Harbor district power distribution network reactive power optimization method and system based on improved multi-objective fund washing algorithm

Abstract

The invention discloses a harbor district power distribution network reactive power optimization method and system based on an improved multi-objective fund washing algorithm. The method mainly comprises three aspects: clustering a large-scale scene into a typical distributed energy output scene by adopting a K-means clustering algorithm; according to different external characteristics of the SST, an operation mode corresponding to the SST is selected, then an equivalent node type corresponding to an SST node is selected according to the operation mode, a self-adaptive fuzzy coefficient, bidirectional guidance and a secondary cross update method are sequentially introduced to improve an algorithm, so that a model is solved, and a scheduling strategy for multiple distributed energy sources is obtained. Through the control, the problems of out-of-limit voltage fluctuation and network loss increase of the harbor district power distribution network caused by grid connection of various distributed energy sources can be effectively solved.

Description

Technical Field

[0001] The invention belongs to the technical field of application of high-voltage and high-power power electronic technology in power systems, and relates to a reactive power optimization method for a port area distribution network. Background Art

[0002] With the high proportion of distributed energy connected to the distribution network, it will inevitably lead to an increase in the difficulty of controlling the power quality and operating costs of the distribution network, especially the voltage stability and network loss. Therefore, it is of great significance to study the reactive power optimization scheduling strategy after the high proportion of distributed energy is integrated into the distribution network. At present, the research on reactive power optimization scheduling of distribution networks using SST and optimization algorithms has attracted widespread attention from scholars at home and abroad. As a new type of power electronic device, the solid-state transformer can directly absorb and emit reactive power without the assistance of other devices, realize the bidirectional transmission of reactive power, and directly regulate the reactive power to ensure its high regulation accuracy and reduce the error in calculation. This is particularly important for the distribution network.

[0003] However, once a single-objective problem is extended to a multi-objective problem, the computational dimension of the algorithm will increase significantly, increasing the search space, so the algorithm itself needs to have strong global search capabilities to prevent local convergence. Summary of the invention

[0004] The purpose of the present invention is to prevent voltage fluctuation from exceeding the limit and network loss from increasing, so as to provide guarantee for the safe and economical operation of the distribution network.

[0005] The technical solution to achieve the purpose of the present invention is:

[0006] According to one aspect of the present application, a reactive power optimization method for a port distribution network based on an improved multi-objective gold panning algorithm is provided, comprising the following steps:

[0007] The K-means clustering algorithm considering spatiotemporal characteristics is used to cluster the output scenarios of different types of energy in the distribution network to obtain typical output scenario data;

[0008] According to the external characteristics of the solid-state transformer SST, the corresponding operation mode is selected, and a reactive power optimization model of the distribution network containing SST is established; the input data of the reactive power optimization model of the distribution network containing SST includes typical output scenario data, so as to constrain the output of distributed energy;

[0009] Call the multi-objective gold mining algorithm MOGRO, and introduce adaptive fuzzy coefficient, two-way guidance and secondary cross update to improve the multi-objective gold mining algorithm MOGRO;

[0010] The improved multi-objective gold-mining algorithm MOGRO is used to solve the reactive power optimization model of the distribution network containing SST, and the reactive power optimization dispatching strategy set of the distribution network is obtained.

[0011] According to one aspect of the present application, the K-means clustering algorithm considering the spatiotemporal characteristics is used to cluster the output scenarios of different types of energy in the distribution network to obtain typical output scenario data, specifically:

[0012] Taking active output as the indicator, k samples are randomly selected as the initial cluster centers; k is a natural number greater than 0;

[0013] Calculate the distance between each sample in the data set and the centers of k clusters, and classify them into the cluster with the smallest distance;

[0014] For each cluster, recalculate the cluster center position;

[0015] Repeat the above steps until the set number of iterations is reached to obtain a typical output scenario that can represent different spatiotemporal characteristics.

[0016] According to one aspect of the present application, the operation mode corresponding to the solid-state transformer SST is selected according to its external characteristics, and a reactive power optimization model of the distribution network including the SST is established, specifically:

[0017] Considering whether the secondary side of the SST is connected to an active network, select the operation mode of constant AC voltage control, constant power control or a combination of the two;

[0018] Select the equivalent node type corresponding to the SST node according to the operation mode;

[0019] A dual-objective optimization model is established with minimizing network loss and voltage deviation as the objective function;

[0020] The model constraints include power balance constraints, distributed energy output constraints, reactive power compensator constraints, OLTC gear constraints, energy storage state of charge constraints and SST capacity constraints.

[0021] According to one aspect of the present application, the multi-objective gold mining algorithm MOGRO is called, and adaptive fuzzy coefficients, two-way guidance and secondary cross update are introduced to improve the multi-objective gold mining algorithm MOGRO, specifically:

[0022] The adaptive fuzzy coefficient C1 is introduced into the optimal update, which changes dynamically from large to small with the iteration, and the search range is narrowed in the later stage;

[0023] C 1 = [ 2 cos ( t − 1 T − 1 ⋅ π 2 ) ⋅ ( r 1 − 0 . 5 ) + 1 ] ⋅ r 2 ;

[0024] t is the current iteration number, T is the maximum iteration number; r1, r2 are random numbers in [0, 1];

[0025] Introducing bidirectional guidance for cooperative updating, using the vector sum of individual position changes in two iterations to guide the update;

[0026] ;

[0027] X i (t+1), X i (t) is the position of the i-th gold digger before and after the update; D3 is the distance coefficient;

[0028] After each iterative update, a secondary crossover operation is performed on the population.

[0029] ;

[0030] Where: pc is the crossover probability, pc1 and pc2 are the crossover start and end probabilities, f is the maximum fitness of any objective function of any two particles, and f average is the average value of the fitness of the objective function corresponding to f, max , f min is the maximum and minimum value of the corresponding objective function fitness;

[0031] Introducing adaptive crossover probability and simulated binary crossover to enhance local search capabilities;

[0032] X worker2 ( t + 1 ) = [ ( 1 + l ) ⋅ X worker1 ( t ) + ( 1 − l ) X worker2 ( t ) ] 2 X worker2 ( t + 1 ) = [ ( 1 − l ) ⋅ X worker1 ( t ) + ( 1 + l ) X worker2 ( t ) ] 2 ;

[0033] ;

[0034] In the formula, mu is the cross coefficient, m is a random number [0, 1], and X worker1 (t), X worker2 (t) for any two gold diggers;

[0035] Among them, the adjustment strategy of the adaptive fuzzy coefficient is:

[0036] In each iteration, the information entropy H(t) of the current population is calculated: H(t)=-sum i p i (t)logp i (t);

[0037] Among them, p i (t) is an individual i The probability of selection is proportional to the number of solutions it dominates;

[0038] Then, let the fuzzy coefficient C be negatively correlated with the entropy H(t): C(t)=C max -kH(t) / logN;

[0039] In the formula, C maxis the maximum fuzzy coefficient, k is the proportional coefficient, and N is the population size.

[0040] According to one aspect of the present application, the improved multi-objective gold panning algorithm MOGRO is used to solve the reactive power optimization model of the distribution network containing SST to obtain a reactive power optimization scheduling strategy set for the distribution network, which is specifically:

[0041] Initialize the individual positions and Archive set of the population; the positions include the output of each distributed energy source and equipment parameters;

[0042] Screen the current population individuals based on non-dominated relationships, update the Archive set based on the crowding degree, and select the global optimal individual;

[0043] If the maximum number of iterations is not reached, the current determining factor, determination coefficient, fuzzy coefficient and distance coefficient are calculated respectively, and the fuzzy coefficient of the optimal update is improved by adaptive fuzzy coefficient, and then the update method of cooperative update is improved by two-way guidance;

[0044] Randomly select one of the three updating methods of improved GRO to update the position, update the corresponding fitness, use binary simulated crossover to perform secondary update, and screen the non-dominated relationship after the update;

[0045] The binary crossover principle of genetic algorithm is used to update the population twice, and non-dominated relationships are screened, and the updated population is mixed with Archive to form a new population;

[0046] Select the reactive power optimization dispatching strategy that meets the actual needs from the Pareto optimal solution of the final Archive set;

[0047] During the solution process, the current population is adaptively gridded every several generations, the target space is divided into several small areas, and then the number of individuals in each grid is counted.

[0048] Grids with too many individuals are clustered, and a representative individual is selected from each cluster to be retained, and the rest are discarded. For grids with too few individuals, new individuals are generated by interpolating the individuals in adjacent grids to fill the grid, and the diversity and uniformity of the population are maintained by dynamically adjusting the grid granularity.

[0049] According to one aspect of the present application, it also includes selecting a reactive power optimization dispatching strategy that meets actual needs from the Pareto optimal solution of the reactive power optimization dispatching strategy set of the distribution network, specifically:

[0050] Construct an evaluation index system, including balance index, robustness index and preference matching index,

[0051] BAL=1-sqrt(sum((F i -mean(F) 2 ) / n); where F i To solve the i The function value on the target, mean(F) is the average value of all target function values, n is the number of targets, and the value range of BAL is [0, 1];

[0052] R=1 / (1+sum(abs(F'-F)) / n); where F and F' are the objective function value vectors of the original solution and the perturbation solution, respectively, and the value range of R is (0, 1];

[0053] PREF=sum(w i ·(F i -F i min ) / (F i max -F i min )); where w i is the weight coefficient of the i-th target, reflecting its relative importance; F i min and F i max are the minimum and maximum values ​​of the i-th objective function respectively;

[0054] Retrieve Pareto non-dominated solutions, extract the eigenvector of each solution, and perform weighted summation of their qualities); select the top a% of solutions as positive samples and the remaining solutions as negative samples to train the machine model; the value range of a is 1-100;

[0055] The newly generated Pareto solutions of the optimization algorithm are evaluated and screened in real time, and the optimal scheduling strategy that meets actual needs is automatically and efficiently selected to obtain the optimization results. The optimization results are then fed back to the optimization algorithm to guide the evolution direction of the population.

[0056] According to the optimization results, the active output, reactive power and energy storage charging and discharging power of various types of distributed energy are coordinated and controlled; the reactive power of various devices including SST and distributed energy inverters is coordinated, and the transformer tap adjustment is coordinated to maintain voltage stability;

[0057] Continuously monitor changes in distribution network load and fluctuations in distributed energy output, and initiate rolling adjustment of reactive power optimization.

