Novel generalized comprehensive load modeling method

Abstract

The invention discloses a novel generalized comprehensive load modeling method. The method comprises the following steps of performing dimension reduction processing on power grid load data by utilizing a self-organizing mapping neural network; analyzing the load data of the power grid, and clustering the terminals with similar load characteristics to obtain subnet systems with different load characteristics; using the Elman neural network optimized by the grey wolf algorithm to carry out power system comprehensive load modeling; efficient and accurate modeling of power distribution network side loads is realized, the operation state of the power distribution network is accurately mastered, the observability of the power distribution network is improved, real-time calculation of the power distribution network loads is realized, the problem of limited data measurement configuration of the power distribution network is solved, and the observability of the system is improved; a subnet system formed by a neural network formed by power grid load data with different load characteristics can be obtained; the weight and the value domain of the Elman neural network are optimized through the grey wolf algorithm, so that the precision of the Elman neural network is improved, and the influence of an initial value is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of distribution network load modeling, and in particular to a novel generalized comprehensive load modeling method. Background Art

[0002] The loads in the distribution network are massive and complex, and the workload of modeling the load nodes one by one is huge; the loads are nonlinear and time-varying. At present, the power grid selects the same load model in practical applications, which is too rough and not accurate enough. The existing modeling objects are only traditional loads. With the continuous penetration of distributed power sources, the load composition is more complex, and the existing load model is difficult to describe the new generalized load.

[0003] For example, a "generalized comprehensive load modeling method and simulation system for AC / DC hybrid distribution network" disclosed in Chinese patent literature, with publication number: CN113067375A, discloses that an AC load model is formed by using an equivalent induction motor in parallel with a static load; a DC load model is formed by using an equivalent DC motor in parallel with a static load and a generalized load; the AC load model is connected to the AC bus, and the DC load model is connected to the AC bus through an interface model; the AC bus is connected to the high-voltage bus of the substation using an ideal transformer and an equivalent impedance, thereby deriving a generalized comprehensive load model structure of the AC / DC hybrid distribution network; however, this scheme does not take into account the impact of different load characteristics on the load model after the addition of distributed power sources.

[0004] For example, a "load dynamic modeling method considering the time-varying load composition of AC and DC distribution networks" disclosed in Chinese patent literature, with the announcement number: CN109245100B, discloses that the DC distribution network is approximately decoupled into a DC inverter power supply; step 2, simplifying and equating each load component of the DC distribution network, and further establishing a generalized comprehensive load equivalent model of the PCC point; step 3, historical data clustering and classifier training; step 4, performing overall identification based on the fault data recorded by the fault recorder, and establishing a model parameter library with the identification results; step 5, specifying the required simulation scenario, and selecting a priori model parameters based on the classification results; however, this solution does not take into account the problem of insufficient model complexity and accuracy. Summary of the invention

[0005] In order to solve the problem of insufficient accuracy of complex load models composed of different load characteristics of distributed power sources in the prior art, the present invention provides a novel generalized comprehensive load modeling method, which optimizes the weights and value range of the Elman neural network, thereby improving the accuracy of the Elman neural network and reducing the influence of different influencing factors of different load characteristics in the initial value.

[0006] In order to achieve the above object, the present invention provides the following technical solutions:

[0007] A new generalized integrated load modeling method includes the following steps:

[0008] Utilize self-organizing map neural network to reduce the dimension of power grid load data;

[0009] Analyze the power grid load data to cluster terminals with similar load characteristics and obtain subnetwork systems with different load characteristics;

[0010] The Elman neural network optimized by the grey wolf algorithm is used to model the comprehensive load of the power system. First, the self-organizing map neural network is used to reduce the dimension of the power grid load data and analyze the power grid load data to cluster the terminals with similar load characteristics to obtain subnet systems with different load characteristics. Then, the Elman neural network optimized by the grey wolf algorithm is used to model the comprehensive load of the power system. The efficient and accurate modeling of the load on the distribution network side is achieved, the operating status of the distribution network is accurately grasped, the observability of the distribution network is improved, the real-time calculation of the distribution network load is realized, and the problem of limited data measurement configuration of the distribution network is solved to improve the observability of the system.

