Active power distribution network real-time random scheduling method and system fused with fuzzy neural network pre-decision

Abstract

The invention discloses an active power distribution network real-time random scheduling method and system fused with fuzzy neural network pre-decision. The method comprises the following steps: re-deriving a power flow model based on an Euler equation; the uncertainty of photovoltaic output and the randomness of loads are described as the uncertainty of corresponding prediction errors, and probability distribution of uncertain variables is obtained through a data driving method; establishing a random optimization power scheduling model of the active power distribution network; a fuzzy neural network is fused to describe the uncertainty of the random optimization scheduling model, and historical data is used to train the random optimization scheduling model; inputting power distribution network information at the current moment, and pre-deciding a scheduling strategy by using the trained neural network; and taking the pre-decision value as an initial value of solver optimization, and then carrying out accurate solving, so that the model solving process is accelerated while the accuracy of a solving result is ensured. According to the method, the defect that errors are caused due to the fact that a power flow model based on second-order cone relaxation possibly violates relaxation conditions is overcome, and the real-time performance of a scheduling strategy is guaranteed.

Description

Technical Field

[0001] The present invention relates to the technical field of power distribution network optimization, and in particular to a real-time random dispatching method and system for an active power distribution network integrating fuzzy neural network pre-decision. Background Art

[0002] In order to alleviate environmental pollution, cope with the crisis of traditional energy, and reduce carbon emissions in the power industry, distributed photovoltaics are being connected to the distribution network in large quantities. In the future, the distribution network will contain a high proportion of new energy and distributed power supply access. In order to cope with the randomness of new energy, with a large number of distributed energy storage connected to the distribution network, the traditional distribution network is transforming into an active distribution network.

[0003] Optimal dispatching of distribution networks is a necessary way to achieve economic and safe operation of distribution networks. The solution of optimal dispatching models will become more frequent. However, as traditional distribution networks are becoming more and more complex, a large number of distributed resources are connected, and the uncertainty of distribution networks increases, the difficulty of solving models increases and the speed of solving them slows down. Establishing effective models and finding efficient solutions are challenges that must be faced. On the one hand, to establish an accurate and practical active distribution network dispatching model, the key lies in using a suitable and reliable power flow model and accurately describing its uncertainty. At present, the power flow equation is widely based on second-order cone relaxation, but it may not correctly reflect the power flow equation before relaxation due to the failure of relaxation conditions, resulting in large errors. On the other hand, it is necessary to develop a solution method with fast solution speed and good real-time performance. In solving optimization problems, traditional solvers may have the disadvantage of slow solution speed. Due to the development of deep neural networks, some scholars have used them to solve optimization problems, but the current neural network-based methods may have the problem of high training cost or deviation between the predicted results and the true value, resulting in the non-convergence of the optimization problem constraints. Summary of the invention

[0004] Purpose of the invention: In view of the increasing uncertainty in the distribution network, it is difficult to directly solve the optimization model when the number of variables increases, and there is a defect that the solution result may be inaccurate. The present invention proposes a real-time random scheduling method and system for active distribution networks that integrates fuzzy neural network pre-decision, which ensures the real-time performance of the model without sacrificing the solution accuracy.

[0005] Technical solution: A real-time random dispatching method for active distribution network integrating fuzzy neural network pre-decision, comprising the following steps:

[0006] Step 1: Re-derive the power flow model of the active distribution network based on the Euler equation, including the network loss equation, active power balance equation, reactive power balance equation, voltage drop equation, and line current constraint. The derived model is equivalent to the original power flow model.

[0007] Step 2: The power prediction values ​​of PV output and load are regarded as deterministic variables, and the prediction error is regarded as an uncertain variable. The actual power of PV output and load is decomposed into power prediction value and prediction error respectively. The uncertainty of PV output and the randomness of load are described as the uncertainty of corresponding prediction error, and the probability distribution of uncertain variables is obtained through data-driven method.

[0008] Step 3: Comprehensively consider the power supply cost of the upper substation, the network loss of the distribution network, the dispatching cost of the distributed energy storage system ESS, and the dispatching cost of the static VAR compensator SVC, and build the objective function of the active distribution network stochastic optimization power dispatching model with the goal of minimizing the cost. Its constraints include power flow constraints, safe operation constraints, substation injection power constraints, photovoltaic unit operation constraints, ESS operation constraints, SVC constraints, and flexible load constraints.

[0009] Step 4: Fusion fuzzy neural network is used to describe the uncertainty of the random optimization power dispatch model of active distribution network, and the fuzzy neural network is trained using historical data. The forward propagation process of the fuzzy neural network includes data input layer, membership function calculation layer, rule generation layer, normalization layer, and output layer. The back propagation process uses the negative gradient descent algorithm to optimize the center point of the membership function, the width vector, and the connection weights of the output layer to obtain the fuzzy neural network parameters.

[0010] Step 5: Input the real-time state data of the active distribution network, including the real-time power prediction value and the data-driven prediction error, into the trained fuzzy neural network, and use the fuzzy neural network to quickly output the dispatch strategy pre-decision;

[0011] Step 6: Use the pre-decision value as the initial value for the solver to find the optimal solution, and then use the solver to accurately solve the active distribution network random optimization power dispatching model in step (3) to obtain the real-time random dispatching strategy of the active distribution network.