[0058] According to one aspect of the present application, clustering the output scenarios of different types of energy in the distribution network using the K-means clustering algorithm considering spatiotemporal characteristics to obtain typical output scenario data also includes:

[0059] Take the membership degrees of each load node to all cluster centers as its characteristic attributes. For the \(i\)-th load node, extract its multi-dimensional attribute vector \(P(i)=[P(i,1),P(i,2),\cdots,P(i,d)]\) including active power load, reactive power load, voltage amplitude and spatial coordinates. Use the fuzzy membership function \(\mu\) to calculate the membership degree of node \(i\) to the \(j\)-th cluster center \(C(j)\), \(\mu(i,j) = 1 / (\sum_{k = 1}^{c}(|P(i)-C(j)| / |P(i)-C(k)|)^{m})\), where \(m\) is the fuzzy weighting exponent and \(c\) is the number of clusters; 2 / (m-1) )

[0060] Based on the Euclidean distance, introduce the attribute weight vector \(W=[w(1),w(2),\cdots,w(d)]\), and construct the weighted distance \(D(i,j)=\sum_{k = 1}^{d}(w(k)\cdot(P(i,k)-C(j,k))^{2})\). Use the gradient descent method to iteratively optimize the attribute weight vector \(W\). Each iteration updates the weight according to \(n(\sum_{i = 1}^{n}\sum_{j = 1}^{c}(\mu(i,j)D(i,j)))\) until convergence, where \(n\) is the number of load nodes; k=1 d (w(k)\cdot(P(i,k)-C(j,k))^{2}) 2 ) i n(\sum_{i = 1}^{n}\sum_{j = 1}^{c}(\mu(i,j)D(i,j))) i=1 n \sum_{i = 1}^{n}\sum_{j = 1}^{c}(\mu(i,j)D(i,j)) j=1 c (\mu(i,j)^{m}D(i,j)) m· )

[0061] When calculating the membership degree \(\mu(i,j)\) of node \(i\) to cluster center \(j\), not only consider the distance factor, but also introduce the mutual influence factor between nodes \(\alpha(i,k)=\exp(-|P(i)-P(k)| / \beta)\) to describe the influence intensity of node \(i\) by other node \(k\). Modify the membership degree calculation formula through \(\mu(i,j)=(1-\sum_{k = 1}^{n}(\alpha(i,k)\cdot\mu(k,j))) / (\sum_{l = 1}^{c}((D(i,j) / D(i,l))^{m}))\), and then perform row and column normalization processing on the membership degree matrix; k≠i \sum_{k = 1}^{n}(\alpha(i,k)\cdot\mu(k,j)) l=1 c ((D(i,j) / D(i,l))^{m}) (2 / (m-1) ))

[0062] Use the gradient descent method to update the cluster centers. Based on the current membership degree matrix, let \(C(j)=\sum_{i = 1}^{n}(\mu(i,j)\cdot P(i)) / \sum_{i = 1}^{n}(\mu(i,j))\) to update the cluster centers. To overcome the limitation of the fixed step size, each iteration updates \(C(j)\) through \(C(j)=C(j)+\eta(\sum_{i = 1}^{n}(\mu(i,j)\cdot P(i)) / \sum_{i = 1}^{n}(\mu(i,j))-C(j))\), where \(\eta\) is the learning rate; i=1 n (\mu(i,j)\cdot P(i)) m / \sum_{i = 1}^{n}(\mu(i,j)) i=1 n (\mu(i,j)) m ) new =C(j)+\eta(\sum_{i = 1}^{n}(\mu(i,j)\cdot P(i)) / \sum_{i = 1}^{n}(\mu(i,j))-C(j)) old-η·∇J(C(j)) adjusts the clustering center, and the step size η is dynamically updated using a decay strategy η = η0·exp(-t / τ), where t is the number of iterations, and the hyperparameters η0 and τ control the initial step size and the decay rate, respectively;

[0063] Automatically determine the optimal number of clusters. Initially, set the number of clusters c equal to the number of load nodes n and perform iterative optimization. After each iteration, calculate the validity index XB(c) of the current clustering: XB(c)=∑ i=1 n ∑ j=1 c (μ(i,j) m ·丨P(i)-C(j)丨 2 ) / (n·min(k≠l)(丨C(k)-C(l)丨 2 ))). If XB(c)>XB(c - 1), then set c = c - 1, merge the two closest clustering centers, and re-iterate; otherwise, set c best = c and terminate the clustering.

[0064] According to one aspect of the present application, the reactive power optimization model of the distribution network includes:

[0065] The distribution network power flow equations containing interval and fuzzy variables

[0066] The power balance equation of node i:

[0067] P Gi -P Li =U i *sum(U j *(G ij cos(theta ij )+B ij s i n(theta ij ))), j = 1:N;

[0068] Q Gi -Q Li =U i *sum(U j *(G ij s i n(theta ij )-B ij cos(theta ij ))), j = 1:N;

[0069] where P Gi , Q Gi are the active and reactive power generation power intervals of node i, respectively, and U i is the triangular fuzzy number of the voltage amplitude of node i;

[0070] The tie-line switch state variable matrix {x} and continuous variable matrix {f} are introduced to couple the network topology optimization with the power flow solution.

[0071] x ij ∈{0, 1}, indicating the on-off state of the tie line between nodes i and j;

[0072] f ij ∈[0,f ij max], indicating the tie line flow;

[0073] Topological variables and power flow variables must satisfy the following constraints:

[0074] f ij -B ij *(theta i -theta j )=0;

[0075] |f ij |≤x ij *f ij max;

[0076] Among them B ij is the tie line susceptance, f ij max is the upper limit of the power flow.

[0077] According to one aspect of the present application, a two-stage optimization strategy is adopted, in which the outer layer is optimized for the topological variables matr i x{x} is combined for optimization. The inner layer performs interval-fuzzy iterative solution of the power flow equations under a given topology. The outer layer uses the taboo search algorithm to minimize the network loss interval width and voltage deviation fuzziness as the optimization goal, and searches for the optimal topology combination matr i x{x}; the inner layer uses the interval Newton-Raphson method combined with the fuzzy expansion operator to iteratively solve the mixed uncertainty power flow equations to obtain the power flow interval and voltage fuzzy distribution;

[0078] The power flow calculation results under different topologies are evaluated from two dimensions: interval width and fuzziness.

[0079] Interval width indicator: W=sum((P imax -P imin ) 2 +(Q imax -Q imin ) 2 , i=1:N) / N;

[0080] Fuzziness index: D = sum((U ic -U ia ) 2 +(Uic -U ib ) 2 , i=1:N) / (2*N);

[0081] The final optimal topology and the corresponding power flow calculation results are selected on the Pareto frontier of the above two indicators.

[0082] According to another aspect of the present application, a reactive power optimization system for a port distribution network based on an improved MOGRO is also provided, comprising:

[0083] at least one processor; and,

[0084] a memory communicatively connected to at least one of the processors; wherein,

[0085] The memory stores instructions that can be executed by the processor, and the instructions are used to be executed by the processor to implement the reactive power optimization method and system for the port distribution network based on the improved multi-objective gold panning algorithm described in the above technical solution.

[0086] Beneficial effect: the present invention can obtain a more accurate dispatching strategy to prevent the problems of voltage fluctuation exceeding the limit and network loss increase. BRIEF DESCRIPTION OF THE DRAWINGS

[0087] Figure 1 It is the IEEE33 node topology diagram.

[0088] Figure 2 This is the K-means clustering algorithm process

[0089] Figure 3 It is a flow chart of the reactive power optimization strategy of distribution network based on the improved multi-objective gold panning algorithm. DETAILED DESCRIPTION

[0090] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments.

[0091] According to one aspect of the present application, a method and system for optimizing reactive power of a port distribution network based on an improved multi-objective gold panning algorithm are provided, comprising the following steps:

[0092] The K-means clustering algorithm considering the spatiotemporal characteristics is used to cluster the output scenarios of different types of energy in the distribution network, and typical output scenario data is obtained to provide a basis for the output constraints of distributed energy sources operating at a constant power factor.

[0093] According to the external characteristics of the solid-state transformer SST, the corresponding operation mode is selected, and a reactive power optimization model of the distribution network including the SST is established; the input data of the reactive power optimization model of the distribution network including the SST includes typical output scenario data;

[0094] Call the multi-objective gold mining algorithm MOGRO, and introduce adaptive fuzzy coefficient, two-way guidance and secondary cross update to improve the multi-objective gold mining algorithm MOGRO;

[0095] The improved multi-objective gold-mining algorithm MOGRO is used to solve the reactive power optimization model of the distribution network containing SST, and the reactive power optimization dispatching strategy set of the distribution network is obtained.

[0096] In this application, the K-means clustering algorithm that considers spatiotemporal characteristics is used to quickly and accurately divide the massive distribution network operation data into several typical scenarios, each of which has similar load demand and new energy output characteristics. Through clustering simplification, not only the scale and frequency of power flow calculations are reduced, but also the typical scenarios obtained are more representative and can cover most of the actual operating conditions. The subsequent reactive power optimization model uses typical scenarios as input, which can reduce the complexity of calculations while increasing the scope of application of the results, and has strong engineering practical value. Improve the accuracy and efficiency of distribution network operation scenario analysis.

[0097] According to the external characteristics of SST, its operation mode is optimized, and the multi-operating characteristics of SST are embedded in the reactive power optimization model of the distribution network, which can give full play to the reactive power regulation ability of SST, dynamically track load and output fluctuations, and improve the voltage quality of the distribution network. Compared with conventional transformers, SST can achieve continuous, accurate and fast reactive power dispatching, providing strong support for the safe and stable operation of the distribution network. Flexible and efficient operation control of solid-state transformers (SSTs) is realized.

[0098] In view of the shortcomings of the existing algorithms, three improvements were made: the adaptive fuzzy coefficient dynamically balances global exploration and local development; the bidirectional guidance enhances the group intelligence and accelerates the convergence process; the quadratic crossover expands the spread of beneficial genes and improves the quality of the solution. These improvements effectively overcome the "curse of dimensionality" challenge faced by multi-objective optimization, and can not only quickly locate the global optimal area in the huge solution space, but also achieve uniform distribution on the non-inferior solution set, thereby obtaining a Pareto frontier with high accuracy and wide distribution. A large number of simulation experiments show that the improved MOGRO is significantly superior to the existing mainstream multi-objective optimization algorithms in terms of solution convergence and diversity. The solution performance of reactive power optimization of distribution networks has been greatly improved.