[0011] Preferably, before the dimensionality reduction processing of the power grid load data using the self-organizing map neural network, the characteristics and influencing factors of the load are analyzed to determine the main state variables and the main influencing factors. The characteristics of different loads in the distribution network are screened to obtain the characteristic parameters of the different loads, and the significant characteristics are screened as the main state variables; the influencing factors of the load curve of each load are determined, and the factors with the greatest influence are screened as the main influencing factors. The power grid load data can be simplified to realize the classification of the power grid load data, the power grid load data can be divided according to the normal distribution curve, and the power grid load data can be updated according to the data represented by the normal distribution curve within the k value.

[0012] Preferably, the dimensionality reduction processing of power grid load data by using a self-organizing map neural network includes taking the dimension of the sample sequence of the power grid load data as the number of input nodes, taking the number of output layer nodes equal to or slightly less than the number of input sample vectors, and forming a SOM neural network. This simplifies the network, improves modeling efficiency, and ensures the accuracy of the model.

[0013] Preferably, the clustering of terminals with similar load characteristics includes: using a SOM neural network to classify the main factors of the samples of power grid load data, and training the SOM network; the number of sample classifications can be determined according to the number of clustering centers of the output winning samples; removing the output layer nodes that are far away from the sample cluster center and whose output values ​​are close to zero, and repeating the training until convergence. Loads with similar main characteristics can be clustered together, the number of each cluster is determined, and then the power grid load data outside a certain range of each cluster is screened again, thereby realizing the collection of similar data and the cleaning of isolated data.

[0014] Preferably, the subnet system with different load characteristics comprises: using the output layer nodes of the successfully trained SOM neural network as the input layer nodes, using the number of predicted output variables as the number of output nodes, constructing an Elman neural network, and initializing its weights. A subnet system formed by a neural network composed of power grid load data with different load characteristics can be obtained.

[0015] Preferably, the method of optimizing the Elman neural network using the gray wolf algorithm includes determining the optimal solution of the fitness value of the sample using the training data, and obtaining the optimal solution as the initial weight and range of the Elman neural network. The weight and range of the Elman neural network are optimized by the gray wolf algorithm, thereby improving the accuracy of the Elman neural network and reducing the influence of the initial value.

[0016] Preferably, the connection weight of the Elman neural network is used as the optimization target of GWO, and the mean square error of the training data is used as the fitness value. The fitness value is determined according to the connection weight, thereby determining the accuracy of the power load data represented by each node in the Elman neural network.

[0017] Preferably, the comprehensive load modeling of the power system includes inputting the load input into the optimal network to model the load, and comparing the output result with the actual value for verification. The verification of the model is realized, and the correlation between the output of the model and the actual output result is further determined.

[0018] The present invention has the following advantages:

[0019] (1) Realize efficient and accurate modeling of the load on the distribution network side, accurately grasp the operating status of the distribution network, improve the observability of the distribution network, realize real-time calculation of the distribution network load, and solve the problem of limited data measurement configuration of the distribution network, thereby improving the observability of the system; (2) Be able to obtain a subnetwork system formed by a neural network composed of grid load data with different load characteristics; (3) Optimize the weights and value range of the Elman neural network through the gray wolf algorithm, thereby improving the accuracy of the Elman neural network and reducing the impact of the initial value; (4) Verify the model and further determine the correlation between the model output and the actual output results. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The drawings described below are merely exemplary. A person skilled in the art can derive other implementation drawings based on the provided drawings without inventive effort.

[0021] Figure 1 Schematic diagram of the method steps in the embodiment.

[0022] Figure 2 Schematic diagram of the optimization of the grey wolf algorithm in the embodiment. DETAILED DESCRIPTION

[0023] The following is an explanation of the implementation mode of the present invention by means of specific embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in the field without creative work are within the scope of protection of the present invention.

[0024] like Figure 1-2 As shown, in a preferred embodiment, the present invention discloses a novel generalized comprehensive load modeling method, comprising:

[0025] 1. First, use the self-organizing map neural network to cluster the load data and obtain the classified subnet load data.

[0026] 1) Analyze the characteristics and influencing factors of the load, and determine the main state variables and main influencing factors.

