Optimized scheduling method, system and equipment for 5G base station and active power distribution network, and medium

Abstract

The invention provides an optimal scheduling method, system and device for a 5G base station and an active power distribution network and a medium, an improved space-time diagram convolutional neural network is established, a diagram convolution method is combined with a gating circulation unit, and the flexible load adjustment potential of the 5G base station is deeply excavated; through an adjacent matrix adaptive correction method, spatial relevance among different power stations is automatically learned based on historical data, the utilization value of key data is improved, and the prediction precision is improved; based on prediction results of the 5G base station load rate, the 5G base station power and the photovoltaic output, calculating the 5G base station energy storage real-time minimum reserve capacity, establishing a target function by taking the optimal operation cost as a target, and decomposing the solving problem of the target function into main and sub problems for solving by using an improved column appointment generation algorithm to obtain a scheduling plan of the 5G base station and the active power distribution network. The time-space relevance of different 5G base station groups is efficiently extracted, and the economical efficiency of each base station group is effectively improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of energy management and relates to an optimization scheduling method, system, equipment and medium. Background Art

[0002] The energy storage capacity of a single macro base station of a 5G base station is 4-16kWh, and the maximum adjustable power is 2-4kW / h. The massive 5G base stations contain huge adjustable resources. The optimized scheduling of 5G base stations and active distribution networks is a key link in the energy conservation and consumption reduction of new infrastructure. 5G base stations can effectively support the safe operation of the power grid and achieve mutual benefit and win-win results between communication operators and power grids by interacting with the power grid and tapping the energy-saving potential. With the construction and development of 5G base stations, the number of 5G base stations has increased year by year, forming a massive distributed flexible resource. However, in the optimized scheduling of 5G base stations and active distribution networks, 5G base stations have problems such as insufficient exploration of load flexible adjustment potential and insufficient ability to interact with the power grid. If a centralized scheduling method is adopted, it will face user privacy protection issues and a huge computing burden. Summary of the invention

[0003] In order to solve the problems described in the background technology, such as insufficient exploration of the flexible load adjustment potential of 5G base stations, insufficient ability to coordinate and interact with the power grid, user privacy protection issues and huge computing burden faced by the use of centralized scheduling methods, the present invention provides an optimized scheduling method, system, equipment and medium for 5G base stations and active distribution networks.

[0004] The method of the present invention comprises:

[0005] Based on the original input data of 5G base stations and active distribution network optimization scheduling, the graph convolution and gated recurrent unit are combined to establish a spatiotemporal graph convolutional neural network. The historical operation data of 5G base stations and active distribution networks are used to learn the relationship between the nodes in the graph structure of the graph convolution layer in the spatiotemporal graph convolutional neural network. The adjacency matrix parameters of the graph convolution layer in the spatiotemporal graph convolutional neural network are adaptively corrected to obtain an improved spatiotemporal graph convolutional neural network.

[0006] The soft attention mechanism and the scaled dot product scoring function are used to calculate the weights of different input features in the original input data of the optimal scheduling of 5G base stations and active distribution networks, and the corrected input features are obtained;

[0007] The corrected input features are input into the improved spatiotemporal graph convolutional neural network, and the output is the predicted load rate of 5G base stations, the predicted power of 5G base stations, and the predicted output of photovoltaics.

[0008] Based on the predicted power of the 5G base station and the backup power time of the 5G base station, the real-time minimum backup capacity of the 5G base station energy storage is calculated;

[0009] According to the operation status of 5G base stations and active distribution networks, establish the objective function and operation constraints with the goal of optimizing the operation cost;

[0010] Based on the real-time minimum backup capacity of 5G base station energy storage, the predicted load rate of 5G base stations, the predicted photovoltaic output and operating constraints, the objective function is solved through the column constraint generation algorithm to obtain the scheduling plan of 5G base stations and active distribution networks. According to the scheduling plan, the 5G base stations and active distribution networks are optimally scheduled.

[0011] Furthermore, the original input data for the optimal scheduling of 5G base stations and active distribution networks include the operation data of 5G base stations, the geographical distribution data of 5G base stations, the meteorological data of 5G base stations, and the historical operation data of 5G base stations and active distribution networks;

[0012] The spatiotemporal graph convolutional neural network model is represented as a topological graph G = (V, E, A), where V represents the point set, E represents the edge set, A is the adjacency matrix of the topological graph G, and the convolution range is node v i Neighborhood B:

[0013] B(v i )={v j |d(v j ,v i )≤D}(1),

[0014] In the formula, d(v j ,v i ) represents the graph node v i to v j The shortest distance; D is the maximum convolution neighborhood; when D = 1, it indicates that the size of the sampling area is node v i The first-order neighborhood of ;

[0015] The forward propagation of the graph convolutional layer is as follows:

[0016]

[0017] In the formula, ReLU refers to the rectified linear unit activation function; X is the feature matrix of the topological graph G; is the adjacency matrix with autocorrelation, that is A is the adjacency matrix of the topological graph G, and I is the identity matrix; for The element in row i and column j; is the adjacency matrix after normalization and symmetry; σ(·) is the sigmoid function; W0 is the weight matrix from the upper layer to the graph convolution layer; W1 is the weight matrix from the graph convolution layer to the lower layer, is the node degree matrix;

[0018] The graph convolution and gated recurrent unit are combined, and the graph convolution algorithm is introduced into the gated recurrent unit to form a spatiotemporal graph convolution unit. A spatiotemporal graph convolution neural network is constructed by stacking multiple spatiotemporal graph convolution units to capture the spatiotemporal dynamic changes of nodes in the graph and conduct deep learning and analysis of spatiotemporal graph data. The forward propagation formula is shown in the following equations (4)-(7):

[0019] u t =σ(W u [f(A,X t ),h t-1 ]+b u ) (4),

[0020] r t =σ(W r [f(A,X t ),h t-1 ]+b r ) (5),

[0021] c t =tanh(W c [f(A,X t ),(r t *h t-1 )]+b c ) (6),

[0022] h t =u t *h t-1 +(1-u t )*c t (7)

[0023] In the formula, u t 、r t 、c t are the internal parameters of the unit, corresponding to the update gate, reset gate, and the state of the current unit at time t; h t is the unit output at time t; W u With b u are the weight matrix and bias matrix of the update gate respectively; W r With b r is the weight matrix and bias matrix of the reset gate; f(A,X t ) is the graph convolution operation on the model input at time t; h t-1 is the hidden state at the previous time t-1; W c is used to calculate the candidate state c t The weight matrix of c is the candidate state c t the associated bias vector;

[0024] The method for adaptively correcting the adjacency matrix parameters of the graph convolution layer in the spatiotemporal graph convolutional neural network is:

[0025] The characteristic matrix X of the topological graph G = (x1, x2, ... x N ) T , define A mn is the parameter of the mth row and nth column of the adjacency matrix, A mn The expression is:

[0026]

[0027] In the formula, x m and x n represents the characteristics of node m and node n, ω is the learnable weight vector in the adjacency matrix parameters, and N is the total number of first-order neighbors of node m;

[0028] Formula (8) uses the ReLU unit to limit the value of the adjacency matrix to non-negative numbers, and uses the softmax form to convert A mn Perform normalization;

[0029] Using the following loss function L graph Study A mn The weight parameters are:

[0030]

[0031] In the formula, λ is the penalty coefficient of the regularization term, F is the norm of the regularization term, and its expression when F-2 regularization is used is:

[0032]

[0033] The total loss function L of the adaptive neural network loss As shown below:

[0034] L loss =L mae +L graph (11)

[0035] Where, L mae is the loss function of the original neural network, using the mean absolute error.

[0036] Furthermore, the method for obtaining the modified input features is:

[0037] When the time period of the input feature is T, its feature weight vector α T for:

[0038]

[0039] Where X is the original input feature; Q is the query vector of the attention mechanism, that is, the predicted target value at the corresponding time; D T is the total length of the input feature; α T Calculated by the soft attention mechanism, the input features are normalized using the scaled dot product scoring function to ensure that the weights of different features are reasonably adjusted, thereby optimizing the input features of the model;

[0040] All input features are standardized to obtain the corrected input features, expressed as:

[0041]

[0042] Where, X ref is the corrected input feature, X mean is the mean of each characteristic component of input X; std is the standard deviation of each characteristic component of input X;

[0043] According to formula (13), the input features are standardized to generate the corrected input feature dataset X ref , which is the basis for subsequent predictions; the corrected input feature dataset X ref The data is input into the improved spatiotemporal graph convolutional neural network model, and the 5G base station load rate, power and photovoltaic output in different time periods are predicted using multi-layer convolution operations and spatiotemporal dependency analysis.

[0044] The 5G base station predicted load rate λ 5G It is expressed as the following formula:

[0045] λ 5G ={λ1(t),λ2(t),…,λ N (t)} (14),

[0046] In the formula, λ N (t) represents the load rate of the Nth 5G base station at time t;

[0047] The 5G base station predicted power P 5G It is expressed as the following formula:

[0048] P 5G = {P1(t),P2(t),…,P N (t)} (15),

[0049] Where P N (t) represents the power of the Nth 5G base station at time t;

[0050] The photovoltaic predicted output P PV It is expressed as the following formula:

[0051] P PV = {P1(t),P2(t),…,P M (t)} (16),

[0052] Where P M (t) represents the power of the Mth 5G base station at time t.

