Lightweight sensitivity-based multi-modal adaptive control method for distribution network edge

Abstract

The lightweight sensitivity-based multi-modal adaptive control method for the edge side of the distribution network of the present invention is based on solving the problem of multi-modal adaptive control of the distribution network when the network parameters are unknown or inaccurate. The present invention can accurately simulate the global sensitivity of the distribution network by effectively refining historical data, and use the global sensitivity of the distribution network to construct a multi-modal adaptive optimization control model on the edge side, effectively avoiding the problem that the operation constraints of the distribution network cannot be established when the network parameters are unknown or inaccurate. At the same time, by constructing a multi-modal selection objective function, the adaptive selection of the operation control mode on the edge side of the distribution network can be achieved, highlighting the difference in control effects of different modes, improving the flexibility of the operation control on the edge side of the distribution network, promoting the reduction of network loss, voltage deviation and line load balancing of the distribution network, and supporting the flexible and economic operation of the edge side of the distribution network under the access of a high proportion of distributed power sources.

Description

Technical Field

[0001] The present invention relates to a multi-mode adaptive control method at the edge of a distribution network, and more particularly to a multi-mode adaptive control method at the edge of a distribution network based on lightweight sensitivity. Background Art

[0002] The large-scale and high-proportion integration of distributed generators (DGs) poses numerous challenges to the stable operation of distribution networks. Due to the intermittent and uncertain nature of DGs, the operational volatility of distribution networks has increased dramatically, and power flows have become more variable. This has severely impacted the economic and safe operation of distribution networks and has led to a greater complexity in their optimization and control methods.

[0003] Edge computing technology enables data to be analyzed and processed locally at the edge of the distribution network close to the data source, thereby significantly reducing the size of the data and increasing the speed of calculations. Edge computing technology and the cloud-edge collaborative operation mode implemented on this basis are highly compatible with the operation and control architecture of the distribution network under a high proportion of distributed power access, and can be used as an implementation solution for flexible operation and control of the distribution network. Cloud-edge collaborative control continues the "layered + partitioned" operation and control concept, effectively avoiding the information transmission pressure brought about by the aggregation of massive information to the distribution network master station, the computational pressure faced by the centralized solution of global optimization problems, and the difficulty and efficiency of solving complex control problems. At the same time, it has an optimization effect comparable to that of centralized methods. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a multi-modal adaptive control method at the edge of the distribution network based on lightweight sensitivity for distribution network operation with a high proportion of distributed power access in order to overcome the shortcomings of the existing technology.

[0005] The technical solution adopted by the present invention is: a multi-modal adaptive control method at the edge of a distribution network based on lightweight sensitivity, comprising the following steps:

[0006] 1) According to the selected distribution network, divide the area of ​​each edge computing device and input the distribution network parameters; input the distribution network N d The historical tide data of a typical day is collected; the total sampling time ΔT and sampling time interval Δt of the training samples are set; the parameters of the dual-channel graph convolutional neural network model are set, including the learning rate γ, the regularization coefficient β, and the number of complete traversals of the training set E;

[0007] 2) Based on the distribution network N in step 1) dBased on the historical tide data of a typical day, a dual-channel graph convolutional neural network model and its training set are built in the cloud. The training of the dual-channel graph convolutional neural network model is completed, and a lightweight sensitivity curve is generated based on the trained dual-channel graph convolutional neural network model.

[0008] 3) Sending the lightweight sensitivity curve in step 2) to the edge computing devices at each edge of the distribution network to construct a sensitivity piecewise linearization constraint;

[0009] 4) Establishing a multi-modal adaptive control model for the edge of the distribution network based on lightweight sensitivity at each edge of the distribution network, including: setting the maximum improvement effect of the operating mode at the edge of the distribution network as the objective function, and considering the sensitivity piecewise linearization constraint, the distribution network safe operation constraint, and the distributed power generation operation constraint respectively;

[0010] 5) Solve the multi-modal adaptive control model of the distribution network edge side based on lightweight sensitivity obtained in step 4) and output the solution results, including: the control mode of the distribution network edge side area and the operation control strategy of each distributed power source.

