Load forecasting method for low voltage substation based on neural network

Abstract

The low-voltage area load prediction method based on neural network disclosed in the present invention specifically includes the following steps: step 1, collecting historical load data of the low-voltage area; step 2, constructing a joint neural network model, and pre-training the joint neural network model; step 3, inputting the historical load data into the trained joint neural network model to obtain a preliminary prediction value; step 4, performing error and pass rate analysis on the preliminary prediction value, if the pass rate meets the requirements, the preliminary prediction value is identified as the final prediction value, if the pass rate does not meet the requirements, return to step 3, and obtain a new preliminary prediction value again. The present invention solves the problem of slow load prediction process and low prediction accuracy of the existing low-voltage area through the combination of multiple algorithms.

Description

Technical Field

[0001] The invention belongs to the technical field of power system distribution network management, and in particular relates to a low-voltage area load prediction method based on a neural network. Background Art

[0002] In recent years, with the improvement of residents' living standards, the number of electricity users has increased exponentially, and the structure of the distribution network has become more and more complex. At the same time, the grid structure of the distribution system has become more and more complex. Complex and multi-source factors make load forecasting of the power system, especially the low-voltage distribution station area, increasingly difficult.

[0003] Load forecasting is the basis of power system planning, scheduling, dispatching and operation, and plays an important supporting role in ensuring the safe and economic operation of the power grid. Load forecasting can be divided into long-term, medium-term, short-term and ultra-short-term forecasts based on the forecast time range. Among them, short-term load forecasting plays an important guiding and regulating role in the daily operation of power grid companies. Accurate forecasting results can arrange the daily production plan of electricity more reasonably. However, short-term load forecasting is currently mostly based on neural networks. When traditional BP neural networks are used to forecast the load of low-voltage substations, the complexity of the low-voltage substation network and the conflict between the learning rate of the algorithm lead to a slow entire forecasting process and low forecasting accuracy. Summary of the invention

[0004] The purpose of the present invention is to provide a low-voltage area load forecasting method based on a neural network, which solves the problems of slow low-voltage area load forecasting process and low forecasting accuracy through the combination of multiple algorithms.

[0005] The technical solution adopted by the present invention is a low-voltage area load forecasting method based on a neural network, which specifically includes the following steps:

[0006] Step 1, collecting historical load data of low voltage areas;

[0007] Step 2, constructing a joint neural network model and pre-training the joint neural network model;

[0008] Step 3, input the historical load data into the trained joint neural network model to obtain a preliminary prediction value;

[0009] Step 4, analyze the error and pass rate of the preliminary prediction value. If the pass rate meets the requirements, the preliminary prediction value will be identified as the final prediction value. If the pass rate does not meet the requirements, return to step 3 to obtain a new preliminary prediction value.

[0010] The present invention is also characterized in that

[0011] Step 3 specifically includes the following steps:

[0012] Step 3.1, preprocess the historical load data to obtain the initial data set;

[0013] Step 3.2, use Tent mapping to initialize the data set, generate uniform sequences, and calculate their fitness;

[0014] Step 3.3, determine whether the fitness meets the requirements, if so, proceed to step 3.8, if not, proceed to step 3.4;

[0015] Step 3.4, update the locations of the discoverer, joiner, and sentinel based on the sparrow search algorithm;

[0016] Step 3.5, perform Gaussian mutation and Tent perturbation to change the position according to the fitness after the position update;

[0017] Step 3.6, input the parameter data after the position change in step 3.5 into the graph convolutional neural network to initialize the parameter data;

[0018] Step 3.7, perform the output of the first layer convolution, the second layer convolution, and the third layer convolution through the graph convolutional neural network, and perform step 3.8 based on the output data;

[0019] Step 3.8, initialize the data, update the weight of the initialized data based on the Grey Wolf Optimization Algorithm, and obtain the preliminary prediction value;

[0020] Step 3.9, judge the accuracy of the preliminary prediction value. If it meets the requirements, output it for error and pass rate analysis. If it does not meet the requirements, return to step 3.6.

