Power distribution network dynamic reconstruction acceleration method and system based on deep learning

Abstract

The invention discloses a deep learning-based power distribution network dynamic reconstruction acceleration method and system. The method comprises the following steps of 1, establishing a power distribution network dynamic reconstruction optimization model; 2, in the power distribution network power flow equation constraint of the power distribution network dynamic reconstruction optimization model, simplifying a power distribution network power flow equation by using a second-order cone relaxation method to form a power distribution network dynamic reconstruction MIQP model; 3, screening a non-key action switch according to historical data of the state of the power distribution network line switch, fixing the state of the non-key action switch, and reducing the number of binary variables; and 4, predicting the state of the key action switch of the power distribution network within a period of time scale by using a deep learning method. According to the method, the state of the key action switch of the power distribution network line in a period of time scale is predicted through the Informer model, so that the number of binary variables is reduced, and the solving time of dynamic reconstruction of the power distribution network is greatly shortened.

Description

Technical Field

[0001] The invention relates to a distribution network dynamic reconstruction acceleration method, belonging to the technical field of power system power grid reconstruction. Background Art

[0002] The dynamic reconstruction of the distribution network optimizes the operation of the distribution system by changing the network topology of the distribution system to change the direction of the power flow, so as to achieve the balance of power loads between different regions, reduce line overloads, improve voltage stability and reduce network losses. At present, the main methods for solving the dynamic reconstruction model of the distribution network are analytical algorithms, heuristic algorithms and genetic algorithms. Although heuristic algorithms and genetic algorithms can reduce the computational complexity to a certain extent, they are inefficient and easily fall into local optimality. Therefore, it is possible to consider using analytical methods to construct mathematical models to complete the solution of the dynamic reconstruction model of the distribution network.

[0003] The dynamic reconstruction model of the distribution network can be constructed as a very complex non-convex nonlinear programming problem, which is difficult to solve with current commercial solvers. For large-scale and long-time scale dynamic reconstruction optimization problems, due to the existence of a large number of binary variables, even if the problem is subjected to second-order cone relaxation and converted into a MIQP (mixed integer quadratic programming) problem, there are still problems of low efficiency and difficulty in solving, and it may even be impossible to solve it with existing solvers. Therefore, accelerating the use of analytical methods to solve the dynamic reconstruction optimization model of the distribution network is an issue that needs to be further addressed. Summary of the invention

[0004] The technical problem to be solved by the present invention is that the use of analytical methods to solve the dynamic reconstruction model of the distribution network is inefficient and difficult to solve.

[0005] In order to solve the above technical problems, the present invention provides a distribution network dynamic reconstruction acceleration method based on deep learning, comprising the following steps:

[0006] Step 1: Establish a dynamic reconstruction optimization model for distribution network;

[0007] Step 2: In the distribution network power flow equation constraint of the distribution network dynamic reconstruction optimization model, the distribution network power flow equation is simplified using the second-order cone relaxation method to form a MIQP model of distribution network dynamic reconstruction. The decision variables of the distribution network dynamic reconstruction model after the second-order cone relaxation include continuous variables and binary variables.

[0008] Step 3: select non-critical action switches according to the historical data of the switch status of the distribution network line, and fix the status of the non-critical action switches to reduce the number of binary variables;

[0009] Step 4: Use deep learning methods to predict the state of key action switches in the distribution network over a period of time, and optimize and solve the dynamic reconstruction optimization model of the distribution network.

[0010] The aforementioned method for accelerating dynamic reconstruction of a distribution network based on deep learning, in step 1, establishes a dynamic reconstruction optimization model of a distribution network, including:

[0011] The objective function is to minimize the network loss:

[0012] (1)

[0013] in, is the set of all branches, is the set of all time points, For branch The current flowing through For branch The resistance of

[0014] The branch power flow equation constraints are:

[0015] (2) (3)

[0016] (4)

[0017] (5)

[0018] in, is the set of all nodes, For Node The voltage, For Node The voltage, , The nodes are Active power injection and reactive power injection of generators, and The nodes are Active power and reactive power injection of the photovoltaic system, and The lines are The active power and reactive power on and The nodes are The active load and reactive load, It is a line The reactance, and Node The set of preceding nodes and subsequent nodes;