[0012] A real-time random dispatching system for active distribution network integrating fuzzy neural network pre-decision, comprising:

[0013] The power flow model re-derivation module is used to re-derive the power flow model of the active distribution network based on the Euler equation, including the network loss equation, active power balance equation, reactive power balance equation, voltage drop equation, and line current constraint. The derived model is equivalent to the original power flow model;

[0014] The uncertainty description module is used to regard the power prediction values ​​of photovoltaic output and load as deterministic variables and the prediction error as uncertain variables, and decompose the actual power of photovoltaic output and load into power prediction value and prediction error respectively. The uncertainty of photovoltaic output and the randomness of load are described as the uncertainty of corresponding prediction error, and the probability distribution of uncertain variables is obtained through data-driven methods;

[0015] The dispatch model construction module is used to comprehensively consider the power supply cost of the upper substation, the network loss of the distribution network, the dispatch cost of the distributed energy storage system ESS, and the dispatch cost of the static VAR compensator SVC, and to construct the objective function of the active distribution network stochastic optimization power dispatch model with the goal of minimizing the cost. Its constraints include power flow constraints, safe operation constraints, substation injection power constraints, photovoltaic unit operation constraints, ESS operation constraints, SVC constraints, and flexible load constraints.

[0016] The fuzzy neural network description and training module is used to integrate the fuzzy neural network to describe the uncertainty of the random optimization power dispatch model of the active distribution network, and use historical data to train the fuzzy neural network. The forward propagation process of the fuzzy neural network includes the data input layer, the membership function calculation layer, the rule generation layer, the normalization layer, and the output layer. The reverse propagation process uses the negative gradient descent algorithm to optimize the center point of the membership function, the width vector, and the connection weights of the output layer to obtain the fuzzy neural network parameters;

[0017] The pre-decision acquisition module is used to input the real-time status data of the active distribution network, including the real-time power prediction value and the data-driven prediction error, into the trained fuzzy neural network, and use the fuzzy neural network to quickly output the scheduling strategy pre-decision;

[0018] The real-time random scheduling strategy acquisition module is used to use the pre-decision value as the initial value for the solver to find the optimal solution, and then use the solver to accurately solve the active distribution network random optimization power scheduling model in the scheduling model construction module to obtain the real-time random scheduling strategy of the active distribution network.

[0019] The present invention also provides a computer device, comprising: one or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and are configured to be executed by the one or more processors, and when the programs are executed by the processors, the steps of the real-time random scheduling method for active distribution networks integrating fuzzy neural network pre-decision are implemented as described above.

[0020] The present invention also provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps of the active distribution network real-time random scheduling method integrating fuzzy neural network pre-decision as described above are implemented.

[0021] Beneficial effects:

[0022] (1) The present invention re-derives the power flow model and accurately describes the uncertain variables of the active distribution network, and establishes an economical and safe random optimization power dispatching model for the active distribution network, which helps dispatchers to optimize the dispatching of active distribution networks with photovoltaic output fluctuations and random load changes.

[0023] (2) The present invention proposes a method of integrating fuzzy neural network pre-decision to efficiently solve the random optimization power scheduling model. The membership function of the fuzzy neural network can be used to characterize the probability distribution of the uncertain variables of random optimization, and the decision variables are pre-determined through the output of the neural network. The pre-decision value is used as the initial point for the solver to find the optimal solution, ultimately achieving the acceleration of the model solution without sacrificing the accuracy of the model solution. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 is a flow chart of the method of the present invention;

[0025] Figure 2 This is the improved IEEE 33-node system network topology diagram;

[0026] Figure 3 This is a comparison chart of the fuzzy neural network's predicted value and actual value of the energy storage charging and discharging power. DETAILED DESCRIPTION

[0027] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0028] Figure 1 The flowchart of the real-time random dispatching method of active distribution network integrating fuzzy neural network pre-decision proposed by the present invention is shown, which includes the following steps:

[0029] Step 1: Re-derive the power flow model based on the Euler equation. The proposed model is equivalent to the original power flow model and can be solved efficiently.

[0030] Step 2: Describe the uncertainty of PV output and the randomness of load as the uncertainty of the corresponding prediction error, and obtain the probability distribution of uncertain variables through data-driven methods;

[0031] Step 3: Based on Step 1 and Step 2, an economical and safe stochastic optimization power dispatch model for active distribution network is established. The objective function comprehensively considers multiple factors, and the constraints include power flow constraints, safe operation constraints, etc.

[0032] Step 4: Fusion fuzzy neural network is used to describe the uncertainty of the random optimization scheduling model in step 3, and historical data is used to train it to obtain neural network parameters;

[0033] Step 5: Input the current distribution network information and use the trained neural network to make a preliminary decision on the dispatching strategy;

[0034] Step 6: Use the pre-decision value as the initial value for the solver to find the optimal solution, and then use the solver to accurately solve the stochastic optimization scheduling model in step (3), thereby accelerating the model solving process while ensuring the accuracy of the solution results, and obtaining a real-time stochastic scheduling strategy for the active distribution network.

