CN113285440B - Low-voltage transformer area loss reduction optimization method - Google Patents

Abstract

The invention discloses a low-voltage transformer area loss reduction optimization method, which comprises the following steps: 1) Establishing a BP neural network model, wherein the BP neural network comprises an input layer, an output layer and one or more hidden layers; 2) Training sample data by utilizing a particle swarm optimization algorithm, and taking output as an initial weight and a threshold of the BP neural network; 3) And randomly selecting a plurality of distribution room data to input into the trained BP neural network, and deriving a calculation value of the BP neural network. According to the invention, the particle swarm algorithm is applied to the optimization of the initial threshold and the weight of the BP neural network according to the characteristics of low-voltage transformer area line loss calculation, so that the precision and the speed of network training are improved.

Description

A Loss Reduction Optimization Method for Low-voltage Platform Area

technical field

The invention relates to the technical field of line loss calculation in a low-voltage station area of a distribution network, in particular to a loss reduction optimization method for a low-voltage station area.

technical background

In the distribution network, the low-voltage station area is a power supply area with a voltage of 0.4kV, which provides electricity guarantee for the majority of residents and enterprises. Line loss refers to the loss of electric energy from the power supply end to the power consumption end of the power system, which is caused by the loss of electric energy through the transmission line during the transmission process. In line loss management, due to the large number of equipment, insufficient management and power theft, the power grid company's precise loss reduction of the power grid is affected.

In the study of low-voltage station areas, the topology of the network is diverse and complex. In the same low-voltage platform area, the load and power also have large fluctuations. Therefore, in the theoretical calculation of line loss, it can only be as close as possible to the actual value of line loss through algorithm optimization, but it can never be the real value.

The line loss rate of the power grid is a key technical and economic indicator that comprehensively reflects the level of power grid planning, production operation, and operation management. Reducing the line loss rate is an important measure for power grid energy conservation and emission reduction, and it is also a need for power grid companies to improve their own competitiveness. In order to control the line loss within a reasonable range, the power grid company needs to establish a line loss management platform to monitor abnormal line loss and reduce the line loss rate.

In the traditional line loss calculation, various electrical data of the power system are used. Including the original structure diagram of the power grid and various electrical operating parameters (such as current, voltage and regular factors, etc.). In the management process of low-voltage station areas, due to the large number of users in the station area, imperfect data collection, and large gaps in line distribution, it is difficult to achieve accurate calculation of theoretical line loss. In line loss management, a lot of manpower and material resources need to be mobilized, and the effect is not good. For example, errors and omissions may occur in manual meter reading. Due to the huge workload, it is difficult for the power grid company to collect the necessary data and information for accurate calculation of line loss, and it is even more difficult to find abnormal line loss and complete line loss reduction processing.

Based on the above status quo, it is necessary to find a suitable intelligent algorithm to quickly and accurately calculate the line loss in the station area, and apply the results to the measures to improve the management level of the power grid such as line loss management.

Contents of the invention

The purpose of the present invention is to overcome the deficiencies in the prior art, propose a kind of low-voltage station area loss reduction optimization method, apply the particle swarm algorithm to optimize the initial threshold and weight of BP neural network according to the characteristics of low-voltage station area line loss calculation, improve Accuracy and speed of network training.

In order to achieve the above-mentioned purpose, a method for optimizing the loss reduction of the low-pressure platform area designed by the present invention is special in that it includes the following steps:

1) set up BP neural network model, described BP neural network comprises an input layer, an output layer and one and above hidden layers;

2) Utilize the particle swarm optimization algorithm to train the sample data, and use the output as the initial weight and threshold of the BP neural network;

3) Randomly select the data of several stations to input the trained BP neural network, and derive the calculated value of the BP neural network.

Preferably, the steps of the particle swarm optimization algorithm include:

(1) Initialize the particle parameters and determine the fitness function;

(2) Evaluate the fitness of each individual in the group;

(3) Determine the individual extremum of each particle and the overall extremum of the particle swarm according to the fitness of the individual;

(4) Update the position and velocity of the particle according to the formula;

(5) Stop updating when the maximum number of iterations is reached, and save the optimal solution;

(6) Take the optimal solution as the initial network weight and threshold, and use the neural network to solve it;

(7) Compare the m optimal solutions obtained by the BP neural network, so as to obtain the overall optimal solution.

