strategy.

6. A power electronic distribution network source-grid-load optimization operation control method as claimed in claim 1, characterized in that: In S4, the regression model is constructed as follows: S41: The historical load data of the distribution network and the output data of the photovoltaic system are standardized and dimensionless as the eigenvalue input, and the optimization result of a single static VAR compensator is used as the output of the model; S42: Build a fully connected neural network including 1 input layer, 2 fully connected layers, and 1 output layer. Use the gradient descent optimization algorithm Adam algorithm to solve the weights and biases in the fully connected neural network. Add a Dropout layer and use L1 regularization. The parameters of the regression model are as follows: Input layer: There are N units, each unit is used to receive data of one feature value; The first fully connected layer: has 64 units and the activation function is the ReLU function; The second fully connected layer: has 64 units and the activation function is the ReLU function; Output layer: The output layer contains one unit that outputs the result of a single SVC.

7. A power electronic distribution network source-grid-load optimization operation control method as claimed in claim 1, characterized in that: The classification model is constructed as follows: Step 41: Output the dispatch result of a single discrete device as a label, form matrix data with the historical load data of the distribution network and the photovoltaic output data, input the matrix data into the convolution layer for convolution operation to extract features; directly use the flattening layer to convert the input data into a one-dimensional vector for subsequent network classification tasks; Step 42: Construct a convolutional neural network model including 1 input layer, 1 convolution layer, 1 flat layer and 2 fully connected layers, and use the RMSprop gradient descent algorithm to iteratively solve the network parameters; The parameters of the classification model are as follows: Input layer: represents load and PV output data as input data and specifies the data size; Convolution layer: The number of convolution kernels is 32, ReLU is used as the activation function, and the convolution kernel size is 3*3; Flattening layer: converts the matrix output by the convolutional layer into a one-dimensional vector for subsequent classification tasks in the fully connected layer; The first fully connected layer: has 64 units and the activation function is ReLU; The second fully connected layer: The number of units is the number of all possible values ​​of a single discrete variable, and the activation function uses the Softmax function for multi-classification tasks.

8. The power electronic distribution network source-grid-load optimization operation control method according to claim 1, characterized in that: In S5, the method for calculating the load data required for day-ahead scheduling using the user load demand daily ratio is as follows: The daily ratio of user load demand is used to describe the user load demand. The daily ratio of load demand is the ratio of various types of user loads to the total load of the node within 24 hours of a day. The load of each node in the distribution network is divided into three categories: commercial load, residential load and industrial load. For a given distribution network structure, the load calculation formula at node i at time point t is as follows: Where: K represents the type of load, 1 represents commercial load, 2 represents residential load, and 3 represents industrial load; p i0 represents the rated load power at node i; r K，i represents the load ratio coefficient of the Kth load to the node i, which varies with different nodes; m K，j is the daily ratio coefficient of the load demand of the Kth load at the jth moment, which changes with time.

9. A power electronic distribution network source-grid-load optimization operation control system, characterized in that: Applying a power electronic distribution network source-grid-load optimization operation control method as described in any one of claims 1-8; It includes a photovoltaic power generation system output prediction module, a power electronic distribution network coordinated optimization control strategy acquisition module, a day-ahead equipment coordinated optimization control model training module, and a distribution network coordinated optimization control strategy output module.

10. The power electronic distribution network source-grid-load optimization operation control system according to claim 9, characterized in that: Photovoltaic power generation system output prediction module: by calculating the maximum information coefficient MIC value between the photovoltaic historical data characteristics and output information, the strongest feature value is found as input, and a fully connected neural network prediction model is built to obtain the photovoltaic output forecast value for the 24 hours before the day; Power electronic distribution network coordinated optimization control strategy acquisition module: build a distribution network coordinated optimization control model based on distributed power sources and various intelligent control devices and solve it by particle swarm algorithm to obtain the optimal control strategy of each device for subsequent training of each device control model; Day-ahead equipment coordination optimization control model training module: The corresponding control model is trained based on the optimal control strategy of each intelligent device based on historical load and photovoltaic data. The continuous adjustable device uses the regression model of the fully connected neural network, and the discrete adjustable device uses the classification model of the convolutional neural network. Distribution network coordinated optimization control strategy output module: The 24-hour load distribution is calculated the day before by the load day ratio method. The control model of each device responds quickly to the obtained load and predicted photovoltaic output to obtain the coordinated optimization of the power electronic distribution network in different time periods.

Images