An Alternating Iterative State Estimation Method for AC / DC Distribution Networks Based on Deep Neural Network Pseudo-Measurement Modeling

Abstract

The present invention discloses an alternating iterative state estimation method for an AC / DC distribution network based on pseudo-measurement modeling using a deep neural network. The method first establishes a steady-state model of a voltage source converter and a real-time measurement model of an AC / DC distribution network; then uses historical measurement data to perform offline training on the DNN to establish a pseudo-measurement model of load node injection power; finally, alternating iterative state estimation is performed on the AC area and the DC area, and during the alternating process, the estimated values ​​of the VSC branch state quantities are exchanged between regions to ensure the consistency of the boundary state quantities. The present invention can not only improve the computational accuracy of the state estimation of the AC / DC distribution network, accurately estimate the state of the AC / DC distribution network in a low-redundancy measurement configuration, but also has the advantages of low computational complexity, fast convergence speed, and easy program expansion.

Description

Technical Field

[0001] The invention relates to a state estimation method for an AC / DC distribution network, and in particular to an alternating iterative state estimation method for an AC / DC distribution network based on deep neural networks (DNN) pseudo-measurement modeling. Background Art

[0002] With the development of distributed power sources and the increasing proportion of DC loads, the advantages of DC distribution networks have received more and more attention. Compared with AC distribution networks, DC distribution networks can reduce power conversion links, reduce network losses, and increase transmission capacity and energy utilization. Considering the long-term development of AC distribution technology, it is difficult for DC distribution networks to completely replace AC distribution networks. Therefore, building an AC / DC hybrid distribution network is an effective way to solve the problem of new energy grid connection and absorption and to be compatible with AC / DC heterogeneous loads. However, this complex hybrid structure poses a challenge to the real-time monitoring system operation status of the distribution management system (DMS).

[0003] State estimation (SE) can provide reliable basic data for DMS. The research on power flow calculation and state estimation algorithm of AC / DC hybrid distribution network can be divided into three categories: (1) Overall method: AC equation and DC equation are solved together and solved in a unified iterative way. (2) Alternating iterative method: AC and DC regions are calculated sequentially. (3) Distributed method: AC and DC regions are calculated in parallel. The overall method can fully consider the power coupling relationship between AC and DC systems and has high estimation accuracy, but the matrix dimension will increase with the expansion of system scale, and the amount of calculation will also increase accordingly. The distributed method needs to configure additional global information calculation modules to realize parallel calculation of each region, and all regions need to be calculated in each iteration. The alternating iterative method can converge as long as the difference of coupling variables is small enough in the sequential calculation process, and the calculation complexity is lower, but it also needs to use measurement redundancy to improve the estimation accuracy. Therefore, the effective application of alternating iterative state estimation technology has certain requirements on the number of real-time measurements. However, the number of real-time measurements in distribution networks is often small, which usually makes it difficult to ensure the observability of the system, and the extensive configuration of measurement devices will greatly increase the economic cost. Therefore, it is worthwhile to study how to ensure the observability of the distribution network when real-time measurement is limited. Establishing a pseudo-measurement model can improve measurement redundancy. The traditional method is to use short-term load forecasting to obtain node injection power, but the random fluctuation of load is large, the prediction accuracy is not high, and the state estimation accuracy is reduced. Summary of the invention

[0004] Purpose of the invention: In view of the above problems, the present invention provides an alternating iterative state estimation method for AC / DC distribution networks based on DNN pseudo-measurement modeling, which can improve the computational accuracy of state estimation of AC / DC distribution networks and accurately estimate the state of AC / DC distribution networks in low-redundancy measurement configuration.

