Intelligent emergency control method and system for power system for data missing

Abstract

The invention relates to a data missing-oriented power system intelligent emergency control method and system, and the method comprises the steps: setting a data missing condition and a fault scene, and setting a constraint condition, a control objective function and a low-voltage load shedding reward function of a power system emergency control strategy; initializing the state of the power system, collecting a dynamic multi-time sequence observation sample set in real time, filling missing data, generating a node feature matrix, and constructing a topological graph of the power system after a fault; performing normalization processing on the node feature matrix, constructing a graph depth reinforcement learning model, performing topological feature extraction and dynamic feature weight adjustment, and generating a low-voltage load shedding emergency control strategy; iteratively updating the model based on a control objective function and a low-voltage load shedding reward function; new dynamic multi-time-sequence observation data are collected in real time, and an updated control strategy is repeatedly generated. According to the invention, the problem of power system emergency control under the condition of data missing is effectively solved, and an economical and efficient power system emergency control strategy is provided.

Description

Technical Field

[0001] The present invention belongs to the technical field of power system control, and in particular relates to a data-missing power system intelligent emergency control method and system. Background Art

[0002] With the expansion of the scale of power systems and the continuous improvement of the penetration rate of new energy, the emergency control methods of power systems have become increasingly complex, and the applicability of traditional control measures has declined. Research on emergency control strategies for power systems needs to be further studied. On the other hand, the synchronous vector measurement unit (PMU) used to evaluate system stability has data quality problems, and a series of problems such as data missing will occur, which will affect the temporary stability assessment of the power system and thus affect the formulation of emergency control strategies. At present, there is a lack of suitable emergency control means for the emergency control of power systems in the case of data missing. In the face of data missing, fault information cannot be obtained, and miscontrol, non-control and over-control may occur, affecting the correctness and economy of emergency control of power systems.

[0003] With the rapid development of artificial intelligence today, we can combine deep reinforcement learning with traditional emergency control strategies to train an emergency control strategy for the power system that can not only cope with data missing situations, but also solve emergency control problems, while being both economical and efficient. Summary of the invention

[0004] The purpose of the present invention is to address the shortcomings of the current power system emergency control technology in the case of data missing, and to provide a power system emergency control method, device and storage medium based on deep reinforcement learning, which can not only fill the PMU missing data caused by various abnormal situations, but also ensure the safe operation of the system, while being both economical and efficient.

[0005] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0006] An intelligent emergency control method for a power system facing data loss includes the following steps:

[0007] Step 1. Set data missing conditions and fault scenarios, and set constraints, control objective functions and low-voltage load reduction reward functions of the power system emergency control strategy under the data missing conditions and fault scenarios;

[0008] Step 2. Initialize the data missing situation and the power system state under the fault scenario, collect dynamic multi-time series observation sample sets in real time, fill in the missing data in the dynamic multi-time series observation sample sets and generate a supplementary matrix, and construct a post-fault power system topology map;

[0009] Step 3. Normalize the augmented matrix, construct the DGFEN-D3QN graph deep reinforcement learning model, perform topological feature extraction and dynamic feature weight adjustment based on the DGFEN-D3QN graph deep reinforcement learning model, and generate the emergency control strategy for low-voltage load shedding;

[0010] Step 4. Based on the control objective function in Step 1 and the low-voltage load shedding reward function, iteratively update the DGFEN-D3QN graph deep reinforcement learning model through reinforcement learning; collect new dynamic multi-temporal observation data in real time, and repeat Steps 2-3 to generate updated control strategies.

[0011] Further, the data missing situation and fault scenario in Step 1 are as follows: all bus data is partially missing before the fault, a bus is randomly selected as the fault bus, and the fault type is three-phase grounding short circuit; after the fault, the data of the fault bus is completely missing, and a line connected to the bus is disconnected when the fault is removed.

[0012] Further, the constraint conditions in Step 1 include:

[0013] Generator dynamic behavior constraints, network coupling constraints, power balance constraints, upper and lower limits of voltage amplitude constraints, and load shedding amount constraints;

[0014] The control objective function is:

[0015] minC all =C L,all +C V,all +C others

[0016]

[0017] where min is the minimum value, C all is the cost function of UVLS control, C L,all is the load shedding cost function, C V,all is the voltage deviation cost function, C others is the cost function of other factors; c L ,c V are the load shedding cumulative cost coefficient and the node voltage deviation cumulative cost coefficient respectively; S i corresponds to the apparent power of node i, V s,i is the normalized value of the sending-end bus voltage, Z l,i is the power line impedance, D is the total load demand value, V min,i is the minimum voltage value of bus node i; 1 p.u. represents the per-unit value of voltage;

[0018] The low-voltage load shedding reward function R rew is:

[0019]

[0020] Among them, -f is the penalty value given to the agent when the system crashes; a, b, c, and d are the coefficients of the low-voltage load shedding reward function, and ΔV ol,i (t) is the deviation of the target voltage value of node i at time t; ΔL i (t) is the load shedding amount of node i at time t; ΔF i (t) is the frequency deviation of node i at time t; ΔA j (t) is the power angle deviation of node j at time t.

[0021] Furthermore, the dynamic multi-temporal observation sample set X in step 2 is as follows:

[0022]

[0023] Among them, respectively represent the node voltage amplitudes of the first node in the selected power system at sampling time t and the previous n - 1 and n - 2 moments;

[0024] respectively represent the node voltage amplitudes of the second node in the selected power system at sampling time t and the previous n - 1 and n - 2 moments;

[0025] respectively represent the node voltage amplitudes of the m-th node in the selected power system at sampling time t and the previous n - 1 and n - 2 moments;

[0026] respectively represent the node load ratios of the first node in the selected power system at sampling time t and the previous n - 1 and n - 2 moments;

[0027] respectively represent the node load ratios of the second node in the selected power system at sampling time t and the previous n - 1 and n - 2 moments;

[0028] respectively represent the node load ratios of the m-th node in the selected power system at sampling time t and the previous n - 1 and n - 2 moments;

[0029] respectively represent the power angles of the first generator set in the selected power system at sampling time t and the previous n - 1 and n - 2 moments;

[0030] respectively represent the power angles of the second generator set in the selected power system at sampling time t and the previous n - 1 and n - 2 moments;

[0031] respectively represent the power angles of the x-th generator set in the selected power system at the sampling moment t and the previous n-1 and n-2 moments;

[0032] respectively represent the node frequencies of the first node in the selected power system at the sampling moment t and the previous n-1 and n-2 moments;

[0033] respectively represent the node frequencies of the second node in the selected power system at the sampling moment t and the previous n-1 and n-2 moments;

[0034] respectively represent the node frequencies of the m-th node in the selected power system at the sampling moment t and the previous n-1 and n-2 moments.