[0053] Furthermore, in the calculation of the real-time minimum backup capacity of the 5G base station energy storage, the predicted power of the 5G base stations in different areas obtained by formula (15) and the minimum backup power time t of the 5G base station are combined. s Combined, we can get the minimum spare capacity of 5G base stations at time t As shown below:

[0054]

[0055] Where P 5G Predicting power for 5G base stations.

[0056] Furthermore, the objective function is expressed as:

[0057]

[0058] In the formula, It represents the cost of short-time load power shortage; Indicates the charging and discharging cost of base station energy storage; represents the photovoltaic output cost;

[0059] The specific expression is as follows:

[0060]

[0061] In the formula, c LD 、c PV They represent the unit cost of load power shortage and the unit cost of photovoltaic power generation respectively; c 5G 、c SH They represent the unit charging and discharging cost and loss cost of 5G base station energy storage respectively; P t ch , P t dis Respectively represent the charge and discharge power at time t; P t PV Represents photovoltaic output; η 5G represents the charging and discharging efficiency of G base station energy storage; P t ΔLD Indicates the load power shortage;

[0062] The operation constraints include 5G base station energy storage charging and discharging constraints, photovoltaic output constraints and power balance constraints;

[0063] The expression of the 5G base station energy storage charging and discharging constraint is:

[0064]

[0065] Where P t ch and P t dis They represent the charging power and discharging power at time t respectively, and They represent the maximum charging power and the maximum discharging power at time t, respectively, and E 5G Indicates the energy storage capacity of 5G base stations, and They represent the minimum and maximum energy storage capacity of the 5G base station respectively;

[0066] The expression of the photovoltaic output constraint is:

[0067]

[0068] Where P t PV Represents photovoltaic output, Indicates the maximum photovoltaic output;

[0069] The expression of the power balance constraint is:

[0070] P t ΔLD =P t ch +P t LD -P t PV -P t dis (twenty two),

[0071] Where P t LD represents the load power of the distribution network at time t, P t ch and P t dis They represent the charging power and discharging power at time t, respectively, P t PV Indicates photovoltaic output.

[0072] Furthermore, the objective function is solved as follows:

[0073] The uncertainty of photovoltaic and load is expressed by set Ω, as shown in the following formula:

[0074]

[0075] In the formula, is the actual output of the photovoltaic system at time t; is the predicted output of the photovoltaic system at time t; Indicates the maximum value of the uncertainty range of the photovoltaic system output; represents the actual load demand at time t; Indicates the maximum value of the load power uncertainty range; represents the predicted load demand at time t;

[0076] Formulas (18)-(22) are expressed as an uncertain two-stage robust optimization model using compact matrix equations. The first stage is to minimize the system operation cost, and the second stage is to obtain the worst operation scenario under the minimum cost. The specific expression is shown in the following formula:

[0077]

[0078] In the formula, c T x represents the optimal objective function of operating cost; d T y represents the worst operating cost when the load loses power; u represents the uncertain variables including photovoltaic and load power; max represents the worst scenario with the highest cost in the optimal scheduling; H, F, and G represent three different coefficient matrices respectively; y| represents the corresponding state variable or decision variable in the column constraint generation algorithm;

[0079] Based on the real-time minimum spare capacity model of 5G base station energy storage, the predicted load rate of 5G base stations, the predicted output of photovoltaic power generation and operation constraints, the column constraint generation algorithm is used to solve the uncertain two-stage robust optimization model. The objective function of the robust optimization model is first decoupled and decomposed into the main problem and sub-problems.

[0080] The main problem is:

[0081]

[0082] In the formula, m represents the number of iterations; y m represents the solution to the subproblem after m iterations.

[0083] The sub-problems are:

[0084]

[0085] By giving the (x,u) value, the minimization problem is transformed into a deterministic linear optimization problem, which is expressed as follows:

[0086]

[0087] Where, d and y represent column vectors; A1, A2, A3 represent m1, m2, m3-column n-row matrices; b1, b2, b3 represent m1, m2, m3-dimensional column vectors;

[0088] Using the duality theory, the inner min model in the sub-problem of the above robust optimization model is transformed into a max model and merged with the outer max model. The transformation method is:

[0089] If the original question is:

[0090]

[0091] The dual form is:

[0092]

[0093] Where z1, z2, z3 means m1, m2, m3 are row vectors;

[0094] According to the principle of duality theory, the dual form is:

[0095]

[0096] Based on formulas (24)-(30), the scheduling plan of the 5G base station and the active distribution network is obtained, and the scheduling plan includes: the scheduling plan of the 5G base station and the output plan of the photovoltaic power generation.

[0097] The present invention proposes an optimization scheduling system for 5G base stations and active distribution networks, including an improved spatiotemporal graph convolutional neural network establishment module, an input feature correction module, a prediction result calculation module, a 5G base station energy storage real-time minimum spare capacity calculation module, an objective function and constraint establishment module, and an optimization scheduling module.

[0098] The improved spatiotemporal graph convolutional neural network establishment module is used to combine graph convolution and gated recurrent units to establish a spatiotemporal graph convolutional neural network based on the original input data of the optimal scheduling of 5G base stations and active distribution networks, use the historical operation data of 5G base stations and active distribution networks to learn the relationship between each node in the graph structure of the graph convolution layer in the spatiotemporal graph convolutional neural network, adaptively correct the adjacency matrix parameters of the graph convolution layer in the spatiotemporal graph convolutional neural network, and obtain an improved spatiotemporal graph convolutional neural network.

[0099] The input feature correction module is used to calculate the weights of different input features in the original input data of the optimal scheduling of 5G base stations and active distribution networks by using a soft attention mechanism and a scaled dot product scoring function to obtain the corrected input features.

[0100] The prediction result calculation module is used to input the corrected input features into the improved spatiotemporal graph convolutional neural network, and output the 5G base station predicted load rate, 5G base station predicted power and photovoltaic predicted output.

[0101] The 5G base station energy storage real-time minimum spare capacity calculation module is used to calculate the 5G base station energy storage real-time minimum spare capacity based on the 5G base station predicted power and the 5G base station backup power time.

[0102] The objective function and constraint establishment module is used to establish an objective function and operating constraints with the goal of optimizing operating costs based on the operating conditions of the 5G base station and the active distribution network.

[0103] The optimization scheduling module is used to solve the objective function through a column constraint generation algorithm based on the real-time minimum spare capacity of the 5G base station energy storage, the predicted load rate of the 5G base station, the predicted photovoltaic output and the operating constraints, to obtain the scheduling plan of the 5G base station and the active distribution network, and optimize the scheduling of the 5G base station and the active distribution network according to the scheduling plan.

[0104] The present invention also proposes an electronic device, comprising: a memory and a processor, the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor implements the above-mentioned optimization scheduling method of 5G base stations and active distribution networks by executing the computer instructions.

[0105] The present invention also proposes a computer-readable storage medium, which stores a computer program, and is characterized in that when the computer program is executed by a processor, it implements the above-mentioned optimization scheduling method for 5G base stations and active distribution networks.

[0106] Compared with the prior art, the present invention establishes an improved spatiotemporal graph convolutional neural network, combines the graph convolution method with the gated recurrent unit, takes into account the advantages of graph convolution in extracting spatial features and the gated recurrent unit in extracting temporal features, explores the spatiotemporal dependence of 5G base stations and photovoltaic output, and deeply explores the potential for flexible load adjustment of 5G base stations; through the adjacency matrix adaptive correction method, automatically learns the spatial correlation between different power stations based on historical data, and weights the input data through the soft attention mechanism and the scaled dot product scoring function to improve the utilization value of key data and improve the prediction accuracy; based on the prediction results of 5G base station load rate, 5G base station power and photovoltaic output, calculates the real-time minimum spare capacity of 5G base station energy storage, establishes the objective function with the goal of optimizing the operating cost, and uses the improved column reduction generation algorithm to decompose the solution problem of the objective function into main sub-problems for solution, and obtains the scheduling plan of 5G base stations and active distribution networks, ensuring efficient extraction of the spatiotemporal correlation of different 5G base station groups and effectively improving the economy of each base station group. The present invention improves the potential for flexible load adjustment of 5G base stations, solves the problems of user privacy protection and huge computational burden faced by centralized scheduling methods, and greatly improves the prediction accuracy of 5G base station load rate, 5G base station power and photovoltaic output, thereby effectively evaluating the real-time dispatchable capacity of each base station energy storage group, achieving efficient and friendly interaction between 5G base stations and active distribution networks, and reducing the operating costs of 5G base stations and active distribution networks. BRIEF DESCRIPTION OF THE DRAWINGS

[0107] Figure 1 The present invention is a flow chart of the method.

[0108] Figure 2 Schematic diagram of the architecture of the improved spatiotemporal graph convolutional neural network.

[0109] Figure 3 This is a system architecture diagram of the present invention.

[0110] Figure 4 This is a schematic diagram of a 5G base station and an active distribution network in Example 5.

[0111] Figure 5 are the adjacency matrices of different methods.

[0112] Figure 6 The prediction results of different methods.

[0113] Figure 7 Predict 5G base station load rates in different regions and the dispatchable capacity of energy storage.

[0114] Figure 8 Schematic diagram of the cluster division of distribution network 5G base stations and photovoltaics.

[0115] Fig. 9 It is the change of power loss of distribution network load.

[0116] Fig.10 A scheduling plan for each cluster.