[0099] After obtaining the reactive power optimal dispatch strategy set of the distribution network, the solution further screens out the optimal reactive power dispatch solution that meets the actual engineering requirements, including various reactive power control means such as DG output, capacitor switching, and tap regulation. This dispatch solution comprehensively considers multiple technical and economic indicators such as minimum network loss and qualified voltage, and can ensure the optimal operation of the distribution network under multiple scenarios and working conditions. In addition, the solution can also be integrated with the distribution automation system to realize the online correction and rolling execution of the optimal dispatch strategy, form a closed-loop control based on real-time feedback, and keep the distribution network in the best operating state continuously under complex and changeable environments. A closed-loop control strategy for reactive power optimization of the distribution network is formed.

[0100] According to one aspect of the present application, the use of the K-means clustering algorithm considering spatio-temporal characteristics to cluster the output scenarios of different types of energy in the distribution network to obtain typical output scenario data, providing a basis for the output constraints of distributed energy operating at a constant power factor, specifically:

[0101] Taking the active power output as an index, randomly select k samples as the initial cluster centers; k is a natural number greater than 0;

[0102] Calculate the distance from each sample in the dataset to the k cluster centers, and classify it into the cluster with the smallest distance.

[0103] For each cluster, recalculate the position of the cluster center.

[0104] Repeat the above steps until the set number of iterations is reached to obtain typical output scenarios that can represent different spatio-temporal characteristics.

[0105] According to one aspect of the present application, the use of the K-means clustering algorithm considering spatio-temporal characteristics to cluster the output scenarios of different types of energy in the distribution network to obtain typical output scenario data, providing a basis for the output constraints of distributed energy operating at a constant power factor, further includes:

[0106] Taking the membership degrees of each load node to all cluster centers as its characteristic attributes, for the i-th load node, extract its multi-dimensional attribute vector P(i)=[P(i,1), P(i,2),..., P(i,d)] including active load, reactive load, voltage amplitude, and spatial coordinates, and use the fuzzy membership function μ to calculate the membership degree of the node i to the j-th cluster center C(j), μ(i,j)=1 / (∑(k = 1:c)(丨P(i)-C(j)丨 / 丨P(i)-C(k)丨) 2 / (m-1) ), where m is the fuzzy weighting exponent and c is the number of clusters;

[0107] Based on the Euclidean distance, introduce the attribute weight vector W = [w(1), w(2),..., w(d)] to construct the weighted distance D(i, j) = ∑ k=1 d (w(k) · (P(i, k) - C(j, k)) 2 ), and use the gradient descent method to iteratively optimize the attribute weight vector W. Each iteration updates the weight according to m i n(∑ i=1 n ∑ j=1 c (μ(i, j) m· D(i, j))), until convergence. n is the number of load nodes;

[0108] When calculating the membership degree μ(i, j) of node i to the clustering center j, not only consider the distance factor, but also introduce the mutual influence factor between nodes α(i, k) = exp(-丨P( i ) - P(k)丨 / β) to describe the intensity of node i affected by other nodes k. Through μ(i, j) = (1 - ∑ k≠i (α(i, k) · μ(k, j))) / (∑ l=1 c ((D(i, j) / D(i, l)) (2 / (m-1) ))), correct the membership degree calculation formula, and then perform row and column normalization processing on the membership degree matrix;

[0109] Use the gradient descent method to update the clustering center. Based on the current membership degree matrix, let C(j) = ∑ i=1 n (μ(i, j) m · P(i)) / ∑ i=1 n (μ(i, j) m ) to update the clustering center. To overcome the limitation of the fixed step size, each iteration adjusts the clustering center through C(j) new = C(j) old - η · ∇J(C(j)). The step size η is dynamically updated using the attenuation strategy η = η0 · exp(-t / τ), where t is the number of iterations, and the hyperparameters η0 and τ control the initial step size and attenuation rate respectively;

[0110] Automatically determine the optimal number of clusters. Initially, set the number of clusters c equal to the number of load nodes n, and perform iterative optimization. After each round of iteration, calculate the validity index XB(c) of the current clustering = ∑ i=1 n ∑ j=1 c (μ(i, j) m · 丨P(i) - C(j)丨2 ) / (n·m i n(k≠l)(|C(k) - C(l)| 2 )), if XB(c) > XB(c - 1), then let c = c - 1, merge the two closest cluster centers, and iterate again: otherwise let c best = c, terminate the clustering.

[0111] In another embodiment of the present application, the clustering process further includes:

[0112] Introduce other attributes reflecting load and new energy output characteristics, including reactive power, power volatility, daily average power interval, and peak-valley ratio. Perform feature engineering on these original attributes, including data cleaning, normalization, dimensionality reduction, etc., and extract a set of comprehensive clustering feature indicators.

[0113] Use the Gap Statistics method to automatically determine the optimal number of clusters K. This method calculates the Gap statistic of clustering under different K values by introducing a reference data set, and selects the K corresponding to the largest Gap value as the optimal number of clusters.

[0114] The specific steps are as follows:

[0115] For a given number of clusters K, use the K-means algorithm to cluster the data set D, and calculate the observed statistic of the clustering result: O K = sum(sum(dist(x i , c k ))), i = 1..n, k = 1..K; 2 )),其中x

[0116] where x i is the i-th sample, c k is the k-th cluster center, and dist is the distance metric function.

[0117] Through Monte Carlo simulation, generate B reference data sets D1, D2,..., D B with the same dimension and number of samples as D.

[0118] For each reference data set D b (b = 1..B), perform the clustering step and calculate its observed statistic O Kb .

[0119] Calculate the Gap statistic:

[0120] Gap(K) = (1 / B)sum(log(O _ Kb)) - log(O _ K), b = 1..B

[0121] Repeat the above steps, try different K values, and draw a curve of Gap(K) changing with K.

[0122] Select the K corresponding to the maximum Gap value as the optimal clustering number K.

[0123] In another embodiment of the present application,

[0124] On the basis of K-means, fuzzy membership and weighted distance metrics are introduced to improve the robustness and adaptability of the algorithm. The improved fuzzy K-means clustering steps are as follows:

[0125] Randomly initialize the membership matrix U, where u ik Represents sample x i The membership of the k-th cluster satisfies the condition:

[0126] 0 <= u ik <= 1, sum(u ik ) = 1, i=1..n, k=1..K*;

[0127] b) Based on the current membership matrix U, calculate the cluster centers:

[0128] c k = sum(u ikm * x i ) / sum(u ik m ), i=1..n, k=1..K*;

[0129] Where m>1 is a fuzzy weighted index that controls the degree of cluster overlap.

[0130] Update the membership matrix using weighted Euclidean distance:

[0131] u ik = 1 / sum((d i st(x i , c k ) / d i st(x i , c j )) 2 / (m-1) ), j=1..K*;

[0132] d i st(x i , c k ) = sqrt(sum(w d * (x id -c kd ) 2)), d=1..D;

[0133] where w d is the weight coefficient of the d-th dimension feature, reflecting its importance, and can be determined by the entropy weight method.

[0134] Until the membership matrix U converges or the maximum number of iterations is reached.

[0135] According to the membership matrix U, each sample x i Assigned to the cluster with the largest membership.

[0136] C i = argmax(u ik ), k=1..K*;

[0137] Using DB I The clustering effect is quantitatively evaluated and optimized by combining the method with the silhouette coefficient.

[0138] DB I The index measures the compactness within clusters and the separation between clusters and is defined as:

[0139] DB I = (1 / K*) * sum(max(R _ i j)), i=1..K*, j!=i;

[0140] R ij = (avg dist (C i ) + avg dist (C j )) / dist(c i , c j );

[0141] Among them C i is the i-th cluster, c i As its center, avg dist is the average distance from the samples in the cluster to the center. The smaller the DBI value, the better the clustering effect.

[0142] The silhouette coefficient measures the similarity between a sample and its cluster and other clusters, and is defined as:

[0143] S i = (b _ i - a _ i ) / max(a _ i , b _ i ),i =1..n

[0144] a i = (1 / |C _ k|) * sum(d i st(x i , x j )), x i , x j Belongs to C k , j!= i ;

[0145] b i = m i n((1 / |C l |) * sum(d i st(x i , x j ))), x j Belongs to C l , l!=k;

[0146] SC = (1 / n) * sum(S i );

[0147] where a i For sample x i The average distance to other samples in the same cluster, b i For x i The minimum average distance to other cluster samples. The silhouette coefficient SC takes values ​​between [-1, 1]. The closer it is to 1, the better the clustering effect. By iteratively adjusting clustering parameters (such as m value, w _ d weight, etc.), continuously optimize DB I and SC indicators until the preset threshold is met or the maximum number of tuning times is reached, a high-quality scene clustering result can be obtained.

[0148] In this embodiment, clustering considering spatiotemporal characteristics has the following technical effects:

[0149] Traditional scenario analysis often only considers the load and output data of a single time section, ignoring the continuity and correlation of the scenario in time and space. However, by introducing spatiotemporal characteristics, this clustering algorithm can capture the evolution of different types of energy output in time and space, and explore the inherent connections between scenarios, thereby obtaining more accurate and comprehensive typical scenario division results. This improves the accuracy of distribution network operation scenario analysis.

[0150] Large-scale distribution networks usually contain a large number of load nodes and power nodes, and their operating scenarios are characterized by high dimensionality, nonlinearity, and strong coupling. Directly performing power flow analysis on the entire scenario not only requires huge computational effort, but may also cause the calculation results to not converge due to the "curse of dimensionality". This clustering algorithm can automatically classify massive scenarios into a few typical modes. The characteristics of the scenarios within each mode are similar, while the characteristics between modes vary significantly. Subsequent power flow calculations and optimization decisions only need to be performed for these typical scenarios, greatly reducing the computational complexity and time cost. The scale and frequency of power flow calculations in distribution networks are reduced.

[0151] In the reactive power optimization of distribution networks, it is necessary to accurately model and characterize the active and reactive outputs of various types of distributed energy sources, especially considering the strong coupling relationship between active and reactive power under constant power factor operation. The typical output scenario data generated by the clustering algorithm can be directly used as the basis for the output constraints of distributed energy sources, avoiding the tedious process of modeling all scenarios one by one. Replacing the full scenario with a typical scenario can significantly improve the scope of application and computational efficiency of the reactive power optimization model while meeting the optimization accuracy requirements. The constraint capability of the output of distributed energy sources under constant power factor operation is enhanced.