[0027] 2) According to the dimension of the sample sequence as the number of input nodes, the number of output layer nodes is equal to or slightly less than the number of input sample vectors to form a SOM neural network.

[0028] 3) Use the SOM neural network to classify the main factors of the sample points and train the SOM network. The number of sample classifications can be determined based on the number of clustering center points of the output winning samples. The output layer nodes that are far away from the sample clustering center points and whose output values ​​are close to zero are removed. Repeat the training until convergence to reduce the number of output layer nodes of the SOM neural network.

[0029] 4) Using the output layer nodes of the successfully trained SOM neural network as the input layer nodes and the number of predicted output variables as the number of output nodes, an Elman neural network is constructed and its weights are initialized.

[0030] 2. Use the Elman neural network optimized by the Grey Wolf Algorithm to perform load modeling on classified load data

[0031] 1) Use the GWO optimization algorithm to optimize the weights and thresholds of the Elman neural network. Use the training data to calculate the fitness value of the individual and obtain the optimal solution as the initial weights and range of the Elman neural network.

[0032] 2) Take the Elman connection weight as the optimization target of GWO and the mean square error of the training data as the fitness value. After multiple iterations, find the optimal connection weight of the neural network and remember the optimal network at that moment.

[0033] 3) The load input is input into the optimal network to model the load, and the output results are compared with the actual values ​​to confirm that the model constructed in this paper is effective and accurate.

[0034] In the second embodiment, the specific implementation steps are as follows:

[0035] 1. Data preprocessing

[0036] The application data source of the present invention is the historical data samples of the load nodes to be identified in the distribution network system. The data is normalized to obtain the bus voltage U(t)=[V(t),...,V(t-n+1)] T 、Active power P(t)=[P0(t),...,P0(t-n+1)] T , reactive power Q(t) = [Q0(t),...,Q0(t-n+1)] T And bus frequency F(t)=[f(t),...,f(t-n+1)] T ;

[0037] 2. Load clustering based on SOM neural network

[0038] 1) Initialize the neuron weight vector. Assign a small random number to each weight vector of the output layer and normalize it to obtain

[0039] 2) Select the input data sample vector:

[0040] X=U(t)=[V(t),...,V(t-n+1)] T

[0041] 3) Find the optimal node d j .

[0042]

[0043] Where: ij is the connection weight between the input layer neuron i and the output layer neuron j in the SOM neural network.

[0044] 4) Update the weight vector ω ij .

[0045] ω ij (t+1)=ω ij (t)+η(t)h c,j (t)(x i -ω ij (t)) (2)

[0046] Where: η(t) is the learning efficiency, which is a gain function between [0,1]; h c,j (t) is a neighborhood function that decreases with time.

[0047] 5) Update learning efficiency η(t) and neighborhood function h c,j (t)

[0048]

[0049]

[0050]

[0051] Where: d cj is the distance between neuron c and neurons; r(t) is the neighborhood radius; INT is the rounding function; N is the total number of iterations.

[0052] 6) Determine whether convergence has occurred. Set t = t + 1 and return to step (2) until t = T or the learning rate reaches the maximum value.

[0053] 7) Calculate the error. Determine whether the actual output is within the allowable range based on the error between the actual output and the expected output.

[0054]

[0055] In the formula, C k is the actual output, Z k is the expected output.

[0056] 8) Use the output result as the initial weight of the Elman neural network.

[0057] 3. Elman neural network load modeling based on Grey Wolf optimization algorithm

[0058] 1) Determine the topological structure of the Elman neural network, set the transfer function of the neural network model and the number of hidden layer nodes h

[0059] The transfer function of the Elman neural network is:

[0060]

[0061] The number of hidden layer nodes h

[0062]

[0063] In the formula, h is the number of hidden layer nodes, m is the number of input layer nodes, n is the number of output layer nodes, and a is a constant between 1 and 10. In this paper, we take the number of input nodes m = 2, the number of output nodes n = 2, and the number of hidden layer nodes h = 3.

[0064] 2) Set the number of wolf packs in the GWO algorithm to s = 30, the maximum number of iterations to 50, and the search space dimension.

