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

Abstract

The application discloses a power distribution network dynamic reconstruction acceleration method and system based on deep learning, which comprises the following steps: step one, establishing a power distribution network dynamic reconstruction optimization model; step two, in the power distribution network dynamic reconstruction optimization model, using a second-order cone relaxation method to simplify the power distribution network power flow equation to form a MIQP model of the power distribution network dynamic reconstruction; step three, according to the historical data of the power distribution network line switch state, screening non-key action switches, and fixing the non-key action switch state to reduce the number of binary variables; step four, using a deep learning method to predict the state of the power distribution network key action switch in a time scale. The application predicts the state of the power distribution network line key action switch in a time scale through an Informer model, so that the number of binary variables is reduced, and the solving time of the power distribution network dynamic reconstruction is greatly reduced.

Description

Technical Field

[0001] This invention relates to a method for accelerating dynamic reconfiguration of distribution networks, belonging to the field of power system grid reconfiguration technology. Background Technology

[0002] Dynamic reconfiguration of distribution networks optimizes the operation of the distribution system by altering its network topology to change power flow direction, thereby achieving load balancing between different areas, reducing line overload, improving voltage stability, and lowering network losses. Currently, the main methods for solving dynamic reconfiguration models of distribution networks include analytical algorithms, heuristic algorithms, and genetic algorithms. While heuristic and genetic algorithms can reduce computational complexity to some extent, they are inefficient and prone to getting trapped in local optima. Therefore, analytical methods can be considered to construct mathematical models to solve dynamic reconfiguration models of distribution networks.

[0003] The dynamic reconfiguration model of a distribution network can be constructed as a highly complex non-convex nonlinear programming problem, which is difficult to solve using current commercial solvers. For large-scale, long-term dynamic reconfiguration optimization problems, due to the large number of binary variables, even after performing second-order cone relaxation to transform it into a MIQP (Mixed Integer Quadratic Programming) problem, inefficiency and difficulty in solving remain, and it may even be impossible to solve using existing solvers. Therefore, accelerating the use of analytical methods to solve the dynamic reconfiguration optimization model of the distribution network is a problem that needs further investigation. Summary of the Invention

[0004] The technical problem to be solved by this invention is that the analytical method for solving the dynamic reconfiguration model of the distribution network is inefficient and difficult to solve.

[0005] To address the aforementioned technical problems, this invention provides a deep learning-based method for accelerating dynamic reconfiguration of distribution networks, comprising the following steps:

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

[0007] Step 2: In the power flow equation constraints of the power flow optimization model of the power flow dynamic reconfiguration, the second-order cone relaxation method is used to simplify the power flow equation of the power flow dynamic reconfiguration to form the MIQP model of the power flow dynamic reconfiguration. The decision variables of the power flow dynamic reconfiguration model of the power flow dynamic reconfiguration after second-order cone relaxation include continuous variables and binary variables.

[0008] Step 3: Screen non-critical operation switches based on historical data of the status of distribution network line switches, and fix the status of non-critical operation switches to reduce the number of binary variables;

[0009] Step four: Use deep learning methods to predict the state of key action switches in the distribution network over a certain time scale, and optimize the solution of the dynamic reconfiguration optimization model of the distribution network.

[0010] The aforementioned deep learning-based method for accelerating dynamic reconfiguration of distribution networks, in step one, establishes a dynamic reconfiguration optimization model for the distribution network, including:

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

[0012] (1)

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

[0014] The constraints for the branch power flow equations are:

[0015] (2)

[0016] (3)

[0017] (4)

[0018] (5)

[0019] in, The set of all nodes. For nodes voltage, For nodes voltage, , These are nodes Active power injection and reactive power injection from the generator. and These are nodes The active and reactive power injection of the photovoltaic system and These are the lines Active power and reactive power on and These are nodes Active and reactive loads, It is a line Reactance, and They are nodes The set of predecessor and successor nodes;

[0020] The voltage and current constraints of the distribution network are:

[0021] (6)

[0022] (7)

[0023] in, , The lines are respectively The minimum and maximum allowable current values, , They are nodes The minimum and maximum allowable voltage values;

[0024] The generator constraints are:

[0025] (8)

[0026] (9)

[0027] in, and These are nodes Maximum active power injection and reactive power injection of the generator;

[0028] The constraints of the photovoltaic system are:

[0029] (10)

[0030] in, For nodes The maximum active power output of a photovoltaic system. The power factor of the photovoltaic system;

[0031] The constraints for the radial structure are:

[0032] (11)

[0033] (12)

[0034] (13)

[0035] (14)

[0036] (15)

[0037] (16)

[0038] in, and These are the larger and smaller positive numbers, respectively. This represents the total number of nodes in the distribution network. For the line The on / off state, and These are the lines Positive and negative signs of trends.