[0005] Technical solution: The technical solution adopted by the present invention is an alternating iterative state estimation method for AC / DC distribution network based on DNN pseudo-measurement modeling, comprising the following steps:

[0006] (1) Establishing an AC / DC hybrid distribution network model, performing power flow calculation based on network parameter information and load curve of the AC / DC distribution network, using an overall method to calculate and simulate the operating state of the system at continuous sampling moments, superimposing random noise on the true value of the power flow to simulate the real-time measurement value, and obtaining the real-time measurement value and the true value of the load at each sampling moment; the establishment of the AC / DC hybrid distribution network model is to add a voltage source converter branch to the admittance matrix of the AC / DC distribution network, and for each AC node, establish two separate deep neural networks, the outputs of which are the load active power and reactive power, and the inputs are the real-time measurement values ​​of the AC distribution network; for each DC node, establish a deep neural network, the output of which is the load active power, and the input is the real-time measurement value of the DC distribution network.

[0007] The network parameter information includes: topological information of the AC / DC distribution network, impedance of each branch, equivalent impedance of the voltage source converter and control mode of the voltage source converter. The control mode of the voltage source converter includes: setting the active power P on the AC side ft，k and AC side reactive power Q ft，k Control; set AC side active power P ft，k And AC side voltage U f，k Control; Constant DC side voltage U d，k and AC side reactive power Q ft，k Control. The real-time measurement values ​​include: AC and DC node voltage amplitude measurement values, AC and DC branch current amplitude measurement values, AC and DC branch power measurement values, and AC and DC node injection power real-time measurement values. The DC node injection power measurement equation is:

[0008]

[0009] Where: P dc，i Inject active power into the DC node; the superscript m indicates the measured value; U dc，i is the node voltage amplitude; G dc，ij is the corresponding element in the DC distribution network admittance matrix; j∈N dc represents the node j of the DC distribution network; VSC k ∈i represents the VSC connected to DC node i k , P d，kFor VSC k Active power flowing to DC node i; υ Pdci is the measurement error.

[0010] (2) Divide the operating data obtained in step (1) into a training set and a test set, use the real-time measurement value of the training set as the input of the deep neural network, and the true value of the load as the target output, train the deep neural network, and use the test set data to test the trained deep neural network. Train the deep neural network so that the final test error is less than a certain threshold. The test error is the difference between the pseudo-measurement of the injected power of the load node at each sampling moment and the true value of the load. If the test error is large, adjust the parameters of the deep neural network, such as the number of network layers, the number of neurons, batch_size, epoch, etc., and retrain until the test error no longer decreases.

[0011] (3) Determine the estimation time t, input the real-time measurement value of the AC distribution network at time t into the deep neural network, obtain the injected power of the AC load node, and input the real-time measurement value of the DC distribution network at time t into the deep neural network, obtain the injected power of the DC load node; the corresponding weight of the obtained node power pseudo-measurement value in the state estimation is 1 / (E max / 3) 2 .

[0012] (4) Using the voltage source converter branch as an extended branch of the AC area and the DC area, the real-time measurement, virtual measurement and load node injection power pseudo-measurement of the extended AC area are input into a weighted least squares estimator for state estimation, wherein the virtual measurement is zero injection node power, and the load node injection power pseudo-measurement is the injection power of the AC node obtained in step (3);

[0013] (5) passing the estimated values ​​of the state quantities of the nodes at both ends of the voltage source converter branch in step (4) to the extended DC region as a pseudo-measurement for state estimation of the region, and inputting the pseudo-measurement together with the real-time measurement, virtual measurement and load node injection power pseudo-measurement of the extended DC region into a weighted least squares estimator for state estimation, wherein the virtual measurement is zero injection node power and the load node injection power pseudo-measurement is the DC node injection power obtained in step (3);

[0014] (6) The estimated values ​​of the state quantities of the nodes at both ends of the voltage source converter branch in the extended DC region are then transferred to the extended AC region as pseudo-measurements for state estimation of the region, and the state estimation of the region is performed again together with the measurements of the extended AC region in step (4).

[0015] (7) Repeat steps (5) and (6) to perform alternating iterative state estimation on the extended AC and DC regions until the difference between the estimated values ​​of the state variables of the nodes at both ends of the voltage source converter branch by the two region estimators is small enough, and finally obtain the estimated values ​​of the state variables of the AC / DC hybrid distribution network at time t.