[0011] The lightweight sensitivity-based multi-modal adaptive control method for the edge side of the distribution network of the present invention is based on solving the problem of multi-modal adaptive control of the distribution network when the network parameters are unknown or inaccurate. The present invention can accurately simulate the global sensitivity of the distribution network by effectively refining historical data, and use the global sensitivity of the distribution network to construct a multi-modal adaptive optimization control model on the edge side, effectively avoiding the problem that the operation constraints of the distribution network cannot be established when the network parameters are unknown or inaccurate. At the same time, by constructing a multi-modal selection objective function, the adaptive selection of the operation control mode on the edge side of the distribution network can be achieved, highlighting the difference in control effects of different modes, improving the flexibility of the operation control on the edge side of the distribution network, promoting the reduction of network loss, voltage deviation and line load balancing of the distribution network, and supporting the flexible and economic operation of the edge side of the distribution network under the access of a high proportion of distributed power sources. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 It is a flow chart of the multi-modal adaptive control method of the edge side of the distribution network based on lightweight sensitivity of the present invention;

[0013] Figure 2 This is the modified IEEE 33-node distribution network example structure diagram;

[0014] Figure 3a This is the system voltage distribution diagram of Scheme I;

[0015] Figure 3b This is the system voltage distribution diagram of Scheme II;

[0016] Figure 4aThis is the system line load rate distribution diagram of Scheme I;

[0017] Figure 4b This is the system line load factor distribution diagram of Scheme II;

[0018] Figure 5 is the result of the control mode selection of each edge side area in Scheme II;

[0019] Figure 6 This is a comparison chart of network losses among Scheme I, Scheme II, and Scheme III. DETAILED DESCRIPTION

[0020] The following describes in detail the multi-modal adaptive control method for the edge side of a distribution network based on lightweight sensitivity of the present invention in conjunction with the embodiments and drawings.

[0021] like Figure 1 As shown, the multi-modal adaptive control method for the edge side of the distribution network based on lightweight sensitivity of the present invention includes the following steps:

[0022] 1) According to the selected distribution network, divide the area of ​​each edge computing device and input the distribution network parameters, including the network topology connection relationship of the distribution network, the access location, capacity and parameters of the distributed power supply; input the distribution network N d The historical tide data of a typical day is collected; the total sampling time ΔT and sampling time interval Δt of the training samples are set; the parameters of the dual-channel graph convolutional neural network model are set, including the learning rate γ, the regularization coefficient β, and the number of complete traversals of the training set E;

[0023] 2) Based on the distribution network N in step 1) d Based on the historical tide data of a typical day, a dual-channel graph convolutional neural network model and its training set are built in the cloud, the dual-channel graph convolutional neural network model is trained, and a lightweight sensitivity curve is generated based on the trained dual-channel graph convolutional neural network model.

[0024] The said construction of a dual-channel graph convolutional neural network model and its training set in the cloud includes determining the number of training samples in the training set, the composition of each training sample, and preprocessing the training samples; wherein,

[0025] The number of training samples in the training set is expressed as follows: The training set of the dual-channel graph convolutional neural network model is based on N d The historical tide data of a typical day is constructed, and the total sampling time is set to ΔT and the sampling time interval is Δt. Therefore, the training set of the dual-channel graph convolutional neural network model contains N d ×ΔT / Δt training samples; determine the composition of each training sample, expressed as:

[0026] Each training sample consists of four matrices: feature information matrix X, topology information matrix A, distributed power reactive output information matrix D, and label matrix Z. Channel 1 of the dual-channel graph convolutional neural network model takes the feature information matrix X and topology information matrix A as input, and channel 2 takes the distributed power output information matrix D as input to fit the label matrix Z.

[0027] The distribution network topology information matrix A is used to represent the connection relationship between nodes. The distribution network topology information matrix A of the nth training sample is n Expressed as:

[0028]

[0029]

[0030] Where, is the topological information matrix A of the nth training sample n The element in the i-th row and j-th column of , where N is the total number of nodes in the distribution network;

[0031] The characteristic information matrix X represents the characteristic information of the current training sample. The characteristic information matrix X of the nth training sample n It is composed of the voltage value of the node at the sampling moment and the reactive power value of the line, and is expressed as:

[0032]

[0033] Where, Represents the feature information matrix X of the nth training sample n The element in the i-th row and k-th column, N is the total number of distribution network nodes, and I is the number of input features of the model;

[0034] The distributed power reactive output information matrix D represents the change of the distributed power reactive output in the distribution network. The distributed power reactive output information matrix D of the nth training sample is n Expressed as:

[0035]

[0036] Where, Represents the distributed generation reactive output information matrix D of the nth training sample n The elements in the gth row are the reactive output changes of g distributed power sources; G is the total number of distributed power sources in the distribution network;

[0037] The label matrix Z represents the label value of the output feature of the training sample, that is, the actual distribution network state change, including the node voltage value change and the line reactive power value change. The label matrix Z of the nth training sample is n Expressed as:

[0038]

[0039] Where, Represents the label matrix Z of the nth training sample n The element in the i-th row and the k-th column; N is the total number of nodes in the distribution network; O is the number of output features;

[0040] The preprocessing of the training samples is expressed as:

[0041] Before model training, each element in the feature information matrix X and label matrix Z of the training samples in the training set is preprocessed;

[0042] Among them, the feature information matrix X of the nth training sample n The preprocessing is expressed as:

[0043]

[0044] Where, is the feature information matrix X of the nth training sample before preprocessing n The kth input feature of the i-th node in , is the feature information matrix X of the nth training sample after preprocessing n The kth input feature of the i-th node in, is the mean of all elements of the kth input feature in the feature information matrix X of the training sample, is the variance of all elements of the kth input feature in the feature information matrix X of the training sample;

[0045] The label matrix Z for the nth training sample n The preprocessing is expressed as:

[0046]

[0047] Where, is the label matrix Z of the nth training sample before preprocessing n The label value of the feature output by the i-th node in , is the label matrix Z of the nth training sample after preprocessing n The label value of the output feature of the i-th node, μ Z is the mean of all elements of the output feature in the label matrix Z of the training sample, δ Z The variance of all elements of the output features in the label matrix Z of the training sample.

[0048] The dual-channel graph convolutional neural network model is expressed as:

[0049]

[0050] F (θ+1) =σ(K (θ) F (θ) +b (θ) ) (9)

[0051] σ(x)=sigmoid(x)=1 / (1+e -x ) (10)

[0052] Where H (θ+1) is the output of the θ+1th hidden layer of channel 1 of the dual-channel graph convolutional neural network model; H (θ) is the output of the θth hidden layer of channel 1 of the dual-channel graph convolutional neural network model; A is the topological information matrix, is the normalized topological information matrix, I N is the N-order identity matrix, N is the total number of distribution network nodes, and N is the total number of distribution network nodes; is a diagonal matrix, for The diagonal elements of the matrix, is the element in row i and column j of the normalized topological information matrix; W (θ) is the weight matrix of the θth layer of channel 1 of the dual-channel graph convolutional neural network model; F (θ+1) is the output of the θ+1th hidden layer of channel 2 of the dual-channel graph convolutional neural network model; F (θ) is the output of the θth hidden layer of channel 2 of the dual-channel graph convolutional neural network model; K (θ) is the weight matrix of the θth layer of channel 2 of the dual-channel graph convolutional neural network model; b (θ) is the bias matrix of the θth layer of channel 2 of the dual-channel graph convolutional neural network model; σ(·) is the nonlinear activation function, and sigmoid(·) is the hyperbolic sine activation function;

[0053] The dual-channel graph convolutional neural network model is trained in a supervised manner, and the mean square error is used as the loss function to measure the fitting error, which is expressed as:

[0054]

[0055] Where, and The fitted value and true value of the feature are output for the nth training sample in the i-th row and the k-th column respectively, N C is the number of training samples; N is the total number of distribution network nodes; O is the number of output features.

[0056] 3) Sending the lightweight sensitivity curve in step 2) to the edge computing devices at each edge of the distribution network to construct a sensitivity piecewise linearization constraint;

[0057] The sensitivity piecewise linearization constraint is to simplify the lightweight sensitivity curve, which is expressed as:

[0058]

[0059] Where ρ represents the node active power P or node reactive power Q, and Δρ is the power change value; The mathematical expression of the piecewise linearized sensitivity function of power ρ, ρ b is the bth turning point of the piecewise linearized sensitivity function, f ρ (ρ b ) is the true sensitivity function at the segment point ρ b The value of α b , β b is an auxiliary variable; b is the turning point index, and B is the total number of turning points.