[0035] In the embodiment of the present invention, the improved IEEE 33-node system is used as the experimental object to explain in detail the specific implementation process of using the method of the present invention to perform random optimization power dispatching of active distribution network. Figure 2 As shown. In this system, 10 nodes are connected to distributed photovoltaics to simulate the uncertainty factors of new energy, and 6 energy storage devices and 10 SVCs are configured to ensure the stable operation of the system. The photovoltaic output is 10%-90% of its rated output, and the load is 80%-100% of the rated load of the IEEE33 node system. Among them, in order to generate the power of photovoltaics and loads at multiple times, the fluctuation range is uniformly sampled as the power prediction value; the typical scenario of power prediction error is obtained by scenario reduction from the historical data of typical similar days, and the historical data is generated by Monte Carlo simulation. Based on the constructed model data, the specific implementation steps of the method of the present invention are as follows:

[0036] (1) The power flow model is re-derived based on the Euler equation, including the network loss equation, active power balance equation, reactive power balance equation, voltage drop equation, and line current constraints:

[0037] a) For the network loss equation, the original network loss function formula is expressed as:

[0038]

[0039] Where: P mn and Q mn are the active and reactive power flowing through line mn, and are respectively the active and reactive losses on line mn, I mn is the current flowing through line mn, R mn and X mn are the resistance and reactance of line mn respectively; U m is the voltage amplitude at node m.

[0040] b) For the active and reactive power balance equations, the active and reactive power balance of node m is expressed as:

[0041]

[0042] Where: are the active power injected into node m by three distributed resources: substation, distributed photovoltaic, and distributed energy storage. are the reactive power injected into node m by substation, distributed photovoltaic and static VAR compensator respectively, are the active and reactive power of the load at node m, P km , Q km are the active and reactive power injected into node m by line km, P mn , Q mn are the active and reactive power flowing out of node m from line mn, A collection of lines.

[0043] c) For the voltage drop equation, the initial definition formula for voltage drop is:

[0044]

[0045] Where: Z mn is the line mn impedance, that is, Z mn =R mn +jX mn ; j is an imaginary unit, V m 、V n is the complex voltage of nodes m and n, that is, V m =|V m |∠θ m , V n =|V n |∠θ n ,|V m |、|V n | are the voltage amplitudes of nodes m and n, respectively, ∠θ m ,∠θ n are the voltage phase angles of nodes m and n respectively; S mn is the apparent power of the line, S mn =P mn +jQ mn ; and S mn and V m The complex conjugate of .

[0046] Since the initial definition formula of voltage drop contains complex terms, it is difficult to directly optimize and solve it. The voltage phase angle is relaxed as follows:

[0047] First, introduce the multiplication of both sides

[0048]

[0049] V m 、Vn , Z mn , S mn Expanded:

[0050]

[0051] ∠θ nm writing According to Euler's formula: e jθ =cosθ+jsinθ, the voltage phase angle difference ∠θ nm After rewriting, we get:

[0052] |V m | 2 -(|V m ||V n |cosθ nm +j|V m ||V n |sinθ nm )=(R mn +jX mn )(P mn -jQ mn )

[0053] Separate the real and imaginary parts:

[0054] |V m ||V n |cosθ mn =|V m | 2 -(R mn P mn +X mn Q mn )

[0055] |V m ||V n |sinθ mn =R mn Q mn -X mn P mn

[0056] Add the squares of the two equations:

[0057] (|V m | 2 -(R mn P mn +X mn Q mn )) 2 +(R mn Q mn -X mn P mn ) 2 =(|Vm ||V n |) 2

[0058] Expand the square term and divide by |V m | 2 :

[0059]

[0060] The above formula is the voltage drop formula after voltage phase angle relaxation. Compared with the commonly used second-order cone relaxation, this method does not need to introduce assumptions, is equivalent to the original power flow equation, and can be efficiently solved using existing nonlinear solvers.

[0061] d) For line current constraints, we have:

[0062]

[0063]

[0064]

[0065] Where: is the upper limit of the current amplitude flowing through line mn.

[0066] (2) Description of uncertainty of active distribution network:

[0067] First, due to the randomness of photovoltaic and load, the distribution network is usually dispatched based on the power prediction data of the two. Therefore, the power prediction data is a deterministic variable. The uncertain variables in the active distribution network are essentially the uncertainty of the prediction error. The actual power of photovoltaic and load is decomposed into its power prediction value and prediction error, which can be expressed as follows:

[0068] P PV =P PV,pre +ξ PV

[0069] P D0 =P D,pre +ξ DP

[0070] Q D0 =Q D,pre +ξ DQ

[0071] Where: P PV , P D0 , Q D0 are the actual values ​​of photovoltaic output and load active and reactive power, respectively. PV,pre , P D,pre , Q D,preare the predicted values ​​of photovoltaic output and load active and reactive power, respectively, PV , DP , DQ are the prediction errors of photovoltaic output and load active and reactive power respectively.

[0072] Secondly, the goal of stochastic optimization is to optimize the mathematical expectation value of the objective function under uncertainty. The uncertain variable in the active distribution network is the uncertainty of the power prediction error. In order to avoid the excessive dimension of the stochastic optimization scheduling model caused by too many historical data scenarios of uncertain variables, resulting in slow model solution or inability to solve directly, the k-means clustering algorithm and the synchronous back-substitution elimination technology based on Kantorovich probability distance are used to reduce the prediction data of similar days of power prediction to obtain typical prediction error scenarios and corresponding probabilities of typical scenarios: first, the k-means clustering algorithm is used to divide the original scenarios into different clusters, and then the synchronous back-substitution elimination algorithm based on Kantorovich distance is used to eliminate the scenario sets in each cluster into the only typical scenario. The final scene set of each cluster is the reduced scene set that can represent the original scene set, so that a small number of typical scenes are used to effectively represent all historical data scenes, taking into account the accuracy and real-time performance of the model solution. Since the scene reduction technology used is an existing technology, the specific details are not repeated here.