[0016] In the alternating iteration method, the objective functions in the AC and DC regions are:

[0017]

[0018]

[0019] Among them, the state quantity x in the extended communication area ac and quantity measurement ac for:

[0020]

[0021] The state quantity x in the extended DC region dc and quantity measurement dc for:

[0022]

[0023] Where: J ac The objective function value of the weighted least squares state estimation in the extended AC region; J dc is the objective function value of the weighted least squares state estimation in the extended DC region; the superscripts (k) and (k+1) are the number of iterations; {x a , x av} is the state variable of the AC distribution network, where x av is the state quantity of the AC node connected to the VSC; x v is the state quantity of the node at the AC outlet of the VSC branch; a Measurement inside the AC distribution network; av Injection power measurement of AC nodes connected to VSC; v The measurement of the VSC branch includes the real-time measurement of branch power or the VSC control setting value; {x d , x dv} is the state variable of the DC distribution network, where x dv is the state quantity of the DC node connected to the VSC; d It is the measurement inside the DC distribution network; dv Injected active power measurement for the DC node connected to the VSC; are the estimated value of the state quantity of the AC node connected to the VSC in the extended DC area and the inverse of its estimated error variance, respectively. In the extended AC region state estimation The weight of are the estimated value of the state quantity of the AC outlet node of the VSC branch in the extended DC area and the inverse of its estimated error variance, respectively. In the extended AC region state estimation The weight of are the estimated value of the state quantity of the AC node connected to the VSC in the extended AC area and the inverse of its estimated error variance, respectively. In the extended DC region state estimation The weight of are the estimated value of the state quantity of the AC outlet node of the VSC branch in the extended AC area and the inverse of its estimated error variance, respectively. In the extended DC region state estimation The weight of .

[0024] Beneficial effects: Compared with the prior art, the present invention has the following advantages: In view of the problems of AC / DC interconnection and low real-time measurement coverage in hybrid distribution networks, the present invention proposes a DNN pseudo-measurement model to assist the AC / DC distribution network in performing alternating iterative state estimation. The method first establishes an equivalent model of the voltage source converter VSC and a real-time measurement model of the AC / DC distribution network; then, the historical measurement data is input into the DNN to model the node injection power, thereby improving the measurement redundancy; finally, the AC area and the DC area are decoupled, and alternating iterative state estimation is achieved between the AC and DC areas by transmitting the state quantity information of the VSC branch. The algorithm proposed in the present invention can not only improve the accuracy of pseudo-measurement and accurately estimate the state of the AC / DC distribution network in a low-redundancy measurement configuration, but also has the advantages of low computational complexity, fast convergence speed, and easy program expansion. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is the VSC equivalent model diagram;

[0026] Figure 2 is the DNN pseudo measurement model diagram of the present invention;

[0027] Figure 3 It is a schematic diagram of the expansion zoning of the AC and DC distribution network;

[0028] Figure 4 It is a flow chart of the method for alternating iterative state estimation of AC / DC distribution network based on DNN pseudo-measurement modeling according to the present invention;

[0029] Figure 5 It is the AC / DC distribution network test system diagram expanded by IEEE33 node system;

[0030] Figure 6 This is a comparison chart of the DNN pseudo-measurement model output;

[0031] Figure 7 This is a comparison chart of the maximum estimation errors of CSE and AISE under different pseudo measurement sets;

[0032] Figure 8 It is a convergence curve diagram of the method of the present invention. DETAILED DESCRIPTION

[0033] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0034] The method for alternating iterative state estimation of an AC / DC distribution network based on DNN pseudo-measurement modeling of the present invention comprises the following steps:

[0035] (1) Establish an AC / DC hybrid distribution network model, perform power flow calculation based on the network parameter information and load curve of the AC / DC distribution network, use the overall method to calculate and simulate the operating status of the system at continuous sampling times, and superimpose random noise on the true value of the power flow to simulate the real-time measurement value, so as to obtain the real-time measurement value and load true value at each sampling time;