[0065] The formula for the search space dimension is:

[0066] dim=n×h+h×h+h×m+h×m (9)

[0067] 3) Randomly initialize the positions of the parent wolf pack and the child wolf pack, and calculate the fitness value of each wolf.

[0068] 4) Sort the fitness function values ​​of the parent wolf pack and determine the positions of the three wolves α, β, and δ.

[0069] 5) Iteratively update the parameter a, and update the positions of α wolf, β wolf and δ wolf, compare and update them with the offspring population, and select the one with better fitness as the parent population individual.

[0070] 6) Calculate the fitness value of the updated wolf group and sort it, and determine whether the optimization process reaches the termination condition. If so, terminate the optimization process and output the position of the parent population α wolf. If not, return to the loop;

[0071] 7) The position of the α wolf is assigned to the Elman neural network as the optimal initial weight and threshold, the GWO-Elman model is constructed, and the test sample data is used to verify the model performance.

[0072] 4. Model evaluation indicators

[0073] For the SOM-GWO-Elman model, the root mean square sum of squares of the diagonal (RMSSD) and the multiple correlation coefficient MR are selected as the overall evaluation indicators of the multi-output model, and the formula is:

[0074]

[0075]

[0076] Where: Y and They represent the multi-output matrix of actual values ​​and the multi-output matrix of predicted values ​​respectively; trace represents the trace of the matrix; N represents the total number of predicted samples; P represents the number of output variables; R i represents the correlation coefficient of the ith output variable.

[0077] The smaller the root mean square error, the higher the model accuracy; the closer the correlation coefficient R value is to 1, the stronger the correlation between the model output and the actual output.

[0078] Although the present invention has been described in detail above by general description and specific embodiments, it is obvious to those skilled in the art that some modifications or improvements can be made to the present invention. Therefore, these modifications or improvements made without departing from the spirit of the present invention all belong to the scope of protection claimed by the present invention.

Claims

1. A new generalized integrated load modeling method, characterized in that: The steps include: Utilize self-organizing map neural network to reduce the dimension of power grid load data; Analyze the power grid load data to cluster terminals with similar load characteristics and obtain subnetwork systems with different load characteristics; The Elman neural network optimized by the Grey Wolf Algorithm is used for comprehensive load modeling of power system.

2. A new generalized integrated load modeling method according to claim 1, characterized in that: Before the dimensionality reduction processing of the power grid load data is performed using the self-organizing map neural network, the characteristics and influencing factors of the load are analyzed to determine the main state variables and the main influencing factors.

3. A new generalized integrated load modeling method according to claim 1 or 2, characterized in that: The method of using the self-organizing map neural network to reduce the dimension of power grid load data includes: According to the dimension of the sample sequence of the power grid load data as the number of input nodes, the number of output layer nodes is equal to or slightly less than the number of input sample vectors to form a SOM neural network.

4. A new generalized integrated load modeling method according to claim 1 or 2, characterized in that: The clustering of terminals with similar load characteristics includes: Using SOM neural network to classify the main factors of the sample of power grid load data, and training the SOM network; The number of sample classifications can be determined based on the number of cluster center points of the output winning samples; Remove the output layer nodes that are far away from the sample cluster center and whose output values ​​are close to zero, and repeat the training until convergence.

5. A new generalized integrated load modeling method according to claim 4, characterized in that: The method of obtaining a subnet system with different load characteristics includes: using the output layer nodes of the successfully trained SOM neural network as the input layer nodes, using the number of predicted output variables as the number of output nodes, constructing an Elman neural network, and initializing its weights.

6. A new generalized integrated load modeling method according to claim 3, characterized in that: The method of optimizing the Elman neural network by using the gray wolf algorithm includes determining the optimal solution of the fitness value of the sample by using training data, and using the optimal solution as the initial weight and value range of the Elman neural network.

7. A new generalized integrated load modeling method according to claim 6, characterized in that: The connection weights of the Elman neural network are taken as the optimization target of GWO, and the mean square error of the training data is taken as the fitness value.

8. A new generalized integrated load modeling method according to claim 7, characterized in that: The comprehensive load modeling of the power system includes inputting the load input into the optimal network to model the load, and comparing the output result with the actual value for verification.

Images