[0117] Fig.11 The figure shows the comparison of the overall operating cost of the system using the random optimization method, the traditional robust optimization method and the present invention. DETAILED DESCRIPTION

[0118] In order to make the technical problems, technical solutions and beneficial effects to be solved by the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0119] Example 1

[0120] The optimization scheduling method of 5G base stations and active distribution networks is shown in the flowchart Figure 1 As shown, the specific steps are described as follows.

[0121] Firstly, based on the original input data of 5G base stations and active distribution network optimization scheduling, the spatiotemporal graph convolutional neural network is established by combining graph convolution and gated recurrent unit. The historical operation data of 5G base stations and active distribution networks are used to learn the relationship between the nodes in the graph structure of the graph convolution layer in the spatiotemporal graph convolutional neural network, and the adjacency matrix parameters of the graph convolution layer in the spatiotemporal graph convolutional neural network are adaptively corrected to obtain an improved spatiotemporal graph convolutional neural network.

[0122] Specifically, the original input data for the optimal scheduling of 5G base stations and active distribution networks includes the operating data of 5G base stations, the geographical distribution data of 5G base stations, the meteorological data of 5G base stations, and the historical operating data of 5G base stations and active distribution networks.

[0123] The architecture diagram of the improved spatiotemporal graph convolutional neural network is as follows Figure 2 As shown in the figure, in the improved spatiotemporal graph convolutional neural network, compared with the traditional convolutional neural network, the spatiotemporal graph convolutional neural network generalizes the convolutional network to the unstructured feature space and can process graph structure data of any shape. The spatiotemporal graph convolutional neural network model is represented as a topological graph G = (V, E, A), where V represents the point set, E represents the edge set, A is the adjacency matrix of the topological graph G, and the convolution range is node v i Neighborhood B:

[0124] B(v i )={v j |d(v j ,v i )≤D}(1),

[0125] In the formula, d(v j ,vi ) represents the graph node v i to v j The shortest distance; D is the maximum convolution neighborhood; when D = 1, it indicates that the size of the sampling area is node v i The first-order neighborhood of ;

[0126] At this time, the forward propagation of the graph convolution layer is as follows:

[0127]

[0128] In the formula, ReLU refers to the rectified linear unit activation function; X is the feature matrix of the topological graph G; is the adjacency matrix with autocorrelation, that is A is the adjacency matrix of the topological graph G, and I is the identity matrix; for The element in row i and column j; is the adjacency matrix after normalization and symmetry; σ(·) is the sigmoid function; W0 is the weight matrix from the upper layer to the graph convolution layer; W1 is the weight matrix from the graph convolution layer to the lower layer, is the node degree matrix.

[0129] The graph convolution method is used to convert the historical load rate and geographical distribution characteristics of 5G base stations into node parameters of the graph. The hidden spatial correlation of different 5G base station power timing curves and photovoltaic output curves is learned based on the graph convolution layer. In order to further extract the timing characteristics of the 5G base station load rate and photovoltaic output curve, the graph convolution and gated recurrent unit are combined, and the graph convolution algorithm is introduced into the gated recurrent unit to form a spatiotemporal graph convolution unit. The spatiotemporal graph convolution neural network is constructed by stacking multiple spatiotemporal graph convolution units to capture the spatiotemporal dynamic changes of nodes in the graph. The deep learning and analysis of spatiotemporal graph data are shown in the following forward propagation formulas (4)-(7):

[0130] u t =σ(W u [f(A,X t ),h t-1 ]+b u ) (4), r t =σ(W r [f(A,X t ),h t-1 ]+b r ) (5),

[0131] c t =tanh(W c [f(A,X t ),(r t *h t-1 )]+b c ) (6), ht =u t *h t-1 +(1-u t )*c t (7)

[0132] In the formula, u t 、r t 、c t are the internal parameters of the unit, corresponding to the update gate, reset gate, and the state of the current unit at time t; h t is the unit output at time t; W u With b u are the weight matrix and bias matrix of the update gate respectively; W r With b r is the weight matrix and bias matrix of the reset gate; f(A,X t ) is the graph convolution operation on the model input at time t; h t-1 is the hidden state at the previous time t-1; W c is used to calculate the candidate state c t The weight matrix of c is the candidate state c t The associated bias vector.

[0133] The network unit of the spatiotemporal graph convolutional neural network can input graph structure data, and extract the spatial correlation of input features through the graph convolution method, and then extract the temporal correlation of input features based on the gated recurrent unit structure. After the extracted features are sent to the fully connected layer and the output layer, the 5G base station load rate and photovoltaic output prediction results can be output. Based on the 5G base station load rate, the 5G base station power prediction result can be obtained.

[0134] The adjacency matrix parameters of the graph convolution layer represent the spatial correlation of different prediction sites. In the prediction of multiple 5G base station load rates, 5G base station power and photovoltaic output, in order to more accurately learn the spatial relationship between the output timing curves of different base stations and photovoltaic power stations, the adjacency matrix parameters are adaptively corrected. During the training process, massive historical data samples are used to learn the relationship between each node in the graph structure, thereby adaptively correcting the adjacency matrix parameters.

[0135] Specifically, the method for adaptively correcting the adjacency matrix parameters of the graph convolution layer in the spatiotemporal graph convolutional neural network is:

[0136] The characteristic matrix X of the topological graph G = (x1, x2, ... x N ) T , define A mn is the parameter of the mth row and nth column of the adjacency matrix, A mn The expression is:

[0137]

[0138] In the formula, x m and x n represents the characteristics of node m and node n, ω is the learnable weight vector in the adjacency matrix parameters, and N is the total number of first-order neighbors of node m.

[0139] Formula (8) uses the ReLU unit to limit the value of the adjacency matrix to non-negative numbers, and uses the softmax form to convert A mn Normalization according to the line;

[0140] Using the following loss function L graph Study A mn The weight parameters are:

[0141]

[0142] In the formula, λ is the penalty coefficient of the regularization term, F is the norm of the regularization term, and its expression when F-2 regularization is used is:

[0143]

[0144] When x m With x n The greater the difference between the output timing curves of the photovoltaic power stations represented, the smaller their connection in space will be, which will make the adjacency matrix A mn The smaller the value of is, the smaller the loss function is. In addition, a regular term is added to the loss function to avoid falling into a zero weight solution. The total loss function L of the adaptive neural network is loss As shown below:

[0145] L loss =L mae +L graph (11)

[0146] Where, L mae is the loss function of the original neural network, using the mean absolute error.

[0147] Then, the soft attention mechanism and the scaled dot product scoring function are used to calculate the weights of different input features in the original input features to obtain the modified input features.

[0148] Specifically, the method for obtaining the corrected input features is:

[0149] When the time period of the input feature is T, its feature weight vector α T for:

[0150]

[0151] Where X is the original input feature; Q is the query vector of the attention mechanism, that is, the predicted target value at the corresponding time; DT is the total length of the input feature; α T Calculated by the soft attention mechanism, the input features are normalized using the scaled dot product scoring function to ensure that the weights of different features are reasonably adjusted, thereby optimizing the input features of the model;

[0152] All input features are standardized to obtain the corrected input features, expressed as:

[0153]

[0154] Where, X ref is the corrected input feature, X mean is the mean of each characteristic component of input X; std is the standard deviation of each characteristic component of input X;

[0155] According to formula (13), the input features are standardized to generate the corrected input feature dataset X ref , which is the basis for subsequent predictions; the corrected input feature dataset X ref The data is input into the improved spatiotemporal graph convolutional neural network model, and the 5G base station load rate, power and photovoltaic output in different time periods are predicted using multi-layer convolution operations and spatiotemporal dependency analysis.

[0156] The predicted load rate of 5G base stations is expressed as follows:

[0157] λ 5G ={λ1(t),λ2(t),…,λ N (t)} (14),

[0158] In the formula, λ N (t) represents the load rate of the Nth 5G base station at time t;

[0159] The predicted power of 5G base station is expressed as follows:

[0160] P 5G = {P1(t),P2(t),…,P N (t)} (15),

[0161] Where P N (t) represents the power of the Nth 5G base station at time t;

[0162] The photovoltaic predicted output is expressed as follows:

[0163] P PV = {P1(t),P2(t),…,P M (t)} (16),

[0164] Where PM (t) represents the power of the Mth 5G base station at time t.

[0165] Next, based on the predicted power of the 5G base station and the backup power time of the 5G base station, the real-time minimum backup capacity of the 5G base station energy storage is calculated.

[0166] The minimum backup power capacity of 5G base station energy storage is related to the power consumption of base station equipment and the minimum backup time. For the power consumption of base station equipment, such as a macro base station with a typical configuration of 3 active antenna units (AAU) and 1 baseband unit (BBU), the base station power consumption includes two parts: dynamic power consumption and static power consumption. The power consumption of a single base station communication equipment can be expressed as: represents the power of the base station at time t; Indicates static power, i.e., power amplifier power in AAU, BBU power, and transmission equipment power; Indicates the maximum dynamic power, that is, the power of the radio frequency unit in the AUU, and its communication load rate with the base station λ t Related.

[0167] Specifically, the predicted power of 5G base stations in different regions obtained by formula (15) and the minimum backup power time t of 5G base stations are s Combined, we can get the minimum spare capacity of 5G base stations at time t As shown below:

[0168]

[0169] Where P 5G Predicting power for 5G base stations.

[0170] Next, based on the operating conditions of 5G base stations and active distribution networks, an objective function and operating constraints with the goal of optimizing operating costs are established.