[0073] (3) Constructing a random power dispatch model for active distribution networks:

[0074] Firstly, considering the power supply cost of the upper substation, the network loss of the distribution network, the dispatching cost of the distributed energy storage system (ESS), and the dispatching cost of the static var compensator (SVC), the objective function of the stochastic optimization power dispatching model of the active distribution network is constructed as follows:

[0075]

[0076] Where: Ω L ,Ω ESS ,Ω SVC ,Ω S They are line set, ESS node set, SVC node set, and typical scene set after data driving; K T , K L , K E , K S are the power supply cost coefficient of substation, network loss coefficient, ESS dispatch coefficient, and SVC dispatch coefficient respectively; P T , P l L , P i,ESS , Q i,SVCis the active power output of the substation, the active power loss of the line, the charging and discharging power of the ESS node, and the reactive power output of the SVC node, where P i,ESS The positive and negative signs represent the charge and discharge status of the ESS, with positive for discharge and negative for charge. π(S) is the probability corresponding to a typical scenario driven by data.

[0077] Secondly, the constraints for safe operation of distribution networks include:

[0078] a) Power flow constraints:

[0079]

[0080]

[0081]

[0082]

[0083]

[0084] Where: P mn,S and Q mn,S are the active and reactive power flowing through line mn in scenario S, and are respectively the active and reactive losses on line mn in scenario S, R mn and X mn are the resistance and reactance of line mn respectively, U m,S , U n,S are the voltage amplitudes of nodes m and n under scenario S, is the active power injected into node m by distributed photovoltaic in scenario S, are the active power injected into node m by the distributed energy storage in the substation, are the reactive power injected into node m by substation, distributed photovoltaic and static VAR compensator respectively, are the active and reactive power of the load at node m, P km,S , Q km,S are the active and reactive power injected into node m by line km under scenario S, P mn,S , Q mn,S are the active and reactive power of line mn flowing out of node m under scenario S, A collection of lines.

[0085] b) Safe operation constraints, including line transmission current constraints, line transmission power constraints and node voltage constraints:

[0086]

[0087] V imin ≤V i,S ≤V i max

[0088] Where: P mn,S , Q mn,S are respectively the active and reactive power transmitted by line mn in scenario S, S mn,max is the upper limit of the current carrying capacity of line mn, V i,S is the voltage amplitude of node i in scenario S, V i max 、V i min are the upper and lower limits of the specified node voltage, is the upper limit of the current amplitude flowing through line mn.

[0089] c) Substation injection power constraints:

[0090]

[0091] Where: P T , Q T Provide active and reactive power for the substation. It is the upper and lower limits of active and reactive power output of substation.

[0092] d) PV unit operation constraints:

[0093]

[0094] Where: P PV,S is the active power output of distributed photovoltaics in scenario S, Q PV , S PV They are the reactive output and capacity of distributed photovoltaics respectively.

[0095] e) ESS operation constraints:

[0096]

[0097]

[0098]

[0099] Where: P i,ESS , is the ESS charging and discharging power, and the upper and lower limits of the charging and discharging power, is the ESS power at adjacent moments, is the upper and lower limits of power, η loss , η ESS They are power loss coefficient and charge and discharge efficiency respectively.

[0100] f) SVC constraints:

[0101]

[0102] Where: Q i,SVC , It is the reactive power compensated by SVC and the upper and lower limits of reactive power compensation.

[0103] g) Flexible load constraints:

[0104] (1-K P )*P D0 ≤P D ≤(1+K P )*P D0

[0105] (1-K Q )*Q D0 ≤Q D ≤(1+K Q )*Q D0

[0106] Where: P D0 , Q D0 is the load active and reactive power calculated according to step (2), K P , K Q is the proportion of active and reactive flexible loads, P D , Q D It is the active and reactive power of the load after flexible load control.

[0107] (4) Fusion fuzzy neural network is used to describe the uncertainty of the stochastic optimization scheduling model in step (3), and historical data is used to train it to obtain neural network parameters:

[0108] Firstly, the forward propagation process of historical data feature information of fuzzy neural network learning stochastic optimization scheduling model includes data input layer, membership function calculation layer, rule generation layer, normalization layer and output layer.

[0109] a) Data input layer: The number of neurons in this layer is the characteristic dimension of the input data, that is, when the characteristic dimension of the data is n, the number of neurons in the input layer is n. The input data includes the predicted photovoltaic active power P PV,pre , predicted load active and reactive power P D,pre , Q D,pre And the prediction error ξ in each typical scenario PV , DP , DQ , the input variables contain uncertainty.