[0035] Furthermore, the step 2 includes:

[0036] Based on the data missing situation and fault scenarios set in step 1 and the dynamic multi-temporal observation sample set X, perform random missing processing on the original matrix X according to the missing rate p to generate a random mask matrix M; the mask matrix M processes the sample set X to generate a missing matrix X';

[0037] Fill in the missing data based on the ADIN algorithm, combine the dynamic noise matrix Z and the hint matrix H, fill in the missing values through the zero-sum game of the generator and the discriminator, and output the augmented matrix

[0038] Based on the augmented matrix and the original topological structure of the power system, construct the post-fault adjacency matrix A and generate the power system topology graph G.

[0039] Furthermore, the objective function of the ADIN algorithm is:

[0040]

[0041] where G is the generator, D is the discriminator, E is the mathematical expectation, λ is the hyperparameter, is the adaptive regularization term.

[0042] Furthermore, the step 3 includes:

[0043] Based on the augmented matrix output in step 2 normalize the node features to generate the normalized feature matrix X";

[0044] Construct the DGFEN-D3QN graph deep reinforcement learning model, extract the topological features of the power system through the graph convolutional network based on the power system topology graph G and the normalized feature matrix X", and dynamically calibrate the feature weights in combination with the DSE module;

[0045] Input the calibrated features into the D3QN network, calculate the Q value through a dual-branch, and select the low-voltage load shedding action strategy corresponding to the maximum Q value.

[0046] Further, step 4 includes:

[0047] Based on the control objective function and the low-voltage load shedding reward function, iteratively update the parameters of the DGFEN-D3QN graph deep reinforcement learning model through reinforcement learning to approximate the Q value to the target value;

[0048] Collect new dynamic multi-temporal observation data in real time, repeat steps 2-3 for dynamic filling and feature extraction, and generate an updated control strategy.

[0049] Further, calculating the Q value through a dual-branch includes: aggregating the state value and the action value function to obtain the Q value corresponding to the action:

[0050] Q(s,a;η,μ,τ) = S(s;η,τ) + A(s,a;η,μ)

[0051] where S(s;η,τ) is the state value function, A(s,a;η,μ) is the action value function, s is the power system state, α is the low-voltage load shedding action, τ is the network parameter of the state value function, η is the parameter of the common neural network convolution; μ is the network parameter of the action advantage value function.

[0052] On the other hand, the present invention provides an intelligent emergency control system for a power system facing data loss, including:

[0053] A fault scenario setting module, which is used to set the data loss situation and fault scenarios, and set the constraint conditions, control objective function and low-voltage load shedding reward function of the power system emergency control strategy under the data loss situation and fault scenarios;

[0054] An augmentation matrix generation module, which is used to initialize the power system state under the data loss situation and fault scenarios, collect a dynamic multi-temporal observation sample set in real time, fill the missing data in the dynamic multi-temporal observation sample set and generate an augmentation matrix, and construct a post-fault power system topology graph;

[0055] A control strategy generation module, which is used to perform normalization processing on the augmentation matrix, construct a DGFEN-D3QN graph deep reinforcement learning model, perform topology feature extraction and dynamic feature weight adjustment based on the DGFEN-D3QN graph deep reinforcement learning model, and generate a low-voltage load shedding emergency control strategy;

[0056] The control strategy update module is used to iteratively update the deep reinforcement learning model of the DGFEN-D3QN graph based on the control objective function and the low-voltage load shedding reward function; collect new dynamic multi-temporal observation data in real time, and repeatedly generate updated control strategies.

[0057] Compared with the prior art, the present invention has the following beneficial effects:

[0058] The present invention can fill and recover data in the case of missing or abnormal PMU data, capture the spatio-temporal relationship of the power system topology structure, extract key features, can cope with power system scenarios under different degrees of data loss, has strong adaptability, and can provide an efficient and economical low-voltage load shedding strategy for power system emergency control. Description of the Drawings

[0059] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following will briefly introduce the drawings required for the implementation examples or the description of the prior art. Obviously, the following drawings are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0060] Figure 1 It is a schematic diagram of the invention method flow.

[0061] Figure 2 It is a schematic diagram of the improved ADIN algorithm flow of the present invention.

[0062] Figure 3 It is a schematic diagram of the improved DGFEN graph convolutional neural network adopted by the present invention.

[0063] Figure 4 It is a schematic diagram of Embodiment 2 of the present invention.

[0064] Figure 5 It is a schematic diagram of the software platform of Embodiment 5 of the present invention. Detailed Embodiments

[0065] In order to make the above objects, features, and advantages of the present application more obvious and understandable, the following will describe the specific embodiments of the present application in detail with reference to the drawings. Many specific details are set forth in the following description to facilitate a thorough understanding of the present application. However, the present application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without departing from the connotation of the present application. Therefore, the present application is not limited by the specific embodiments disclosed below.

[0066] Embodiment 1

[0067] Such as Figure 1As shown, this embodiment provides an intelligent emergency control method for a power system facing data loss, including the following steps:

[0068] Step 1. Set the data loss situation and fault scenario, and set the constraint conditions, control objective function, and low-voltage load shedding reward function of the power system emergency control strategy under the data loss situation and fault scenario;

[0069] The data loss situation and fault scenario in Step 1 are as follows: Before the fault, partial data of all busbars are missing. Randomly select one busbar as the fault busbar, and the fault type is three-phase grounding short circuit; after the fault, the data of the fault busbar is completely missing, and one line connected to the busbar is disconnected when the fault is cleared.

[0070] The constraint conditions in Step 1 include:

[0071] Generator dynamic behavior constraints, network coupling constraints, power balance constraints, voltage amplitude upper and lower limit constraints, and load shedding amount constraints;

[0072] Among them, the dynamic behavior constraint relationship of the generator and its controller:

[0073] x t = f(x t , y t , d t , a t )

[0074] Network coupling constraint relationship among generators, transmission branches, and loads

[0075] g(x t , y t , d t , a t ) = 0

[0076] x t is the rotational speed and angular displacement of the generator rotor, y t represents the voltage of the power grid node, d t represents the interference in the power grid, a t represents low-voltage load shedding;

[0077] Power balance constraint relationship:

[0078]

[0079] m is the total number of system nodes, x is the number of load nodes, p is the number of generator nodes, P i is the injected active power of node i, P L,j is the active power demand of load node j, P G,k is the active power output of generator node k. Q i is the injected reactive power of node i, QL,j is the reactive power demand of load node j, Q G,k is the reactive power output of generator node k.