[0171] Specifically, the expression of the objective function is:

[0172]

[0173] In the formula, It represents the cost of short-time load power shortage; Indicates the charging and discharging cost of base station energy storage; represents the photovoltaic output cost;

[0174] The specific expression is as follows:

[0175]

[0176] In the formula, c LD 、c PVThey represent the unit cost of load power shortage and the unit cost of photovoltaic power generation respectively; c 5G 、c SH They represent the unit charging and discharging cost and loss cost of 5G base station energy storage respectively; P t ch , P t dis Respectively represent the charge and discharge power at time t; P t PV Represents photovoltaic output; η 5G represents the charging and discharging efficiency of G base station energy storage; P t ΔLD Indicates that the load is short of power.

[0177] Operation constraints include 5G base station energy storage charging and discharging constraints, photovoltaic output constraints, and power balance constraints;

[0178] The expression of 5G base station energy storage charging and discharging constraint is:

[0179]

[0180] Where P t ch and P t dis They represent the charging power and discharging power at time t respectively, and They represent the maximum charging power and the maximum discharging power at time t, respectively, and E 5G Indicates the energy storage capacity of 5G base stations, and They represent the minimum and maximum energy storage capacity of the 5G base station respectively;

[0181] The expression of photovoltaic output constraint is:

[0182]

[0183] Where P t PV Represents photovoltaic output, Indicates the maximum photovoltaic output.

[0184] The expression of power balance constraint is:

[0185] P t ΔLD =P t ch +P t LD -P t PV -P t dis (twenty two),

[0186] Where P t LD represents the load power of the distribution network at time t, P t ch and P t dis They represent the charging power and discharging power at time t, respectively, P t PV Indicates photovoltaic output.

[0187] Finally, based on the real-time minimum backup capacity of 5G base station energy storage, the predicted load rate of 5G base stations, the predicted photovoltaic output and operating constraints, the objective function is solved through the column constraint generation algorithm to obtain the scheduling plan of 5G base stations and active distribution networks. The 5G base stations and active distribution networks are then optimized according to the scheduling plan.

[0188] Specifically, the solution method of the objective function is:

[0189] The uncertainty of photovoltaic and load is expressed by set Ω, as shown in the following formula:

[0190]

[0191] In the formula, is the actual output of the photovoltaic system at time t; is the predicted output of the photovoltaic system at time t; Indicates the maximum value of the uncertainty range of the photovoltaic system output; represents the actual load demand at time t; Indicates the maximum value of the load power uncertainty range; Represents the predicted load demand at time t.

[0192] Formulas (18)-(22) are expressed as an uncertain two-stage robust optimization model using compact matrix equations. The first stage is to minimize the system operation cost, and the second stage is to obtain the worst operation scenario under the minimum cost. The specific expression is shown in the following formula:

[0193]

[0194] In the formula, c T x represents the optimal objective function of operating cost; d T y represents the worst operating cost when the load loses power; u represents the uncertain variables including photovoltaic and load power; max represents the worst scenario with the highest cost in the optimal scheduling; H, F, and G represent three different coefficient matrices respectively; y| represents the corresponding state variable or decision variable in the column constraint generation algorithm.

[0195] Based on the real-time minimum spare capacity model of 5G base station energy storage, the predicted load rate of 5G base stations, the predicted output of photovoltaic power generation and operation constraints, the column constraint generation algorithm is used to solve the uncertain two-stage robust optimization model. The objective function of the robust optimization model is first decoupled and decomposed into the main problem and sub-problems.

[0196] The main problem is:

[0197]

[0198] In the formula, m represents the number of iterations; y m represents the solution to the subproblem after m iterations.

[0199] The sub-problems are:

[0200]

[0201] By giving the (x,u) value, the minimization problem is transformed into a deterministic linear optimization problem, which is expressed as follows:

[0202]

[0203] Where, d and y represent column vectors; A1, A2, A3 represent m1, m2, m3-column n-row matrices; b1, b2, b3 represent m1, m2, m3-dimensional column vectors;

[0204] Using the duality theory, the inner min model in the sub-problem of the above robust optimization model is transformed into a max model and merged with the outer max model. The transformation method is:

[0205] If the original question is:

[0206]

[0207] The dual form is:

[0208]

[0209] Where z1, z2, z3 means m1, m2, m3 are row vectors;

[0210] According to the principle of duality theory, the dual form is:

[0211]

[0212] Based on the above formulas (24)-(30), the scheduling plan of the 5G base station and the active distribution network is obtained, and the scheduling plan includes: the scheduling plan of the 5G base station and the output plan of photovoltaics.

[0213] Example 2

[0214] The optimized dispatching system of 5G base stations and active distribution networks has an architecture diagram as shown below: Figure 3 As shown, it consists of an improved spatiotemporal graph convolutional neural network establishment module, an input feature correction module, a prediction result calculation module, a 5G base station energy storage real-time minimum spare capacity calculation module, an objective function and constraint establishment module, and an optimization scheduling module.

[0215] An improved spatiotemporal graph convolutional neural network establishment module is used to establish a spatiotemporal graph convolutional neural network based on the original input data of optimal scheduling of 5G base stations and active distribution networks, combines graph convolution and gated recurrent units to establish a spatiotemporal graph convolutional neural network, uses historical operation data of 5G base stations and active distribution networks to learn the relationship between each node in the graph structure of the graph convolution layer in the spatiotemporal graph convolutional neural network, and adaptively corrects the adjacency matrix parameters of the graph convolution layer in the spatiotemporal graph convolutional neural network to obtain an improved spatiotemporal graph convolutional neural network.

[0216] Specifically, the original input data for the optimal scheduling of 5G base stations and active distribution networks includes the operating data of 5G base stations, the geographical distribution data of 5G base stations, the meteorological data of 5G base stations, and the historical operating data of 5G base stations and active distribution networks.

[0217] The architecture diagram of the improved spatiotemporal graph convolutional neural network is as follows Figure 2 As shown in the figure, in the improved spatiotemporal graph convolutional neural network, compared with the traditional convolutional neural network, the spatiotemporal graph convolutional neural network generalizes the convolutional network to the unstructured feature space and can process graph structure data of any shape. The spatiotemporal graph convolutional neural network model is represented as a topological graph G = (V, E, A), where V represents the point set, E represents the edge set, A is the adjacency matrix of the topological graph G, and the convolution range is node v i Neighborhood B:

[0218] B(v i )={v j |d(v j ,v i )≤D}(1),

[0219] In the formula, d(v j ,v i ) represents the graph node v i to v j The shortest distance; D is the maximum convolution neighborhood; when D = 1, it indicates that the size of the sampling area is node v i The first-order neighborhood of ;

[0220] At this time, the forward propagation of the graph convolution layer is as follows:

[0221]

[0222] In the formula, ReLU refers to the rectified linear unit activation function; X is the feature matrix of the topological graph G; is the adjacency matrix with autocorrelation, that is A is the adjacency matrix of the topological graph G, and I is the identity matrix; for

[0223] The element in row i and column j; is the adjacency matrix after normalization and symmetry; σ(·) is the sigmoid function; W0 is the weight matrix from the upper layer to the graph convolution layer; W1 is the weight matrix from the graph convolution layer to the lower layer, is the node degree matrix.

[0224] The graph convolution method is used to convert the historical load rate and geographical distribution characteristics of 5G base stations into node parameters of the graph. The hidden spatial correlation of different 5G base station power timing curves and photovoltaic output curves is learned based on the graph convolution layer. In order to further extract the timing characteristics of the 5G base station load rate and photovoltaic output curve, the graph convolution and gated recurrent unit are combined, and the graph convolution algorithm is introduced into the gated recurrent unit to form a spatiotemporal graph convolution unit. The spatiotemporal graph convolution neural network is constructed by stacking multiple spatiotemporal graph convolution units to capture the spatiotemporal dynamic changes of nodes in the graph. The deep learning and analysis of spatiotemporal graph data are shown in the following forward propagation formulas (4)-(7):

[0225] u t =σ(W u [f(A,X t ),h t-1 ]+b u ) (4), r t =σ(W r [f(A,X t ),h t-1 ]+b r ) (5),

[0226] c t =tanh(W c [f(A,X t ),(r t *h t-1 )]+b c ) (6), h t =u t *h t-1 +(1-u t )*c t (7)

[0227] In the formula, u t 、r t 、c tare the internal parameters of the unit, corresponding to the update gate, reset gate, and the state of the current unit at time t; h t is the unit output at time t; W u With b u are the weight matrix and bias matrix of the update gate respectively; W r With b r is the weight matrix and bias matrix of the reset gate; f(A,X t ) is the graph convolution operation on the model input at time t; h t-1 is the hidden state at the previous time t-1; W c is used to calculate the candidate state c t The weight matrix of c is the candidate state c t The associated bias vector.

[0228] The network unit of the spatiotemporal graph convolutional neural network can input graph structure data, and extract the spatial correlation of input features through the graph convolution method, and then extract the temporal correlation of input features based on the gated recurrent unit structure. After the extracted features are sent to the fully connected layer and the output layer, the 5G base station load rate and photovoltaic output prediction results can be output. Based on the 5G base station load rate, the 5G base station power prediction result can be obtained.

[0229] The adjacency matrix parameters of the graph convolution layer represent the spatial correlation of different prediction sites. In the prediction of multiple 5G base station load rates, 5G base station power and photovoltaic output, in order to more accurately learn the spatial relationship between the output timing curves of different base stations and photovoltaic power stations, the adjacency matrix parameters are adaptively corrected. During the training process, massive historical data samples are used to learn the relationship between each node in the graph structure, thereby adaptively correcting the adjacency matrix parameters.