[0060] 4) A multi-modal adaptive control model based on lightweight sensitivity is established at each edge of the distribution network, including: setting the maximum improvement effect of the operating mode at the edge of the distribution network as the objective function, considering the sensitivity piecewise linearization constraint, the distribution network safe operation constraint, and the distributed power supply operation constraint respectively; wherein:

[0061] The objective function is to maximize the effect F of the distribution network edge side operation mode improvement, which can be expressed as:

[0062]

[0063] Where, The flag indicating that the edge region e is running mode C at time t, is the optimized operating cost of the edge area e corresponding to the operating mode C at time t, is the operating cost of the edge region e before optimization at time t;

[0064] Since the edge region e can only work in one operating mode at time t, the mode selection constraint is expressed as:

[0065]

[0066] Where, Ω M represents the control mode set; Ω E represents the edge region set; LC, VC, and LB represent the network loss control mode, voltage deviation control mode, and load balancing control mode, respectively; It indicates that the edge area e selects the operating mode C at time t; T represents the total number of time sections;

[0067] The network loss control mode LC is used to reduce the active power loss of the system and the operating cost of the network loss control mode. Expressed as:

[0068]

[0069] Where, represents the set of all branches in the edge area e of the distribution network; r ij is the resistance value of branch ij; c LC is the unit cost of network loss; I ij represents the current amplitude on branch ij, which is calculated using formula (16):

[0070]

[0071] Where, P ij , Q ij Respectively represent the active power and reactive power on the distribution network branch ij; V i represents the voltage amplitude of node i;

[0072] When the system network parameters are not available, the network loss control mode operation cost Calculated using formula (17).

[0073]

[0074] The voltage deviation control mode VC is used to reduce the voltage deviation of the system. The operating cost of the voltage deviation control mode is Expressed as:

[0075]

[0076]

[0077] Where, represents the set of all nodes in the edge region e of the distribution network; is the unit cost of voltage deviation at node i; is the voltage deviation cost of node i in the edge region e; V op 、 They represent the lower limit and upper limit of the dead zone of the voltage control of the distribution network node respectively; P i is the active power of the load on node i;

[0078] Introducing auxiliary variables Linearize equation (19) and express it as:

[0079]

[0080] Auxiliary variables The constraints are expressed as:

[0081]

[0082] The load balancing control mode LB is used to reduce the load imbalance of the system. The load balancing control mode operation cost Expressed as:

[0083]

[0084]

[0085] Where, represents the set of all branches in the edge area e of the distribution network; is the set of downstream nodes of branch ij; is the unit cost of load imbalance of branch ij; is the load imbalance cost of branch ij in edge region e; Indicates the maximum current value of branch ij; Represents the current control threshold of branch ij; I ij represents the current amplitude on branch ij; P m is the load active power of node m;

[0086] Introducing auxiliary variables Linearize equation (22) and express it as:

[0087]

[0088] Auxiliary variables The constraints are expressed as:

[0089]

[0090]

[0091] The distribution network safe operation constraint is expressed as:

[0092] V i min ≤V i ≤V i max (27)

[0093]

[0094] Where V i is the voltage amplitude of node i; V i max and V i min are the upper and lower limits of the voltage safety operation of node i; I ij Represents the current amplitude on branch ij; is the maximum current value of branch ij.

[0095] The distributed power supply operation constraints are expressed as:

[0096]

[0097] Where, and are the active and reactive outputs of the distributed generation at node i; P i max and P i min are the upper and lower limits of active output of distributed generation at node i respectively; and are the upper and lower limits of reactive power output of distributed generation at node i respectively; is the access capacity of the distributed power inverter at node i.

[0098] 5) Solve the multi-modal adaptive control model of the distribution network edge side based on lightweight sensitivity obtained in step 4) and output the solution results, including: the control mode of the distribution network edge side area and the operation control strategy of each distributed power source.

[0099] Examples are given below.

[0100] For this embodiment, the modified IEEE 33-node distribution network example structure is as follows: Figure 2 As shown, first input the distribution network topology connection relationship, distributed power supply access location, capacity and parameters, predicted distributed power supply, load output curve and historical flow data of the distribution network. Detailed parameters are shown in Table 1 and Table 2. Node 12, node 13, node 17, node 21, node 30 and node 32 are connected to a photovoltaic system with an active capacity of 100kW; node 16, node 18, node 22, node 31 and node 33 are connected to a wind turbine with an active capacity of 200kW; the upper and lower limits of the voltage fluctuation range are 1.05 and 0.95 respectively, and the upper limit of the line load rate fluctuation range is 20%; set the number of typical days N d =100, the total sampling time of the training sample ΔT = 24h, and the sampling time interval Δt = 5min.