[0036] (2) Divide the operating data obtained in step (1) into a training set and a test set, use the real-time measurement value of the training set as the input of the deep neural network, and the load true value as the target output, train the deep neural network, and use the test set data to test the trained deep neural network to obtain the corresponding test error, which is the difference between the pseudo-measurement of the injected power of the load node at each sampling moment and the load true value. If the test error is large, adjust the parameters of the deep neural network, such as the number of network layers, the number of neurons, batch_size, epoch, etc., and retrain until the test error no longer decreases;

[0037] (3) Determine the estimated time t, input the real-time measurement value of the AC distribution network at time t into the deep neural network to obtain the injected power of the AC load node, and input the real-time measurement value of the DC distribution network at time t into the deep neural network to obtain the injected power of the DC load node;

[0038] (4) Using the voltage source converter branch as an extended branch of the AC area and the DC area, the real-time measurement, virtual measurement and load node injection power pseudo-measurement of the extended AC area are input into a weighted least squares estimator for state estimation, wherein the virtual measurement is zero injection node power, and the load node injection power pseudo-measurement is the injection power of the AC node obtained in step (3);

[0039] (5) passing the estimated values ​​of the state quantities of the nodes at both ends of the voltage source converter branch in step (4) to the extended DC region as a pseudo-measurement for state estimation of the region, and inputting the pseudo-measurement together with the real-time measurement, virtual measurement and load node injection power pseudo-measurement of the extended DC region into a weighted least squares estimator for state estimation, wherein the virtual measurement is zero injection node power and the load node injection power pseudo-measurement is the DC node injection power obtained in step (3);

[0040] (6) The estimated values ​​of the state quantities of the nodes at both ends of the voltage source converter branch in the extended DC region are then transferred to the extended AC region as pseudo-measurements for state estimation of the region, and the state estimation of the region is performed again together with the measurements of the extended AC region in step (4).

[0041] (7) Repeat steps (5) and (6) to perform alternating iterative state estimation on the extended AC and DC regions until the difference between the estimated values ​​of the state variables of the nodes at both ends of the voltage source converter branch by the two region estimators is small enough, and finally obtain the estimated values ​​of the state variables of the AC / DC hybrid distribution network at time t.

[0042] The following is a detailed description of several important parts:

[0043] 1. AC / DC hybrid distribution network modeling

[0044] The VSC converter station is a key facility of the hybrid distribution network, which can realize the integration and interconnection of the AC distribution network and the DC distribution network. Ignoring the filter loss, the VSC is composed of converter transformers, converter reactors, converter bridges and other components, which can be equivalent to the form of an ideal VSC in series with an equivalent AC branch, such as Figure 1 shown.

[0045] Figure 1 In: k is the serial number of VSC; P ft，k , Q ft，k is the active and reactive power flowing from the AC system to the VSC; U f，k ∠θ f，k is the voltage phasor of the AC system node connected to the VSC; R v，k is the equivalent resistance of the VSC branch, X v，k is the equivalent reactance; U t，k ∠θ t，k is the voltage phasor of the converter station node; P in，k , Q in，k U is the active and reactive power flowing from the converter station node to the VSC; d，k is the voltage of the DC node connected to VSC; P d，k is the active power flowing from VSC to the DC system.

[0046] The power flow equation on the AC system side is:

[0047]

[0048]

[0049] Where: g v，k 、b v，k are the conductance and susceptance of the equivalent AC branch, that is, g v，k +jb v，k =1 / (R v，k +jX v，k );θ ft，k =θ f，k -θ t，k is the phase angle difference between the two ends of the branch.

[0050] The active power flowing into the VSC from the converter station node is:

[0051]

[0052] Where: θ tf，k =θ t，k -θ f，k .

[0053] Since the loss of VSC can be represented by the equivalent resistance R v，k is equivalent, so Figure 1 The DC power in is equal to the active power flowing into the ideal VSC.

[0054] P d，k =P in，k

[0055] The operating state of the AC / DC hybrid distribution network is related to the control mode of the VSC: (1) The AC side active power P ft，k and AC side reactive power Q ft，k ; (2) Fixed AC side active power P ft，k And AC side voltage U f，k ; (3) Constant DC side voltage U d，k and AC side reactive power Q ft，k The control variable values ​​of VSC are often known, so they can be used as pseudo-measurements to improve measurement redundancy and be given a larger weight.