[0080] Constraints on upper and lower limits of voltage magnitude and load shedding amount:

[0081] V min,i ≤V i ≤V max

[0082] 0 ≤ ΔP i ≤ ΔP max

[0083] i = 1, 2, …, m, where V i is the voltage magnitude of node i, V min,i and V max,i are the upper and lower limits of the voltage magnitude of node i respectively. ΔP i is the load shedding amount of node i, ΔP max,i is the upper limit of the load shedding amount of node i.

[0084] The control objective function of the low-voltage load shedding control strategy designed in Step 1 is:

[0085] minC all = C L,all + C V,all + C others

[0086]

[0087] where, min is the minimum value, C all is the cost function of UVLS control, C L,all is the load shedding cost function, C V,all is the voltage deviation cost function, C others is the cost function of other factors; c L and c V are the cumulative load shedding cost coefficient and the cumulative node voltage deviation cost coefficient respectively; S i corresponds to the apparent power of node i, V s,i is the normalized value of the voltage of the sending-end bus, Z l,i is the power line impedance, D is the total load demand value, V min,i is the minimum voltage value of bus node i; 1 p.u. represents the per-unit value of voltage;

[0088] The low-voltage load shedding reward function R rew is:

[0089]

[0090] Among them, -f is the penalty value given to the agent when the system crashes; a, b, c, and d are the coefficients of the low-voltage load shedding reward function, and ΔV ol,i (t) is the deviation of the target voltage value of the voltage at node i at time t; ΔL i (t) is the load shedding amount of node i at time t; ΔF i (t) is the frequency deviation of node i at time t; ΔA j (t) is the power angle deviation of node j at time t.

[0091] In this embodiment, the selected emergency control type in step 1 is low-voltage load shedding, which operates in rounds and includes four control actions: 0 (no action), 1 (load shedding by 1%), 2 (load shedding by 2.5%), and 3 (load shedding by 5%); it is necessary to ensure that the load ratio of the action nodes remains within the range of [0.7, 1] during the action process;

[0092] In this embodiment, the data missing situation and the scale of the fault scenario are m, and the new energy penetration rate is 60%. The data missing situation is set as follows: the PMU data missing rate during the operation of all buses is 5%, and the PMU data of the fault buses is completely missing after the fault,

[0093] Step 2. Initialize the power system state under the data missing situation and the fault scenario, collect the dynamic multi-temporal observation sample set in real time, fill in the missing data in the dynamic multi-temporal observation sample set and generate an augmentation matrix, and construct the post-fault power system topology diagram;

[0094] The collection of the dynamic multi-temporal observation sample set is specifically defined as:

[0095] The observation sample of the i-th node of the selected power system at time t includes the node voltage amplitude, load ratio, and frequency value at the previous n - 1 moments; the power angle of the j-th generator set at time t includes the previous n - 1 moments, which can be expressed as:

[0096]

[0097] Among them, and only when there is a generator set at this node, there is power angle information in the sampling set; respectively represent the node voltage amplitudes of the i-th node of the new power system topology at the previous n - 1 moments before the sampling time t, the previous n - 2 moments before the sampling time t, and the sampling time t, respectively represent the node load ratios of the i-th node of the new power system topology at the previous n - 1 moments before the sampling time t, the previous n - 2 moments before the sampling time t, and the sampling time t; respectively represent the power angles of the generator sets of the j-th node of the new power system topology at the previous n - 1 moments before the sampling time t, the previous n - 2 moments before the sampling time t, and the sampling time t; respectively represent the node frequencies of the i-th node of the new power system topology at the (n - 1)-th moment before the sampling moment t, the (n - 2)-th moment before the sampling moment t, and the sampling moment t;

[0098] It is necessary to form a sample set with all the node observation states in the power system topology:

[0099]

[0100] where respectively represent the node voltage amplitudes of the first node in the selected power system at the sampling moment t and the (n - 1)-th and (n - 2)-th moments before;

[0101] respectively represent the node voltage amplitudes of the second node in the selected power system at the sampling moment t and the (n - 1)-th and (n - 2)-th moments before;

[0102] respectively represent the node voltage amplitudes of the m-th node in the selected power system at the sampling moment t and the (n - 1)-th and (n - 2)-th moments before;

[0103] respectively represent the node load ratios of the first node in the selected power system at the sampling moment t and the (n - 1)-th and (n - 2)-th moments before;

[0104] respectively represent the node load ratios of the second node in the selected power system at the sampling moment t and the (n - 1)-th and (n - 2)-th moments before;

[0105] respectively represent the node load ratios of the m-th node in the selected power system at the sampling moment t and the (n - 1)-th and (n - 2)-th moments before;

[0106] respectively represent the power angles of the first generator set in the selected power system at the sampling moment t and the (n - 1)-th and (n - 2)-th moments before;

[0107] respectively represent the power angles of the second generator set in the selected power system at the sampling moment t and the (n - 1)-th and (n - 2)-th moments before;

[0108] respectively represent the power angles of the x-th generator set in the selected power system at the sampling moment t and the (n - 1)-th and (n - 2)-th moments before;

[0109] respectively represent the node frequencies of the first node in the selected power system at the sampling moment t and the (n - 1)-th and (n - 2)-th moments before;

[0110] respectively represent the node frequencies of the second node in the selected power system at the sampling moment t and the previous n - 1 and n - 2 moments;

[0111] respectively represent the node frequencies of the m - th node in the selected power system at the sampling moment t and the previous n - 1 and n - 2 moments.