[0019] The voltage and current constraints of the distribution network are:

[0020] (6)

[0021] (7)

[0022] in, , Line The minimum and maximum values ​​of the current allowed, , Node The voltage allows minimum and maximum values;

[0023] The generator constraints are:

[0024] (8)

[0025] (9)

[0026] in, and The nodes are Maximum active power injection and reactive power injection of the upper generator;

[0027] The constraints of the photovoltaic system are:

[0028] (10)

[0029] in, For Node The maximum active power output of the photovoltaic system, is the power factor of the photovoltaic system;

[0030] The radial structure constraints are:

[0031] (11)

[0032] (12)

[0033] (13)

[0034] (14)

[0035] (15)

[0036] (16)

[0037] in, and are the larger positive number and the smaller positive number set respectively. is the total number of nodes in the distribution network, For Line The switch state, and The lines are Positive and negative signs of the upper trend.

[0038] In the aforementioned method for accelerating dynamic reconstruction of distribution network based on deep learning, in step 2, during the process of simplifying the power flow equation of distribution network, let , , and They are and The conjugate of , the power flow equation is expressed as:

[0039] (17)

[0040] (18)

[0041] (19)

[0042] (20)

[0043] (twenty one)

[0044] (twenty two).

[0045] In the aforementioned method for accelerating dynamic reconstruction of distribution networks based on deep learning, in step four, the deep learning method selects an Informer model, and the Informer model includes an encoder and a decoder. The encoder part is composed of multiple stacked attention blocks, and the output of each attention block passes through a one-dimensional convolution layer, an activation layer, and a pooling layer in sequence.

[0046] In the aforementioned method for accelerating dynamic reconstruction of distribution networks based on deep learning, in step 4, in the improved self-attention mechanism of the Informer model, The attention formula for a query vector query is:

[0047] (twenty three)

[0048] (twenty four)

[0049] (25)

[0050] in, represents the attention function, is the query vector, is the set of all key vectors, is the set of all value vectors, Represents the query vector With key vector The degree of correlation or similarity between Represents the multi-head attention mechanism for a given query vector The probability distribution of the calculated attention weights, Express expectations, Represents the value vector elements, is an exponential function, is the dimension of the vector, is the key vector set elements, is the key vector set elements.

[0051] In the aforementioned method for accelerating dynamic reconstruction of distribution networks based on deep learning, in step 4, the probability distribution is estimated The KL divergence with the uniform distribution is expressed as:

[0052] (26)

[0053] (27)

[0054] in, represents uniform distribution, represents the length of the key vector set, is the query vector Vector collection with keys A measure of the sparsity between .

[0055] The aforementioned method for accelerating dynamic reconstruction of distribution networks based on deep learning, if the probability distribution is estimated The difference between the uniform distribution and the previous The attention scores of the dominant query vectors are:

[0056] (28)

[0057] Where Q is a sparse matrix containing only the first The leading query vectors are queries.

[0058] The aforementioned distribution network dynamic reconstruction acceleration method based on deep learning also includes:

[0059] Step 5: Adaptively adjust the binary variables of the fixed distribution network dynamic reconstruction optimization model to determine whether the solution is feasible. If the solution is feasible, optimize the distribution network dynamic reconstruction optimization model.

[0060] The aforementioned method for accelerating the dynamic reconstruction of distribution networks based on deep learning is feasible to solve if the prediction result is within the set threshold range, and the solver completes the solution. If the prediction result is not within the set threshold range, the binary variable is adjusted to be fixed.

[0061] In the aforementioned method for accelerating dynamic reconstruction of distribution networks based on deep learning, in step four, in the deep learning method, the model used for long time series data prediction is a long short-term memory neural network or a Transformer model.

[0062] A computer device / equipment / system comprises a memory, a processor and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the above method.