Preferably, the update formula of particle velocity and position in the particle swarm optimization algorithm is:

X _i is the variable of the i-th particle position, k is the number of updates, pbest represents the best current position of each particle itself, gbest represents the best position of all particles in the entire group in history, r ₁ is from 0 to 1 Random number, c ₁ is the learning factor, w is the inertia factor, V _i ^k represents the current speed.

Preferably, the hidden layer nodes of the BP neural network are 5 layers, with 25 weights and 6 thresholds in total.

Preferably, the station area data includes station area user properties, line type, distribution transformer capacity, station area user capacity ratio, single-phase user total proportion, station area average load rate, station area average power factor, station area average busbar Voltage, total shape factor of the station area, average three-phase unbalance coefficient of the station area, proportion of total power consumption of single-phase users, power supply distance, and average user power supply length.

Preferably, the calculation formula of the minimum mean square error of the BP neural network is:

SN is the total number of samples, N is the dimension of the output vector, a _ij is the first output value a _ij , and t _ij is the target value t _ij .

Preferably, the calculation method of the average load rate of the station area is: (total daily power supply/24)/distribution transformer capacity*100%.

Preferably, the calculation method of the average three-phase unbalance coefficient ε _at of the station area is:

Among them: I _At is the A-phase current, the unit is: A, a total of n collection points a day; I _Bt is the B-phase current, the unit is: A, a total of n collection points a day; I _Ct is the C-phase current, the unit is: A, There are n collection points in one day.

Preferably, the power supply distance refers to the distance between the distribution transformer and the load center, and the average user power supply length refers to the average distance between all users and the load center.

Preferably, the user properties and line types in the station area data are discrete categorical variables, and the distribution transformer capacity, station area user capacity ratio, and station area line loss rate are continuous variables.

In order to meet the quality requirements of the neural network algorithm on the sample data, the beneficial effects of a loss reduction optimization method for low-voltage platform areas proposed by the present invention are as follows:

1. Through the analysis of the distribution characteristics of the sample data and the correlation analysis of the characteristic indicators, the variables and original data that can be used for the next step of machine learning are selected.

2. The particle swarm optimization algorithm has low computational complexity, guarantees the optimal solution with a high probability, and overcomes the local optimal defect of the BP algorithm. The improved particle swarm optimization algorithm established in MATLAB optimizes the BF neural network model, which verifies that the model has the characteristics of high precision.

Description of drawings

Fig. 1 is a schematic diagram of a topological equivalent model of a low-pressure station area.

Figure 2 is a schematic diagram of the original data in an Excel table.

Figure 3 is a schematic diagram of the analysis of the distribution characteristics of the sample data.

Fig. 4 is a schematic diagram of the structure of the BP neural network model.

detailed description

The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

The specific steps of a loss reduction optimization method for a low-pressure platform area proposed by the present invention are as follows:

1) set up BP neural network model, described BP neural network comprises an input layer, an output layer and one and above hidden layers;

2) Utilize the particle swarm optimization algorithm to train the sample data, and use the output as the initial weight and threshold of the BP neural network;

3) Randomly select the data of several stations to input the trained BP neural network, and derive the calculated value of the BP neural network.

The equivalent topological model of the low-voltage station area is shown in Figure 1. When the network topology is unknown and the required data is insufficient, a virtual power distribution network can be established. It is assumed that the distribution transformer and each user are connected to this network to form a virtual power distribution network. The network topology structure, the connection point of all loads in this structure is the virtual load center point. The original data of the low-voltage station area includes distribution transformer side data and user side data. The clustering indicators initially selected through the collected data are shown in Table 1.

Table 1 Distribution transformer side data and user side data

The specific meaning of the data in the table is as follows in order of precedence:

User property values: 1 urban network, 2 rural network, 0 others; calculation method: according to the statistics of urban and rural network attributes of low-voltage users in the station area: if the number of urban network users is large, the result is 1; if the number of rural network users is large , the result is 2, and if there are as many urban-rural networks or more empty ones, the result is 0.