[0110] b) Membership function calculation layer: Each neuron represents a fuzzy quantity. The membership function is used to calculate the probability that each neuron in the input layer belongs to the fuzzy quantity. The number of neurons is the number of fuzzy conditions that the input variables may constitute. All neurons constitute a fuzzy set. This layer is used to describe the probability of uncertain input variables. Since the range of uncertain variables in the random optimization scheduling model proposed in step (3) is the set of all typical scenarios, the number of fuzzy classifications of input variables using this layer is the number of typical scenarios, and the number of nodes in this layer is the number of typical scenarios; the membership function uses a Gaussian function, as shown below:

[0111]

[0112] Where: u ij is the output of the neuron, x i is the input data, c ij is the center point of the membership function, b ij is the width vector of the membership function, and m is the number of fuzzy classifications of the input, which is the number of typical scenarios in this paper.

[0113] c) Rule generation layer, each neuron represents a fuzzy rule. The output of this layer is the applicability of each rule, which is calculated using the following formula:

[0114] α j =Πu ij ,j=1,2,...,m

[0115] d) Normalization layer, normalizes the fitness of each rule in the rule generation layer using the following formula:

[0116]

[0117] e) The output layer performs weighted processing to obtain the final output. The calculation formula is:

[0118]

[0119] Where: y i is the output variable, w ij is the corresponding weight, and k is the number of output variables.

[0120] Then, the back propagation process of the historical data feature information of the fuzzy neural network learning stochastic optimization scheduling model uses the negative gradient descent algorithm to optimize the center point of the membership function, the width vector, and the connection weights of the output layer.

[0121] a) Calculate the negative gradient of the error E with respect to the output layer as follows:

[0122]

[0123] b) Calculate the negative gradient of the error E with respect to the normalization layer, the rule generation layer, and the membership function layer in turn, and obtain the negative gradient of the error E with respect to the membership function layer:

[0124]

[0125] c) Error E versus weight w ij 、Center point c ij and width vector b ij The negative gradient of is:

[0126]

[0127] d) The final parameter adjustment algorithm is:

[0128]

[0129] Where: η, ξ, ψ are the learning rates of the corresponding parameters, and r is the number of rounds of neural network learning.

[0130] (5) Input information reflecting the real-time status of the distribution network, i.e., real-time power prediction data P PV,pre , P D,pre , Q D,pre And the typical scene error ξ obtained by data driving PV , DP , DQ , using the trained neural network in step (4) to quickly output the estimated value of the scheduling variable Make preliminary decisions on scheduling strategies.

[0131] (6) Using the mature nonlinear solver IPOPT, the pre-decision value is used as the initial point for the solver to find the optimal solution, and the scheduling variables are accelerated to ensure the accuracy of the results, thus realizing the real-time solution of the random optimization scheduling model proposed in step (3).

[0132] In order to illustrate the effectiveness of the proposed scenario-based stochastic optimization power dispatching method for active distribution networks, the number of typical scenarios extracted from historical data is 5, 10, 20, and 30, and the objective function values ​​at 20 moments are finally solved as shown in Table 1. As can be seen from Table 1, when the number of typical scenarios is 10, 20, and 30, the objective function values ​​at 20 moments are basically stable, and the difference in the objective function values ​​is basically within ±100. It can be foreseen that if the number of typical scenarios is increased, the objective function value will also be stable in the stable range, indicating that the typical scenarios constructed from historical data by the proposed method can effectively characterize uncertain variables.

[0133] Table 1 The impact of the number of typical scenarios on the objective function

[0134]

[0135]

[0136] In order to verify the performance of the proposed fusion fuzzy neural network pre-decision solving method, the original problem is solved directly using the solver as a comparison object; at the same time, in order to reduce the random influence of computer performance on the calculation time, 20 time sections are solved, and then the total solution time is obtained. The simulation results under different typical scene numbers are shown in Table 2. In Table 2, T0 is the time required for direct use of the solver, T1 is the time required for the second stage correction of the solver, T2 is the fuzzy neural network pre-decision time, and T 总 is the total time required for the fusion fuzzy neural network solution method, and α is the solution speed acceleration ratio, that is, the acceleration effect of the proposed method compared with the direct solution.

[0137] Table 2 Comparison of computing time under different typical scenarios

[0138]

[0139] As can be seen from Table 2, when the proposed model is directly solved by the solver, each solution takes 249.023 seconds when the number of scenarios is 30, and the time scale of real-time scheduling is 5 minutes, which barely meets the requirements, but the time margin is small. Considering the time required for power prediction, communication delay, etc., its real-time performance needs to be further improved; the solution method integrating fuzzy neural network pre-decision takes 159.827 seconds when the number of scenarios is 30, which better meets the requirements of real-time scheduling. At the same time, the acceleration effect of the proposed method is more than 30% under each typical scenario number, especially when the number of typical scenarios is 30, the proposed method can shorten the solution time significantly, and the total time for solving 20 power scheduling is shortened by 1760 seconds, which can effectively improve the solution speed of the model and better meet the real-time performance requirements under the current environment. In addition, as the number of typical scenarios increases, the complexity of the original optimization problem increases, which leads to a significant increase in the time required to solve the problem directly using the solver. However, the time required for the first-stage pre-decision-making using the fuzzy neural network remains basically unchanged. Although the time required for the second-stage correction using the solver also increases, overall, the acceleration ratio is slightly improved, which shows that the proposed method is more effective in improving the efficiency of solving complex systems.