[0230] Specifically, the method for adaptively correcting the adjacency matrix parameters of the graph convolution layer in the spatiotemporal graph convolutional neural network is:

[0231] The characteristic matrix X of the topological graph G = (x1, x2, ... x N ) T , define A mn is the parameter of the mth row and nth column of the adjacency matrix, A mn The expression is:

[0232]

[0233] In the formula, x m and x n represents the characteristics of node m and node n, ω is the learnable weight vector in the adjacency matrix parameters, and N is the total number of first-order neighbors of node m.

[0234] Formula (8) uses the ReLU unit to limit the value of the adjacency matrix to non-negative numbers, and uses the softmax form to convert A mn Normalization according to the line;

[0235] Using the following loss function L graph Study A mn The weight parameters are:

[0236]

[0237] In the formula, λ is the penalty coefficient of the regularization term, F is the norm of the regularization term, and its expression when F-2 regularization is used is:

[0238]

[0239] When x m With x n The greater the difference between the output timing curves of the photovoltaic power stations represented, the smaller their connection in space will be, which will make the adjacency matrix A mn The smaller the value of is, the smaller the loss function is. In addition, a regular term is added to the loss function to avoid falling into a zero weight solution. The total loss function L of the adaptive neural network is loss As shown below:

[0240] L loss =L mae +L graph (11)

[0241] Where, L mae is the loss function of the original neural network, using the mean absolute error.

[0242] The input feature correction module uses the soft attention mechanism and the scaled dot product scoring function to calculate the weights of different input features in the original input data of the optimal scheduling of 5G base stations and active distribution networks, and obtains the corrected input features.

[0243] Specifically, the method for obtaining the corrected input features is:

[0244] When the time period of the input feature is T, its feature weight vector α T for:

[0245]

[0246] Where X is the original input feature; Q is the query vector of the attention mechanism, that is, the predicted target value at the corresponding time; D T is the total length of the input feature; α T Calculated by the soft attention mechanism, the input features are normalized using the scaled dot product scoring function to ensure that the weights of different features are reasonably adjusted, thereby optimizing the input features of the model;

[0247] All input features are standardized to obtain the corrected input features, expressed as:

[0248]

[0249] Where, X ref is the corrected input feature, X mean is the mean of each characteristic component of input X; std is the standard deviation of each characteristic component of input X;

[0250] According to formula (13), the input features are standardized to generate the corrected input feature dataset X ref , which is the basis for subsequent predictions; the corrected input feature dataset X ref The data is input into the improved spatiotemporal graph convolutional neural network model, and the 5G base station load rate, power and photovoltaic output in different time periods are predicted using multi-layer convolution operations and spatiotemporal dependency analysis.

[0251] The predicted load rate of 5G base stations is expressed as follows:

[0252] λ 5G ={λ1(t),λ2(t),…,λ N (t)} (14),

[0253] In the formula, λ N (t) represents the load rate of the Nth 5G base station at time t;

[0254] The predicted power of 5G base station is expressed as follows:

[0255] P 5G = {P1(t),P2(t),…,P N (t)} (15),

[0256] Where P N (t) represents the power of the Nth 5G base station at time t;

[0257] The photovoltaic predicted output is expressed as follows:

[0258] P PV = {P1(t),P2(t),…,P M (t)} (16),

[0259] Where P M (t) represents the power of the Mth 5G base station at time t.

[0260] The 5G base station energy storage real-time minimum backup capacity calculation module is used to calculate the 5G base station energy storage real-time minimum backup capacity based on the 5G base station predicted power and the 5G base station backup power time.

[0261] The minimum backup power capacity of 5G base station energy storage is related to the power consumption of base station equipment and the minimum backup time. For the power consumption of base station equipment, such as a macro base station with a typical configuration of 3 active antenna units (AAU) and 1 baseband unit (BBU), the base station power consumption includes two parts: dynamic power consumption and static power consumption. The power consumption of a single base station communication equipment can be expressed as: represents the power of the base station at time t; P J 5G Indicates static power, i.e., power amplifier power in AAU, BBU power, and transmission equipment power; Indicates the maximum dynamic power, that is, the power of the radio frequency unit in the AUU, and its communication load rate with the base station λ t Related.

[0262] Specifically, the predicted power of 5G base stations in different regions obtained by formula (15) and the minimum backup power time t of 5G base stations are s Combined, we can get the minimum spare capacity of 5G base stations at time t As shown below:

[0263]

[0264] Where P 5G Predicting power for 5G base stations.

[0265] The objective function and constraint establishment module is used to establish the objective function and operation constraints with the goal of optimizing the operation cost according to the operation conditions of the 5G base station and the active distribution network.

[0266] Specifically, the expression of the objective function is:

[0267]

[0268] In the formula, It represents the cost of short-time load power shortage; Indicates the charging and discharging cost of base station energy storage; Represents the photovoltaic output cost; its specific expression is shown in the following formula:

[0269]

[0270] In the formula, c LD 、c PV They represent the unit cost of load power shortage and the unit cost of photovoltaic power generation respectively; c 5G 、cSH They represent the unit charging and discharging cost and loss cost of 5G base station energy storage respectively; P t ch , P t dis Respectively represent the charge and discharge power at time t; P t PV Represents photovoltaic output; η 5G represents the charging and discharging efficiency of G base station energy storage; P t ΔLD Indicates that the load is short of power.

[0271] The operation constraints include 5G base station energy storage charging and discharging constraints, photovoltaic output constraints and power balance constraints;

[0272] The expression of the 5G base station energy storage charging and discharging constraint is:

[0273]

[0274] Where P t ch and P t dis They represent the charging power and discharging power at time t respectively, and They represent the maximum charging power and the maximum discharging power at time t, respectively, and E 5G Indicates the energy storage capacity of 5G base stations, and They represent the minimum and maximum energy storage capacity of the 5G base station respectively;

[0275] The expression of the photovoltaic output constraint is:

[0276]

[0277] Where P t PV Represents photovoltaic output, Indicates the maximum photovoltaic output

[0278] The expression of the power balance constraint is:

[0279] P t ΔLD =P t ch +P t LD -P t PV -P t dis (twenty two),

[0280] Where P t LDrepresents the load power of the distribution network at time t, P t ch and P t dis They represent the charging power and discharging power at time t, respectively, P t PV Indicates photovoltaic output.

[0281] The optimization scheduling module is used to solve the objective function through the column constraint generation algorithm based on the real-time minimum spare capacity of 5G base station energy storage, the predicted load rate of 5G base station, the predicted photovoltaic output and the operating constraints, and obtain the scheduling plan of 5G base station and active distribution network, and optimize the scheduling of 5G base station and active distribution network according to the scheduling plan.

[0282] Specifically, the solution method of the objective function is:

[0283] The uncertainty of photovoltaic and load is expressed by set Ω, as shown in the following formula:

[0284]

[0285] In the formula, is the actual output of the photovoltaic system at time t; is the predicted output of the photovoltaic system at time t; Indicates the maximum value of the uncertainty range of the photovoltaic system output; represents the actual load demand at time t; Indicates the maximum value of the load power uncertainty range; Represents the predicted load demand at time t.

[0286] Formulas (18)-(22) are expressed as an uncertain two-stage robust optimization model using compact matrix equations. The first stage is to minimize the system operation cost, and the second stage is to obtain the worst operation scenario under the minimum cost. The specific expression is shown in the following formula:

[0287]

[0288] In the formula, c T x represents the optimal objective function of operating cost; d T y represents the worst operating cost when the load loses power; u represents the uncertain variables including photovoltaic and load power; max represents the worst scenario with the highest cost in the optimal scheduling; H, F, and G represent three different coefficient matrices respectively; y| represents the corresponding state variable or decision variable in the column constraint generation algorithm.

[0289] Based on the real-time minimum spare capacity model of 5G base station energy storage, the predicted load rate of 5G base stations, the predicted output of photovoltaic power generation and operation constraints, the column constraint generation algorithm is used to solve the uncertain two-stage robust optimization model. The objective function of the robust optimization model is first decoupled and decomposed into the main problem and sub-problems.

[0290] The main problem is:

[0291]

[0292] In the formula, m represents the number of iterations; y m represents the solution to the subproblem after m iterations.

[0293] The sub-problems are:

[0294]

[0295] By giving the (x,u) value, the minimization problem is transformed into a deterministic linear optimization problem, which is expressed as follows:

[0296]

[0297] Where, d and y represent column vectors; A1, A2, A3 represent m1, m2, m3-column n-row matrices; b1, b2, b3 represent m1, m2, m3-dimensional column vectors;

[0298] Using the duality theory, the inner min model in the sub-problem of the above robust optimization model is transformed into a max model and merged with the outer max model. The transformation method is:

[0299] If the original question is:

[0300]

[0301] The dual form is:

[0302]

[0303] Where z1, z2, z3 means m1, m2, m3 are row vectors;

[0304] According to the principle of duality theory, the dual form is:

[0305]

[0306] Based on the above formulas (24)-(30), the scheduling plan of the 5G base station and the active distribution network is obtained, and the scheduling plan includes: the scheduling plan of the 5G base station and the output plan of photovoltaics.

[0307] Example 3

[0308] An electronic device comprises: a memory and a processor, the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor implements the optimized scheduling method for 5G base stations and active distribution networks described in Example 1 and the optimized scheduling system for 5G base stations and active distribution networks described in Example 2 by executing the computer instructions.