Line type value: 1 overhead line, 2 cable, 3 mixed, 0 others; calculation method: according to the line input method corresponding to the user's power supply: if there are many overhead lines, it is 1, and cables (including direct buried cables, overhead cables, Cable tray, cable tunnel, cable pipe well) is 2 if the result is mixed, 3 if the result is mixed, and 0 for others.

Values of distribution variable type: 1 public variable, 2 special variable; temporarily the system only has public variable data, and this index is temporarily not included.

Calculation method of user capacity ratio in station area: (total capacity of single-phase users + total capacity of three-phase users)/capacity of distribution transformer.

Calculation method for the total proportion of single-phase users: total capacity of single-phase users / (total capacity of single-phase users + total capacity of three-phase users).

Calculation method for the total proportion of three-phase users: total capacity of three-phase users/(total capacity of single-phase users + total capacity of three-phase users).

The calculation method of the average load rate in the station area: (total daily power supply/24)/distribution transformer capacity*100%. It can be checked with average power (average power/distribution transformer capacity*100%).

The calculation method of the maximum load rate in the station area: maximum power/distribution transformer capacity*100%.

The calculation method of the total shape coefficient of the station area: the daily load shape coefficient is used to characterize the change curve of the daily load of the station area, which is defined as the ratio K of the root mean square current to the average current, and the calculation formula is:

When the collected data is n data points a day, considering the load three-phase unbalance, the calculation formula is:

Among them: I _At is the A-phase current, the unit is: A, a total of n points a day; I _Bt is the B-phase current, the unit is: A, a total of n points a day; I _Ct is the C-phase current, the unit is: A, A total of n points a day;

Calculation formula of average three-phase unbalance coefficient in station area

Considering the calculation of the daily line loss in the station area, the daily average three-phase unbalance coefficient is used to characterize the unbalanced load of the distribution transformer in the station area in a single day.

The calculation formula is:

Among them: I _At is the A-phase current, the unit is: A, a total of n collection points a day; I _Bt is the B-phase current, the unit is: A, a total of n collection points a day; I _Ct is the C-phase current, the unit is: A, a total of n collection points a day;

Proportion of total electricity consumption of single-phase users: total daily electricity consumption of single-phase users/(total daily electricity consumption of single-phase users + total daily electricity consumption of three-phase users).

Load center: the center point of the user's geographic coordinates.

Power supply distance: the distance between the distribution transformer and the load center.

Average user power supply length: the average distance between all users and the load center.

The original distribution transformer side data and user side data of the low-voltage station area cannot be directly used for machine learning, and blank values and meaningless data need to be removed by preprocessing. Analyze the distribution characteristics of the preprocessed data to further understand the data of the low-pressure station area, and remove a few variables with high correlation and low importance through the correlation analysis and importance analysis among all variables, and initially select them for neural network training Variables.

The original data in this embodiment includes 16 variables, the basic information of which is shown in Table 1, wherein, user properties, line types are discrete category variables, distribution transformer capacity, station area user capacity ratio and station area line loss rate, etc. is a continuous variable. Figure 2 is a screenshot of the data in an Excel table. The original data includes relevant indicators of 19324 station areas. The line loss rate of the station area in the data is distributed from -100% to 100%. Since the line loss rate of the station area less than 0 has no practical significance, the less than 0 and greater than 20 are excluded. The relevant data of the line loss rate of % station area is obtained, and the data of 18720 station areas are obtained for one day.

Import the sample data into the data mining software Weka to analyze the distribution characteristics. The basic distribution is shown in Figure 3.

BP (back propagation) neural network is a feed-forward neural network with three or more layers, and it is the most widely used machine learning algorithm. The basic principle is to continuously modify the weights and thresholds of the network to obtain the output closest to the target value by backpropagating the error between the network output value and the target value. The most basic BP neural network includes an input layer, an output layer and one or more hidden layers, as shown in Figure 4. In Fig. 4, x ₁ , x ₂ ... x _m are input variables, y ₁ , y ₂ ... y _p are hidden layer neurons, O ₁ , O ₂ ... O _m are output variables. w _ij is the weight from the i-th input variable to the j-th neuron, which represents the importance of the input variable, and the higher the weight, the more important it is. w _jk is the weight from the jth neuron to the kth output variable. b _ij and b _jk are thresholds, and only when the sum of the inputs exceeds the threshold will the output change.