[0112] Specifically, the obtained observation sample set X in this embodiment i is a set of topological observation samples at time t and its previous 9 historical moments, which includes the voltage amplitude, load ratio, and frequency value of the node. If the node is connected to a generator set, the power angle value will be included; if it is not connected to a generator set, there will be no reactive power angle information in its sampling set:

[0113]

[0114] Among them, there is power angle information in the sampling set if and only if the node has a generator set; respectively represent the voltage amplitude, load ratio, power angle, and frequency of the sampling node i at the 9 moments before the sampling moment t; respectively represent the voltage amplitude, load ratio, power angle, and frequency of the sampling node i at the 8 moments before the sampling moment t; respectively represent the voltage amplitude, load ratio, power angle, and frequency of the sampling node i at the sampling moment t;

[0115] For the entire system topology, its sampling set X is:

[0116]

[0117] Among them, respectively represent the node voltage amplitudes of the power system topology at the 1st, 2nd, and 500th nodes at the 9 moments before the sampling moment t;

[0118] respectively represent the node load ratios of the power system topology at the 1st, 2nd, and 500th nodes at the 9 moments before the sampling moment t;

[0119] respectively represent the power angles of the power system topology at the 1st, 2nd, and 100th generator sets at the 9 moments before the sampling moment t;

[0120] respectively represent the node frequencies of the power system topology at the 1st, 2nd, and 500th nodes at the 9 moments before the sampling moment t;

[0121] respectively represent the node voltage amplitudes of the power system topology at the 1st, 2nd, and 500th nodes at the 8 moments before the sampling moment t;

[0122] respectively represent the node load ratios of the power system topology at the 1st, 2nd, and 500th nodes in the 8 instants before sampling time t;

[0123] respectively represent the power angles of the power system topology at the 1st, 2nd, and 100th generator sets in the 8 instants before sampling time t;

[0124] respectively represent the node frequencies of the power system topology at the 1st, 2nd, and 500th nodes in the 8 instants before sampling time t;

[0125] respectively represent the node voltage amplitudes of the power system topology at the 1st, 2nd, and 500th nodes at sampling time t;

[0126] respectively represent the node load ratios of the power system topology at the 1st, 2nd, and 500th nodes in the 8 instants before sampling time t;

[0127] respectively represent the power angles of the power system topology at the 1st, 2nd, and 100th generator sets at sampling time t;

[0128] respectively represent the node frequencies of the power system topology at the 1st, 2nd, and 500th nodes at sampling time t;

[0129] Perform data missing processing on the 10×(1500 + 100) multi-temporal topology observation state matrix X s obtained in step 2:

[0130] Generate a random mask matrix M matrix with the same dimension as the original matrix, and each element M ij (i = 1, 2,..., 10; j = 1, 2,..., 1600) follows a uniform distribution on [0, 1].

[0131] Set the relevant missing rate p = 0.05, and then determine the mask matrix M according to the missing rate p. When M ij ≤0.05, M ij = 0, indicating that the data at this position is missing; when M ij > p, M ij = 1, indicating that the information at this position is retained; the values of the mask matrix corresponding to the faulty busbar index after the fault time are 0;

[0132] X′ ij = X ij ⊙M ij

[0133] X ijis an element in the original matrix, X' ij The element after missing value processing.

[0134] As Figure 2 shown, after normalizing the obtained data, it is passed to the ADIN model algorithm in step 3, and the generator is used to generate the generation matrix (Regardless of whether the corresponding PMU data is generated or not, data will be generated in this way); and the data augmentation matrix that outputs the sample values replacing the missing data part in the input data matrix with the generation matrix In addition, the input of the ADIN model also needs to introduce the noise matrix Z;

[0135]

[0136] where X1 is the result of normalizing the input matrix data, and its dimension is the same as that of the input matrix; z ij is an element in the noise matrix; ⊙ is the Hadamard product operation, indicating that the corresponding elements of two matrices are multiplied;

[0137] The ADIN algorithm includes the hint matrix H, where B is a random matrix containing only 0 and 1, and its dimension is the same as that of the input matrix. The relationship between the number of 0 and 1 is related to the hint rate;

[0138] H = B⊙M + 0.5(1 - B)

[0139] The deep learning objective function of the ADIN algorithm is:

[0140]

[0141] G is the generator, D is the discriminator. The generator needs to generate values as close to the real data as possible so that the discriminator cannot distinguish, while the discriminator needs to correctly distinguish which are real; based on this, zero-sum game training is carried out; λ is a hyperparameter, is the adaptive regularization term. a i represents an element in the D prediction matrix, b i represents the element in M corresponding to a i ; after deep learning, the final predicted augmented matrix is obtained

[0142] Step 3. Normalize the augmented matrix, construct the DGFEN-D3QN graph deep reinforcement learning model, and perform topological feature extraction and dynamic feature weight adjustment based on the DGFEN-D3QN graph deep reinforcement learning model to generate the low-voltage load shedding emergency control strategy;

[0143] The power system topology graph G generated in step 3 mainly includes the power system topology graph of the power system topology adjacency matrix A, and the information of adjacent buses is considered when analyzing topological features;

[0144] As shown Figure 3 in the figure, the augmented matrix after data recovery according to the ADIN model Construct a power system topology diagram based on the adjacency matrix A of the power system topology structure:

[0145]

[0146] where v i and v j represent the i-th node and the j-th node of the power system topology respectively, i = 1, 2,..., 500; j = 1, 2,..., 500;

[0147] Normalize the obtained node topology feature matrix:

[0148]

[0149] where i is the row index of the augmented matrix after data recovery and j is its column index; x″ i,j is the element in the (i + 1)-th and (j + 1)-th rows of the augmented matrix X″ s after normalization; and are the maximum and minimum values of the (i + 1)-th row of the augmented matrix before preprocessing;

[0150] Extract its topology features according to the power system topology structure, and its layer-by-layer propagation rules are:

[0151]

[0152] H 0 = X″ s

[0153]

[0154] where X″ s is the multi-temporal topology feature matrix of the power system, A dj is the node adjacency matrix, X″ s and A dj are the inputs of the neural network model f(X″ s , A dj ); σ is the non-linear activation function, I N is the identity matrix, is 's degree matrix; H l is the matrix of the l-th layer, and W is the neural network weight.

[0155] Adaptive calibration of features is performed through the DSE module, and the output feature matrix of the l-th layer GCN is H l , whose dimension is 10×1600, n is the number of sampling moments, and d is the feature dimension; then, the squeezing operation of the DSE module is performed:

[0156]

[0157] The channel descriptor z c has a dimension of 1600×1, represents the i-th element of the c-th channel in H l .

[0158] The excitation operation is performed:

[0159]

[0160] z” = δ(z')

[0161]

[0162] s = F ex (z, W) = σ(g(z, w)) = σ(W2δ(W1z))

[0163] r = 16 is the reduction parameter of the DSE module, z′ is the intermediate result of the channel descriptor transformation, and its dimension is

[0164] z″ is the result of z′ activated by Relu, and its dimension remains unchanged at s is the channel weight, and its dimension is 1600×1.

[0165] Finally, the reweighting operation is performed:

[0166]

[0167] Among them, represents the element in the matrix after feature recalibration , represents the element obtained from the matrix H before feature recalibration l , s c is the element corresponding to the pre-calibration matrix in the channel weight s; the dimension of the matrix after feature calibration is the same as that of H l .