[0152] The load of the distribution network has obvious regional distribution characteristics, and the load change trends in adjacent areas are often similar. This clustering algorithm innovatively introduces the mutual influence factor between nodes, quantitatively describes the proximity of load nodes in geographical topology and its impact on clustering, which helps to identify typical load groups in the same area and optimize the power supply partitioning and reactive compensation points of the distribution network. The spatial topological association between load nodes is mined.

[0153] In traditional K-means clustering, the number of clusters K needs to be set manually and remains unchanged during the iteration process, which largely depends on experience and prior knowledge. This clustering algorithm uses the Xie-Beni indicator to automatically determine the optimal number of clusters and adaptively adjusts the cluster size during the iteration process, overcoming the blindness of manual settings. In addition, the algorithm also introduces attribute weight vectors and learning rate decay strategies, which can adaptively adjust the importance of each attribute and the update step size according to the characteristics of data distribution, effectively improving clustering performance. Adaptive optimization of clustering parameters is achieved.

[0154] Distribution network scenario data usually presents non-uniform and non-convex distribution characteristics, and the data density in different regions varies greatly. Traditional K-means clustering uses Euclidean distance to measure sample similarity, which easily leads to over-fine clustering in areas with high data density and over-coarse clustering in areas with sparse data. This clustering algorithm uses a method combining fuzzy membership and weighted distance, which can dynamically adjust its membership and distance weight according to the local density of sample distribution, thereby obtaining density-sensitive clustering results. The adaptability of clustering results to non-uniform data distribution is improved.

[0155] According to one aspect of the present application, the operation mode corresponding to the solid-state transformer SST is selected according to its external characteristics, and a reactive power optimization model of the distribution network including the SST is established, specifically:

[0156] Considering whether the secondary side of the SST is connected to an active network, select the operation mode of constant AC voltage control, constant power control or a combination of the two;

[0157] Select the equivalent node type corresponding to the SST node according to the operation mode;

[0158] A dual-objective optimization model is established with minimizing network loss and voltage deviation as the objective function;

[0159] The model constraints include power balance constraints, distributed energy output constraints, reactive power compensator constraints, OLTC gear constraints, energy storage state of charge constraints and SST capacity constraints.

[0160] According to one aspect of the present application,

[0161] Optimization variable: Q Gi : Node i Reactive power compensation capacity, continuous variable; b ij : The switching state of branch ij, 0-1 integer variable;

[0162] tap k : Transformer k tap position, discrete variable; Q DG m : DG m Reactive power output, continuous variable. Q SST : Reactive output of SST, continuous variable.

[0163] Objective function: m i n F = w1sum(P lossi j ) + w2sum((V devi ) 2 );

[0164] P lossi j = G ij * (V i 2 + V j 2 - 2V i V j *cos(theta ij ));

[0165] V devi = (V i-V spec ) 2 ;

[0166] Where w1 and w2 are the weight coefficients of the two sub-goals, P lossi j is the branch active power loss, V devi is the node voltage deviation.

[0167] Constraints:

[0168] Node power flow balance constraints:

[0169] P Gi -P Li = V i * sum(V j * (G ij cos(theta ij ) + B ij s i n(theta ij )));

[0170] Q Gi -Q Li = V i * sum(V j * (G ij s i n(theta ij ) - B ij cos(theta ij )));

[0171] Voltage amplitude constraint: V min <= V i <= V max ;

[0172] Line flow constraint: |P ij | <= b ij * P ij max ;|Q ij | <= b ij * Q ij max ;

[0173] Reactive power compensation constraint: Q Ci min <= Q Ci <= Q Ci max ;

[0174] Transformer tap constraints:

[0175] tapm i n <= tap k <= tap max ;

[0176] DG reactive power output constraint: Q DG m min <= Q DG m <= Q DG m max ;

[0177] SST reactive power output constraint: Q SST min <= Q SST <= Q SST max ;

[0178] P SST 2 + Q SST 2 <= (S SST max ) 2 ;

[0179] According to whether the secondary side of the SST is connected to an active distribution network and the voltage / power control requirements, a rule-based hierarchical mode switching logic is designed.

[0180] First level: determine the properties of the distribution network to which the SST is connected;

[0181] If it is a passive network: execute constant voltage amplitude control mode (QV);

[0182] If it is an active network:

[0183] If voltage stability is the main concern: Execute constant voltage amplitude control mode (QV);

[0184] If power flow balance is the main consideration: Execute fixed active and reactive power control mode (PQ);

[0185] The second level: further refine the control parameters in the selected control mode;

[0186] If QV mode:

[0187] If the voltage at the feeder end is too high: switch to the inductive operation area first to absorb reactive power;

[0188] If the voltage at the feeder end is too low: switch to the capacitive operation area first to generate reactive power;

[0189] Dynamically adjust reactive power output Q _SST, so that the feeder terminal voltage V is maintained at [V min , V max ]Inside;

[0190] If PQ mode:

[0191] According to the real-time active load demand P of the distribution network load and reactive load demand Q load ;

[0192] Dynamically adjust the active output P of SST _ SST and reactive power output Q _ SST, which meets:

[0193] P SST + P DG =P load + P loss

[0194] Q SST + Q DG + Q C = Q load + Q loss ;

[0195] Where P DG and Q DG Help DG, Q C For other reactive power compensation capacity;

[0196] The constructed mixed integer nonlinear programming (MINLP) model is solved by using the decomposition-coordination optimization strategy.

[0197] Heuristic algorithms based on hill climbing and tabu search for integer decision variables, including:

[0198] Branch switching status b ij ; Transformer tap position tap k ;

[0199] Heuristic rules are constructed based on indicators such as network loss sensitivity and voltage deviation.

[0200] Thus, several feasible solutions of network topology and transformer tap locations are obtained.

[0201] By fixing the integer variables, the original MINLP problem is transformed into a nonlinear programming (NLP) problem.

[0202] For each set of integer feasible solutions, PDIPM is used to solve the NLP subproblems and obtain the optimal solutions of continuous variables, including:

[0203] Reactive power compensation capacity Q Ci ; DG reactive power output QDG m ; SST reactive power output Q SST ;

[0204] Compare the objective function values ​​of all feasible solutions and select the global optimal solution as the final dispatching strategy for reactive power optimization.

[0205] According to one aspect of the present application, the multi-objective gold mining algorithm MOGRO is called, and adaptive fuzzy coefficients, two-way guidance and secondary cross update are introduced to improve the multi-objective gold mining algorithm MOGRO, specifically:

[0206] The adaptive fuzzy coefficient C1 is introduced into the optimal update, which changes dynamically from large to small with the iteration, and the search range is narrowed in the later stage;

[0207] C 1 = [ 2 cos ( t − 1 T − 1 ⋅ π 2 ) ⋅ ( r 1 − 0 . 5 ) + 1 ] ⋅ r 2 ;

[0208] t is the current iteration number, T is the maximum iteration number; r1, r2 are random numbers in [0, 1];

[0209] Introducing bidirectional guidance for cooperative updating, using the vector sum of individual position changes in two iterations to guide the update;

[0210] ;

[0211] X i (t+1), X i (t) is the position of the i-th gold digger before and after the update; D3 is the distance coefficient;

[0212] After each iterative update, a secondary crossover operation is performed on the population.

[0213] ;

[0214] Where: pc is the crossover probability, pc1 and pc2 are the crossover start and end probabilities, f is the maximum fitness of any objective function of any two particles, and f average is the average value of the fitness of the objective function corresponding to f, max , f min is the maximum and minimum value of the corresponding objective function fitness;

[0215] Introducing adaptive crossover probability and simulated binary crossover to enhance local search capabilities;

[0216] X worker2 ( t + 1 ) = [ ( 1 + l ) ⋅ X worker1 ( t ) + ( 1 − l ) X worker2 ( t ) ] 2 X worker2 ( t + 1 ) = [ ( 1 − l ) ⋅ X worker1 ( t ) + ( 1 + l ) X worker2 ( t ) ] 2 ;

[0217] ;

[0218] In the formula, mu is the cross coefficient, m is a random number [0, 1], and X worker1 (t), X worker2 (t) for any two gold diggers;

[0219] Among them, the adjustment strategy of the adaptive fuzzy coefficient is:

[0220] In each iteration, the information entropy H(t) of the current population is calculated: H(t)=-sum i p i (t)logp i (t);

[0221] Among them, p i (t) is an individual i The probability of selection is proportional to the number of solutions it dominates;

[0222] Then, let the fuzzy coefficient C be negatively correlated with the entropy H(t): C(t)=C max -kH(t) / logN;

[0223] In the formula, C max is the maximum fuzzy coefficient, k is the proportional coefficient, and N is the population size.

[0224] According to one aspect of the present application, the improved multi-objective gold panning algorithm MOGRO is used to solve the reactive power optimization model of the distribution network containing SST to obtain a reactive power optimization scheduling strategy set for the distribution network, which is specifically:

[0225] Initialize the individual positions and Archive set of the population; the positions include the output of each distributed energy source and equipment parameters;

[0226] Screen the current population individuals based on non-dominated relationships, update the Archive set based on the crowding degree, and select the global optimal individual;

[0227] If the maximum number of iterations is not reached, the current determining factor, determination coefficient, fuzzy coefficient and distance coefficient are calculated respectively, and the fuzzy coefficient of the optimal update is improved by adaptive fuzzy coefficient, and then the update method of cooperative update is improved by two-way guidance;

[0228] Randomly select one of the three updating methods of improved GRO to update the position, update the corresponding fitness, use binary simulated crossover to perform secondary update, and screen the non-dominated relationship after the update;

[0229] The binary crossover principle of genetic algorithm is used to update the population twice, and non-dominated relationships are screened, and the updated population is mixed with Archive to form a new population;

[0230] Select the reactive power optimization dispatching strategy that meets the actual needs from the Pareto optimal solution of the final Archive set;

[0231] During the solution process, the current population is adaptively gridded every several generations, the target space is divided into several small areas, and then the number of individuals in each grid is counted.

[0232] Grids with too many individuals are clustered, and a representative individual is selected from each cluster to be retained, and the rest are discarded. For grids with too few individuals, new individuals are generated by interpolating the individuals in adjacent grids to fill the grid, and the diversity and uniformity of the population are maintained by dynamically adjusting the grid granularity.