The BP neural network needs to set the initial weight and threshold before training. If the gap between the first output value a _ij and the target value t _ij is not small enough, the neural network will use the minimum mean square error MSE of the two as the error signal to be backpropagated by the output layer, and the connection weights and thresholds will be corrected one by one.

The formula for calculating the minimum mean square error (MSE) is as follows:

It can be seen from the formula that MSE is a function of weight and threshold. In the formula, SN is the total number of samples, and N is the output vector dimension. Next, the BP neural network will perform the next training according to the new weight and threshold until the gap between the output value a _ij and the target value t _ij is small enough. Although BP neural networks are widely used in practical research and applications, there are no guiding standards for the number of neuron nodes in the hidden layer of the network, the type of activation function, and the selection of initial weights and thresholds, which lacks theoretical support.

The training results of BP neural network are shown in Table 2 below.

Table 2 BP neural network training results

In the table, E _c is the percentage of sample relative error; the number of iterations is the sum of the number of iterations of various types. It can be seen from Table 2 that as the training error increases, the number of iterations decreases, more and more stations fall into the area where the relative error is less than 1%, and fewer and fewer targets fall into the relative error Greater than 5% of the area.

The data of 10 station areas were randomly selected and put into the trained BP neural network, and the calculated value of the BP model was exported and compared with the real line loss rate. The results are shown in Table 3.

Table 3 Errors of test results of BP neural network model

The maximum absolute error of BP neural network model is 0.63%, and the maximum relative error is 0.135%.

This embodiment adopts the BP model neural network, which has 4 input terminals and 1 output terminal, corresponding to 4 characteristic parameters and line loss rate respectively. The total number of weights in the network should be equal to or less than the sample size, so that the resulting mathematical model is relatively stable. Therefore, in this embodiment, it is more appropriate to have 5 hidden layer nodes, and the network structure is 4 input variables-5 hidden layers-1 output variable, with a total of 25 weights and 6 thresholds. A total of 10381 samples are selected for the calculation example, and each sample data contains one independent variable and one dependent variable. They are the average power factor of the station area, the load shape factor, the proportion of single-phase capacity in the station area and the ratio of user capacity in the station area, and the output is the line loss rate.

Particle swarm optimization algorithm (PSO) is not sensitive to the scaling of design variables, and only has few algorithm parameters, so it is an efficient global search algorithm. In PSO there are two important features: velocity and position. It is updated by some definite rule formula, and reaches the optimal point after continuous updating. The following two formulas are update formulas for speed and position.

X _i is the variable of the i-th particle position, k is the number of updates, pbest represents the best current position of each particle itself, gbest represents the best position of all particles in the entire group in history, r ₁ is from 0 to 1 Random number, c ₁ is the learning factor, w is the inertia factor, V _i ^k represents the current speed.

The process of the weight and threshold of the PSO-BP neural network is to set the minimum mean square error MSE of the network as the fitness function, and use the optimization result of PSO as the initial weight and threshold of the neural network.

The main steps of the PSO-BP neural network are as follows:

(1) Initialize the particle parameters and determine the fitness function;

(2) Evaluate the fitness of each individual in the group;

(3) Determine the individual extremum of each particle and the overall extremum of the particle swarm according to the fitness of the individual;

(4) Update the position and velocity of the particle according to the formula;

(5) Stop updating when the maximum number of iterations is reached, and save the optimal solution;

(6) Use these optimal solutions as initial network weights and thresholds for neural network solution;

(7) Compare the m optimal solutions obtained by the BP neural network, so as to obtain the overall optimal solution.

The experiment of neural network algorithm combined with particle swarm optimization (PSO-BP): Firstly, the parameter setting of the algorithm is given. The population size of the training algorithm based on particle swarm optimization is s=30, the inertia weight W takes a random value, the acceleration factor c1=c2=2, the weight is a variable in the [-1,1] interval, and the hidden layer nodes of the neural network are determined by the empirical formula The number is 5, and the algorithm stop condition is the maximum number of iterations (4000) or the error precision is 0.0001.