[0039] In the aforementioned deep learning-based method for accelerating dynamic reconfiguration of distribution networks, during the simplification of the distribution network power flow equations in step two, let... , , and They are and The conjugate of is then expressed as:

[0040] (17)

[0041] (18)

[0042] (19)

[0043] (20)

[0044] (twenty one)

[0045] (twenty two).

[0046] In the aforementioned deep learning-based method for accelerating dynamic reconfiguration of power distribution networks, in step four, the deep learning method uses the Informer model, which includes an encoder and a decoder. The encoder consists of multiple stacked attention blocks, and the output of each attention block passes through a one-dimensional convolutional layer, an activation layer, and a pooling layer in sequence.

[0047] In the aforementioned deep learning-based method for accelerating dynamic reconfiguration of distribution networks, in step four, the improved self-attention mechanism of the Informer model... The attention formula for each query vector is:

[0048] (twenty three)

[0049] (twenty four)

[0050] (25)

[0051] in, Represents the attention function. It is a query vector. It is the set of all key vectors. It is the set of all value vectors. Represents the query vector With key vector The degree of correlation or similarity between them This indicates that in a multi-head attention mechanism, given a query vector... The calculated probability distribution of attention weights, Expressing expectations, Represents the first value vector One element, It is an exponential function. It is the dimension of the vector. It is the first key vector set One element, It is the first key vector set Each element.

[0052] In the aforementioned deep learning-based method for accelerating dynamic reconfiguration of distribution networks, step four involves estimating the probability distribution. The KL divergence between a uniform distribution and a homogeneous distribution is expressed as:

[0053] (26)

[0054] (27)

[0055] in, Indicates uniform distribution. This represents the length of the key vector set. It is a query vector With key vector set A measure of sparsity between them.

[0056] The aforementioned deep learning-based method for accelerating dynamic reconfiguration of distribution networks, if the probability distribution is estimated... The difference between the uniform distribution and the set value is greater than the set value, and each key vector key and the previous The attention scores for the dominant query vectors are:

[0057] (28)

[0058] Where Q is a sparse matrix containing only the first few elements. The dominant query vectors are queries.

[0059] The aforementioned method for accelerating dynamic reconfiguration of distribution networks based on deep learning also includes:

[0060] Step 5: Adaptively adjust the binary variables of the fixed distribution network dynamic reconfiguration optimization model, determine whether the solution is feasible, and if the solution is feasible, optimize the distribution network dynamic reconfiguration optimization model.

[0061] The aforementioned deep learning-based method for accelerating dynamic reconfiguration of power distribution networks is feasible if the prediction result is within a set threshold range, and the solution is completed by the solver. If the prediction result is outside the set threshold range, the binary variables are adjusted and fixed.

[0062] In the aforementioned deep learning-based method for accelerating dynamic reconfiguration of power distribution networks, in step four, the model used for long-term series data prediction in the deep learning method is a long short-term memory neural network or a Transformer model.

[0063] A computer device / apparatus / system includes 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-described method.

[0064] The beneficial effects achieved by this invention are as follows: First, the method of this invention forms a distribution network dynamic reconfiguration MIQP optimization model based on second-order cone relaxation. Second, by screening non-critical action switches and directly fixing their states, a deep learning method is used to predict the state of critical action switches within a certain time scale to reduce the number of binary variables and significantly reduce the optimization solution time of distribution network dynamic reconfiguration. Attached Figure Description

[0065] Figure 1 This is a flowchart of the deep learning-based method for accelerating dynamic reconfiguration of power distribution networks in Embodiment 1 of the present invention;

[0066] Figure 2 This is a schematic diagram showing the proportion of normally closed circuit switches in Embodiment 1 of the present invention;

[0067] Figure 3 This is a schematic diagram of the prediction and training errors of the Informer model in Embodiment 1 of the present invention;

[0068] Figure 4a This is a schematic diagram of the prediction effect of the key action switch in Embodiment 1 of the present invention;

[0069] Figure 4b This is a schematic diagram of the direct prediction effect in Embodiment 1 of the present invention;

[0070] Figure 5 This is a diagram illustrating the effect of using LSTM to predict key action switches in Embodiment 3 of the present invention.