[0063] The beneficial effects achieved by the present invention are as follows: the method of the present invention first forms a MIQP optimization model for dynamic reconstruction of distribution networks based on second-order cone relaxation, and secondly, by screening non-critical action switches and directly fixing their states, a deep learning method is used to predict the states of critical action switches within a time scale to reduce the number of binary variables, thereby greatly reducing the optimization solution time for dynamic reconstruction of distribution networks. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 is a flow chart of a method for accelerating dynamic reconstruction of a distribution network based on deep learning in Example 1 of the present invention;

[0065] Figure 2 Schematic diagram of the proportion of normally closed circuit switches in Embodiment 1 of the present invention;

[0066] Figure 3 It is a schematic diagram of the prediction and training errors of the Informer model in Example 1 of the present invention;

[0067] Figure 4a This is a schematic diagram of the prediction effect of the key action switch in Example 1 of the present invention;

[0068] Figure 4b This is a schematic diagram of the direct prediction effect in Example 1 of the present invention;

[0069] Figure 5 This is a diagram showing the effect of using LSTM to predict key action switches in Example 3 of the present invention;

[0070] Figure 6 It is a structural diagram of the Informer model in Example 1 of the present invention;

[0071] Figure 7-Figure 10 1 is a schematic diagram of prediction errors of different switch numbers at 4h, 8h, 11h and 13h in Example 1 of the present invention;

[0072] Fig.11 This is a schematic diagram of the distribution network reconstruction results at 1-11h, 22h, and 23h in Example 2 of the present invention;

[0073] Fig.12 This is a schematic diagram of the distribution network reconstruction result at 12h in Example 2 of the present invention;

[0074] Fig.13 This is a schematic diagram of the distribution network reconstruction result at 13h in Example 2 of the present invention;

[0075] Fig.14 This is a schematic diagram of the distribution network reconstruction result at 14h in Example 2 of the present invention;

[0076] Fig.15 This is a schematic diagram of the distribution network reconstruction result at 15h in Example 2 of the present invention;

[0077] Fig.16 This is a schematic diagram of the distribution network reconstruction result at 16h in Example 2 of the present invention;

[0078] Fig.17 This is a schematic diagram of the distribution network reconstruction result at 17h in Example 2 of the present invention;

[0079] Fig.18 This is a schematic diagram of the distribution network reconstruction result at 18h in Example 1 of the present invention;

[0080] Fig.19 This is a schematic diagram of the distribution network reconstruction result at 19h in Example 1 of the present invention;

[0081] Fig. 20 This is a schematic diagram of the distribution network reconstruction result at 20h in Example 1 of the present invention;

[0082] Fig.21 It is a schematic diagram of the distribution network reconstruction result at 21h in Example 1 of the present invention. DETAILED DESCRIPTION

[0084] The present invention is further described in detail below in conjunction with the accompanying drawings.

[0085] Example 1

[0086] like Figure 1 As shown, this embodiment provides a distribution network dynamic reconstruction acceleration method based on deep learning, comprising the following steps:

[0087] Step 1: Establish a dynamic reconstruction optimization model for distribution network;

[0088] Step 2: In the distribution network power flow equation constraint of the distribution network dynamic reconstruction optimization model, the distribution network power flow equation is simplified using the second-order cone relaxation method to form a MIQP model of distribution network dynamic reconstruction. The decision variables of the distribution network dynamic reconstruction model after the second-order cone relaxation include continuous variables and binary variables.

[0089] Step 3: select non-critical action switches according to the historical data of the switch status of the distribution network line, and fix the status of the non-critical action switches to reduce the number of binary variables;

[0090] Step 4: Use deep learning methods to predict the state of key action switches in the distribution network over a period of time, and optimize and solve the dynamic reconstruction optimization model of the distribution network.

[0091] In step 1, a distribution network dynamic reconstruction optimization model is established, including:

[0092] The objective function is to minimize the network loss:

[0093] (1)

[0094] in, is the set of all branches, is the set of all time points, For branch The current flowing through For branch The resistance of

[0095] The branch power flow equation constraints are:

[0096] (2) (3)

[0097] (4)

[0098] (5)

[0099] in, is the set of all nodes, For Node The voltage, For Node The voltage, , The nodes are Active power injection and reactive power injection of generators, and The nodes are Active power and reactive power injection of the photovoltaic system, and The lines are The active power and reactive power on and The nodes are The active load and reactive load, It is a line The reactance, and Node The set of preceding nodes and subsequent nodes;

[0100] The voltage and current constraints of the distribution network are:

[0101] (6)

[0102] (7)

[0103] in, , Line The minimum and maximum values ​​of the current allowed, , Node The voltage allows minimum and maximum values;

[0104] The generator constraints are:

[0105] (8)

[0106] (9)

[0107] in, and The nodes are Maximum active power injection and reactive power injection of the upper generator;

[0108] The constraints of the photovoltaic system are:

[0109] (10)

[0110] in, For Node The maximum active power output of the photovoltaic system, is the power factor of the photovoltaic system;

[0111] The radial structure constraints are:

[0112] (11)

[0113] (12)

[0114] (13)

[0115] (14)

[0116] (15)

[0117] (16)

[0118] in, and are the larger positive number and the smaller positive number set respectively. is the total number of nodes in the distribution network, For Line The switch state, and The lines are Positive and negative signs of the upper trend.

[0119] Formulas (2)-(5) are the constraints of the distribution network power flow equation, (6)-(10) are the constraints of the safe operation of the distribution network system, and (11)-(16) use the big M method to describe the line radial constraints. , and This problem becomes a mixed integer non-convex constraint, which is difficult to solve directly using existing commercial solvers. Therefore, the present invention adopts a second-order cone relaxation method to achieve relaxation of non-convex constraints.

[0120] In step 2, in the distribution network power flow equation constraint condition of the distribution network dynamic reconstruction optimization model, the distribution network power flow equation is simplified using the second-order cone relaxation method to form a MIQP model of distribution network dynamic reconstruction, including:

[0121] Since there are non-convex terms in the above distribution network dynamic reconstruction optimization model, it is difficult to solve the model. The existing commercial solvers cannot solve it directly. The second-order cone relaxation method can be used to relax the non-convex constraints. , , and They are and The conjugate of , the power flow equation is expressed as:

[0122] (17)

[0123] (18)

[0124] (19)

[0125] (20)

[0126] (twenty one)

[0127] (twenty two)

[0128] in, and Node Voltage, circuit The square of the current flowing through. The decision variables of the dynamic reconstruction model of the distribution network after second-order cone relaxation are continuous variables and binary variables. The continuous variables include nodes Active power injection from generators , reactive power injection ,line Active power on and reactive power ,line The square of the current flowing ,node The square of the voltage , binary variables include line Switching status ,line A positive sign of being on trend and reverse sign , which makes the original mixed-integer non-convex nonlinear model become a mixed-integer quadratic programming MIQP problem. For large-scale MIQP problems, due to the existence of a large number of integer variables, it is very time-consuming or even impossible to solve using existing solvers such as Gurobi and COPT. The solution time for the 96-point dynamic reconstruction of the IEEE 33-node distribution network takes about 150 seconds to complete. For larger-scale distribution networks, the solution time of dynamic reconstruction is greatly increased, and it may even be impossible to solve using existing commercial solvers.

[0129] In step 3, non-critical action switches are selected based on the historical data of the switch status of the distribution network line, and the status of the non-critical action switches is fixed to reduce the number of binary variables;

[0130] The switches of multiple lines in the dynamic reconstruction of the distribution network are often in a closed state. Therefore, by screening non-critical action switches based on the collected historical data and directly fixing the states of non-critical action switches, the number of switch states predicted by deep learning can be reduced, and the accuracy of the prediction can be improved.

[0131] Taking the IEEE 33-node distribution network system as an example, a total of 7997 historical data sets were collected. There are 37 lines in total, and the proportion of normally closed switches is Figure 2As shown in the figure, the red dotted line is the 0.98 reference line. The lines with a normally closed switch state ratio higher than 0.98 are selected as non-critical action switches and fixed directly, which reduces the corresponding binary variables generated by 17 line switches in total.

[0132] If non-critical action switches are not selected, all line switches are predicted directly, and the prediction error is as follows: Figure 4b As shown on the side, compared with Figure 4a If only the key action switches are predicted, the error of direct prediction will be greatly increased, affecting the optimality and feasibility of the optimization result obtained by fixing the variable state of the optimization model with the prediction results.