[0309] Example 4

[0310] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the optimization scheduling method for 5G base stations and active distribution networks described in Example 1 and the optimization scheduling system for 5G base stations and active distribution networks described in Example 2.

[0311] Example 5

[0312] In order to better demonstrate the technical effect of the present invention, the traditional method and the present invention are used to conduct comparative analysis on the optimal scheduling of a 5G base station and an active distribution network. The Python compilation environment is used in the analysis process, the model is built on PyTorch, and the improved IEEE33 node is used for simulation verification.

[0313] The schematic diagram of a 5G base station and active distribution network is as follows: Figure 4 As shown, the active distribution network is hereinafter referred to as the distribution network. 5G base stations with energy storage are set up at nodes 2-5 and 10-21. Five base stations are set up under each node. Photovoltaic power station 4 with an access capacity of 0.5MW is added to nodes 8 and 11. The equipment parameters of the 5G base station are shown in Table 1.

[0314] Table 1 5G base station equipment parameters

[0315]

[0316] 80% of the days of the input data are randomly selected as the training set, and the rest of the data is the test set. In the improved spatiotemporal graph convolutional neural network of this embodiment, the spatiotemporal graph convolution layer uses one hidden layer with 64 units; the fully connected layer contains 32 units, and a dropout layer with a dropout rate of 0.2 is added after the fully connected layer. The Adam algorithm is used as the optimization algorithm for training, and the initial learning rate is 0.001; the prediction time step is 1h, the data resolution is 5min, and the model lookback period is 24h.

[0317] The expression for evaluating the prediction result based on the root mean square error RMSE and the mean absolute error MAE between the prediction result and the true value is:

[0318]

[0319] Where T is the total length of the evaluation period, yt is the true value, is the model prediction value.

[0320] In order to better reflect the effect of the proposed method on the spatial characteristics of base stations and photovoltaics, two current traditional methods and the present invention were tested on a cloudy day sample in July in the test set, and two spatiotemporal graph convolutional neural networks with fixed adjacency matrix parameters were selected for comparison. The two traditional methods are: the first method directly limits the adjacency matrix parameters of base stations and photovoltaics within a certain distance to 1, and other adjacency matrix parameters to 0, referred to as the distance method; the second method constructs the adjacency matrix based on the K nearest neighbor algorithm, referred to as the K nearest neighbor method.

[0321] The adjacency matrices of different methods are as follows Figure 5 As shown in the figure, the thickness of the connecting line indicates the strength of the spatial characteristic coupling relationship between the 5G base station and the photovoltaic power plant. The strength of the coupling relationship is mainly represented by the size of the adjacency matrix. Figure 5 It can be seen that compared with the distance-based adjacency matrix, the adjacency weight calculated by the K nearest neighbor method takes into account the relationship between the historical photovoltaic output and weather data of each station, has a large change in the adjacency weight parameter structure, and has a higher prediction accuracy; the adjacency matrix adaptive correction method proposed in the present invention directly learns the influence of different adjacency matrix parameters on the prediction target from historical data, which can more accurately describe the relationship between each photovoltaic power station and further reduce the prediction error.

[0322] The prediction results are as follows Figure 6 As shown in Table 2, the RMSE and MAE on the full test set are shown in Table 2. Figure 6 Compared with the prediction results in Table 2, it can be concluded that the prediction results of the present invention are closest to the true value and have the highest prediction accuracy.

[0323] Table 2 Prediction errors under different methods

[0324]

[0325] Figure 7 The load rate prediction of 5G base stations in different regions and the dispatchable capacity of energy storage are shown. The different reserve capacities of each base station are mainly due to differences in base station load characteristics and energy storage backup hours. Figure 7 It can be seen that the backup power time of the base station accessed by node 13 is longer than that of the base station accessed by node 5. Therefore, the minimum backup capacity is significantly increased and the dispatchable capacity range is significantly reduced.

[0326] According to the adjacency relationship between the base station and photovoltaic accessed by the node obtained by the present invention, the distribution network is divided into 5G base station + photovoltaic clusters, as shown in the schematic diagram Figure 8By considering the spatiotemporal connection relationship between 5G base stations and photovoltaics, aggregating the adjustable capacity of 5G base station energy storage and photovoltaics, it is possible to effectively support the load power outage between distribution network clusters and between cluster distribution networks. The load power outage changes in the distribution network are shown in the figure below. Fig. 9 shown.

[0327] Fig.10 The scheduling plan of each cluster based on the present invention is shown. Fig.10 It can be seen that during the period of 00:00-05:00, the load demand is small, there is almost no load power outage, the photovoltaic power does not work, and the electricity price is low during this period. The 5G base station energy storage equipment is charged, and the cluster purchases electricity from the distribution network to maintain the load demand in each cluster. During the period of 07:00-18:00, the photovoltaic output of clusters 1 and 2 increases to meet the demand for short-term large-capacity power outages during the peak load period, and the energy storage equipment also begins to release power. In this stage with higher electricity prices, the increased power will be sold to obtain economic benefits. During the period of 21:00-23:00, the energy storage equipment of each cluster is charged, the load power shortage during this period is alleviated, and the amount of electricity obtained by each cluster from the distribution network is reduced accordingly.

[0328] Fig.11 The overall operating cost of the system under the stochastic optimization method (SO), the traditional robust optimization method (RO) and the present invention is compared. Fig.11 It can be seen from the overall trend that the system operation cost based on the present invention is relatively small, especially in the time periods of 14:00, 19:00, and 22:00, the RO method and the SO method generate relatively high operation costs, mainly because there is no accurate assessment of the aggregation capacity of the cluster itself, which increases the operation cost.

[0329] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the application can adopt the form of complete hardware embodiment, complete software embodiment, or the embodiment in combination with software and hardware. Moreover, the application can adopt the form of the computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code. The scheme in the embodiment of the present application can be implemented in various computer languages, for example, object-oriented programming language Java, C++, Python and literal scripting language JavaScript, etc.

[0330] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0331] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0332] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0333] Although the preferred embodiments of the present application have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present application.

[0334] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

Claims

The optimization scheduling method of 1.5G base station and active distribution network is characterized in that: include: Based on the original input data of 5G base stations and active distribution network optimization scheduling, the graph convolution and gated recurrent unit are combined to establish a spatiotemporal graph convolutional neural network. The historical operation data of 5G base stations and active distribution networks are used to learn the relationship between the nodes in the graph structure of the graph convolution layer in the spatiotemporal graph convolutional neural network. The adjacency matrix parameters of the graph convolution layer in the spatiotemporal graph convolutional neural network are adaptively corrected to obtain an improved spatiotemporal graph convolutional neural network. The soft attention mechanism and the scaled dot product scoring function are used to calculate the weights of different input features in the original input data of the optimal scheduling of 5G base stations and active distribution networks, and the corrected input features are obtained; The corrected input features are input into the improved spatiotemporal graph convolutional neural network, and the output is the predicted load rate of 5G base stations, the predicted power of 5G base stations, and the predicted output of photovoltaics. Based on the predicted power of the 5G base station and the backup power time of the 5G base station, the real-time minimum backup capacity of the 5G base station energy storage is calculated; According to the operation status of 5G base stations and active distribution networks, establish the objective function and operation constraints with the goal of optimizing the operation cost; Based on the real-time minimum backup capacity of 5G base station energy storage, the predicted load rate of 5G base stations, the predicted photovoltaic output and operating constraints, the objective function is solved through the column constraint generation algorithm to obtain the scheduling plan of 5G base stations and active distribution networks. According to the scheduling plan, the 5G base stations and active distribution networks are optimally scheduled.

2. The optimization scheduling method of 5G base stations and active distribution networks according to claim 1 is characterized in that: In the improved spatiotemporal graph convolutional neural network, the original input data for optimizing the scheduling of 5G base stations and active distribution networks includes the operation data of 5G base stations, the geographical distribution data of 5G base stations, the meteorological data of 5G base stations, and the historical operation data of 5G base stations and active distribution networks; The spatiotemporal graph convolutional neural network model is represented as a topological graph G = (V, E, A), where V represents the point set, E represents the edge set, A is the adjacency matrix of the topological graph G, and the convolution range is node v i Neighborhood B: B(v i )={v j |d(v j ,v i )≤D}(1), In the formula, d(v j ,v i ) represents the graph node v i to v j The shortest distance; D is the maximum convolution neighborhood; when D = 1, it indicates that the size of the sampling area is node v i The first-order neighborhood of ; The forward propagation of the graph convolutional layer is as follows: In the formula, ReLU refers to the rectified linear unit activation function; X is the feature matrix of the topological graph G; is the adjacency matrix with autocorrelation, that is A is the adjacency matrix of the topological graph G, and I is the identity matrix; for The element in row i and column j; is the adjacency matrix after normalization and symmetry; σ(·) is the sigmoid function; W0 is the weight matrix from the upper layer to the graph convolution layer; W1 is the weight matrix from the graph convolution layer to the lower layer, is the node degree matrix; The graph convolution and gated recurrent unit are combined, and the graph convolution algorithm is introduced into the gated recurrent unit to form a spatiotemporal graph convolution unit. A spatiotemporal graph convolution neural network is constructed by stacking multiple spatiotemporal graph convolution units to capture the spatiotemporal dynamic changes of nodes in the graph and conduct deep learning and analysis of spatiotemporal graph data. The forward propagation formula is shown in the following equations (4)-(7): u t =σ(W u [f(A,X t ),h t-1 ]+b u ) (4), r t =σ(W r [f(A,X t ),h t-1 ]+b r ) (5), c t =tanh(W c [f(A,X t ),(r t *h t-1 )]+b c ) (6), h t =u t *h t-1 +(1-u t )*c t (7), In the formula, u t 、r t 、c t are the internal parameters of the unit, corresponding to the update gate, reset gate, and the state of the current unit at time t; h t is the unit output at time t; W u With b u are the weight matrix and bias matrix of the update gate respectively; W r With b r is the weight matrix and bias matrix of the reset gate; f(A,X t ) is the graph convolution operation on the model input at time t; h t-1 is the hidden state at the previous time t-1; W c is used to calculate the candidate state c t The weight matrix of c is the candidate state c t the associated bias vector; The method for adaptively correcting the adjacency matrix parameters of the graph convolution layer in the spatiotemporal graph convolutional neural network is: The characteristic matrix X of the topological graph G = (x1, x2, ... x N ) T , define A mn is the parameter of the mth row and nth column of the adjacency matrix, A mn The expression is: In the formula, x m and x n represents the characteristics of node m and node n, ω is the learnable weight vector in the adjacency matrix parameters, and N is the total number of first-order neighbors of node m; Formula (8) uses the ReLU unit to limit the value of the adjacency matrix to non-negative numbers, and uses the softmax form to convert A mn Perform normalization; Using the following loss function L graph Study A mn The weight parameters are: In the formula, λ is the penalty coefficient of the regularization term, F is the norm of the regularization term, and its expression when F-2 regularization is used is: The total loss function L of the adaptive neural network loss As shown below: L loss =L mae +L graph (11), Where, L mae is the loss function of the original neural network, using the mean absolute error.