[0071] Figure 6 This is a structural diagram of the Informer model in Embodiment 1 of the present invention;

[0072] Figures 7-10 This is a schematic diagram of the prediction error for different switch numbers at times 4h, 8h, 11h, and 13h in Embodiment 1 of the present invention.

[0073] Figure 11 This is a schematic diagram of the power distribution network reconfiguration results at times 1-11h, 22h, and 23h in Embodiment 2 of the present invention;

[0074] Figure 12 This is a schematic diagram of the distribution network reconfiguration results at 12 hours in Embodiment 2 of the present invention;

[0075] Figure 13 This is a schematic diagram of the distribution network reconfiguration results at time 13h in Embodiment 2 of the present invention;

[0076] Figure 14 This is a schematic diagram of the distribution network reconfiguration results at 14 hours in Embodiment 2 of the present invention;

[0077] Figure 15 This is a schematic diagram of the distribution network reconfiguration results at 15 hours in Embodiment 2 of the present invention;

[0078] Figure 16 This is a schematic diagram of the distribution network reconfiguration results at time 16h in Embodiment 2 of the present invention;

[0079] Figure 17 This is a schematic diagram of the distribution network reconfiguration results at time 17h in Embodiment 2 of the present invention;

[0080] Figure 18 This is a schematic diagram of the distribution network reconfiguration results at 18 hours in Embodiment 1 of the present invention;

[0081] Figure 19 This is a schematic diagram of the distribution network reconfiguration results at time 19h in Embodiment 1 of the present invention;

[0082] Figure 20 This is a schematic diagram of the distribution network reconfiguration results at time 20h in Embodiment 1 of the present invention;

[0083] Figure 21 This is a schematic diagram of the power distribution network reconfiguration results at time 21h in Embodiment 1 of the present invention. Detailed Implementation

[0084] The present invention will now be described in further detail with reference to the accompanying drawings.

[0085] Example 1

[0086] like Figure 1 As shown, this embodiment provides a deep learning-based method for accelerating dynamic reconfiguration of distribution networks, including the following steps:

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

[0088] Step 2: In the power flow equation constraints of the power flow optimization model of the power flow dynamic reconfiguration, the second-order cone relaxation method is used to simplify the power flow equation of the power flow dynamic reconfiguration to form the MIQP model of the power flow dynamic reconfiguration. The decision variables of the power flow dynamic reconfiguration model of the power flow dynamic reconfiguration after second-order cone relaxation include continuous variables and binary variables.

[0089] Step 3: Screen non-critical operation switches based on historical data of the status of distribution network line switches, and fix the status of non-critical operation switches to reduce the number of binary variables;

[0090] Step four: Use deep learning methods to predict the state of key action switches in the distribution network over a certain time scale, and optimize the solution of the dynamic reconfiguration optimization model of the distribution network.

[0091] In step one, a dynamic reconfiguration optimization model for the distribution network is established, including:

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

[0093] (1)

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

[0095] The constraints for the branch power flow equations are:

[0096] (2) (3)

[0097] (4)

[0098] (5)

[0099] in, The set of all nodes. For nodes voltage, For nodes voltage, , These are nodes Active power injection and reactive power injection from the generator. and These are nodes The active and reactive power injection of the photovoltaic system and These are the lines Active power and reactive power on and These are nodes Active and reactive loads, It is a line Reactance, and They are nodes The set of predecessor and successor nodes;

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

[0101] (6)

[0102] (7)

[0103] in, , The lines are respectively The minimum and maximum allowable current values, , They are nodes The minimum and maximum allowable voltage values;

[0104] The generator constraints are:

[0105] (8)

[0106] (9)

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

[0108] The constraints of the photovoltaic system are:

[0109] (10)

[0110] in, For nodes The maximum active power output of a photovoltaic system. The power factor of the photovoltaic system;

[0111] The constraints for the radial structure are:

[0112] (11)

[0113] (12)

[0114] (13)

[0115] (14)

[0116] (15)

[0117] (16)

[0118] in, and These are the larger and smaller positive numbers, respectively. This represents the total number of nodes in the distribution network. For the line The on / off state, and These are the lines Positive and negative signs of trends.