[0101] In order to fully verify the advanced nature of the method of the present invention, in this embodiment, the following three schemes are adopted for comparative analysis:

[0102] Solution I: No optimization, obtain the initial state of the distribution network;

[0103] Scheme II: Using a graph convolutional neural network to fit voltage sensitivity and line reactive sensitivity, the proposed lightweight sensitivity-based multi-modal adaptive control method at the edge of the distribution network is used to select the edge-side operating control mode and optimize the edge-side control strategy.

[0104] Scheme III: Optimizing distributed power control strategies based on a centralized approach.

[0105] The comparison of the optimization results of Scheme I, Scheme II and Scheme III is shown in Table 3, and the comparison of the system voltage distribution of Scheme I and Scheme II is shown in Figure 3a 、 Figure 3b , the system line load rate distribution of Scheme I and Scheme II is shown in Figure 4a 、 Figure 4b , the results of the operation mode selection of each edge side area are shown in Figure 5 , the network loss comparison of scheme I, scheme II and scheme III is shown in Figure 6 shown.

[0106] The computer hardware environment for performing optimization calculations is an Intel(R) Core(TM) i7-9750H CPU with a main frequency of 2.60 GHz and a memory of 16 GB; the software environment is the Windows 10 operating system.

[0107] Comparing Schemes I and II, it can be seen that without optimization, the system suffers from issues such as node voltage over-limit, excessive network losses, and uneven line load distribution. By optimizing the output of distributed power sources using the lightweight sensitivity-based multimodal adaptive control method at the edge of the distribution network, the present invention can fully leverage the control potential of distributed power sources, reduce system voltage fluctuations, lower system network losses, minimize system load imbalance, and improve system performance.

[0108] From the comparison between Schemes II and III, it can be seen that the lightweight sensitivity-based multi-modal adaptive control method at the edge of the distribution network of the present invention can effectively realize the rapid formulation of the distributed power supply control strategy and obtain an optimization effect similar to that of the centralized method.

[0109] Table 1 Load access location and power of IEEE 33-node example

[0110]

[0111] Table 2 IEEE 33-bus line parameters

[0112]

[0113]

[0114] Table 3 Comparison of simulation results under different control strategies

[0115]