[0056] The real-time measurement of AC distribution network includes node voltage amplitude U ac，i , branch current amplitude I ac，ij , branch power P ac，ij , Q ac，ij and node injection power P ac，i , Q ac，iWhen the VSC branch is added to the admittance matrix of the AC distribution network and the VSC branch and the AC distribution network are considered as a whole, the measurement equation of the AC part of the hybrid distribution network is the same as that of the pure AC distribution network.

[0057] The real-time measurement of DC distribution network includes node voltage amplitude U dc，i , branch current amplitude I dc，ij , branch active power P dc，ij and node injected active power P dc，i Among them, the measurement equation of the DC node injection power is:

[0058]

[0059] Where: superscript m represents the measured value; U dc，i is the node voltage amplitude; j∈N dc represents the node j of the DC distribution network; G dc，ij is the corresponding element in the DC distribution network admittance matrix; VSC k ∈i represents the VSC connected to DC node i k , P d，k For VSC k Active power flowing to DC node i; υ Pdci is the measurement error.

[0060] 2. DNN Pseudo-Measurement Model

[0061] DNN is a neural network with multiple hidden layers. The hidden layer and output layer are composed of many neurons. Each neuron uses the connection weights and biases to linearly combine the output data of the previous layer, and then generates output through the activation function and passes it to the neurons in the next layer. The layer-by-layer transmission of data is to repeat the above process until the final output is generated. If there are a large number of data samples with known inputs and outputs, the DNN can be trained so that the DNN can fit the relationship between inputs and outputs. The training process is to combine the output and target output of the DNN into a loss function, and use the backpropagation and gradient descent algorithms to adjust the parameters of the DNN.

[0062] The present invention uses DNN to construct a pseudo-measurement model of load power. For each AC node, two separate DNNs are established, and the outputs are load active and reactive power, and the inputs are real-time measurement values ​​of the AC distribution network; for each DC node, a DNN is established, and the output is load active power, and the input is real-time measurement values ​​of the DC distribution network. The output of the DNN is compared with the actual node load power, and the difference is the output error of the pseudo-measurement model. The specific steps of offline training are as follows: Figure 2 shown.

[0063] In the online application stage, the trained DNN can quickly generate pseudo-measurements of node injection power during real-time measurement updates. The maximum relative error of DNN output power is E max According to the 3-sigma principle, the weight of pseudo-measurement is 1 / (E max / 3) 2 Using the pseudo-measurement of node injected power generated by DNN together with the real-time measurement for online state estimation can meet both the redundancy requirement and the update rate requirement.

[0064] 3.WLS algorithm

[0065] Select the AC node voltage amplitude, phase angle and DC node voltage amplitude as the state variable x = [U ac ,θ,U dc ]. The measurement equation for state estimation can be expressed as:

[0066] z=h(x)+υ

[0067] Where: z is a vector of measurement values, including real-time measurement, zero-injection virtual measurement, and pseudo measurement; h(x) is the nonlinear mapping from x to z; υ is the measurement error vector.

[0068] WLS estimation can be regarded as solving an optimization problem with nonlinear equality constraints. The objective function is to minimize the sum of squares of weighted residuals. The optimization model is:

[0069]

[0070] Where: J is the objective function value; is the estimated value of the measurement vector; R -1 is the measurement weight matrix, which usually takes the inverse of the measurement error variance as the diagonal element.

[0071] 4. Alternating Iterative State Estimation of AC / DC Distribution Network

[0072] The AC / DC distribution network consists of three parts: the AC area, the DC area, and the VSC branch. The VSC branch is the connecting line of the AC / DC area. By expanding the AC area and the DC area, the hybrid distribution network is divided into two parts, such as Figure 3 As shown in the figure: {x a , x av} is the state variable of the AC distribution network, where x av is the state quantity of the AC node connected to the VSC; x v is the state quantity of the VSC converter station node; {x d , x dv} is the state variable of the DC distribution network, where x dv is the state quantity of the DC node connected to the VSC; aMeasurement inside the AC distribution network; av Injection power measurement of AC nodes connected to VSC; v Measurement of VSC branches, including real-time branch power measurement or VSC control setting values; d It is the measurement inside the DC distribution network; dv It is the injected active power measurement of the DC node connected to the VSC.