[0119] Formulas (2)-(5) are the power flow equation constraints of the distribution network, (6)-(10) are the safety operation constraints of the distribution network system, and (11)-(16) use the Big M method to describe the radial constraints of the lines. Due to the binary variables... , and This problem becomes a mixed integer non-convex constraint, which is difficult to solve directly using existing commercial solvers. Therefore, this invention uses a second-order cone relaxation method to relax the non-convex constraint.

[0120] In step two, the power flow equations of the distribution network dynamic reconfiguration optimization model are simplified using a second-order cone relaxation method to form the MIQP model for distribution network dynamic reconfiguration, including:

[0121] Because the aforementioned dynamic reconfiguration optimization model for the distribution network contains non-convex terms, making the model difficult to solve, existing commercial solvers cannot directly solve it. A second-order cone relaxation method can be used to relax the non-convex constraints, allowing... , , and They are and The conjugate of is then expressed as:

[0122] (17)

[0123] (18)

[0124] (19)

[0125] (20)

[0126] (twenty one)

[0127] (twenty two)

[0128] in, and They are nodes voltage, circuit The square of the current flowing through the distribution network, after second-order cone relaxation, has decision variables that include continuous variables and binary variables. Continuous variables include nodes. Active power injection from generator Reactive power injection ,line active power and reactive power ,line The square of the current flowing through ,node The square of the voltage The binary variables include the line Switch status ,line A positive sign of being on trend and reverse sign This transforms the original mixed-integer nonconvex nonlinear model into a mixed-integer quadratic programming (MIQP) problem. For large-scale MIQP problems, due to the large number of integer variables, solving them using existing solvers such as Gurobi and COPT is very time-consuming, or even impossible. The solution time for the 96-point dynamic reconfiguration of the IEEE 33-node distribution network takes about 150 seconds to complete. For even larger-scale distribution networks, the solution time for dynamic reconfiguration increases significantly, and it may even be impossible to complete the solution using existing commercial solvers.

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

[0130] In dynamic reconfiguration of distribution networks, the switches of multiple lines are often in the closed state. Therefore, by filtering non-critical switches based on collected historical data and fixing their states directly, the number of switch states predicted by deep learning can be reduced, while improving the accuracy of prediction.

[0131] Taking the IEEE 33-bus distribution network system as an example, a total of 7,997 historical datasets were collected, covering 37 lines. Normally closed switches accounted for a significant portion of the datasets. Figure 2As shown in the figure, the red dashed line is the 0.98 reference line. Circuits 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 circuit switches.

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

[0133] In step four, the state of key action switches in the distribution network over a certain time scale is predicted using a deep learning method, wherein the deep learning method selected is the Informer model.

[0134] This embodiment uses deep learning methods to predict the status of key operation switches of distribution network lines over a time scale, in order to fix binary variables, i.e., long-term series data prediction.

[0135] The Informer model is commonly used for long-term series data prediction. The Informer model has the following three improvements: it proposes the ProbSparse Self Attention mechanism, which can filter out a portion of the more important query vectors, greatly reducing the time and space complexity of computation; it uses attention distillation operation, which reduces the dimensionality and number of network parameters through convolution and pooling operations, thereby reducing the overall space complexity and facilitating the reception of long-term input sequences; and it uses a generative decoder, which can obtain all prediction results in just one step, improving the speed of long-term series prediction while avoiding the propagation of accumulated errors.

[0136] The structure of the Informer model is as follows: Figure 6As shown, the model includes an encoder and a decoder. The encoder consists of multiple stacked attention blocks, and the output of each block sequentially passes through a one-dimensional convolutional layer, an activation layer, and a pooling layer. Pooling layers reduce the feature dimensionality, a process known as feature distillation. After multiple operations, a final reduced feature map is obtained, improving the algorithm's robustness. The Informer model also stacks multiple layers. For each subsequent stack, the input length is halved at the beginning of the embedding process using a one-dimensional convolutional layer. Simultaneously, to maintain the same output dimensionality, the second stack reduces one attention block. Finally, the feature maps from multiple stacks are concatenated to obtain the encoder's final output. The Informer model uses a generative decoder. Unlike traditional decoders, generative decoders generate all predicted outputs at once, while traditional Transformer models feed the output of the previous step into the decoder to obtain the output of the next step. This means each step can only output data for one time step. Generative decoders effectively improve prediction speed and reduce prediction errors.