[0133] In step 4, a deep learning method is used to predict the state of key action switches of the distribution network within a time scale, and the deep learning method uses an Informer model;

[0134] This embodiment uses a deep learning method to predict the key action switch status of the distribution network line within a time scale to fix the binary variables, that is, long time series data prediction.

[0135] The Informer model is a commonly used model for long time series data prediction. The Informer model has the following three improvements: ProbSparse Self Attention is proposed, which can filter out a more important part of the query vector (Query), greatly reducing the time complexity and space complexity of the calculation; using the attention distillation operation, through convolution and pooling operations to reduce the dimension and network parameters, reduce the overall space complexity, and help receive long sequence input; using the generative decoder, all prediction results can be obtained in one step, which improves the speed of long sequence prediction and avoids the cumulative error propagation.

[0136] The structure of the Informer model is as follows Figure 6As shown in the figure, it includes an encoder and a decoder. The encoder part consists of multiple stacked attention blocks. The output of each block will pass through a one-dimensional convolution layer, an activation layer, and a pooling layer in sequence. The feature dimension is reduced by the pooling layer, that is, the feature distillation operation. After multiple operations, the final reduced representation feature map can be obtained, which improves the robustness of the algorithm. The Informer model also stacks multiple stacked layers. For the next stacked layer, when embedding at the beginning, the input length is directly reduced by half through the one-dimensional convolution layer. At the same time, in order to make the output dimension the same, the second stacked layer reduces one attention block; finally, the feature maps obtained by multiple stacked layers are spliced ​​to obtain the final output of the encoder. The Informer model uses a generative decoder. Unlike traditional decoders, the generative decoder generates all prediction outputs at once, while the traditional Transformer model puts the output of the previous step into the decoder and then obtains the output of the next step. In this way, each step can only output data for one time step. The generative decoder can effectively improve the prediction speed and reduce the prediction error.

[0137] The improved self-attention mechanism adopts the same optimization strategy for each head in the multi-head self-attention mechanism. The result after dot product follows a long-tail distribution, that is, the dot product results of a small number of query vectors and key vectors dominate. Based on this, The attention formula for a query vector query is:

[0138] (twenty three)

[0139] (twenty four)

[0140] (25)

[0141] in, represents the attention function, is the query vector, is the set of all key vectors, is the set of all value vectors, Represents the query vector With key vector The degree of correlation or similarity between Represents the multi-head attention mechanism for a given query vector The probability distribution of the calculated attention weights, Express expectations, Represents the value vector elements, is an exponential function, is the dimension of the vector, is the key vector set elements, is the key vector set elements.

[0142] Since we need to get a dominant result, we compare it with the uniform distribution and estimate the probability distribution. KL divergence with uniform distribution:

[0143] (26)

[0144] (27)

[0145] in, represents uniform distribution, represents the length of the key vector set, is the query vector Vector collection with keys A measure of the sparsity between .

[0146] If the estimated probability distribution If the difference between the uniform distribution and the Dominant dot product result, each key vector key and the previous The attention scores of the dominant query vectors are:

[0147] (28)

[0148] Where Q is a sparse matrix containing only the first The leading query vectors are queries.

[0149] like Figure 7 As shown in Figure 1, the z-axis is the prediction error of the key action switch. Since the prediction variable is a binary variable, when the absolute value of the prediction error is greater than 0.5, it indicates a prediction error. Taking 4h, 8h, 11h and 13h as examples, the change trend of the prediction error with the switch number is shown in Figure 1. Figure 7-Figure 10 As shown, there are two switch state prediction errors at 11h and 13h, but the number of positive peaks and negative peaks is consistent, which meets the necessary conditions for the radial structure constraint of the distribution network.

[0150] Example 2

[0151] On the basis of Example 1, it also includes:

[0152] Step 5: Adaptively adjust the binary variables of the fixed distribution network dynamic reconstruction optimization model to determine whether the solution is feasible. If the solution is feasible, quickly optimize and solve the distribution network dynamic reconstruction optimization model.

[0153] If all variables are fixed, the original MIQP problem will be transformed into a QP problem, and the solution speed will be greatly improved. However, there will be certain errors in the results predicted by the Informer model, which makes the final solution infeasible. Therefore, this embodiment uses adaptive fixed binary variables.