[0073] Since the injected power of the nodes connected to the VSC in the AC and DC distribution networks is affected by the VSC branches, the VSC branches are considered as the common overlapping areas after the area expansion. v As a shared measurement in the AC and DC regions, the state and quantity measurements in the extended AC region are:

[0074]

[0075] The state quantities and quantity measurements in the extended DC region are:

[0076]

[0077] z ac Only with x ac About dc Only with x dc Therefore, the AC region and the DC region can be decoupled and estimated. Considering the power coupling relationship between the two, the AC (DC) region {x av , x v The estimated value of} is used as the pseudo measurement in the extended DC (AC) region. Therefore, the objective functions in the AC and DC regions are:

[0078]

[0079]

[0080] Where: J ac The objective function value of the weighted least squares state estimation in the extended AC region; J dc is the objective function value of the weighted least squares state estimation in the extended DC region; the superscripts (k) and (k+1) are the number of iterations; are the estimated value of the state quantity of the AC node connected to the VSC in the extended DC area and the inverse of its estimated error variance, respectively. In the extended AC region state estimation The weight of are the estimated value of the state quantity of the AC outlet node of the VSC branch in the extended DC area and the inverse of its estimated error variance, respectively. In the extended AC region state estimation The weight of are the estimated value of the state quantity of the AC node connected to the VSC in the extended AC area and the inverse of its estimated error variance, respectively. In the extended DC region state estimation The weight of are the estimated value of the state quantity of the AC outlet node of the VSC branch in the extended AC area and the inverse of its estimated error variance, respectively. In the extended DC region state estimation The weight of .

[0081] The state estimation is performed alternately in the AC and DC regions until the difference between the VSC branch state quantities estimated by the two region estimators is less than the convergence threshold. The implementation steps are as follows: Figure 4 The technologies not mentioned in the present invention are all referred to the prior art.

[0082] 5. Example Analysis

[0083] The example tested in this invention is an AC / DC distribution network constructed by an IEEE 33-node system. The real-time measurement configuration is as follows: Figure 5 The equivalent parameters and control methods of VSC are shown in Table 1. The result of the overall method power flow calculation is taken as the true value, and the random error of normal distribution is added to the true value of the power flow to form the measurement value. The standard deviation of the real-time measurement error is simulated to be 1%.

[0084] Table 1 VSC parameters and control methods

[0085]

[0086] Considering only real-time measurement and virtual measurement, the measurement redundancy of the system is 0.76. After adding pseudo-measurement of node injection power, the redundancy is increased to 1.52, ensuring observability. The present invention constructs a DNN with two hidden layers. The operating state of the system under continuous time sections is simulated through a certain load curve. The 1st to 6000th time sections are used as training sets, the 6001st to 6300th time sections are used as test sets, and 2% of the training set is divided as a verification set. The specific training steps are as follows: Figure 2 As shown in the figure, the average error of the output power of the test set is less than 4%, which reflects the accuracy of the pseudo-measurement model. The trained DNN is used to generate pseudo-measurement values ​​of load node power for the 6301-6310 time sections. The comparison between the DNN output active power and the true value of AC node 1 and DC node 34 is shown in the figure. Figure 6 As shown. Figure 6 It can be seen that the DNN pseudo-measurement model is able to track the changing trend of node power and is suitable for situations with large load fluctuations.