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

[0138] (twenty three)

[0139] (twenty four)

[0140] (25)

[0141] in, Represents the attention function. It is a query vector. It is the set of all key vectors. It is the set of all value vectors. Represents the query vector With key vector The degree of correlation or similarity between them This indicates that in a multi-head attention mechanism, given a query vector... The calculated probability distribution of attention weights, Expressing expectations, Represents the first value vector One element, It is an exponential function. It is the dimension of the vector. It is the first key vector set One element, It is the first key vector set Each element.

[0142] Because a dominant result is required, a comparison with a uniform distribution is made to estimate the probability distribution. KL divergence with uniform distribution:

[0143] (26)

[0144] (27)

[0145] in, Indicates uniform distribution. This represents the length of the key vector set. It is a query vector With key vector set A measure of sparsity between them.

[0146] If the probability distribution is estimated If the difference between the uniform distribution and the set value is greater than the set value, then Dominant dot product result, each key vector key and the first The attention scores for the dominant query vectors are:

[0147] (28)

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

[0149] like Figure 7 As shown, the z-axis represents the prediction error of the critical action switch. Since the prediction variable is a binary variable, a prediction error is indicated when the absolute value of the prediction error is greater than 0.5. Taking 4h, 8h, 11h, and 13h as examples, the trend of prediction error with switch number is as follows. Figures 7-10 As shown, two prediction errors occurred at 11h and 13h, but the number of positive peaks and negative peaks were the same, which satisfies the necessary condition of the radial structure constraint of the distribution network.

[0150] Example 2

[0151] Based on Example 1, it also includes:

[0152] Step 5: Adaptively adjust the binary variables of the fixed distribution network dynamic reconfiguration optimization model, determine whether the solution is feasible, and if the solution is feasible, perform rapid optimization of the distribution network dynamic reconfiguration optimization model.

[0153] If all variables are fixed, the original MIQP problem will be transformed into a QP problem, greatly improving the solution speed. However, the results predicted using the Informer model have certain errors, making the final solution infeasible. Therefore, this embodiment uses adaptive fixed binary variables.

[0154] Adaptive optimization with fixed variables means that, to ensure the feasibility of the solution, if the prediction result is within a set threshold range, the solution is feasible and the solver completes the solution; if the prediction result is outside the set threshold range, the binary variables need to be adjusted and fixed. Table 1 shows the solution effect of applying different threshold values ​​to the prediction results of the deep learning model on a certain test instance.

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

[0156] Table 1. Comparison of Optimality and Acceleration for Different Thresholds

[0157] Predicted value less than Predicted value 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] As can be seen from Table 1, both threshold 1 and threshold 2 can significantly reduce the number of binary variables and significantly accelerate the solution. At the same time, the objective function value only increases by less than 0.1MVar. To ensure feasibility, threshold 1 is used in this embodiment.

[0159] Table 2 Comparison of Solution Results with All Variables Fixed

[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 This invention 1004 1291 0 2.9256 1.36 contrast -96.82% -72.71% -100% +0.0017 -99.04%

[0161] Table 3 Comparison of Solution Results with 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 This invention 1054 1116 7 1.9508 4.81 contrast -96.65% -92.89% -99.64% +0.0013 -96.67%

[0163] Based on the model's predictions, the adaptive fixed-variable optimization model can completely transform the MIQP problem into a QP problem when all variables are fixed, as shown in Table 2. The acceleration effect when some binary variables are fixed is shown in Table 3. The dynamic reconstruction results with all variables fixed are as follows: Figure 11-21 As shown, prediction errors only occurred at times 13h and 15h, but the feasibility of the solution was still guaranteed. Meanwhile, the objective function value remained almost unchanged, and the solution time was accelerated from hundreds of seconds to several seconds, achieving an acceleration of 99.04%, thus verifying the technical effectiveness of the invention.

[0164] Example 3

[0165] In step four, the model used for long-term series data prediction in deep learning methods is a Long Short-Term Memory (LSTM) neural network. LSTM has numerous parameters, high computational cost, and slow training and inference speeds. Furthermore, LSTM relies on the preceding and following states in the sequence, exhibiting dependencies between time steps. Error accumulation and propagation limit its performance in processing long-term series. The results of using LSTM to predict key action switches are shown below. Figure 5 As shown, due to the cumulative effect of error, the prediction error will increase over time. The excessive deviation of the output result at 16-18h affects the feasibility and optimality of the final solution.

[0166] Alternatively, the Transformer model may be used for long-term series data prediction.

[0167] A computer device / apparatus / system includes 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-described method.