[0140] Figure 3 Taking the energy storage system installed at node 12 as an example, the scatter plot between the pre-decision value and the true value of the fuzzy neural network in the test set is shown with 20 typical scenarios and 100 samples. Figure 3 The horizontal axis is the true value, and the vertical axis is the pre-decision value. It can be seen that the data points are roughly located near the straight line with a slope of 1, indicating that the pre-decision value is close to the true value, which shows that the fuzzy neural network has good pre-decision performance.

[0141] In summary, the method of the present invention proposes a real-time random scheduling method for active distribution networks that integrates fuzzy neural network pre-decision, performs real-time optimization scheduling on active distribution networks with uncertainty, and verifies the effectiveness of the proposed method through example analysis.

[0142] Based on the same technical concept as the method embodiment, the present invention also provides a real-time random dispatching system for active distribution network integrating fuzzy neural network pre-decision, comprising:

[0143] The power flow model re-derivation module is used to re-derive the power flow model of the active distribution network based on the Euler equation, including the network loss equation, active power balance equation, reactive power balance equation, voltage drop equation, and line current constraint. The derived model is equivalent to the original power flow model;

[0144] The uncertainty description module is used to regard the power prediction values ​​of photovoltaic output and load as deterministic variables and the prediction error as uncertain variables, and decompose the actual power of photovoltaic output and load into power prediction value and prediction error respectively. The uncertainty of photovoltaic output and the randomness of load are described as the uncertainty of corresponding prediction error, and the probability distribution of uncertain variables is obtained through data-driven methods;

[0145] The dispatch model construction module is used to comprehensively consider the power supply cost of the upper substation, the network loss of the distribution network, the dispatch cost of the distributed energy storage system ESS, and the dispatch cost of the static VAR compensator SVC, and to construct the objective function of the active distribution network stochastic optimization power dispatch model with the goal of minimizing the cost. Its constraints include power flow constraints, safe operation constraints, substation injection power constraints, photovoltaic unit operation constraints, ESS operation constraints, SVC constraints, and flexible load constraints.

[0146] The fuzzy neural network description and training module is used to integrate the fuzzy neural network to describe the uncertainty of the random optimization power dispatch model of the active distribution network, and use historical data to train the fuzzy neural network. The forward propagation process of the fuzzy neural network includes the data input layer, the membership function calculation layer, the rule generation layer, the normalization layer, and the output layer. The reverse propagation process uses the negative gradient descent algorithm to optimize the center point of the membership function, the width vector, and the connection weights of the output layer to obtain the fuzzy neural network parameters;

[0147] The pre-decision acquisition module is used to input the real-time status data of the active distribution network, including the real-time power prediction value and the data-driven prediction error, into the trained fuzzy neural network, and use the fuzzy neural network to quickly output the scheduling strategy pre-decision;

[0148] The real-time random scheduling strategy acquisition module is used to use the pre-decision value as the initial value for the solver to find the optimal solution, and then use the solver to accurately solve the active distribution network random optimization power scheduling model in the scheduling model construction module to obtain the real-time random scheduling strategy of the active distribution network.

[0149] It should be understood that the real-time random dispatching system of the active distribution network integrating fuzzy neural network pre-decision in the embodiment of the present invention can implement all the technical solutions in the above-mentioned method embodiment, and the functions of its various functional modules can be specifically implemented according to the methods in the above-mentioned method embodiments. The specific implementation process can refer to the relevant description in the above-mentioned embodiments, which will not be repeated here.

[0150] The present invention also provides a computer device, comprising: one or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and are configured to be executed by the one or more processors, and when the programs are executed by the processors, the steps of the real-time random scheduling method for active distribution networks integrating fuzzy neural network pre-decision are implemented as described above.

[0151] The present invention also provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps of the active distribution network real-time random scheduling method integrating fuzzy neural network pre-decision as described above are implemented.

[0152] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, devices (systems), computer equipment or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0153] The present invention is described with reference to a flowchart of a method according to an embodiment of the present invention. It should be understood that each process in the flowchart and a combination of processes in the flowchart 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 process in the flowchart. Figure 1 A device that specifies functions in a process or multiple processes.

[0154] 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 function specified in a process or multiple processes.

[0155] 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 The steps of a specified function in a process or multiple processes.

Claims

1. A real-time random dispatching method for active distribution network integrating fuzzy neural network pre-decision, characterized in that: The following steps are involved: Step 1: Re-derive the power flow model of the active distribution network based on the Euler equation, including the network loss equation, active power balance equation, reactive power balance equation, voltage drop equation, and line current constraint. The derived model is equivalent to the original power flow model. Step 2: The power prediction values ​​of PV output and load are regarded as deterministic variables, and the prediction error is regarded as an uncertain variable. The actual power of PV output and load is decomposed into power prediction value and prediction error respectively. The uncertainty of PV output and the randomness of load are described as the uncertainty of corresponding prediction error, and the probability distribution of uncertain variables is obtained through data-driven method. Step 3: Comprehensively consider the power supply cost of the upper substation, the network loss of the distribution network, the dispatching cost of the distributed energy storage system ESS, and the dispatching cost of the static VAR compensator SVC, and build the objective function of the active distribution network stochastic optimization power dispatching model with the goal of minimizing the cost. Its constraints include power flow constraints, safe operation constraints, substation injection power constraints, photovoltaic unit operation constraints, ESS operation constraints, SVC constraints, and flexible load constraints. Step 4: Fusion fuzzy neural network is used to describe the uncertainty of the random optimization power dispatch model of active distribution network, and the fuzzy neural network is trained using historical data. The forward propagation process of the fuzzy neural network includes data input layer, membership function calculation layer, rule generation layer, normalization layer, and output layer. The back propagation process uses the negative gradient descent algorithm to optimize the center point of the membership function, the width vector, and the connection weights of the output layer to obtain the fuzzy neural network parameters. Step 5: Input the real-time state data of the active distribution network, including the real-time power prediction value and the data-driven prediction error, into the trained fuzzy neural network, and use the fuzzy neural network to quickly output the dispatch strategy pre-decision; Step 6: Use the pre-decision value as the initial value for the solver to find the optimal solution, and then use the solver to accurately solve the active distribution network random optimization power dispatching model in step (3) to obtain the real-time random dispatching strategy of the active distribution network.