[0087] In order to further verify the superiority of the algorithm proposed in the present invention, the overall method and alternating iterative method state estimation of AC and DC distribution networks are carried out based on the DNN pseudo-measurement model and the traditional prediction method. Among them, the traditional prediction method generally refers to the use of short-term load forecasting to obtain power, and its uncertainty is much greater than that of the supervisory control and data acquisition (SCADA) system. Therefore, the present invention simulates its measurement error as a Gaussian distribution with 10% and 30% uncertainty. CSE represents the overall method, AISE represents the alternating iterative method, p10% represents that the pseudo-measurement error obeys the Gaussian distribution with 10% uncertainty, p30% represents that the pseudo-measurement error obeys the Gaussian distribution with 30% uncertainty, and dnn represents that the pseudo-measurement is generated by DNN. The case test is divided into the following 6 cases: 1) CSEp10%; 2) AISep10%; 3) CSEp30%; 4) AISep30%; 5) CSEdnn; 6) AISEdnn.

[0088] The comparison of estimation errors is Figure 7 As shown in the figure, the results are the average values ​​of 100 consecutive time sections. Figure 7 It can be seen that after the introduction of the DNN pseudo-measurement model, the estimation accuracy of the AC state quantity of CSE and AISE is improved, and the estimation accuracy of the DC state quantity is slightly improved. In addition, the estimation error of the AISE proposed in the present invention is almost the same as that of the CSE, which has higher accuracy.

[0089] Taking the difference between the estimated VSC branch state quantities in the AC and DC regions as the unbalanced quantity, the convergence curve of AISEdnn in a single time section is as follows: Figure 8 As shown. Figure 8 It can be seen that the algorithm proposed in the present invention has good convergence.

[0090] Set the convergence threshold to 10 -3 . Let DNN1 and DNN2 be the pseudo-measurement models of active and reactive power of AC nodes, respectively, and DNN3 be the pseudo-measurement model of active power of DC nodes. The calculation time of each part is shown in Table 2. It should be noted that the training of DNN is carried out offline, and the trained model is directly used in the online application stage without repeated training. In addition, when the real-time measurement is updated, each DNN can generate pseudo-measurements in parallel. As can be seen from Table 2, for Figure 5 In the test system, the total calculation time of AISEdnn is less than 0.5s, and the calculation efficiency can meet the requirements of distribution network state estimation.

[0091] Table 2 Statistical results of calculation time

[0092]