[0021] The preprocessing of historical load data in step 3.1 specifically includes supplementing missing values, processing outliers and normalizing the historical load data in sequence;

[0022] The missing value supplement is specifically shown in the following formula:

[0023]

[0024] where y n and n+1 are the load values ​​at the corresponding time, y n+k is a missing value, and 0 <k<i;

[0025] The outlier processing is divided into a horizontal processing method and a vertical processing method;

[0026] The normalization process is shown in the following formula:

[0027]

[0028] Among them, X *is the data after normalization; X min is the minimum value in the dataset; X max is the maximum value in the dataset.

[0029] In step 3.2, the Tent mapping is used to initialize the dataset as shown in the following formula:

[0030]

[0031] where, x n+j is the (n + 1)-th individual in the population, x n is the n-th individual in the population, and a is a constant in the interval (0, 1);

[0032] Calculate its fitness as shown in the following formula:

[0033]

[0034] where, x i is the position of different sparrows; α i is the distance from the original position.

[0035] In step 3.4, the position update of the discoverer is shown in the following formula:

[0036]

[0037] where, is the position information matrix of the i-th individual in the j-th dimension at the t-th iteration; α is a random number in the interval [0, 1]; iter max is the maximum number of iterations; Q is a random number following a normal distribution; L is a matrix of all 1s; R2 is the warning value, a random number in the interval [0, 1]; ST is the safety value, a random number in the interval [0.5, 1]; R2 < ST means the discoverer is in a safe position, and R2 > ST means the discoverer is in a dangerous position;

[0038] The position update of the joiner is shown in the following formula:

[0039]

[0040] where, is the position information matrix of the i-th individual in the j-th dimension at the t-th iteration; is the position information matrix of the worst individual at the t-th iteration; is the position information matrix of the best discoverer at the t-th iteration; Q is a random number following a normal distribution; A is a matrix randomly assigned 1 or -1, A is the position information matrix of the worst individual at the t-th iteration; is the position information matrix of the best discoverer at the t-th iteration; Q is a random number following a normal distribution; A is a matrix randomly assigned 1 or -1, A + = A T (AA T ) -1 ; L is a matrix of all 1s; n is the number of joiners, To be located at the rear of the joiner population;

[0041] The position update of the sentinel is shown as follows:

[0042]

[0043] in, is the position information matrix of the jth dimension of the ith individual in the tth iteration; is the position information matrix of the optimal individual in the tth iteration; β is a random number that obeys a normal distribution with a mean of 0 and a variance of 1; K is a random number in the interval [-1,1]; ε is a very small constant; f i is the current individual fitness, f g is the optimal fitness, f W is the worst fitness.

[0044] The position change by the Gaussian mutation is shown as follows:

[0045] x i ′=x i (1+N(0,1)),f i <f avg ;

[0046] Among them, x i is the current individual location information; x i ′ is the individual position information after the change; N(0,1) is a random number that conforms to the standard normal distribution; f i is the current individual fitness; f avg is the average fitness of the population;

[0047] The position change through the Tent disturbance is shown as follows:

[0048] x i ′=x i (1+Tent(0,1)),f i >f avg ;

[0049] Among them, x i is the current individual location information; x i ′ is the individual position information after the change; Tent(0,1) is the random number in the Tent chaotic sequence in the interval (0,1); f i is the current individual fitness; f avg is the average fitness of the population.

[0050] Step 3.7 The first layer convolution output is as follows:

[0051]

[0052] Among them, W1 is the weight matrix in the second graph convolution layer, and b1 is the bias vector of the second graph convolution layer;

[0053] The second layer convolution output is as follows:

[0054] H (2) =ReLu(A″H (1) W2+b2);

[0055] Among them, W2 is the weight matrix in the second graph convolution layer, and b2 is the bias vector of the second graph convolution layer;

[0056] The third layer convolution output is shown as follows:

[0057] H (3) =ReLu(H (2) W3+b 31 );

[0058] Among them, W3 is the weight matrix in the second graph convolution layer, and b3 is the bias vector of the second graph convolution layer.