[0168] Combining the acquisition of power topology structure features in step 3 and the operation of DSE module compression excitation and re-calibration of features in step 4, the layer-by-layer propagation rule of the GCN layer is:

[0169]

[0170] After stimulating and compressing the graph neural network convolution, the deep reinforcement learning process of D3QN utilizes the extracted node features, divides the reinforcement learning process into two branches to proceed simultaneously, and based on parameters ξ and ξ′, performs action selection and policy evaluation based on the current state respectively:

[0171] Step 4. Based on the control objective function in Step 1 and the under-voltage load shedding reward function, the model DGFEN-D3QN graph deep reinforcement learning model is iteratively updated through reinforcement learning; new dynamic multi-temporal observation data is collected in real time, and Steps 2-3 are repeated to generate an updated control strategy.

[0172] According to the safety and stability calculation technical specifications of the State Grid Corporation, the voltage recovery standard time in seconds is set to 10s and the voltage recovery standard value is set to 0.8; combined with the emergency control constraint conditions established in Step 1 and the under-voltage load shedding control objective, a reward function is designed:

[0173]

[0174] ∑ i ΔV ol,i (t) = min{V ol,i (t) - 0.8, 0}, t c < t < t c + 10

[0175] Among them, R rew is the reward function for the under-voltage load shedding control process; system collapse means that the system frequency deviation exceeds the standard range within the specified time, or the power angle difference between any two generator sets is greater than the control requirement within the specified time, or the system voltage level does not recover to the target voltage value within the specified time. At this time, a maximum penalty value of -1000 is given to the agent; a = 52, b = 45, c = 23, d = 15 are the optimal reward function coefficients for multiple experiments, ΔV ol,i is the deviation of the node i voltage from the target voltage value; ΔL i is the node load shedding amount; ΔF i (t) is the node frequency deviation; ΔA i (t) is the node power angle deviation; t c is the fault occurrence time;

[0176] According to the value function Q(s t , a; ξ t ) representing the current state of the current value network, action selection is performed based on parameter ξ; where q represents the given under-voltage load shedding strategy t, at the power system state s t at time t, the selected action a, and the state-action value function under the action is obtained based on the action; then the action strategy a rew corresponding to the maximum state-action Q value is found; t,max ;

[0177] After that, the target value network is used to obtain the parameter-updated q-value function based on the parameter ξ′, and the optimal policy a obtained by the current value network is evaluated t,max :

[0178] a t,max = argmax Q(s t+1 , a; ξ t )

[0179] F t = R t+1 + γQ(s t+1 , argmax Q(s t+1 , a; ξ t ); ξ t )

[0180]

[0181] where R t is the sum of the reward values accumulated from time t;

[0182] After that, the obtained Q value is input into the reinforcement learning process, and the Q network is divided into two branches: the current Q network and the competitive Q network. The state value function S(s; η, τ) is only related to the power system state s, τ is the network parameter of the state value function, and η is the parameter of the common neural network convolution. The action value function A(s, a; η, μ) is related to both the system state and action, and μ is the network parameter of the action advantage value function.

[0183] Aggregating the state value and the action value function gives the Q value corresponding to the action:

[0184] Q(s, a; η, μ, τ) = S(s; η, τ) + A(s, a; η, μ)

[0185] After the above process, the output value of the main network is approximated to the target Q value, and then the action corresponding to the maximum Q value is selected to obtain the optimal low-voltage load shedding control strategy.

[0186] This embodiment mainly aims at the intelligent emergency control strategy for PMU data loss. Compared with the prior art in the case of power system data loss in emergency control, the present invention can fill and recover data for PMU data loss or anomalies, capture the spatio-temporal relationship of the power system topology, extract key features, can handle power system scenarios under different degrees of data loss, has strong adaptability, and can provide an efficient and economic low-voltage load shedding strategy for power system emergency control.

[0187] Embodiment 2:

[0188] In the operation of the power system, affected by factors such as extreme weather and equipment aging, various faults are likely to occur. For example, when hit by a strong typhoon, a large number of transmission lines trip due to the pulling of strong winds and the erosion of heavy rain. This instantly destroys the power grid structure in a local area, forces the power transmission path to change, and many power grid components face active power flow shocks far exceeding their rated capacity. Some old transformers, due to the aging of internal insulation and the decline of heat dissipation performance caused by long-term operation, under this overload pressure, the oil temperature rises sharply, triggering the protection device to automatically disconnect, resulting in power supply interruption in the associated areas. At the same time, PMUs located in these fault areas lose some data acquisition sources, resulting in the loss of key data such as voltage and current phasors.

[0189] To address the problem of data loss, a specially designed data recovery algorithm is adopted. This algorithm first deeply mines and analyzes the historical data accumulated by the power system for a long time, especially the variation laws of data at different nodes in past similar fault scenarios. For example, study the voltage and current data fluctuations of multiple substations under different seasons, different load levels, and different fault degrees, and establish an accurate data correlation model through data mining techniques. When facing the missing part of the real-time data, make full use of this model and the information of other complete data nodes to fill in the missing values. For example, for the missing voltage amplitude of the bus node in a substation located on the edge of the city, the algorithm will not only refer to the real-time voltage change trend of the bus nodes of other substations in the same city, but also deeply analyze the historical correlation data of other relevant electrical quantities in this substation, and use a complex mathematical calculation model to comprehensively calculate a most reasonable predicted value of the voltage amplitude, thus effectively realizing data recovery and gain, and providing a relatively complete and reliable data basis for subsequent in-depth fault analysis and precise control.