2. The method according to claim 1, characterized in that The step 1 specifically includes: For the network loss equation, the original network loss function formula is expressed as: Where: P mn and Q mn are the active and reactive power flowing through line mn respectively, and are respectively the active and reactive losses on line mn, I mn is the current flowing through line mn, R mn and X mn are the resistance and reactance of line mn respectively, U m is the voltage amplitude of node m; For the active power balance equation and reactive power balance equation, the active power balance and reactive power balance of node m are expressed as: Where: are the active power injected into node m by three distributed resources: substation, distributed photovoltaic, and distributed energy storage. are the reactive power injected into node m by substation, distributed photovoltaic and static VAR compensator respectively, are the active and reactive power of the load at node m, P km , Q km are the active and reactive power injected into node m by line km, P mn , Q mn are the active and reactive power flowing out of node m from line mn, is a collection of lines; For the voltage drop equation, the initial definition formula for voltage drop is: Where: Z mn is the line mn impedance, that is, Z mn =R mn +jX mn ; V m 、V n is the complex voltage of nodes m and n, that is, V m =|V m |∠θ m , V n =|V n |∠θ n ; S mn is the apparent power of the line, S mn =P mn +jQ mn ; and S mn and V m The conjugate complex number of ; rewrite the voltage phase angle difference using Euler's formula, and derive the voltage drop formula after the voltage phase angle is relaxed by separating the real and imaginary parts and performing mathematical operations: Line current constraints: Where: is the upper limit of the current amplitude flowing through line mn.

3. The method according to claim 1, characterized in that: In step 2, the probability distribution of uncertain variables is obtained by a data-driven method, including: The k-means clustering algorithm and the synchronous back-elimination technique based on probability distance are used to reduce the forecast data of similar days of power forecast and obtain the typical scenario of forecast error.

4. The method according to claim 1, characterized in that: In step 3, the objective function of the stochastic optimization power dispatch model of the active distribution network is as follows: Where: Ω L ,Ω ESS ,Ω SVC ,Ω S They are line set, ESS node set, SVC node set, and typical scene set after data driving; K T , K L , K E , K S are the power supply cost coefficient of substation, network loss coefficient, ESS dispatch coefficient, and SVC dispatch coefficient respectively; P T , P l L , P i,ESS , Q i,SVC is the active power output of the substation, the active power loss of the line, the charging and discharging power of the ESS node, and the reactive power output of the SVC node, where P i,ESS The positive and negative signs represent the charge and discharge status of the ESS, with positive for discharge and negative for charge. π(S) is the probability corresponding to a typical scenario driven by data.

5. The method according to claim 1, characterized in that The power flow constraint is expressed as: Where: P mn,S and Q mn,S are the active and reactive power flowing through line mn in scenario S, and are respectively the active and reactive losses on line mn in scenario S, R mn and X mn are the resistance and reactance of line mn respectively, U m,S , U n,S are the voltage amplitudes of nodes m and n under scenario S, is the active power injected into node m by distributed photovoltaic in scenario S, are the active power injected into node m by the distributed energy storage in the substation, are the reactive power injected into node m by substation, distributed photovoltaic and static VAR compensator respectively, are the active and reactive power of the load at node m, P km,S , Q km,S are the active and reactive power injected into node m by line km under scenario S, P mn,S , Q mn,S are the active and reactive power of line mn flowing out of node m under scenario S, is a collection of lines; The safe operation constraints include line transmission current constraints, line transmission power constraints and node voltage constraints: In i min ≤V i,S ≤V i max Where: P mn,S , Q mn,S are the active and reactive power transmitted by line mn in scenario S, S mn,max is the upper limit of the current carrying capacity of line mn, V i max 、V i min are the upper and lower limits of the specified node voltage, is the upper limit of the current amplitude flowing through line mn; The substation injection power constraint is expressed as: Where: P T , Q T Provide active and reactive power for the substation. It is the upper and lower limits of active and reactive power output of the substation; The photovoltaic unit operation constraints are expressed as: Where: P PV,S is the active power output of distributed photovoltaics in scenario S, Q PV , S PV They are the reactive output and capacity of the photovoltaic unit respectively; The ESS operation constraints are expressed as: Where: P i,ESS , is the ESS charging and discharging power, and the upper and lower limits of the charging and discharging power, is the ESS power at adjacent moments, is the upper and lower limits of power, η loss , η ESS They are power loss coefficient and charge and discharge efficiency respectively. The SVC constraint is expressed as: Where: Q i,SVC , It is the reactive power compensated by SVC and the upper and lower limits of reactive power compensation. The flexible load constraint is expressed as: (1-K P )*P D0 ≤P D ≤(1+K P )*P D0 (1-K Q )*Q D0 ≤Q D ≤(1+K Q )*Q D0 Where: P D0 , Q D0 is the load active and reactive power calculated in step 2, K P , K Q is the proportion of active and reactive flexible loads, P D , Q D It is the active and reactive power of the load after flexible load control.