[0059] In step 3.8, the weight of the initialized data is updated by first updating the nonlinear convergence factor of the initialized data based on the gray wolf optimization algorithm, as shown in the following formula:

[0060]

[0061] Among them, K1 and K2 are adjustment coefficients, and their value range is (0,1); t is the current number of iterations; T is the maximum number of iterations;

[0062] Then update the data dynamic weight, as shown in the following formula:

[0063]

[0064] in, is the current iteration position of different gray wolf individuals; W 1,2,31 is the weight corresponding to the wolf

[0065] Finally, the wolf pack position is updated to obtain the preliminary prediction value, as shown in the following formula;

[0066]

[0067] in, is the position of the gray wolf at the t+1 iteration.

[0068] The error analysis in step 4 specifically includes average error analysis and root mean square error analysis;

[0069] The average error analysis is specifically shown in the following formula:

[0070]

[0071] Among them, x predicted is the predicted value; x real is the true value;

[0072] The root mean square error analysis is specifically shown in the following formula:

[0073]

[0074] Among them, x predicted is the predicted value; x real is the true value.

[0075] The qualified rate analysis described in step 4 is specifically shown in the following formula:

[0076]

[0077] Where r is the qualified rate of load forecasting; N is the number of relative error points in the forecast results;

[0078] When the relative error point number N in the prediction result is less than 5%, the prediction value is deemed to be the final prediction value.

[0079] The beneficial effects of the present invention are:

[0080] The low-voltage substation load forecasting method based on neural network of the present invention adds graph convolutional neural network, sparrow search algorithm and gray wolf optimization algorithm on the basis of the existing traditional BP neural network, which solves the problem that when performing existing load forecasting for low-voltage substations, the complexity of the low-voltage substation network and the conflict between the learning rate of the algorithm lead to the slowness of the entire forecasting process. The rapidity of solving multiple objectives and the method of finding the optimal solution in the algorithm can meet the real-time and accuracy requirements of the load forecasting for low-voltage substations.

[0081] It is conducive to the load and operation planning of low-voltage substations, improving the reliable and stable power supply of the distribution network system, guiding the transformation and construction of low-voltage residential power grids, and maximizing the economic investment benefits. At the same time, it can realize the reasonable planning of the operation of the distribution system, which is also of positive significance to the capacity planning of substations and the adjustment of transformer load rates, and is conducive to the grid planning and construction of the entire distribution system, and realizes the interactive power supply of each substation. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] Figure 1 It is a flow chart of the low voltage area load prediction method based on neural network of the present invention. DETAILED DESCRIPTION

[0083] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0084] Example 1

[0085] like Figure 1 As shown, the low voltage area load prediction method based on neural network of the present invention specifically includes the following steps:

[0086] Step 1, collecting historical load data of low voltage areas;

[0087] Step 2, constructing a joint neural network model and pre-training the joint neural network model;

[0088] Step 3, input the historical load data into the trained joint neural network model to obtain a preliminary prediction value;

[0089] Step 4, analyze the error and pass rate of the preliminary prediction value. If the pass rate meets the requirements, the preliminary prediction value will be identified as the final prediction value. If the pass rate does not meet the requirements, return to step 3 to obtain a new preliminary prediction value.

[0090] Furthermore, the joint neural network model in step 2 refers to improving the shortcomings of the basic BP neural network algorithm. At the same time, the sparrow search algorithm (SSA) and sparrow search algorithm models, the graph convolutional neural network algorithm model based on the graph convolutional neural network (GCN) and the grey wolf model based on the grey wolf optimization (GWO) are added.

[0091] As the most widely used neural network algorithm, BP neural network has very good applicability for nonlinear prediction problems. However, when using BP neural network for load forecasting, it should be noted that the establishment of low-voltage area load forecasting model should pay special attention to its identity when selecting samples, so that the accuracy of load forecasting results can be improved through network training. Therefore, it is necessary to repair and adjust the existing BP neural network.

[0092] For a typical AMN neuron model, the model has m inputs, namely P1, P2, P3...P m , when the next layer of network is connected, each input is assigned a weight value w according to its characteristics. The value of the weight value w depends on the degree of correlation between the input and the result. Therefore, the output expression of the model is as follows:

[0093] a=f(wp+b);

[0094] Where f represents the transfer function of the input / output relationship.