Claims

1. A multi-modal adaptive control method for the edge side of a distribution network based on lightweight sensitivity, characterized in that: The steps include: 1) According to the selected distribution network, divide the area of ​​each edge computing device and input the distribution network parameters; input the distribution network N d The historical tide data of a typical day is collected; the total sampling time ΔT and sampling time interval Δt of the training samples are set; the parameters of the dual-channel graph convolutional neural network model are set, including the learning rate γ, the regularization coefficient β, and the number of complete traversals of the training set E; 2) Based on the distribution network N in step 1) d Based on the historical tide data of a typical day, a dual-channel graph convolutional neural network model and its training set are constructed in the cloud, the dual-channel graph convolutional neural network model is trained, and a lightweight sensitivity curve is generated based on the trained dual-channel graph convolutional neural network model; the dual-channel graph convolutional neural network model is expressed as: F (θ+1) =σ(K (θ) F (θ) +b (θ) ) (9) σ(x)=sigmoid(x)=1 / (1+e -x ) (10) In the formula, H (θ+1) is the output of the θ+1th hidden layer of channel 1 of the dual-channel graph convolutional neural network model; H (θ) is the output of the θth hidden layer of channel 1 of the dual-channel graph convolutional neural network model; A is the topological information matrix, is the normalized topological information matrix, I N is the N-order identity matrix, N is the total number of distribution network nodes, and N is the total number of distribution network nodes; is a diagonal matrix, for The diagonal elements of the matrix, is the element in row i and column j of the normalized topological information matrix; W (θ) is the weight matrix of the θth layer of channel 1 of the dual-channel graph convolutional neural network model; F (θ+1) is the output of the θ+1th hidden layer of channel 2 of the dual-channel graph convolutional neural network model; F (θ) is the output of the θth hidden layer of channel 2 of the dual-channel graph convolutional neural network model; K (θ) is the weight matrix of the θth layer of channel 2 of the dual-channel graph convolutional neural network model; b (θ) is the bias matrix of the θth layer of channel 2 of the dual-channel graph convolutional neural network model; σ(·) is the nonlinear activation function, and sigmoid(·) is the hyperbolic sine activation function; The dual-channel graph convolutional neural network model is trained in a supervised manner, and the mean square error is used as the loss function to measure the fitting error, which is expressed as: Where, and The fitted value and true value of the feature are output for the nth training sample in the i-th row and the k-th column respectively, N C is the number of training samples; N is the total number of distribution network nodes; O is the number of output features; 3) Send the lightweight sensitivity curve in step 2) to the edge computing devices at each edge of the distribution network to construct a sensitivity piecewise linearization constraint; the sensitivity piecewise linearization constraint refers to simplifying the lightweight sensitivity curve, which is expressed as: a b ≤β b +b b-1 ,b=2,3,…,B-1 0≤α b ≤1,β b ∈{0,1} Where ρ represents the node active power P or node reactive power Q, and Δρ is the power change value; The mathematical expression of the piecewise linearized sensitivity function of power ρ, ρ b is the bth turning point of the piecewise linearized sensitivity function, f ρ (ρ b ) is the true sensitivity function at the segment point ρ b The value of α b , β b is an auxiliary variable; b is the turning point index, and B is the total number of turning points; 4) Establishing a multi-modal adaptive control model for the edge of the distribution network based on lightweight sensitivity at each edge of the distribution network, including: setting the maximum improvement effect of the operating mode at the edge of the distribution network as the objective function, and considering the sensitivity piecewise linearization constraint, the distribution network safe operation constraint, and the distributed power generation operation constraint respectively; 5) Solve the multi-modal adaptive control model of the distribution network edge side based on lightweight sensitivity obtained in step 4) and output the solution results, including: the control mode of the distribution network edge side area and the operation control strategy of each distributed power source.

2. The light-weight sensitivity-based multi-modal adaptive control method for the edge side of the distribution network according to claim 1, characterized in that: The distribution network parameters described in step 1) include the network topology connection relationship of the distribution network, the access location, capacity and parameters of the distributed power supply.

3. The light-weight sensitivity-based multi-modal adaptive control method for the edge side of a distribution network according to claim 1, characterized in that: Step 2) constructs a dual-channel graph convolutional neural network model and its training set in the cloud, including determining the number of training samples in the training set, the composition of each training sample, and preprocessing the training samples; wherein, The number of training samples in the training set is expressed as follows: The training set of the dual-channel graph convolutional neural network model is based on N d The historical tide data of a typical day is constructed, and the total sampling time is set to ΔT and the sampling time interval is Δt. Therefore, the training set of the dual-channel graph convolutional neural network model contains N d ×ΔT / Δt training samples; determine the composition of each training sample, expressed as: Each training sample consists of four matrices: feature information matrix X, topology information matrix A, distributed power reactive output information matrix D, and label matrix Z. Channel 1 of the dual-channel graph convolutional neural network model takes the feature information matrix X and topology information matrix A as input, and channel 2 takes the distributed power output information matrix D as input to fit the label matrix Z. The distribution network topology information matrix A is used to represent the connection relationship between nodes. The distribution network topology information matrix A of the nth training sample is n Expressed as: Where, is the topological information matrix A of the nth training sample n The element in the i-th row and j-th column of , where N is the total number of nodes in the distribution network; The characteristic information matrix X represents the characteristic information of the current training sample. The characteristic information matrix X of the nth training sample n It is composed of the voltage value of the node at the sampling moment and the reactive power value of the line, and is expressed as: Where, Represents the feature information matrix X of the nth training sample n The element in the i-th row and k-th column, N is the total number of distribution network nodes, and I is the number of input features of the model; The distributed power reactive output information matrix D represents the change of the distributed power reactive output in the distribution network. The distributed power reactive output information matrix D of the nth training sample is n Expressed as: Where, Represents the distributed generation reactive output information matrix D of the nth training sample n The elements in the gth row are the reactive output changes of g distributed power sources; G is the total number of distributed power sources in the distribution network; The label matrix Z represents the label value of the output feature of the training sample, that is, the actual distribution network state change, including the node voltage value change and the line reactive power value change. The label matrix Z of the nth training sample is n Expressed as: Where, Represents the label matrix Z of the nth training sample n The element in the i-th row and the k-th column; N is the total number of nodes in the distribution network; O is the number of output features; The preprocessing of the training samples is expressed as: Before model training, each element in the feature information matrix X and label matrix Z of the training samples in the training set is preprocessed; Among them, the feature information matrix X of the nth training sample n The preprocessing is expressed as: Where, is the feature information matrix X of the nth training sample before preprocessing n The kth input feature of the i-th node in, is the feature information matrix X of the nth training sample after preprocessing n The kth input feature of the i-th node in, is the mean of all elements of the kth input feature in the feature information matrix X of the training sample, is the variance of all elements of the kth input feature in the feature information matrix X of the training sample; The label matrix Z for the nth training sample n The preprocessing is expressed as: Where, is the label matrix Z of the nth training sample before preprocessing n The label value of the feature output by the i-th node in , is the label matrix Z of the nth training sample after preprocessing n The label value of the output feature of the i-th node, μ Z is the mean of all elements of the output feature in the label matrix Z of the training sample, δ Z The variance of all elements of the output features in the label matrix Z of the training sample.