[0233] In another embodiment of the present application, a distribution estimation (EDA) mechanism is introduced during the solution process to enhance population diversity. The specific process is as follows:

[0234] In each iteration, according to the fitness value of the individual, the Gaussian mixture model (GMM) is used to estimate its distribution:

[0235] p(x) = sum(w k * N(x|mu k , S i gma k )), k=1..K;

[0236] where w k is the mixing coefficient of the kth Gaussian component, mu k and S i gma k is its mean vector and covariance matrix.

[0237] New individuals are sampled from the estimated GMM distribution and merged with the original population, keeping the population size unchanged.

[0238] Adaptively adjust the EDA sampling ratio to dynamically balance exploration and utilization. In the early stage, the focus is on exploration, and the sampling ratio is high; in the later stage, the focus is on utilization, and the sampling ratio is reduced.

[0239] According to the shape characteristics of the Pareto front, the population selection and update strategies are improved.

[0240] In the non-dominated sorting, the individual's Pareto rank, crowding distance and super volume contribution are comprehensively considered.

[0241] NDS (non-dominated sorting), CD (crowding distance), and VC (hypervolume contribution) are quantified as dimensionless indicators in the interval [0, 1].

[0242] Construct a comprehensive evaluation function to balance the three indicators: score = w1'*NDS + w2'*CD + w3'*VC;

[0243] Dynamically adjust the weight coefficients w1', w2', w3' to take into account the convergence and diversity of the solution.

[0244] The scores of the individuals in the population are sorted, and the individuals with the highest ranking are retained as the parent population of the new generation.

[0245] Adaptively adjust the key parameters of the MOGRO algorithm. Based on the fuzzy coefficient C, an adaptive weight coefficient alpha is further introduced to enable C to adaptively switch between global search and local search.

[0246] C(t) = C min + (C max - C min ) * exp(-alpha(t) * t / T max );

[0247] alpha(t) = alpha0 + k * (SSD(t-1) - SSD(t)) / SSD(t);

[0248] Among them, alpha0 and k are control parameters, and SSD is a population spatial distribution metric that measures the degree of discreteness of the population.

[0249] The p-best mechanism is introduced to record the individual's best historical position, and cross it with the current position with a certain probability to update the next generation of individual positions. This helps individuals escape from local traps.

[0250] c) Based on the bidirectional guidance mechanism, adaptive step length control is further introduced.

[0251] v ij (t) = w1 * v ij (t-1) + w2 * (gbest ij -x ij (t-1)) + w3 * (pbest ij -x ij (t-1));

[0252] w1= w max - (w max -w min ) * t / T max ;

[0253] w2= w min + (w max-w min ) * t / T max ;

[0254] w3 = c3 * rand();

[0255] where w max and w min are the maximum and minimum values ​​of the inertia weight, respectively, usually w max =0.9, w min =0.4. c3 is the pbest bootstrap coefficient, which can be a small positive number such as 0.1.

[0256] Here, a linearly decreasing inertia weight w1 is used, so that the algorithm focuses on global search in the early stage and local search in the later stage, balancing the convergence speed and the quality of the solution. _ 3 takes a random small value, which plays a role of perturbation and helps to jump out of the local trap. The weight w2 of the gbest item increases with the number of iterations, which accelerates the convergence of the population. Reasonable adjustment of the size of w1, w2, and w3 can make the population adaptively switch between global search and local development, improving the algorithm's optimization ability.

[0257] In another embodiment of the present application, when calculating the dominance relationship between the newly generated individuals and the individuals in the Archive set, the concept of ε-non-domination is introduced.

[0258] If x i Dominate x j , and for each goal:

[0259] f k (x i ) <= f k (x j ) - ε, k=1, 2,...,M;

[0260] Then x i ε-dominates x j ε is a preset small positive number, representing the significance level.

[0261] When updating the Archive set, individuals that are ε-nondominated are deleted first, which can retain weak non-dominated solutions to a certain extent and improve the coverage of the solution set on the Pareto frontier.

[0262] When the Archive set overflows, individuals are deleted based on their crowding distance, ε-dominance relationship, and the time of last addition. Individuals with small crowding distance, ε-dominance, and the most recently added are deleted first. i ve sets of diversity and uniform distribution.

[0263] According to one aspect of the present application, it also includes selecting a reactive power optimization dispatching strategy that meets actual needs from the Pareto optimal solution of the reactive power optimization dispatching strategy set of the distribution network, specifically:

[0264] Construct an evaluation index system, including balance index, robustness index and preference matching index,

[0265] BAL=1-sqrt(sum((F i -mean(F) 2 ) / n); where F i To solve the function value on the i-th target, mean(F) is the average value of all target function values, n is the number of targets, and the value range of BAL is [0, 1];

[0266] R=1 / (1+sum(abs(F'-F)) / n); where F and F' are the objective function value vectors of the original solution and the perturbation solution, respectively, and the value range of R is (0, 1];

[0267] PREF=sum(w i ·(F i -F i min ) / (F i max -F i min )); where w i is the weight coefficient of the i-th target, reflecting its relative importance; F i min and F i max are the minimum and maximum values ​​of the i-th objective function respectively;

[0268] Retrieve Pareto non-dominated solutions, extract the eigenvector of each solution, and perform weighted summation of their qualities); select the top a% of solutions as positive samples and the remaining solutions as negative samples to train the machine model; the value range of a is 1-100;

[0269] The newly generated Pareto solutions of the optimization algorithm are evaluated and screened in real time, and the optimal scheduling strategy that meets actual needs is automatically and efficiently selected to obtain the optimization results. The optimization results are then fed back to the optimization algorithm to guide the evolution direction of the population.

[0270] According to the optimization results, the active output, reactive power and energy storage charging and discharging power of various types of distributed energy are coordinated and controlled; the reactive power of various devices including SST and distributed energy inverters is coordinated, and the transformer tap adjustment is coordinated to maintain voltage stability;

[0271] Continuously monitor changes in distribution network load and fluctuations in distributed energy output, and initiate rolling adjustment of reactive power optimization.

[0272] According to one aspect of the present application, a distribution network reactive power optimization model includes:

[0273] Power Flow Equations for Distribution Networks with Interval and Fuzzy Variables

[0274] The power balance equation of node i is:

[0275] P Gi -P Li =U i *sum(U j *(G ij cos(theta ij )+B ij s i n(theta ij ))), j=1:N;

[0276] Q Gi -Q Li =U i *sum(U j *(G ij s i n(theta ij )-B ij cos(theta ij ))), j=1:N;

[0277] Where P Gi , Q Gi Node i The active and reactive power generation range, U i is the triangular fuzzy number of the voltage amplitude of node i;

[0278] The tie-line switch state variable matrix {x} and continuous variable matrix {f} are introduced to couple the network topology optimization with the power flow solution.

[0279] x ij ∈{0, 1}, indicating a node i The on / off status of the contact line between and j;

[0280] f ij ∈[0,f ij max], indicating the tie line flow;

[0281] Topological variables and power flow variables must satisfy the following constraints:

[0282] f ij -B ij *(theta i-theta j )=0;

[0283] |f ij |≤x ij *f ij max;

[0284] Among them B ij is the tie line susceptance, f ij max is the upper limit of the power flow.

[0285] According to one aspect of the present application, a two-stage optimization strategy is adopted, in which the outer layer is optimized for the topological variables matr i x{x} is combined for optimization. The inner layer performs interval-fuzzy iterative solution to the power flow equations under a given topology. The outer layer uses the taboo search algorithm to minimize the network loss interval width and voltage deviation fuzziness as the optimization goal, and searches for the optimal topology combination matrix{x}. The inner layer uses the interval Newton-Raphson method combined with the fuzzy expansion operator to iteratively solve the mixed uncertainty power flow equations and obtain the power flow interval and voltage fuzzy distribution.

[0286] The power flow calculation results under different topologies are evaluated from two dimensions: interval width and fuzziness.

[0287] Interval width indicator: W=sum((P imax -P imin ) 2 +(Q imax -Q imin ) 2 , i=1:N) / N;

[0288] Fuzziness index: D = sum((U ic -U ia ) 2 +(U ic -U ib ) 2 , i=1:N) / (2*N);

[0289] The final optimal topology and the corresponding power flow calculation results are selected on the Pareto frontier of the above two indicators.

[0290] In another embodiment of the present application, all non-dominated solutions are extracted from the final Archive set and recorded as the Pareto optimal solution set PS.

[0291] For each solution S in PS i ( i =1, 2, ..., N), calculate its normalized fitness value in the objective function space:

[0292] floss (S i ) = (f loss max -f loss (S i )) / (f loss max -f loss m i n );

[0293] f dev (S i ) = (f dev max -f dev (S i )) / (f dev max -f dev m i n );

[0294] where f loss (S i ) and f dev (S i ) are the solutions S i The corresponding network loss and voltage deviation value, f loss max 、f loss m i n 、f dev max 、f dev min They are the maximum and minimum values ​​of network loss and voltage deviation in PS respectively.

[0295] Perform weighted summation on multiple normalized objective function values ​​to obtain solution S i The comprehensive fitness value of:

[0296] f i t(S i ) = w loss *f loss (S i ) + w dev *f dev (S i );

[0297] where w loss and w dev is the weight coefficient of the two objectives, reflecting the decision preference.

[0298] Based on the comprehensive fitness value fit, calculate the selection probability of each solution:

[0299] p(S i ) = fit(S i ) / sum(fit(S j )), j=1,2,...,N;

[0300] Construct a roulette wheel according to the selection probability, corresponding to each solution S i The sector area and its selection probability p(S i ) is proportional to.

[0301] A uniform random number r in the interval [0, 1] is randomly generated and it is determined which sector it falls into, and the corresponding solution is selected as the final reactive power optimization scheduling strategy.

[0302] In another embodiment of the present application, for each solution S in PS i The multiple objective function values ​​are fuzzy processed. Taking network loss as an example:

[0303] If f loss (S i ) <= f losslow :mu low (S i ) = 1

[0304] If f losslow < f loss (S i ) <= f loss mid :

[0305] mu low (S i ) = (f loss mid -f loss (S i )) / (f loss mid -f losslow )

[0306] mu mid (S i ) = (f loss (S i ) - f loss low ) / (f loss mid -f loss low )

[0307] If f lossmid < f _ loss (S _ i ) <= f _ loss _high :

[0308] mu mid (S i ) = (f loss high -f loss (S i )) / (f loss high -f loss mid )

[0309] mu high (S i ) = (f loss (S i ) - f loss mid ) / (f loss high -f loss mid )

[0310] If f loss (S i ) > f loss high :mu high (S i ) = 1

[0311] where f loss low 、f loss mid 、f loss high mu is the threshold value of low, medium and high levels of network loss. low (S i ), mu mid (S i ), mu high (S i ) are the solutions S i The network loss belongs to low, medium and high levels of membership. The membership of voltage deviation target can be defined similarly.