[0154] Adaptive optimization of fixed variables means that, in order to ensure the feasibility of the solution, if the prediction result is within the set threshold range, the solution is feasible and the solver completes the solution. If the prediction result is not within the set threshold range, the binary variable needs to be adjusted to fix it. The solution effect of applying different set thresholds to the prediction results of the deep learning model on a certain test instance is shown in Table 1.

[0155] Adaptive optimization of fixed variables means that in order to ensure the feasibility of the solution, if the result of deep learning prediction of the line switch state is within a certain threshold range, the prediction is considered accurate and the line switch state is fixed; if the prediction result is not within the threshold range, it is considered a prediction error, and the line switch state is not fixed, and it is still given to the solver as a quantity to be solved.

[0156] Table 1 Comparison of optimality and acceleration of different thresholds

[0157] The predicted value is less than The predicted value is greater than Number of binary variables Objective function value / MVar Solution time / s Threshold 1 0.1 0.9 88 1.8733 12.23 Threshold 2 0.2 0.8 52 1.8937 10.37 Threshold 3 0.3 0.7 incomprehensible incomprehensible incomprehensible Direct solution - - 1956 1.8719 255.79

[0158] It can be seen from Table 1 that both threshold 1 and threshold 2 can significantly reduce the number of binary variables and significantly speed up the solution. At the same time, the objective function value only increases by less than 0.1 MVar. To ensure feasibility, this embodiment adopts threshold 1.

[0159] Table 2 Comparison of solution effects of fixing all variables

[0160] method Constraints Number of variables Number of binary variables Objective function value / MVar Solution time / s Direct solution 31623 15726 1998 2.9239 142.00 The present invention 1004 1291 0 2.9256 1.36 contrast -96.82% -72.71% -100% +0.0017 -99.04%

[0161] Table 3 Comparison of solution effects of fixed variables

[0162] method Constraints Number of variables Number of binary variables Objective function value / MVar Solution time / s Direct solution 31497 15698 1970 1.9495 144.25 The present invention 1054 1116 7 1.9508 4.81 contrast -96.65% -92.89% -99.64% +0.0013 -96.67%

[0163] According to the prediction results of the model, the binary variables of the adaptive fixed optimization model can completely transform the MIQP problem into a QP problem when all variables are fixed. The acceleration effect is shown in Table 2, and the acceleration effect of fixing some binary variables is shown in Table 3. The dynamic reconstruction results of fixing all variables are shown in Table 3. Figure 11-21 As shown in the figure, only at 13h and 15h, the prediction error occurs, but the feasibility of the solution is still guaranteed. At the same time, the objective function value remains almost unchanged, and the solution time can be accelerated from hundreds of seconds to a few seconds, with an acceleration effect of up to 99.04%, which verifies the technical effect of the present invention.

[0164] Example 3

[0165] In step 4, in the deep learning method, the model used for long time series data prediction is the long short-term memory neural network (LSTM). LSTM has many parameters, high computational overhead, slow training and inference speed, and LSTM depends on the previous and next states in the sequence. There is a dependency between time steps, and error accumulation and propagation limit its performance in processing long time series. The results of using LSTM to predict key action switches are as follows: Figure 5 As shown in the figure, due to the error accumulation effect, the prediction error will become larger and larger over time. At 16-18h, the output result deviation is too large, affecting the feasibility and optimality of the final solution.

[0166] Or the model used for long time series data prediction is the Transformer model.

[0167] A computer device / equipment / system comprises a memory, a processor and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the above method.

[0168] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks 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 produce 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.

[0169] 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.

[0170] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0171] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A distribution network dynamic reconstruction acceleration method based on deep learning, characterized in that: The following steps are involved: Step 1: Establish a dynamic reconstruction optimization model for distribution network; Step 2: In the distribution network power flow equation constraint of the distribution network dynamic reconstruction optimization model, the distribution network power flow equation is simplified using the second-order cone relaxation method to form a MIQP model of distribution network dynamic reconstruction. The decision variables of the distribution network dynamic reconstruction model after the second-order cone relaxation include continuous variables and binary variables. Step 3: select non-critical action switches according to the historical data of the switch status of the distribution network line, and fix the status of the non-critical action switches to reduce the number of binary variables; Step 4: Use deep learning methods to predict the state of key action switches in the distribution network over a period of time, and optimize and solve the dynamic reconstruction optimization model of the distribution network.