3. The optimization scheduling method of 5G base stations and active distribution networks according to claim 1 is characterized in that: The method for obtaining the modified input features is: When the time period of the input feature is T, Its feature weight vector α T for: Where X is the original input feature; Q is the query vector of the attention mechanism, that is, the predicted target value at the corresponding time; D T is the total length of the input feature; α T Calculated by the soft attention mechanism, the input features are normalized using the scaled dot product scoring function to ensure that the weights of different features are reasonably adjusted, thereby optimizing the input features of the model; All input features are standardized to obtain the corrected input features, expressed as: Where, X ref is the corrected input feature, X mean is the mean of each characteristic component of input X; std is the standard deviation of each characteristic component of input X; According to formula (13), the input features are standardized to generate the corrected input feature dataset X ref ; The corrected input feature dataset X ref The data is input into the improved spatiotemporal graph convolutional neural network model, and the 5G base station load rate, power and photovoltaic output in different time periods are predicted using multi-layer convolution operations and spatiotemporal dependency analysis. The 5G base station predicted load rate λ 5G It is expressed as the following formula: l 5G ={λ1(t),λ2(t),…,λ N (t)} (14), In the formula, λ N (t) represents the load rate of the Nth 5G base station at time t; The 5G base station predicted power P 5G It is expressed as the following formula: P 5G ={P1(t),P2(t),…,P N (t)} (15), Where P N (t) represents the power of the Nth 5G base station at time t; The photovoltaic predicted output P PV It is expressed as the following formula: P PV ={P1(t),P2(t),…,P M (t)} (16), Where P M (t) represents the power of the Mth 5G base station at time t.

4. The optimization scheduling method of 5G base stations and active distribution networks according to claim 3 is characterized in that: In the calculation of the real-time minimum backup capacity of the 5G base station energy storage, the predicted power of the 5G base stations in different areas obtained by formula (15) and the minimum backup power time t of the 5G base station are combined. s Combined, we can get the minimum spare capacity of 5G base stations at time t As shown below: Where P 5G Predicting power for 5G base stations.

5. The method for optimizing the scheduling of 5G base stations and active distribution networks according to claim 4 is characterized in that: The expression of the objective function is: In the formula, It represents the cost of short-time load power shortage; Indicates the charging and discharging cost of base station energy storage; represents the photovoltaic output cost; The specific expression is as follows: In the formula, c LD 、c PV They represent the unit cost of load power shortage and the unit cost of photovoltaic power generation respectively; c 5G 、c SH They represent the unit charging and discharging cost and loss cost of 5G base station energy storage respectively; Respectively represent the charging and discharging power at time t; Indicates photovoltaic output; η 5G Indicates the charging and discharging efficiency of G base station energy storage; Indicates the load power shortage; The operation constraints include 5G base station energy storage charging and discharging constraints, photovoltaic output constraints and power balance constraints; The expression of the 5G base station energy storage charging and discharging constraint is: In the formula, and They represent the charging power and discharging power at time t respectively, and They represent the maximum charging power and the maximum discharging power at time t, respectively, and E 5G Indicates the energy storage capacity of 5G base stations, and They represent the minimum and maximum energy storage capacity of the 5G base station respectively; The expression of the photovoltaic output constraint is: In the formula, Represents photovoltaic output, Indicates the maximum photovoltaic output; The expression of the power balance constraint is: In the formula, represents the load power of the distribution network at time t, and They represent the charging power and discharging power at time t respectively, Indicates photovoltaic output.

6. The method for optimizing the scheduling of 5G base stations and active distribution networks according to claim 5, characterized in that: The solution method of the objective function is: The uncertainty of photovoltaic and load is expressed by set Ω, as shown in the following formula: In the formula, is the actual output of the photovoltaic system at time t; is the predicted output of the photovoltaic system at time t; Indicates the maximum value of the uncertainty range of the photovoltaic system output; represents the actual load demand at time t; Indicates the maximum value of the load power uncertainty range; represents the predicted load demand at time t; Formulas (18)-(22) are expressed as an uncertain two-stage robust optimization model using compact matrix equations. The first stage is to minimize the system operation cost, and the second stage is to obtain the worst operation scenario under the minimum cost. The specific expression is shown in the following formula: In the formula, c T x represents the optimal objective function of operating cost; d T y represents the worst operating cost when the load loses power; u represents the uncertain variables including photovoltaic and load power; max represents the worst scenario with the highest cost in the optimal scheduling; H, F, and G represent three different coefficient matrices respectively; Based on the real-time minimum spare capacity model of 5G base station energy storage, the predicted load rate of 5G base stations, the predicted output of photovoltaic power generation and operation constraints, the column constraint generation algorithm is used to solve the uncertain two-stage robust optimization model. The objective function of the robust optimization model is first decoupled and decomposed into the main problem and sub-problems. The main problem is: In the formula, m represents the number of iterations; y m represents the solution to the subproblem after m iterations. The sub-problems are: By giving the (x,u) value, the minimization problem is transformed into a deterministic linear optimization problem, which is expressed as follows: Where, d and y represent column vectors; A1, A2, A3 represent m1, m2, m3-column n-row matrices; b1, b2, b3 represent m1, m2, m3-dimensional column vectors; Using the duality theory, the inner min model in the sub-problem of the above robust optimization model is transformed into a max model and merged with the outer max model. The transformation method is: If the original question is: The dual form is: Where z1, z2, z3 means m1, m2, m3 are row vectors; According to the principle of duality theory, the dual form is: Based on formulas (24)-(30), the scheduling plan of the 5G base station and the active distribution network is obtained, and the scheduling plan includes: the scheduling plan of the 5G base station and the output plan of the photovoltaic power generation. The optimized dispatching system of 7.5G base stations and active distribution networks is characterized by: It includes an improved spatiotemporal graph convolutional neural network establishment module, an input feature correction module, a prediction result calculation module, a 5G base station energy storage real-time minimum spare capacity calculation module, an objective function and constraint establishment module, and an optimization scheduling module; The improved spatiotemporal graph convolutional neural network establishment module is used to combine graph convolution and gated recurrent units to establish a spatiotemporal graph convolutional neural network based on the original input data of the optimal scheduling of 5G base stations and active distribution networks, use the historical operation data of 5G base stations and active distribution networks to learn the relationship between each node in the graph structure of the graph convolution layer in the spatiotemporal graph convolutional neural network, and adaptively correct the adjacency matrix parameters of the graph convolution layer in the spatiotemporal graph convolutional neural network to obtain an improved spatiotemporal graph convolutional neural network; The input feature correction module is used to calculate the weights of different input features in the original input data of the 5G base station and the active distribution network optimization scheduling by using a soft attention mechanism and a scaled dot product scoring function to obtain the corrected input features; The prediction result calculation module is used to input the corrected input features into the improved spatiotemporal graph convolutional neural network, and output the 5G base station predicted load rate, 5G base station predicted power and photovoltaic predicted output; The 5G base station energy storage real-time minimum spare capacity calculation module is used to calculate the 5G base station energy storage real-time minimum spare capacity according to the 5G base station predicted power and the 5G base station backup power time; The objective function and constraint establishment module is used to establish an objective function and operation constraint conditions with the goal of optimizing the operation cost according to the operation conditions of the 5G base station and the active distribution network; The optimization scheduling module is used to solve the objective function through a column constraint generation algorithm based on the real-time minimum spare capacity of the 5G base station energy storage, the predicted load rate of the 5G base station, the predicted photovoltaic output and the operating constraints, to obtain the scheduling plan of the 5G base station and the active distribution network, and optimize the scheduling of the 5G base station and the active distribution network according to the scheduling plan.