[0168] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0169] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0170] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0171] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A deep learning-based method for accelerating dynamic reconfiguration of distribution networks, characterized in that, Includes the following steps: Step 1: Establish a dynamic reconfiguration optimization model for the distribution network; Step 2: In the power flow equation constraints of the power flow optimization model of the power flow dynamic reconfiguration, the second-order cone relaxation method is used to simplify the power flow equation of the power flow dynamic reconfiguration to form the MIQP model of the power flow dynamic reconfiguration. The decision variables of the power flow dynamic reconfiguration model of the power flow dynamic reconfiguration after second-order cone relaxation include continuous variables and binary variables. Step 3: Screen non-critical operation switches based on historical data of the status of distribution network line switches, and fix the status of non-critical operation switches to reduce the number of binary variables; Step four: Use deep learning methods to predict the state of key action switches in the distribution network over a certain time scale, and optimize the solution of the dynamic reconfiguration optimization model of the distribution network.

2. The method for accelerating dynamic reconfiguration of distribution networks based on deep learning according to claim 1, characterized in that, In step one, a dynamic reconfiguration optimization model for the distribution network is established, including: The objective function is to minimize network loss: （1）； in, The set of all branches, For the set of all time points, branch road The current flowing through it, branch road The resistance; The constraints for the branch power flow equations are: （2）； （3）； （4）； （5）； in, The set of all nodes. For nodes voltage, For nodes voltage, , These are nodes Active power injection and reactive power injection from the generator. and These are nodes The active and reactive power injection of the photovoltaic system and These are the lines Active power and reactive power on and These are nodes Active and reactive loads, It is a line Reactance, and They are nodes The set of predecessor and successor nodes; The voltage and current constraints of the distribution network are: （6）； （7）； in, , The lines are respectively The minimum and maximum allowable current values, , They are nodes The minimum and maximum allowable voltage values; The generator constraints are: （8）； （9）； in, and These are nodes Maximum active power injection and reactive power injection of the generator; The constraints of the photovoltaic system are: （10）； in, For nodes The maximum active power output of a photovoltaic system. The power factor of the photovoltaic system; The constraints for the radial structure are: （11）； （12）； （13）； （14）； （15）； （16）； in, and These are the larger and smaller positive numbers, respectively. This represents the total number of nodes in the distribution network. For the line The on / off state, and These are the lines Positive and negative signs of trends.

3. The method for accelerating dynamic reconfiguration of distribution networks based on deep learning according to claim 2, characterized in that, In step two, during the simplification of the power flow equations of the distribution network, let , , and They are and The conjugate of is then expressed as: （17）； （18）； （19）； （20）； （21）； （22）。 4. The method for accelerating dynamic reconfiguration of distribution networks based on deep learning according to claim 1, characterized in that, In step four, the deep learning method uses the Informer model, which includes an encoder and a decoder. The encoder consists of multiple stacked attention blocks, and the output of each attention block passes through a one-dimensional convolutional layer, an activation layer, and a pooling layer in sequence.

5. The method for accelerating dynamic reconfiguration of distribution networks based on deep learning according to claim 4, characterized in that, In step four, in the improved self-attention mechanism of the Informer model, the first... The attention formula for each query vector is: （23）； （24）； （25）； in, Represents the attention function. It is a query vector. It is the set of all key vectors. It is the set of all value vectors. Represents the query vector With key vector The degree of correlation or similarity between them This indicates that in a multi-head attention mechanism, given a query vector... The calculated probability distribution of attention weights, Expressing expectations, Represents the first value vector One element, It is an exponential function. It is the dimension of the vector. It is the first key vector set One element, It is the first key vector set Each element.

6. The method for accelerating dynamic reconfiguration of distribution networks based on deep learning according to claim 5, characterized in that, In step four, the probability distribution is estimated. The KL divergence between a uniform distribution and a homogeneous distribution is expressed as: （26）； （27）； in, Indicates uniform distribution. This represents the length of the key vector set. It is a query vector With key vector set A measure of sparsity between them.

7. The method for accelerating dynamic reconfiguration of distribution networks based on deep learning according to claim 6, characterized in that, If the probability distribution is estimated The difference between the uniform distribution and the set value is greater than the set value, and each key vector key and the previous The attention scores for the dominant query vectors are: （28）； Where Q is a sparse matrix containing only the first few elements. The dominant query vectors are queries.

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

9. The method for accelerating dynamic reconfiguration of distribution networks 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 outside the set threshold range, the binary variables are adjusted and 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 1-7.

Images