2. According to a method for accelerating dynamic reconstruction of distribution network based on deep learning according to claim 1, it is characterized in that: In step 1, a distribution network dynamic reconstruction optimization model is established, including: The objective function is to minimize the network loss: （1）； in, is the set of all branches, is the set of all time points, For branch The current flowing through For branch The resistance of The branch power flow equation constraints are: （2）； （3）； （4）； （5）； in, is the set of all nodes, For Node The voltage, For Node The voltage, , The nodes are Active power injection and reactive power injection of generators, and The nodes are Active power and reactive power injection of the photovoltaic system, and The lines are The active power and reactive power on and The nodes are The active load and reactive load, It is a line The reactance, and Node The set of preceding nodes and subsequent nodes; The voltage and current constraints of the distribution network are: （6）； （7）； in, , Line The minimum and maximum values ​​of the current allowed, , Node The voltage allows minimum and maximum values; The generator constraints are: （8）； （9）； in, and The nodes are Maximum active power injection and reactive power injection of the upper generator; The constraints of the photovoltaic system are: （10）； in, For Node The maximum active power output of the photovoltaic system, is the power factor of the photovoltaic system; The radial structure constraints are: （11）； （12）； （13）； （14）； （15）； （16）； in, and are the larger positive number and the smaller positive number set respectively. is the total number of nodes in the distribution network, For Line The switch state, and The lines are Positive and negative signs of the upper trend.

3. According to a method for accelerating dynamic reconstruction of distribution network based on deep learning according to claim 2, it is characterized in that: In step 2, when simplifying the distribution network power flow equation, let , , and They are and The conjugate of , the power flow equation is expressed as: （17）； （18）； （19）； （20）； （21）； （22）。 4. According to a method for accelerating dynamic reconstruction of distribution network based on deep learning according to claim 1, it is characterized in that: In step 4, the deep learning method selects the Informer model, which includes an encoder and a decoder. The encoder part is composed of multiple stacked attention blocks, and the output of each attention block passes through a one-dimensional convolution layer, an activation layer, and a pooling layer in sequence.

5. According to a method for accelerating dynamic reconstruction of distribution network based on deep learning according to claim 4, it is characterized in that: In step 4, in the improved self-attention mechanism of the Informer model, The attention formula for a query vector query is: （23）； （24）； （25）； in, represents the attention function, is the query vector, is the set of all key vectors, is the set of all value vectors, Represents the query vector With key vector The degree of correlation or similarity between Represents the multi-head attention mechanism for a given query vector The probability distribution of the calculated attention weights, Express expectations, Represents the value vector elements, is an exponential function, is the dimension of the vector, is the key vector set elements, is the key vector set elements.

6. According to a method for accelerating dynamic reconstruction of distribution network based on deep learning according to claim 5, it is characterized in that: In step 4, the probability distribution is estimated The KL divergence with the uniform distribution is expressed as: （26）； （27）； in, represents uniform distribution, represents the length of the key vector set, is the query vector Vector collection with keys A measure of the sparsity between .

7. The method for accelerating dynamic reconstruction of a distribution network based on deep learning according to claim 6, characterized in that: If the estimated probability distribution The difference between the uniform distribution and the previous The attention scores of the dominant query vectors are: （28）； Where Q is a sparse matrix containing only the first The leading query vectors are queries.

8. A distribution network dynamic reconstruction acceleration method based on deep learning according to any one of claims 1 to 7, characterized in that: Also includes: Step 5: Adaptively adjust the binary variables of the fixed distribution network dynamic reconstruction optimization model to determine whether the solution is feasible. If the solution is feasible, optimize the distribution network dynamic reconstruction optimization model.

9. The method for accelerating dynamic reconstruction of a distribution network based on deep learning according to claim 8 is characterized in that: If the prediction result is within the set threshold range, the solution is feasible and the solver completes the solution. If the prediction result is not within the set threshold range, the binary variable is adjusted to be fixed.

10. A computer device / equipment / system comprising a memory, a processor and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 7.

Images