[0190] When constructing the DGFEN model, the actual topology of the power system is referred to. Among them, the nodes cover each substation, important bus nodes, and key power equipment such as large generators and high-voltage circuit breakers. The edges accurately represent the electrical connection relationships between these components, such as the connection directions of transmission lines and the connection methods between buses and various equipment, and then a detailed adjacency matrix is determined. The power system operation data after data recovery processing is used as the input feature matrix and input into the model for topological feature extraction. After the output features of each layer of the neural network, the DSE module is connected. The squeezing operation of the DSE module will perform global information integration for the data of specific feature channels of each node. For example, comprehensive analysis is carried out on the voltage feature channel data of all nodes to obtain a descriptor that can reflect the distribution of voltage features in the entire power system. The excitation operation generates weights for each feature channel through a series of complex matrix operations and activation function processing. These weights can accurately reflect the internal correlations between different electrical quantity feature channels, such as the dynamic correlation weights between voltage feature and current feature channels under different fault scenarios. Finally, a reweighting operation is carried out. According to the generated weights, element-wise multiplication is performed on the previous feature matrix to complete the recalibration of the features, greatly enhancing the neural network's ability to extract topological features of the power system, so as to more sensitively capture the subtle changes in the system state when a fault occurs. When integrating the D3QN algorithm later, the features processed by DGFEN are used as the input state representation of the D3QN network, and the action space of D3QN is set as a combination of various refined control strategies for power system faults. In terms of the load adjustment strategy, not only can the load of different substations be accurately increased or decreased, but also the optimal ratio and order of load increase and decrease can be determined according to the real-time needs of the system. In terms of equipment isolation operation, it can intelligently select which faulty equipment to isolate first to minimize the risk of fault spread based on the type of faulty equipment, the severity of the fault, and the impact on the overall system. For the switching control of standby power supplies, the optimal switching time and switching order can be determined according to the power supply and demand balance of the current system, the distribution of important loads, etc. Based on the current input power system state, the D3QN network accurately calculates the value of each action with the help of its value estimation network and advantage estimation network, and then selects the action with the highest value as the decision output. At the same time, a comprehensive and scientific reward function is defined to evaluate the effect of the actions taken. This reward function comprehensively considers many key factors, such as whether the system can resume stable operation in the shortest time after adopting the control strategy, whether the impact of the control operation on power supply reliability is within an acceptable range, whether it strictly meets the various safety constraints of the power system, and the control cost, etc.If the system can quickly return to stability after the control strategy is implemented, has minimal impact on power supply reliability, fully meets safety constraints, and has a low control cost, a high reward will be given; conversely, if the system is still in an unstable state, the power supply reliability has dropped significantly, safety constraints are violated, or the control cost is too high, a low reward will be given or even severe punishment will be imposed. A large amount of historical power system fault data and simulated fault data covering various types and different severities are used to deeply train the constructed DGFEN-D3QN model. These data comprehensively cover various complex state changes from the normal operating state before the fault occurs to the process of fault development, and have all undergone data missing simulation, recovery, and feature extraction processing. During the training process, according to the principle of the D3QN algorithm, the network parameters are continuously updated with the help of reward feedback, gradually optimizing the decision-making ability of the model so that it can accurately select appropriate fault response strategies when facing the real-time state of the power system.

[0191] Use the power system monitoring and control server as shown in Figure 4 to run the above model as the electronic device 90. This server is equipped with multiple high-performance processors. These processors adopt advanced multi-core architectures and cache technologies and can quickly process complex matrix operations, large-scale network parameter updates, and other computational tasks in the DGFEN-D3QN model. In terms of the memory 92, there is a large-capacity random access memory (RAM) 921 for temporarily storing the massive data and frequently updated model parameters during the operation of the model. The cache memory 922 further accelerates the data reading speed, and the read-only memory (ROM) 923 stores some basic system boot programs and fixed key parameters. In addition, the server also includes a high-speed bus connecting different system components. Among them, the data bus has an ultra-high data transmission bandwidth and is responsible for quickly transmitting data between various components. The address bus can accurately locate the data storage location, and the control bus efficiently coordinates the operations of each component. The server also has rich and powerful program modules 924, including an operating system optimized specifically for the server environment, a power system monitoring application program, a fault diagnosis and analysis program module, and a data storage and management program module, etc.

[0192] The electronic device 90 can communicate with external devices 95 through the I / O interface 94. The network adapter 96 can connect to the network so that the server can communicate with other power system devices at high speed and stability, realizing real-time interaction and sharing of data.

[0193] The model and related power system monitoring and control application programs are stored on a computer-readable storage medium, such as a large-capacity hard disk. The hard disk uses redundant array technology to ensure the security and reliability of data storage, and is used to long-term store program codes and a large amount of data. When the server starts up, the program codes are quickly loaded from the hard disk into the RAM921 for running. In addition, a portable disk can also be used for backup and transfer of program codes and data, which is convenient for flexible deployment and debugging in different servers or test environments. At the same time, some key program codes and core parameters can also be stored in the read-only memory (ROM) 923 to ensure that the basic running functions can be quickly restored in case of abnormal power-off or hardware failure of the system.

[0194] Embodiment 3

[0195] This embodiment provides a power system intelligent emergency control system for data loss, including:

[0196] A fault scenario setting module, which is used to set the data loss situation and fault scenarios, and set the constraint conditions, control objective function and low-voltage load shedding reward function of the power system emergency control strategy under the data loss situation and fault scenarios;

[0197] An augmented matrix generation module, which is used to initialize the power system state under the data loss situation and fault scenarios, collect real-time dynamic multi-temporal observation sample sets, fill in the missing data in the dynamic multi-temporal observation sample sets and generate an augmented matrix, and construct a post-fault power system topology diagram;

[0198] A control strategy generation module, which is used to normalize the augmented matrix, construct a DGFEN-D3QN graph deep reinforcement learning model, perform topology feature extraction and dynamic feature weight adjustment based on the DGFEN-D3QN graph deep reinforcement learning model, and generate a low-voltage load shedding emergency control strategy;

[0199] A control strategy update module, which is used to iteratively update the DGFEN-D3QN graph deep reinforcement learning model based on the control objective function and the low-voltage load shedding reward function through reinforcement learning; collect new dynamic multi-temporal observation data in real time, and repeatedly generate updated control strategies.

[0200] Embodiment 4

[0201] The entire power system fault response solution is implemented in the form of a program product, and the program code it includes can be written in various mainstream programming languages and can run on different terminal devices according to actual needs. For example, it can run partially on a local monitoring terminal for real-time data visualization and simple fault warning prompts, and run completely on a remote data center server for comprehensive fault analysis and precise control decision-making, so as to achieve all-round and multi-level response to power system faults, effectively ensure the safe and stable operation of the power system, and have strong adaptability and high reliability in the face of various complex and changeable fault situations, significantly reducing the losses caused by faults and greatly improving the overall operation efficiency of the power system.

[0202] The protection scope of the present invention is not limited to the above embodiments. Professionals in the field of power system technology can make various reasonable modifications and innovative deformations to the model architecture, algorithm details, and device hardware composition according to the actual situation. As long as these modifications and deformations fall within the scope of the claims of the present invention and its equivalent technologies, the intention of the present invention also includes these modifications and deformations.

[0203] Embodiment 5

[0204] As Figure 5 shown, in this embodiment, a software platform named "PowerSysGuard" is constructed to efficiently implement the intelligent emergency control method for power systems facing data loss.

[0205] The software platform architecture adopts a hierarchical architecture design, including a data acquisition layer, a data processing layer, a model training and management layer, and a control decision layer.