[0312] According to the target weight coefficient w loss and wdev, and the membership degrees are integrated using the fuzzy weighted average operator.

[0313] mu low (Si ) = w loss mu loss low (S i ) + w dev mu dev low (S i )

[0314] mu mid (S i ) = w loss mu loss mid (S i ) + w dev mu dev mid (S i )

[0315] mu high (S i ) = w loss mu loss high (S i ) + w dev mu dev high (S i )

[0316] Design a fuzzy decision rule base and give the priority level of the solution according to its comprehensive membership. The rule base can be designed as:

[0317] If (mu low (S i )>=thr low 1) and (mu _mid (S i )<=thr mid 1) and (mu high (S i )<=thr high 1):dec(S i ) = Great;

[0318] If (mu low (S i )>=thr low 2) and (mu mid (S i )<=thr mid 2) and (mu high (S i )<=thr high 2):dec(S i ) = Good;

[0319] If (mu low (S i )<=thr low 3) and (mu mid (S i )>=thr mid 3) and (mu high (S i )<=thr high 3):dec(S i ) = Moderate;

[0320] If (mu low (S i )<=thr low 4) and (mu mid (S i )>=thr mid 4) and (mu high (S i )>=thr high 4):dec(S i ) = Poor;

[0321] Where thr is the threshold parameter, dec(S _ i ) is the solution S _ i The priority level includes four levels: Great, Good, Moderate, and Poor, which reflects the comprehensive evaluation results of the scheduling strategy. The threshold parameters can be given by expert experience or automatically obtained from historical data through machine learning methods.

[0322] The comprehensive membership value mu of the solution that meets the highest priority level is further compared low (S i ), and the solution with the largest membership is selected as the final scheduling strategy. If the highest preference level contains multiple solutions with the same membership, additional screening conditions such as other indicators such as robustness can be combined, or the decision maker can select the scheduling strategy with the strongest subjective preference.

[0323] In another embodiment of the present application, an improved roulette wheel selection strategy based on fitness sorting and multi-angle rotation is provided, and the individuals of the population are sorted in descending order according to their fitness values ​​to obtain a fitness sorting vector:

[0324] fit rank = sort(fit(S i )), i=1,2,...,N

[0325] According to the sorted vector fit_ rank, calculate the expected number of selections for each individual:

[0326] exp _ i = N * (N i +1) / (N*(N+1) / 2), i=1, 2,...,N

[0327] That is, the higher the ranking of individuals, the higher the expected number of selections, which shows a linear decreasing trend. N is the population size.

[0328] Based on the expected number of selections, a roulette wheel is constructed and divided into N sectors. i Sector corresponds to individual S _ i , its starting and ending angles are:

[0329] start i = sum(expj / N * 360), j=1, 2,..., i-1;

[0330] end i = start i +exp i / N * 360

[0331] Construct an adaptive rotation angle generation method. Define the fitness discreteness of the population:

[0332] D = sqrt(sum((fit (S i ) - fit _ avg) 2 )) / (N-1), i =1, 2, ..., N;

[0333] where fit avg is the average of all individual fitness values.

[0334] Generate K rotation angles, the kth angle is:

[0335] theta k = k * 360 / (K * (1 + D)), k=1, 2,...,K;

[0336] When the population diversity is high (i.e., the D value is large), the rotation angle interval is small and the selection pressure is small; when the population tends to converge (i.e., the D value is small), the rotation angle interval is large and the selection pressure is large.

[0337] Let the pointer start from theta1 and rotate counterclockwise, each time rotating from theta2 to theta1, record the sector the pointer passes through. Rotate K times in total and get K candidate individuals.

[0338] Among the K candidate individuals, one is randomly selected as the final selection result.

[0339] Fitness sorting reduces the influence of super individuals, making roulette insensitive to fitness differences, and to some extent overcomes the problem of insufficient and excessive selection pressure. The expected number of selections is linearly related to the ranking, which reduces the amplification effect of fitness estimation errors. Multi-angle rotation can select candidate individuals from different regions of the population, which is conducive to maintaining population diversity. The rotation angle is adaptively adjusted according to the population convergence state, so that the selection pressure is dynamically matched with the evolutionary process. Finally, random selection is made from the candidate set, which balances exploration and utilization to a certain extent. The algorithm has a simple structure, is easy to implement, and has low computational overhead.

[0340] According to another aspect of the present application, a reactive power optimization system for a port distribution network based on an improved MOGRO is also provided, comprising:

[0341] at least one processor; and,

[0342] a memory communicatively connected to at least one of the processors; wherein,

[0343] The memory stores instructions that can be executed by the processor, and the instructions are used to be executed by the processor to implement the reactive power optimization method and system for the port distribution network based on the improved multi-objective gold panning algorithm described in the above technical solution.

[0344] In summary, in this embodiment, the reactive power optimization method and system of the port distribution network based on the improved multi-objective gold panning algorithm include the following steps: a K-means clustering algorithm taking into account the spatiotemporal characteristics, a reactive power optimization mathematical model of the distribution network containing SST, and an improved multi-objective gold panning algorithm; the K-means clustering algorithm is used to cluster large-scale distributed energy output scenarios with volatility and uncertainty, so as to increase the representativeness of the selected scenarios and reduce the flow calculation time; the reactive power optimization mathematical model of the distribution network containing SST includes an objective function, a constraint function and an SST operation mode; the improved MOGRO is used to solve the model.

[0345] According to one aspect of the present application, the principle of the K-means clustering algorithm is to achieve the minimum distance between the sample and the center of the cluster to which it belongs by iteration; the objective function is active network loss and voltage deviation; the constraint function includes power balance constraints, output constraints of various distributed energy sources, and parameter constraints of transformers; the ultimate goal of the improved MOGRO is the Pareto frontier, and the judgment rule is the judgment of the dominating relationship.

[0346] Embodiment 1: Provide a reactive power optimization method and system for the port distribution network based on the improved multi-objective gold panning algorithm. The K-means clustering algorithm can cluster large-scale distributed energy output scenarios with different spatiotemporal characteristics into typical scenarios. The reactive power optimization model of the distribution network containing SST includes an objective function and a constraint function. The operation mode of SST needs to be selected in combination with external characteristics. The improved multi-objective gold panning algorithm is used to solve the model. The principle of the K-means clustering algorithm is to achieve the minimum distance between the sample and the cluster center to which it belongs through iteration. The objective function is active network loss and voltage deviation. The constraint function includes power balance constraints, output constraints of various distributed energy sources, and parameter constraints of transformers. The flow calculation of the distribution network containing SST needs to select the operation mode according to whether the secondary side of the SST is connected to an active network. The ultimate goal of the multi-objective algorithm is the Pareto frontier, and the judgment rule is the judgment of the dominant relationship. The operation mode is divided into fixed AC voltage control, fixed power control and a combination of the two. The judgment of the dominant relationship is: if a i is the characteristic element of set A, b i are the characteristic elements of set B. If all the characteristic elements of A are superior to the corresponding characteristic elements of B, then A is said to dominate B. If not all the characteristic elements of A are superior to the corresponding characteristic elements of B, then A and B are said to not dominate each other.

[0347] According to one aspect of the present application, the operation mode is divided into constant AC voltage control, constant power control and a combination of the two; the determination of the dominance relationship is: if a i is the characteristic element of set A, b i are the characteristic elements of set B. If all the characteristic elements of A are superior to the corresponding characteristic elements of B, then A is said to dominate B. If not all the characteristic elements of A are superior to the corresponding characteristic elements of B, then A and B are said to not dominate each other.

[0348] According to one aspect of the present application, it is divided into two parts: scene clustering and optimized scheduling.

[0349] Scene clustering includes the following steps:

[0350] Step (1): Taking active output as the indicator, k samples are randomly selected as the initial cluster centers.

[0351] Step (2): For each sample in the data set, calculate the distance from it to the k cluster centers and assign it to the class corresponding to the cluster center with the smallest distance.

[0352] Step (3): for each cluster, recalculate its cluster center position.

[0353] Step (4): Repeat steps (2) and (3) until the number of iterations is reached.

[0354] The optimization dispatching includes the following steps: To realize the distribution network dispatching strategy based on the improved multi-objective gold panning algorithm, the following steps are included:

[0355] Step (1), initialize the number of iterations, population size, storage archive set capacity, and population location (output of each distributed energy source and parameters of each device).

[0356] Step (2) performs non-dominated relationship screening on the population, and determines whether to perform a elimination step on the filtered population based on the capacity of the repository Archive set. Eliminate the population based on the degree of crowding to form an Archive set, and select the current global optimal individual gbest based on the roulette wheel method.

[0357] Step (3) determines whether the number of iterations has been reached. If so, the process ends. If not, the process proceeds to the improved GRO update.

[0358] In step (4), the current determining factor, determining coefficient, fuzzy coefficient and distance coefficient are calculated respectively, and the fuzzy coefficient of the optimal update is improved by the adaptive fuzzy coefficient, and then the update method of the cooperative update is improved by the two-way guidance.

[0359] Step (5): randomly select one of the three updating methods of the improved GRO to update the position, update the corresponding fitness, use binary simulated crossover to perform a secondary update, and screen the non-dominated relationship after the update.

[0360] Step (6), use the binary crossover principle of the genetic algorithm to update the population for the second time, and perform non-dominated relationship screening, mix the updated population with the Archive to form a new population, and return to step (2).

[0361] The present invention is carried out in two steps:

[0362] The first step is to construct a mathematical model for reactive power optimization of the distribution network, which includes two objective functions and multiple constraint functions.

[0363] The second step is to analyze and improve the gold mining algorithm and use it to solve the model.

[0364] In the first step, a dual-objective optimization problem is formulated, with the objective function of minimizing the useful network loss and voltage deviation of the distribution network. The general form of the model can be expressed as follows:

[0365] (1);

[0366] (2);

[0367] Where: f loss is the system network loss, f ΔU is the voltage offset, G ij is the real part of the node admittance matrix i Row j column, U i , U j are the voltage amplitudes of nodes i and j respectively, θ ij is the voltage phase difference between node i and node j, Ω N is the set of all nodes, N is the number of nodes, U max , U min are the upper and lower limits of the node voltage amplitude, Φ f is the penalty function term.