4. The light-weight sensitivity-based multi-modal adaptive control method for the edge side of a distribution network according to claim 1, characterized in that: In step 4), the objective function is to set the maximum improvement effect F of the distribution network edge side operation mode, which can be expressed as: Where, The flag indicating that the edge region e is running mode C at time t, is the optimized operating cost of the edge area e corresponding to the operating mode C at time t, is the operating cost of the edge region e before optimization at time t; Since the edge region e can only work in one operating mode at time t, the mode selection constraint is expressed as: Where, Ω M represents the control mode set; Ω E represents the edge region set; LC, VC, and LB represent the network loss control mode, voltage deviation control mode, and load balancing control mode, respectively; It indicates that the edge area e selects the operating mode C at time t; T represents the total number of time sections; The network loss control mode LC is used to reduce the active power loss of the system and the operating cost of the network loss control mode. Expressed as: Where, represents the set of all branches in the edge area e of the distribution network; r ij is the resistance value of branch ij; c LC is the unit cost of network loss; I ij represents the current amplitude on branch ij, which is calculated using formula (16): Where, P ij , Q ij Respectively represent the active power and reactive power on the distribution network branch ij; V i represents the voltage amplitude of node i; When the system network parameters are not available, the network loss control mode operation cost Calculated using formula (17): The voltage deviation control mode VC is used to reduce the voltage deviation of the system. The operating cost of the voltage deviation control mode is Expressed as: Where, represents the set of all nodes in the edge region e of the distribution network; is the unit cost of voltage deviation at node i; is the voltage deviation cost of node i in the edge region e; V o p 、 They represent the lower limit and upper limit of the dead zone of voltage control at the distribution network node respectively; P i is the active power of the load on node i; Introducing auxiliary variables Linearize equation (19) and express it as: Auxiliary variables The constraints are expressed as: The load balancing control mode LB is used to reduce the load imbalance of the system. The load balancing control mode operation cost Expressed as: Where, represents the set of all branches in the edge area e of the distribution network; is the set of downstream nodes of branch ij; is the unit cost of load imbalance of branch ij; is the load imbalance cost of branch ij in edge region e; Indicates the maximum current value of branch ij; Represents the current control threshold of branch ij; I ij represents the current amplitude on branch ij; P m is the load active power of node m; Introducing auxiliary variables Linearize equation (22) and express it as: Auxiliary variables The constraints are expressed as:

5. The light-weight sensitivity-based multi-modal adaptive control method for the edge side of a distribution network according to claim 1, characterized in that: The distribution network safe operation constraint described in step 4) is expressed as: Where V i is the voltage amplitude of node i; and are the upper and lower limits of the voltage safety operation of node i; I ij Represents the current amplitude on branch ij; is the maximum current value of branch ij.

6. The light-weight sensitivity-based multi-modal adaptive control method for the edge side of a distribution network according to claim 1, characterized in that: The distributed power supply operation constraints described in step 4) are expressed as: Where, and are the active and reactive outputs of the distributed generation at node i respectively; and are the upper and lower limits of active output of distributed generation at node i respectively; and are the upper and lower limits of reactive power output of distributed generation at node i respectively; is the access capacity of the distributed power inverter at node i.

Images