Claims

1. A method for alternating iterative state estimation of AC / DC distribution networks based on deep neural network pseudo-measurement modeling, characterized in that: The following steps are involved: (1) Establish an AC / DC hybrid distribution network model, perform power flow calculation based on the network parameter information and load curve of the AC / DC distribution network, use the overall method to calculate and simulate the operating status of the system at continuous sampling times, and superimpose random noise on the true value of the power flow to simulate the real-time measurement value, so as to obtain the real-time measurement value and load true value at each sampling time; (2) Dividing the operating data obtained in step (1) into a training set and a test set, using the real-time measurement value of the training set as the input of the deep neural network and the true load value as the target output, training the deep neural network, and using the test set data to test the trained deep neural network; (3) Determine the estimated time t, input the real-time measurement value of the AC distribution network at time t into the deep neural network to obtain the injected power of the AC load node, and input the real-time measurement value of the DC distribution network at time t into the deep neural network to obtain the injected power of the DC load node; (4) Using the voltage source converter branch as an extended branch of the AC area and the DC area, the real-time measurement, virtual measurement and load node injection power pseudo-measurement of the extended AC area are input into a weighted least squares estimator for state estimation, the virtual measurement of the extended AC area is zero injection node power, and the load node injection power pseudo-measurement of the extended AC area is the injection power of the AC load node obtained in step (3); (5) passing the estimated values ​​of the state quantities of the nodes at both ends of the voltage source converter branch in step (4) to the extended DC region as pseudo-measurements for state estimation of the extended DC region, inputting the pseudo-measurements for state estimation of the extended DC region together with the real-time measurement, virtual measurement and pseudo-measurements of load node injection power of the extended DC region into a weighted least squares estimator for state estimation, wherein the virtual measurement of the extended DC region is zero injection node power, and the pseudo-measurement of load node injection power of the extended DC region is the injection power of the DC load node obtained in step (3); (6) The estimated values ​​of the state quantities of the nodes at both ends of the voltage source converter branch in the extended DC region are then transferred to the extended AC region as pseudo measurements for state estimation of the region, and the state estimation of the region is performed again together with the measurements of the extended AC region in step (4); (7) Repeat steps (5) and (6) to perform alternating iterative state estimation on the extended AC and DC regions until the difference between the estimated values ​​of the state variables of the nodes at both ends of the voltage source converter branch by the two region estimators is sufficiently small, and finally obtain the estimated values ​​of the state variables of the AC / DC hybrid distribution network at time t; The establishment of the AC / DC hybrid distribution network model is to add a voltage source converter branch to the AC / DC distribution network admittance matrix, and for each AC node, two separate deep neural networks are established, the outputs are the load active power and reactive power, and the input is the real-time measurement value of the AC distribution network; for each DC node, a deep neural network is established, the output is the load active power, and the input is the real-time measurement value of the DC distribution network; The network parameter information in step (1) includes: topological information of the AC / DC distribution network, impedance of each branch, equivalent impedance of the voltage source converter, and control mode of the voltage source converter; The real-time measurement values ​​described in step (1) include: AC and DC node voltage amplitude measurement values, AC and DC branch current amplitude measurement values, AC and DC branch power measurement values, and AC and DC node injection power real-time measurement values; In the alternating iterative state estimation described in step (7), the objective functions in the AC and DC regions are: Among them, the state quantity x in the extended communication area ac and quantity measurement ac for: The state quantity x in the extended DC region dc and quantity measurement dc for: Where: J ac The objective function value of the weighted least squares state estimation in the extended AC region; J dc is the objective function value of the weighted least squares state estimation in the extended DC region; the superscripts (k) and (k+1) are the number of iterations; {x a ,x av } is the state variable of the AC distribution network, where x av is the state quantity of the AC node connected to the VSC; x v is the state quantity of the node at the AC outlet of the VSC branch; a Measurement inside the AC distribution network; av Injection power measurement of AC nodes connected to VSC; v The measurement of the VSC branch includes the real-time measurement of branch power or the VSC control setting value; {x d ,x dv } is the state variable of the DC distribution network, where x dv is the state quantity of the DC node connected to the VSC; d It is the measurement inside the DC distribution network; dv Injected active power measurement for the DC node connected to the VSC; are the estimated value of the state quantity of the AC node connected to the VSC in the extended DC area and the inverse of its estimated error variance, respectively. In the extended AC region state estimation The weight of are the estimated value of the state quantity of the AC outlet node of the VSC branch in the extended DC area and the inverse of its estimated error variance, respectively. In the extended AC region state estimation The weight of are the estimated value of the state quantity of the AC node connected to the VSC in the extended AC area and the inverse of its estimated error variance, respectively. In the extended DC region state estimation The weight of are the estimated value of the state quantity of the AC outlet node of the VSC branch in the extended AC area and the inverse of its estimated error variance, respectively. In the extended DC region state estimation The weight of .

2. The method for alternating iterative state estimation of AC / DC distribution networks based on deep neural network pseudo-measurement modeling according to claim 1, characterized in that: The control method of the voltage source converter includes: setting the AC side active power P ft,k and AC side reactive power Q ft,k Control; set AC side active power P ft,k And AC side voltage U f,k Control; Constant DC side voltage U d,k and AC side reactive power Q ft,k control.

3. The method for alternating iterative state estimation of AC / DC distribution networks based on deep neural network pseudo-measurement modeling according to claim 1, characterized in that: The DC node injection power is calculated by the DC node injection power measurement equation, which is: Where: P dc,i Inject active power into the DC node; the superscript m indicates the measured value; U dc,i is the node voltage amplitude; G dc,ij is the corresponding element in the DC distribution network admittance matrix; j∈N dc represents the node j of the DC distribution network; VSC k ∈i represents the VSC connected to DC node i k , P d,k For VSC k Active power flowing to DC node i; υ Pdci is the measurement error.

4. The method for alternating iterative state estimation of AC / DC distribution networks based on deep neural network pseudo-measurement modeling according to claim 1, characterized in that: The corresponding weight of the injected power of the AC load node obtained in step (3) in the state estimation is 1 / (E max / 3) 2 , E max is the maximum relative error of DNN output power.

Images