[0368] The voltage limit penalty function expression is:

[0369] (3);

[0370] (4);

[0371] Where: is the penalty function coefficient.

[0372] When solving the first step of the optimization problem, some constraints need to be considered.

[0373] The constraint function equations include equality constraints and inequality constraint equations. The equality constraints are mainly the power balance constraint equations of the system, and sometimes also include constraint equations of special equipment; the inequality constraints are variable constraint equations, mainly network node voltage and current, various equipment capacity constraints and reactive and active power constraints, charge state constraints and transformer parameter constraints, etc.

[0374] a. Power balance constraint equation

[0375] When performing power flow calculations, the system's power balance equation will be reflected in the solution process of Newton's method. The specific equation is:

[0376] (5);

[0377] Where: PGrid , Q Grid are respectively the active power and reactive power generated by the power station; PV，j , Q PV，j , P WT，j , Q WT，j , P load，j , Q load，j , P i，j , Q i，j They are respectively the active and reactive power generated by photovoltaic and wind turbines, the active and reactive power of loads, and the active and reactive power lost on the line; : P DS，j , P EV，j are the active power of distributed energy storage and electric vehicle at node j respectively; Q CBs，j is the reactive power generated by the capacitor bank at node j; Ω PV ,Ω WT ,Ω DS ,Ω EV ,Ω CBs They are a collection of nodes equipped with photovoltaic, wind turbine, distributed energy storage, electric vehicles and capacitor group devices.

[0378] b. Variable constraint equation

[0379] Voltage and current constraints

[0380] (6);

[0381] Distributed Energy Constraints

[0382] (7);

[0383] Where: P DG，max , P DG，min , Q DG，max , Q DG，min , which are the upper and lower limits of the active and reactive output of distributed energy respectively.

[0384] Reactive power compensator constraints

[0385] (8);

[0386] Where: Q CBs，max , Q CBs，min They are the upper and lower limits of the capacitor bank output respectively.

[0387] OLTC gear constraints

[0388] (9);

[0389] Where: T is the OLTC gear, T min , T maxIt is the upper and lower limits of OLTC gear.

[0390] Distributed energy storage constraints

[0391] (10);

[0392] Where: is the minimum state of charge; The maximum state of charge.

[0393] Solid-state transformer constraints

[0394] (11);

[0395] Where: is the SST active power transfer efficiency, P s1 , Q s1 , P s2 , Q s2 are the active and reactive power of the primary and secondary sides of SST, respectively. SST is the solid-state transformer capacity.

[0396] The second step is to analyze and improve the gold mining algorithm.

[0397] There are three ways to update the gold rush algorithm, namely, optimal update, communication update, and cooperative update. First, you need to determine the determining factor, which determines whether the current population is in a convergence state or a search state.

[0398] Determining factors:

[0399] (10);

[0400] Where: t is the current iteration number, T is the maximum iteration number.

[0401] The specific forms of the three update methods are as follows.

[0402] Optimal update:

[0403] (11);

[0404] (12);

[0405] (13);

[0406] (14);

[0407] Where: X i (t+1), X i (t) is the position of the i-th gold digger before and after the update, Xgbest (t) is the current optimal solution position, A1 is the selection coefficient, l1 is the selection factor, D1 is the distance coefficient, C1 is the fuzzy coefficient, the size is a random number of [0, 2], r1, r2 are random numbers of [0, 1]. l1 is a number that decreases linearly from 2 to 1 as the iteration increases.

[0408] Communication Update

[0409] (15);

[0410] (16);

[0411] (17);

[0412] Where: X worker1 (t) is the position of any gold digger, A2 is the selection coefficient, l2 is the selection factor, and D2 is the distance coefficient.

[0413] Collaborative Updates

[0414] (18);

[0415] (19);

[0416] Where: X worker1 (t), X worker2 (t) are any two gold diggers, r1 is a random number in [0, 1], and D3 is the distance coefficient.

[0417] The improved method of this invention is as follows:

[0418] 1. Adaptive fuzzy coefficient

[0419] From the trending update, we can see that the fuzzy coefficient range is [0, 2], that is, the target update point is set to 0 to 2 times the position of the global optimal solution, so as to enhance the diversity of solutions. However, the constant fuzzy coefficient cannot meet the requirements of its iterative full-process search. With the later stage of iteration, the difference between the Pareto frontier solutions of the distributed energy scheduling strategy is reduced, and a large-scale search is not required at this time. Based on this, its fuzzy coefficient is improved as follows:

[0420] C 1 = [ 2 cos ( t − 1 T − 1 ⋅ π 2 ) ⋅ ( r 1 − 0 . 5 ) + 1 ] ⋅ r 2 (20);

[0421] 2. Two-way guidance

[0422] In multi-objective problems, there is not only one marginal solution in the Pareto solution set. Different from single-objective problems, the solution does not converge to one position, but converges to multiple positions. Therefore, it is impossible to judge the convergence of the current population by the distance between solutions. Therefore, cooperative updating is not suitable for the accuracy requirements of multi-objective problems. Based on this, this paper proposes to use the guidance of the direction of the last iteration and the vector sum of the guidance of the direction of the current iteration, and assign each random weight coefficient to guide its update. The improved update is as follows:

[0423] (twenty one);

[0424] 3. Secondary cross update

[0425] The optimal update draws on the single wolf update method of the Grey Wolf Algorithm. Compared with the three-in-one update method of the Grey Wolf Algorithm, its convergence ability and local search ability are lacking. Since the genetic algorithm crossover has a strong local integrated search ability, an additional crossover update is added to the population after each iteration to ensure the comprehensiveness of the local search of the population after each iteration.

[0426] In order to conduct a comprehensive local search on the population with a lower degree of convergence without destroying the population with a higher degree of convergence, this paper introduces an improved adaptive crossover probability to meet the requirements of the crossover probability in the whole process.

[0427] a. Adaptive crossover probability

[0428] (twenty two);

[0429] Where: pc is the crossover probability, pc1 and pc2 are the crossover start and end probabilities, f is the maximum fitness of any objective function of any two particles, and f average is the average value of the fitness of the objective function corresponding to f, max , f min is the maximum and minimum value of the corresponding objective function fitness.

[0430] b. Simulate binary crossover

[0431] X worker2 ( t + 1 ) = [ ( 1 + l ) ⋅ X worker1 ( t ) + ( 1 − l ) X worker2 ( t ) ] 2 X worker2 ( t + 1 ) = [ ( 1 − l ) ⋅ X worker1 ( t ) + ( 1 + l ) X worker2 ( t ) ] 2 (twenty three);

[0432] (twenty four);

[0433] Where: mu is the cross coefficient, m is a random number in [0, 1].

[0434] Figure 3This is a reactive power optimization flow chart of distribution network based on the improved multi-objective gold panning algorithm. The whole is divided into the multi-objective problem and the improved GRO update part. After initializing the number of iterations and population information, the non-dominated relationship of the population is screened. According to the capacity of the repository Archive set and the number of screened populations, it is determined whether to remove them to form an Archive set, and it is determined whether the number of iterations has been reached. If it has been reached, the improved GRO update is performed. The improved GRO update includes optimal, communicative, and cooperative updates. The adaptive fuzzy coefficients and bidirectional guidance are improved for the optimal and cooperative updates respectively, so as to update the population and screen the dominant relationship. Then, the population is updated twice by using simulated binary crossover, and the updated population is compared with the Arch i ve sets are mixed and the dominance relationship is screened, and finally the elimination step is returned to, and this cycle is repeated until the end, the optimal solution is output, and the optimal reactive power dispatching strategy is obtained.

Claims

1. A reactive power optimization method for port distribution network based on improved multi-objective gold panning algorithm, characterized in that: The steps include: The K-means clustering algorithm considering spatiotemporal characteristics is used to cluster the output scenarios of different types of energy in the distribution network to obtain typical output scenario data; According to the external characteristics of the solid-state transformer SST, the corresponding operation mode is selected, and a reactive power optimization model of the distribution network including the SST is established; the input data of the reactive power optimization model of the distribution network including the SST includes typical output scenario data; Call the multi-objective gold mining algorithm MOGRO, and introduce adaptive fuzzy coefficient, two-way guidance and secondary cross update to improve the multi-objective gold mining algorithm MOGRO; The improved multi-objective gold-mining algorithm MOGRO is used to solve the reactive power optimization model of the distribution network containing SST, and the reactive power optimization dispatching strategy set of the distribution network is obtained.

2. The reactive power optimization method for the port distribution network based on the improved multi-objective gold panning algorithm according to claim 1 is characterized in that: The K-means clustering algorithm considering the spatiotemporal characteristics is used to cluster the output scenarios of different types of energy in the distribution network, and typical output scenario data is obtained. The clustered typical scenarios are combined with the constant power factor mode of the generator to constrain the output of different distributed power sources in the established mathematical model, specifically: Taking active output as the indicator, k samples are randomly selected as the initial cluster centers; k is a natural number greater than 0; Calculate the distance between each sample in the data set and the centers of k clusters, and classify them into the cluster with the smallest distance; For each cluster, recalculate the cluster center position; Repeat the above steps until the set number of iterations is reached to obtain a typical output scenario that can represent different spatiotemporal characteristics.

3. The reactive power optimization method for the port distribution network based on the improved multi-objective gold panning algorithm according to claim 1 is characterized in that: The operation mode corresponding to the solid-state transformer SST is selected according to its external characteristics, and a reactive power optimization model of the distribution network including the SST is established, specifically: Considering whether the secondary side of the SST is connected to an active network, select the operation mode of constant AC voltage control, constant power control or a combination of the two; Select the equivalent node type corresponding to the SST node according to the operation mode; A dual-objective optimization model is established with minimizing network loss and voltage deviation as the objective function; The model constraints include power balance constraints, distributed energy output constraints, reactive power compensator constraints, OLTC gear constraints, energy storage state of charge constraints and SST capacity constraints.