[0206] The data acquisition layer establishes communication connections with various sensors and monitoring devices (such as PMUs) in the power system, and uses standard communication protocols (such as IEC61850, etc.) to collect real-time power system operation data, including node voltage, load ratio, power angle, and frequency values, etc., and transmits the data to the data processing layer at a set time interval (such as once every 10 milliseconds).

[0207] The data processing layer conducts preliminary cleaning and verification on the received data, removing outliers and error data. For data missing situations, the ADIN algorithm consistent with the core patent method is applied for processing. First, according to the historical trends and correlation analysis of the data, a hint matrix H is generated, where the 0 and 1 distributions of the random matrix B are set based on the missing rules and characteristics of past data to better guide the generator G to generate values close to real data. At the same time, an adaptive noise matrix Z is introduced, and its noise intensity is dynamically adjusted according to the data fluctuation conditions to enhance the robustness of the algorithm. During the data filling process, by comparing and verifying with the historical data stored in the local database, the accuracy of the filled data is ensured. Subsequently, the processed data is constructed into a supplementary matrix and combined with the power system topology structure information to generate a topology graph G, which is transmitted to the model training and management layer.

[0208] The model training and management layer is responsible for the training, updating, and storage management of the DGFEN-D3QN model. In the training stage, samples are extracted from a large amount of historical fault data and simulated fault data, and training sets and test sets are constructed according to the method in the patent. High-performance computing resources (such as GPU clusters) are used to accelerate the model training process. During the training process, according to different power system operating conditions and fault scenarios, the hyperparameters of the model (such as learning rate, discount factor, etc.) are dynamically adjusted to optimize the model performance. After training, the model parameters are stored in a distributed file system, and the model is evaluated and updated regularly to ensure that the model has good adaptability to newly emerging fault types and power system changes.

[0209] The control decision-making layer receives the current power system status data from the data processing layer in real time, inputs it into the trained DGFEN-D3QN model, and obtains control decision-making suggestions. During the decision-making process, based on the reward function design principle in the patent, combined with the current system operation indicators (such as voltage stability, frequency deviation, etc.) and control constraint conditions, the value of each control action is accurately calculated. For example, in the face of a voltage sag fault, considering factors such as the voltage recovery requirements of different nodes, the impact of load adjustment on system stability, and the load shedding cost, the optimal low-voltage load shedding strategy and equipment control operations are selected, and the decision result is sent to the corresponding execution equipment through the power system communication network.

[0210] The development language is selected as Python as the main development language, which is convenient for implementing complex algorithm logics and model construction. At the same time, part of the underlying data processing and computing modules with extremely high performance requirements, such as the core matrix operation part in the ADIN algorithm, are written in C++. The interaction between Python and C++ code is carried out through Python's ctypes library to give full play to the high-efficiency performance advantages of C++. The database uses a MySQL relational database to store the historical operation data of the power system, topological structure information, and intermediate results during model training, etc. Utilize its powerful transaction processing ability and data query function to ensure data consistency and efficient retrieval. For the storage of large-scale model parameters and frequent read and write operations, the HBase distributed database is combined and used. Utilize its scalability and high concurrent read and write performance to meet the requirements of model training and management.

[0211] The communication between each module inside the software platform adopts RabbitMQ based on the message queue to achieve asynchronous communication, ensuring the high efficiency of data processing and the stability of the system. For the communication with external power system devices, follow the power industry standard communication protocols such as IEC61850 to achieve reliable data transmission and interaction.

[0212] The "PowerSysGuard" software platform is deployed on a high-performance server cluster. The servers are configured with multi-core CPUs, large-capacity memories (such as 128GB and above), and high-speed storage devices, and are equipped with professional GPU acceleration cards to accelerate model calculations. During operation, each module of the software platform is monitored in real time through the monitoring system, including key indicators such as CPU usage, memory occupancy, and network traffic. Once an abnormal situation is detected, such as a performance bottleneck or communication failure in a certain module, the system automatically triggers an early warning mechanism and performs fault recovery operations according to preset strategies, such as automatically restarting the faulty module, switching to the standby communication link, etc., to ensure the continuous and stable operation of the software platform and provide reliable intelligent emergency control services for the power system. Through the "PowerSysGuard" software platform constructed in this embodiment, it is possible to effectively integrate the technical methods in the patent, realize the intelligent emergency control of the power system in the case of data loss, and improve the safety and stability of the power system.

[0213] As described above, it is only a preferred specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by the present application should be covered within the protection scope of the present application.

[0214] It should be understood that the parts not elaborated in detail in this specification belong to the prior art.

[0215] It should be understood that the above description of the preferred embodiment is relatively detailed, and it should not be considered as a limitation to the protection scope of the present invention patent. Under the inspiration of the present invention, those of ordinary skill in the art can also make substitutions or deformations without departing from the protection scope defined by the claims of the present invention, and all of them fall within the protection scope of the present invention. The scope of protection claimed by the present invention shall be subject to the appended claims.

Claims

1. An intelligent emergency control method for power systems facing data loss, characterized in that, It includes the following steps: Step 1. Set the data missing situation and fault scenarios, and set the constraint conditions, control objective function, and low-voltage load shedding reward function of the power system emergency control strategy under the data missing situation and fault scenarios; Step 2. Initialize the power system state under the data missing situation and fault scenarios, collect real-time dynamic multi-temporal observation sample sets, fill in the missing data in the dynamic multi-temporal observation sample sets and generate an augmented matrix, and construct a post-fault power system topology graph; Step 3. Normalize the augmented matrix, construct a DGFEN-D3QN graph deep reinforcement learning model, perform topology feature extraction and dynamic feature weight adjustment based on the DGFEN-D3QN graph deep reinforcement learning model, and generate a low-voltage load shedding emergency control strategy; Step 4. Based on the control objective function and low-voltage load shedding reward function in Step 1, iteratively update the DGFEN-D3QN graph deep reinforcement learning model through reinforcement learning; collect new real-time dynamic multi-temporal observation data, and repeat Steps 2-3 to generate an updated control strategy.

2. The intelligent emergency control method for a power system facing data loss according to claim 1, characterized in that The data missing situation and fault scenarios in Step 1 are as follows: Before the fault, all bus data is partially missing. Randomly select one bus as the fault bus, and the fault type is three-phase ground short circuit; after the fault, the data of the fault bus is completely missing, and a line connected to the bus is disconnected when the fault is removed.