Table 4 shows the errors of the calculation results after training with the PSO-BP neural network model and the GA-BP neural network model respectively.

Table 4 Comparison of PSO-BP neural network and GA-BP neural network

The maximum absolute error of the GA-BP model is 0.3%, and the maximum absolute error of the PSO-BP model is 0.39%, which is 52% and 38% lower than that of the BP model, which is 0.63%. The maximum relative error of the GA-BP model is 0.093%, and the maximum absolute error of the PSO-BP model is 0.107%, which are respectively reduced by 31% and 20.7% compared with the maximum relative error of the BP model which is 0.135%. Comparing the BP model with the GA-BP model and the PSO-GA model, it can be found that the calculation accuracy of the neural network optimized by both GA and PSO is greatly improved. The number of iterations of the BP neural network optimized by the particle swarm optimization algorithm is obviously better than that of the genetic algorithm in the case of the same magnitude of precision. The result is higher than the precision of BP neural network optimized by genetic algorithm.

Finally, it should be noted that the above specific embodiments are only used to illustrate the technical solution of the patent and not to limit it. Although the patent has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the patent can be Modifications or equivalent replacements without departing from the spirit and scope of the technical solution of this patent shall be covered by the claims of this patent.

Claims (1)

1. A method for optimization of loss reduction in low-pressure platform area, characterized in that, comprising the steps of: 1) set up BP neural network model, described BP neural network comprises an input layer, an output layer and one and above hidden layers; 2) Utilize the particle swarm optimization algorithm to train the sample data, and use the output as the initial weight and threshold of the BP neural network; 3) Randomly select the well-trained described BP neural network of several station area data input, the calculated value of described BP neural network is derived; The station area data includes station area user nature, line type, distribution transformer capacity, station area user capacity ratio, single-phase user total proportion, station area average load rate, station area average power factor, station area average bus voltage, station area The overall shape factor of the area, the average three-phase unbalance coefficient of the station area, the proportion of the total power consumption of single-phase users, the power supply distance, and the average user power supply length; The calculation formula of the minimum mean square error of the BP neural network is:

SN is the total number of samples, N is the output vector dimension, a _ij is the first output value a _ij , t _ij is the target value t _ij ; The calculation method of the average load rate of the station area is: (total daily power supply/24)/distribution transformer capacity*100%; The calculation method of the average three-phase unbalance coefficient ε _at in the station area is:

Among them: I _At is the A-phase current, the unit is: A, a total of n collection points a day; I _Bt is the B-phase current, the unit is: A, a total of n collection points a day; I _Ct is the C-phase current, the unit is: A, a total of n collection points a day; The power supply distance refers to the distance between the assigned transformer and the load center, and the average user power supply length refers to the average distance between all users and the load center; The steps of the particle swarm optimization algorithm include: (1) Initialize the particle parameters and determine the fitness function; (2) Evaluate the fitness of each individual in the group; (3) Determine the individual extremum of each particle and the overall extremum of the particle swarm according to the fitness of the individual; (4) Update the position and velocity of the particle according to the formula; (5) Stop updating when the maximum number of iterations is reached, and save the optimal solution; (6) Take the optimal solution as the initial network weight and threshold, and use the neural network to solve it; (7) compare m optimal solutions obtained by BP neural network, so as to obtain the overall optimal solution; The update formula of particle velocity and position in the particle swarm optimization algorithm is:

X _i is the variable of the i-th particle position, k is the number of updates, pbest represents the best current position of each particle itself, gbest represents the best position of all particles in the entire group in history, r ₁ is from 0 to 1 Random number, c ₁ is the learning factor, w is the inertia factor, V _i ^k represents the current speed; The hidden layer nodes of the BP neural network are 5 layers, with 25 weights and 6 thresholds in total; In the station area data, user nature and line type are discrete category variables, and distribution transformer capacity, station area user capacity ratio and station area line loss rate are continuous variables.

Images