6. The method according to claim 1, characterized in that The forward propagation process of the fuzzy neural network includes: a) Data input layer: The number of neurons in this layer is the characteristic dimension of the input data. That is, when the characteristic dimension of the data is n, the number of neurons in the input layer is n. The input data includes the predicted photovoltaic active power P PV,pre , predicted load active and reactive power P D,pre and Q D,pre And the prediction error ξ in each typical scenario PV , DP , DQ , the input variables contain uncertainty; b) Membership function calculation layer: Each neuron represents a fuzzy quantity. The membership function is used to calculate the probability that each neuron in the input layer belongs to the fuzzy quantity. The number of neurons is the number of fuzzy conditions that the input variables may constitute. All neurons constitute a fuzzy set. The number of nodes in this layer is the number of typical scenarios. The membership function uses the Gaussian function, as shown below: Where: u ij is the output of the neuron, x i is the input data, c ij is the center point of the membership function, b ij is the width vector of the membership function, m is the number of fuzzy classifications of the input, that is, the number of typical scenes; c) Rule generation layer, each neuron represents a fuzzy rule. The output of this layer is the applicability of each rule, which is calculated using the following formula: a j =Where? ij ,j=1,2,...,m d) Normalization layer, normalizes the fitness of each rule in the rule generation layer using the following formula: e) The output layer performs weighted processing to obtain the final output. The calculation formula is: Where: y i is the output variable, w ij is the corresponding weight, and k is the number of output variables.

7. The method according to claim 6, characterized in that The back propagation process of the fuzzy neural network includes: a) Calculate the negative gradient of the error E with respect to the output layer as follows: b) Calculate the negative gradient of the error E with respect to the normalization layer, the rule generation layer, and the membership function layer in turn, and obtain the negative gradient of the error E with respect to the membership function layer: c) Error E versus weight w ij 、Center point c ij and width vector b ij The negative gradient of is: d) The final parameter adjustment algorithm is: Where: η, ξ, ψ are the learning rates of the corresponding parameters, and r is the round of neural network learning.

8. A real-time random dispatching system for active distribution network integrating fuzzy neural network pre-decision, characterized in that: include: The power flow model re-derivation module is used to re-derive the power flow model of the active distribution network based on the Euler equation, including the network loss equation, active power balance equation, reactive power balance equation, voltage drop equation, and line current constraint. The derived model is equivalent to the original power flow model; The uncertainty description module is used to regard the power prediction values ​​of photovoltaic output and load as deterministic variables and the prediction error as uncertain variables, and decompose the actual power of photovoltaic output and load into power prediction value and prediction error respectively. The uncertainty of photovoltaic output and the randomness of load are described as the uncertainty of corresponding prediction error, and the probability distribution of uncertain variables is obtained through data-driven methods; The dispatch model construction module is used to comprehensively consider the power supply cost of the upper substation, the network loss of the distribution network, the dispatch cost of the distributed energy storage system ESS, and the dispatch cost of the static VAR compensator SVC, and to construct the objective function of the active distribution network stochastic optimization power dispatch model with the goal of minimizing the cost. Its constraints include power flow constraints, safe operation constraints, substation injection power constraints, photovoltaic unit operation constraints, ESS operation constraints, SVC constraints, and flexible load constraints. The fuzzy neural network description and training module is used to integrate the fuzzy neural network to describe the uncertainty of the random optimization power dispatch model of the active distribution network, and use historical data to train the fuzzy neural network. The forward propagation process of the fuzzy neural network includes the data input layer, the membership function calculation layer, the rule generation layer, the normalization layer, and the output layer. The reverse propagation process uses the negative gradient descent algorithm to optimize the center point of the membership function, the width vector, and the connection weights of the output layer to obtain the fuzzy neural network parameters; The pre-decision acquisition module is used to input the real-time status data of the active distribution network, including the real-time power prediction value and the data-driven prediction error, into the trained fuzzy neural network, and use the fuzzy neural network to quickly output the scheduling strategy pre-decision; The real-time random scheduling strategy acquisition module is used to use the pre-decision value as the initial value for the solver to find the optimal solution, and then use the solver to accurately solve the active distribution network random optimization power scheduling model in the scheduling model construction module to obtain the real-time random scheduling strategy of the active distribution network.

9. A computer device, characterized in that: include: one or more processors; Memory; And one or more programs, wherein the one or more programs are stored in the memory and are configured to be executed by the one or more processors, and when the programs are executed by the processors, the steps of the real-time random scheduling method for active distribution network integrating fuzzy neural network pre-decision as described in any one of claims 1-7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the real-time random scheduling method of an active distribution network integrating fuzzy neural network pre-decision are implemented as described in any one of claims 1 to 7.

Images