8. The optimization dispatching system of 5G base station and active distribution network according to claim 7, characterized in that: In the improved spatiotemporal graph convolutional neural network establishment module, the original input data for optimizing the scheduling of 5G base stations and active distribution networks includes the operation data of 5G base stations, the geographical distribution data of 5G base stations, the meteorological data of 5G base stations, and the historical operation data of 5G base stations and active distribution networks; The spatiotemporal graph convolutional neural network model is represented as a topological graph G = (V, E, A), where V represents the point set, E represents the edge set, A is the adjacency matrix of the topological graph G, and the convolution range is node v i Neighborhood B: B(v i )={v j |d(v j ,v i )≤D} (1), In the formula, d(v j ,v i ) represents the graph node v i to v j The shortest distance; D is the maximum convolution neighborhood; when D = 1, it indicates that the size of the sampling area is node v i The first-order neighborhood of ; At this time, the forward propagation of the graph convolutional layer is as follows: In the formula, ReLU refers to the rectified linear unit activation function; X is the feature matrix of the topological graph G; is the adjacency matrix with autocorrelation, that is A is the adjacency matrix of the topological graph G, and I is the identity matrix; for The i-th row and j-th column element in ; σ(·) is the sigmoid function; W0 is the weight matrix from the upper layer to the graph convolution layer; W1 is the weight matrix from the graph convolution layer to the lower layer, is the node degree matrix; The graph convolution and gated recurrent unit are combined, and the graph convolution algorithm is introduced into the gated recurrent unit to form a spatiotemporal graph convolution unit. A spatiotemporal graph convolution neural network is constructed by stacking multiple spatiotemporal graph convolution units to capture the spatiotemporal dynamic changes of nodes in the graph and conduct deep learning and analysis of spatiotemporal graph data. The forward propagation formula is shown in the following equations (4)-(7): u t =σ(W u [f(A,X t ),h t-1 ]+b u ) (4), r t =σ(W r [f(A,X t ),h t-1 ]+b r ) (5), c t =tanh(W c [f(A,X t ),(r t *h t-1 )]+b c ) (6), h t =u t *h t-1 +(1-u t )*c t (7), In the formula, u t 、r t 、c t are the internal parameters of the unit, corresponding to the update gate, reset gate, and the state of the current unit at time t; h t is the unit output at time t; W u With b u are the weight matrix and bias matrix of the update gate respectively; W r With b r is the weight matrix and bias matrix of the reset gate; f(A,X t ) is the graph convolution operation on the model input at time t; h t-1 is the hidden state at the previous time t-1; W c is used to calculate the candidate state c t The weight matrix of c is the candidate state c t the associated bias vector; The method for adaptively correcting the adjacency matrix parameters of the graph convolution layer in the spatiotemporal graph convolutional neural network is: The characteristic matrix X of the topological graph G = (x1, x2, ... x N ) T , define A mn is the parameter of the mth row and nth column of the adjacency matrix, A mn The expression is: In the formula, x m and x n represents the characteristics of node m and node n, ω is the learnable weight vector in the adjacency matrix parameters, and N is the total number of first-order neighbors of node m; Formula (8) uses the ReLU unit to limit the value of the adjacency matrix to non-negative numbers, and uses the softmax form to convert A mn Perform normalization; Using the following loss function L graph Study A mn The weight parameters are: In the formula, λ is the penalty coefficient of the regularization term, F is the norm of the regularization term, and its expression when F-2 regularization is used is: The total loss function L of the adaptive neural network loss As shown below: L loss =L mae +L graph (11), Where, L mae is the loss function of the original neural network, using the mean absolute error.

9. The optimization dispatching system of 5G base station and active distribution network according to claim 7, characterized in that: In the input feature correction module, the method for obtaining the corrected input feature is: When the time period of the input feature is T, Its feature weight vector α T for: In the formula, X is the original input feature; Q is the query vector of the attention mechanism, that is, the predicted target value at the corresponding time; D T is the total length of the input feature; α T Calculated by the soft attention mechanism, the input features are normalized using the scaled dot product scoring function to ensure that the weights of different features are reasonably adjusted, thereby optimizing the input features of the model; All input features are standardized to obtain the corrected input features, expressed as: Where, X ref is the corrected input feature, X mean is the mean of each characteristic component of input X; std is the standard deviation of each characteristic component of input X; According to formula (13), the input features are standardized to generate the corrected input feature dataset X ref , which is the basis for subsequent predictions; the corrected input feature dataset X ref The data is input into the improved spatiotemporal graph convolutional neural network model, and the 5G base station load rate, power and photovoltaic output in different time periods are predicted using multi-layer convolution operations and spatiotemporal dependency analysis. The 5G base station predicted load rate λ 5G It is expressed as the following formula: l 5G ={λ1(t),λ2(t),…,λ N (t)} (14), In the formula, λ N (t) represents the load rate of the Nth 5G base station at time t; The 5G base station predicted power P 5G It is expressed as the following formula: P 5G ={P1(t),P2(t),…,P N (t)} (15), Where P N (t) represents the power of the Nth 5G base station at time t; The photovoltaic predicted output P PV It is expressed as the following formula: P PV ={P1(t),P2(t),…,P M (t)} (16), Where P M (t) represents the power of the Mth 5G base station at time t.

10. The optimization dispatching system of 5G base station and active distribution network according to claim 9, characterized in that: In the 5G base station energy storage real-time minimum backup capacity calculation module, the predicted power of 5G base stations in different areas obtained by formula (15) and the minimum backup power time t of the 5G base station are calculated. s Combined, we can get the minimum spare capacity of 5G base stations at time t As shown below: Where P 5G Predicting power for 5G base stations.

11. The optimization dispatching system of 5G base station and active distribution network according to claim 10, characterized in that: In the objective function and constraint establishment module, the expression of the objective function is: In the formula, It represents the cost of short-time load power shortage; Indicates the charging and discharging cost of base station energy storage; Represents the photovoltaic output cost; its specific expression is shown in the following formula: In the formula, c LD 、c PV They represent the unit cost of load power shortage and the unit cost of photovoltaic power generation respectively; c 5G 、c SH They represent the unit charging and discharging cost and loss cost of 5G base station energy storage respectively; Respectively represent the charging and discharging power at time t; Indicates photovoltaic output; η 5G Indicates the charging and discharging efficiency of 5G base station energy storage; Indicates the load power shortage; The operation constraints include 5G base station energy storage charging and discharging constraints, photovoltaic output constraints and power balance constraints; The expression of the 5G base station energy storage charging and discharging constraint is: In the formula, and They represent the charging power and discharging power at time t respectively, and They represent the maximum charging power and the maximum discharging power at time t, respectively, and E 5G Indicates the energy storage capacity of 5G base stations, and They represent the minimum and maximum energy storage capacity of the 5G base station respectively; The expression of the photovoltaic output constraint is: In the formula, Represents photovoltaic output, Indicates the maximum photovoltaic output; The expression of the power balance constraint is: In the formula, represents the load power of the distribution network at time t, and They represent the charging power and discharging power at time t respectively, Indicates photovoltaic output.

12. The optimization dispatching system of 5G base station and active distribution network according to claim 11, characterized in that: In the optimization scheduling module, the method for solving the objective function is: The uncertainty of photovoltaic and load is expressed by set Ω, as shown in the following formula: In the formula, is the actual output of the photovoltaic system at time t; is the predicted output of the photovoltaic system at time t; Indicates the maximum value of the uncertainty range of the photovoltaic system output; represents the actual load demand at time t; Indicates the maximum value of the load power uncertainty range; represents the predicted load demand at time t; Formulas (18)-(22) are expressed as an uncertain two-stage robust optimization model using compact matrix equations. The first stage is to minimize the system operation cost, and the second stage is to obtain the worst operation scenario under the minimum cost. The specific expression is shown in the following formula: In the formula, c T x represents the optimal objective function of operating cost; d T y represents the worst operating cost when the load loses power; u represents the uncertain variables including photovoltaic and load power; max represents the worst scenario with the highest cost in the optimal scheduling; H, F, and G represent three different coefficient matrices respectively; Based on the real-time minimum spare capacity model of 5G base station energy storage, the predicted load rate of 5G base stations, the predicted output of photovoltaic power generation and operation constraints, the column constraint generation algorithm is used to solve the uncertain two-stage robust optimization model. The objective function of the robust optimization model is first decoupled and decomposed into the main problem and sub-problems. The main problem is: In the formula, m represents the number of iterations; y m represents the solution to the subproblem after m iterations. The sub-problems are: By giving the (x,u) value, the minimization problem is transformed into a deterministic linear optimization problem, which is expressed as follows: Where, d and y represent column vectors; A1, A2, A3 represent m1, m2, m3-column n-row matrices; b1, b2, b3 represent m1, m2, m3-dimensional column vectors; Using the duality theory, the inner min model in the sub-problem of the above robust optimization model is transformed into a max model and merged with the outer max model. The transformation method is: If the original question is: The dual form is: Where z1, z2, z3 means m1, m2, m3 are row vectors; According to the principle of duality theory, the dual form is: Based on formulas (24)-(30), the scheduling plan of the 5G base station and the active distribution network is obtained, and the scheduling plan includes: the scheduling plan of the 5G base station and the output plan of the photovoltaic power generation.

13. An electronic device, characterized in that: include: A memory and a processor, wherein the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor implements the optimization scheduling method for 5G base stations and active distribution networks as described in any one of claims 1 to 6 by executing the computer instructions.

14. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the optimization scheduling method for a 5G base station and an active distribution network as described in any one of claims 1 to 6 is implemented.

Images