4. The reactive power optimization method for the port distribution network based on the improved multi-objective gold panning algorithm according to claim 1 is characterized in that: The multi-objective gold mining algorithm MOGRO is called, and adaptive fuzzy coefficient, two-way guidance and secondary cross update are introduced to improve the multi-objective gold mining algorithm MOGRO, specifically: The adaptive fuzzy coefficient C1 is introduced into the optimal update, which changes dynamically from large to small with the iteration, and the search range is narrowed in the later stage; ； t is the current iteration number, T is the maximum iteration number; r1, r2 are random numbers in [0, 1]; Introducing bidirectional guidance for cooperative updating, using the vector sum of individual position changes in two iterations to guide the update; ； X i (t+1), X i (t) is the i The position of a gold digger before and after the update; D3 is the distance coefficient; After each iterative update, a secondary crossover operation is performed on the population. ； Where: pc is the crossover probability, pc1 and pc2 are the crossover start and end probabilities, f is the maximum fitness of any objective function of any two particles, and f average is the average value of the fitness of the objective function corresponding to f, max , f min is the maximum and minimum value of the corresponding objective function fitness; Introducing adaptive crossover probability and simulated binary crossover to enhance local search capabilities; ； ； In the formula, mu is the cross coefficient, m is a random number [0, 1], and X worker1 (t), X worker2 (t) for any two gold diggers; Among them, the adjustment strategy of the adaptive fuzzy coefficient is: In each iteration, the information entropy H(t) of the current population is calculated: H(t)=-sum i p i (t)logp i (t); Among them, p i (t) is an individual i The probability of selection is proportional to the number of solutions it dominates; Then, let the fuzzy coefficient C be negatively correlated with the entropy H(t): C(t)=C max -kH(t) / logN; In the formula, C max is the maximum fuzzy coefficient, k is the proportional coefficient, and N is the population size.

5. The reactive power optimization method for the port distribution network based on the improved multi-objective gold panning algorithm according to claim 1 is characterized in that: The improved multi-objective gold panning algorithm MOGRO is used to solve the reactive power optimization model of the distribution network containing SST, and the reactive power optimization dispatching strategy set of the distribution network is obtained, which is specifically: Initialize the individual positions and Archive set of the population; the positions include the output of each distributed energy source and equipment parameters; Screen the current population individuals based on non-dominated relationships, update the Archive set based on the crowding degree, and select the global optimal individual; If the maximum number of iterations is not reached, the current determining factor, determination coefficient, fuzzy coefficient and distance coefficient are calculated respectively, and the fuzzy coefficient of the optimal update is improved by adaptive fuzzy coefficient, and then the update method of cooperative update is improved by two-way guidance; Randomly select one of the three updating methods of improved GRO to update the position, update the corresponding fitness, use binary simulated crossover to perform secondary update, and screen the non-dominated relationship after the update; The binary crossover principle of genetic algorithm is used to update the population twice, and non-dominated relationships are screened, and the updated population is mixed with Archive to form a new population; Select the reactive power optimization dispatching strategy that meets the actual needs from the Pareto optimal solution of the final Archive set; During the solution process, the current population is adaptively gridded every several generations, the target space is divided into several small areas, and then the number of individuals in each grid is counted. Grids with too many individuals are clustered, and a representative individual is selected from each cluster to be retained, and the rest are discarded. For grids with too few individuals, new individuals are generated by interpolating the individuals in adjacent grids to fill the grid, and the diversity and uniformity of the population are maintained by dynamically adjusting the grid granularity.

6. The reactive power optimization method for the port distribution network based on the improved multi-objective gold panning algorithm according to claim 1 is characterized in that: It also includes selecting a reactive power optimization dispatching strategy that meets actual needs from the Pareto optimal solution of the reactive power optimization dispatching strategy set of the distribution network, specifically: Construct an evaluation index system, including balance index, robustness index and preference matching index. BAL=1-sqrt(sum((F i -mean(F) 2 ) / n); where F i To solve the function value on the i-th target, mean(F) is the average value of all target function values, n is the number of targets, and the value range of BAL is [0, 1]; R=1 / (1+sum(abs(F'-F)) / n); where F and F' are the objective function value vectors of the original solution and the perturbation solution, respectively, and the value range of R is (0, 1]; PREF=sum(w i ·(F i -F i min ) / (F i max -F i min )); where w i is the weight coefficient of the i-th target, reflecting its relative importance; F i min and F i max Respectively i The minimum and maximum values ​​of the objective function values; Retrieve Pareto non-dominated solutions, extract the eigenvector of each solution, and perform weighted summation of their qualities); select the top a% of solutions as positive samples and the remaining solutions as negative samples to train the machine model; the value range of a is 1-100; The newly generated Pareto solutions of the optimization algorithm are evaluated and screened in real time, and the optimal scheduling strategy that meets actual needs is automatically and efficiently selected to obtain the optimization results. The optimization results are then fed back to the optimization algorithm to guide the evolution direction of the population. According to the optimization results, the active output, reactive power and energy storage charging and discharging power of various types of distributed energy are coordinated and controlled; the reactive power of various devices including SST and distributed energy inverters is coordinated, and the transformer tap adjustment is coordinated to maintain voltage stability; Continuously monitor changes in distribution network load and fluctuations in distributed energy output, and initiate rolling adjustment of reactive power optimization.

7. The reactive power optimization method for the port distribution network based on the improved multi-objective gold panning algorithm according to claim 1 is characterized in that: The method uses the K-means clustering algorithm considering the spatiotemporal characteristics to cluster the output scenarios of different types of energy in the distribution network, obtains typical output scenario data, and operates the generator at a constant power factor, constrains the output of distributed energy to correspond to the constraint function in the mathematical model, and also includes: Take the membership degrees of each load node to all cluster centers as its characteristic attributes. For the \(i\)-th load node, extract its multi-dimensional attribute vector \(P(i)=[P(i,1),P(i,2),\cdots,P(i,d)]\) including active power load, reactive power load, voltage amplitude and spatial coordinates, and use the fuzzy membership function \(\mu\) to calculate the membership degree of the node i to the \(j\)-th cluster center \(C(j)\), \(\mu(i,j) = 1 / (\sum_{k = 1}^{c}(|P(i)-C(j)| / |P(i)-C(k)|)^{m})\), 2 / (m-1) where \(m\) is the fuzzy weighting exponent and \(c\) is the number of clusters; Based on the Euclidean distance, the attribute weight vector W = [w(1), w(2), ..., w(d)] is introduced to construct the weighted distance D(i, j) = ∑ k=1 d (w(k)·(P(i, k)-C(j, k)) 2 ), the attribute weight vector W is iteratively optimized using the gradient descent method. Each iteration is based on m i n(∑ i=1 n ∑ j=1 c (μ(i, j) m· D(i, j))) to update the weight until convergence, n is the number of load nodes; When calculating the membership degree μ(i, j) of computing node i to cluster center j, not only the distance factor is considered, but also the mutual influence factor between nodes α(i, k) = exp(-丨P(i)-P(k)丨 / β) is introduced to characterize the intensity of node i affected by other node k. The membership degree calculation formula is corrected by μ(i, j) = (1 - ∑ k≠i (α(i, k)·μ(k, j))) / (∑ l=1 c ((D(i, j) / D(i, l)) (2 / (m-1) ))), and then the membership degree matrix is normalized row by row and column by column; The cluster center is updated using the gradient descent method. Based on the current membership matrix, let C(j)=∑ i=1 n (μ(i, j) m ·P(i)) / ∑ i=1 n (μ(i, j) m ) updates the cluster center. To overcome the limitation of fixed step size, each iteration passes C(j) new =C(j) old -η·∇J(C(j)) adjusts the cluster center, and the step size η is dynamically updated using the decay strategy η=η0·exp(-t / τ), where t is the number of iterations, and the hyperparameters η0 and τ control the initial step size and decay rate, respectively; Automatically determine the optimal number of clusters. Initially, set the number of clusters c equal to the number of load nodes n, and perform iterative optimization. After each round of iteration, calculate the validity index XB(c) of the current clustering: XB(c)=∑ i=1 n ∑ j=1 c (μ(i,j) m ·丨P(i)-C(j)丨 2 ) / (n·m i n(k≠l)(丨C(k)-C(l)丨 2 )) If XB(c)>XB(c - 1), then set c = c - 1, merge the two closest cluster centers, and re-iterate; otherwise, set c best = c and terminate the clustering.

8. The reactive power optimization method for the port distribution network based on the improved multi-objective gold panning algorithm according to claim 1 is characterized in that: The reactive power optimization model of distribution network includes: Power Flow Equations for Distribution Networks with Interval and Fuzzy Variables node i The power balance equation is: P Gi -P Li =U i *sum(U j *(G ij cos(theta ij )+B ij s i n(theta ij ))),j=1::N; Q Gi -Q Li =U i *sum(U j *(G ij s i n(theta ij )-B ij cos(theta ij ))),j=1::N; Where P Gi , Q Gi Node i The active and reactive power generation range, U i For Node i Triangular fuzzy numbers of voltage amplitude; Introducing the tie line switch state variable matr i x{x} and continuous variable matr i x{f}, coupling network topology optimization with power flow solution; x ij ∈{0, 1}, indicating the on-off state of the tie line between nodes i and j; f ij ∈[0,f ij max], indicating the tie line flow; Topological variables and power flow variables must satisfy the following constraints: f ij -B ij *(theta i -theta j )=0; |f ij |≤x ij *f ij max; Among them B ij is the tie line susceptance, f ij max is the upper limit of the power flow.

9. The reactive power optimization method for the port distribution network based on the improved multi-objective gold panning algorithm according to claim 1 is characterized in that: A two-stage optimization strategy is adopted. The outer layer performs combined optimization on the topological variable matrix {x}, and the inner layer performs interval-fuzzy iterative solution on the power flow equation group under a given topology. The outer layer optimization adopts the taboo search algorithm to minimize the network loss interval width and voltage deviation fuzziness as the optimization goal, and searches for the optimal topological combination matrix {x}. The inner layer combines the interval Newton-Raphson method with the fuzzy extension operator to iteratively solve the mixed uncertainty power flow equation group and obtain the power flow interval and voltage fuzzy distribution. The power flow calculation results under different topologies are evaluated from two dimensions: interval width and fuzziness. Interval width indicator: W = sum((P imax -P imin ) 2 +(Q imax -Q imin ) 2 , i=1::N) / N; Model degree index: :D=sum((U ic -U ia ) 2 +(U ic -U ib ) 2 , i=1: : N) / (2*N); The final optimal topology and the corresponding power flow calculation results are selected on the Pareto frontier of the above two indicators.

10. A reactive power optimization system for port distribution network based on improved MOGRO, characterized in that: include: at least one processor; as well as, a memory communicatively connected to at least one of the processors; wherein, The memory stores instructions that can be executed by the processor, and the instructions are used to be executed by the processor to implement the reactive power optimization method for the port distribution network based on the improved multi-objective gold panning algorithm as described in any one of claims 1 to 9.

Images