3. An intelligent emergency control method for a power system facing data loss according to claim 2, characterized in that The constraint conditions in Step 1 include: Generator dynamic behavior constraints, network coupling constraints, power balance constraints, upper and lower limits of voltage amplitude constraints, and load shedding amount constraints; The control objective function is: minC all = C L,all + C V,all + C others where min is the minimum value, C all is the cost function for UVLS control, C L,all is the load shedding cost function, C V,all is the voltage deviation cost function, C others is the cost function for other factors; c L , c V are the cumulative cost coefficients of load shedding and node voltage deviation respectively; S i is the apparent power corresponding to node i, V s,i is the normalized value of the sending-end bus voltage, Z l,i is the power line impedance, D is the total load demand value, V min,i is the minimum voltage value of bus node i; 1 p.u. represents the per-unit value of voltage; Low-voltage load shedding reward function R rew is as follows: Among them, -f is the penalty value given to the agent when the system crashes; a, b, c, and d are the coefficients of the low-voltage load shedding reward function, and ΔV ol,i (t) is the deviation of the target voltage value of node i at time t; ΔL i (t) is the load shedding amount of node i at time t; ΔF i (t) is the frequency deviation of node i at time t; ΔA j (t) is the power angle deviation of node j at time t.

4. An intelligent emergency control method for a power system facing data loss according to claim 1, characterized in that The dynamic multi-temporal observation sample set X in Step 2 is: Among them, respectively represent the node voltage amplitudes of the first node in the selected power system at the sampling time t and the previous n - 1 and n - 2 moments; respectively represent the node voltage amplitudes of the second node in the selected power system at the sampling moment t and the previous n - 1 and n - 2 moments; respectively represent the node voltage amplitudes of the m-th node in the selected power system at the sampling moment t and the previous n-1 and n-2 moments; respectively represent the node load ratios of the first node in the selected power system at the sampling moment t and the previous n-1 and n-2 moments; respectively represent the node load ratios of the second node in the selected power system at the sampling moment t and the previous n - 1 and n - 2 moments; respectively represent the node load proportion of the m-th node in the selected power system at the sampling moment t and the previous n-1 and n-2 moments; respectively represent the power angles of the first generator unit in the selected power system at the sampling moment t and the previous n-1 and n-2 moments; respectively represent the power angles of the second generator set in the selected power system at the sampling moment t and the previous n - 1 and n - 2 moments; respectively represent the power angles of the x-th generator set in the selected power system at the sampling moment t and the previous n-1 and n-2 moments; respectively represent the node frequencies of the first node in the selected power system at the sampling time t and the previous n - 1 and n - 2 times; respectively represent the node frequencies of the second node in the selected power system at the sampling time t and the previous n-1 and n-2 times; respectively represent the node frequencies of the m-th node in the selected power system at the sampling moment t and at the previous n-1 and n-2 moments.

5. An intelligent emergency control method for a power system facing data loss according to claim 1, characterized in that Step 2 includes: Based on the data missing situation and fault scenarios set in Step 1 and the dynamic multi-temporal observation sample set X, perform random missing processing on the original matrix X according to the missing rate p to generate a random mask matrix M; the mask matrix M processes the sample set X to generate a missing matrix X'; Fill in the missing data based on the ADIN algorithm, combine the dynamic noise matrix Z and the hint matrix H, fill in the missing values through the zero-sum game of the generator and the discriminator, and output the augmented matrix Based on the augmented matrix and the original topological structure of the power system, construct the post-fault adjacency matrix A and generate the power system topology graph G.

6. The intelligent emergency control method for a power system facing data loss according to claim 5, characterized in that, The objective function of the ADIN algorithm is as follows: Among them, G is the generator, D is the discriminator, E is the mathematical expectation, and λ is the hyperparameter. is the adaptive regularization term.

7. An intelligent emergency control method for a power system facing data loss according to claim 6, characterized in that, Step 3 includes: Augmentation matrix output based on Step 2 Normalize the node features to generate a canonical feature matrix X''; Construct a DGFEN-D3QN graph deep reinforcement learning model, extract the topology features of the power system through a graph convolutional network based on the power system topology graph G and the canonical feature matrix X", and dynamically calibrate the feature weights in combination with the DSE module; Input the calibrated features into the D3QN network, calculate the Q value through a two-branch method, and select the low-voltage load shedding action strategy corresponding to the maximum Q value.

8. An intelligent emergency control method for a power system facing data loss according to claim 7, characterized in that Step 4 includes: Based on the control objective function and low-voltage load shedding reward function, iteratively update the parameters of the DGFEN-D3QN graph deep reinforcement learning model through reinforcement learning to make the Q value approach the target value; Collect new real-time dynamic multi-temporal observation data, and repeat Steps 2-3 for dynamic filling and feature extraction to generate an updated control strategy.

9. An intelligent emergency control method for a power system facing data loss according to claim 8, characterized in that, The calculation of the Q value through a two-branch method includes: aggregating the state value and the action value function to obtain the Q value corresponding to the action: Q(s,a;η,μ,τ) = S(s;η,τ) + A(s,a;η,μ) Among them, S(s; η, τ) is the state value function, A(s, a; η, μ) is the action value function, s is the power system state, α is the under-voltage load shedding action, τ is the network parameter of the state value function, η is the parameter of the common neural network convolution; μ is the network parameter of the action advantage value function.

10. An intelligent emergency control system for power systems facing data loss, characterized in that, It includes: A fault scenario setting module, which is used to set the data missing situation and fault scenarios, and set the constraint conditions, control objective function and under-voltage load shedding reward function of the power system emergency control strategy under the data missing situation and fault scenarios; An augmented matrix generation module, which is used to initialize the power system state under the data missing situation and fault scenarios, collect real-time dynamic multi-temporal observation sample sets, fill in the missing data in the dynamic multi-temporal observation sample sets and generate an augmented matrix, and construct a post-fault power system topology graph; A control strategy generation module, which is used to normalize the augmented matrix, construct a DGFEN-D3QN graph deep reinforcement learning model, perform topology feature extraction and dynamic feature weight adjustment based on the DGFEN-D3QN graph deep reinforcement learning model, and generate an under-voltage load shedding emergency control strategy; A control strategy update module, which is used to iteratively update the DGFEN-D3QN graph deep reinforcement learning model based on the control objective function and the under-voltage load shedding reward function through reinforcement learning; collect new dynamic multi-temporal observation data in real time, and repeatedly generate updated control strategies; The intelligent emergency control system for a power system facing data loss is used to execute the steps in the intelligent emergency control method for a power system facing data loss according to any one of claims 1-9.

Images