[0095] As mentioned above, the BP neural network corrects the weights and thresholds of the network through the negative gradient direction. The specific process is as follows: First, the weight and iteration matrix X of the BP neural network are formed. i , i is the number of iterations, X k Denotes the matrix of the kth iteration, let g k is the gradient of the function, g k is the network learning rate, and the following formula is obtained:

[0096] X k+1 =X k -a k g k ;

[0097] A conventional three-layer BP network is used for simulation. Assume that the input layer contains i nodes and the hidden layer contains x nodes. i , the output layer node is y m , set the initial thresholds of the hidden layer and the output layer to be b m and b n , the network weights between the input layer, output layer and hidden layer nodes are ω ml and ω mn .

[0098] Then we have:

[0099] The output of the hidden layer node is:

[0100] y m =f(∑ω ml x l -b m )=f(net m );

[0101] The output layer node output is:

[0102] y m =f(∑ω mn y m -b n )=f(net n );

[0103] Assume the expected value of the output layer node is t n , the error function is:

[0104]

[0105] The derivative of the error function with respect to the weights of the hidden layer and the output layer is:

[0106]

[0107] The output layer node error is:

[0108] δ n =-(t n -Z n )×f′(net n );

[0109] Derivative the weights of the input layer and the hidden layer according to the error function:

[0110]

[0111] The hidden layer node error is:

[0112]

[0113] Assume the weight correction value is Δω ml and Δω mn , because Δω ml and Δω mn are proportional to the error function along the gradient descent, then:

[0114]

[0115]

[0116] The entire repair process is now completed.

[0117] Furthermore, when pre-training the joint neural network model, in order to better illustrate the relationship between load and complex meteorological changes, the correlation analysis method is used to study the correlation between daily meteorological conditions and load fluctuations. i , X i are the actual load value and the i-th type of meteorological variable respectively, and the calculation formula obtained according to the Pearson correlation coefficient method is as follows:

[0118]

[0119] Among them, r is the correlation coefficient between weather and load, a positive value indicates positive correlation, and a negative value indicates negative correlation; i is different categories of weather

[0120] Combining weather data with other information after correlation analysis, the following points are included:

[0121] Meteorological factors: The most important meteorological factor affecting load forecasting is temperature. Temperature changes cause residents to use a large amount of cooling and heating equipment.

[0122] Economic factors: The impact of economic factors on load is long-term and far-reaching. The economy is an important factor driving the development of electricity. With economic development and improved living standards, people's demand for electricity and quality requirements are increasing. Economic development drives the popularization of electrification. The increasing degree of electrification leads to an increase in electricity consumption, and the load on the power grid will naturally become higher and higher.

[0123] Time factor: The influence of time factor on the load of distribution network area is mainly manifested in the following aspects:

[0124] 1. Seasonal changes, the length of day and night and temperature changes in each season bring about regularity in load, and load fluctuations are more obvious;

[0125] Second, the alternation of working days and non-working days, and the alternation of weeks brought about by time, make the load fluctuate greatly;

[0126] 3. Legal holidays will cause irregular fluctuations in load. This is mainly reflected in the fact that people will choose to travel during legal holidays, which will lead to a smooth load fluctuation in the substation area.

[0127] Electricity price factor: Whether to consider it in load forecasting depends on the length of the forecast time. For example, it does not need to be considered in short-term load forecasting.

[0128] Rainfall factor: The impact of rainfall on load characteristics is mainly manifested in two aspects. First, the temperature drop will cause a drop in temperature in a short period of time, that is, it will lead to a decrease in the load of residents in summer, and an increase in heating use in winter, which increases the volatility of the load and brings negative impacts. Second, with the rapid development of hydropower stations, heavy rainfall has enabled areas rich in small hydropower to obtain more hydropower supply, reducing the grid supply load, which will also bring negative impacts on load forecasting.

[0129] The impact of rainfall on load is mainly through rainfall level, number of rainfall days, and the cumulative and lag effects brought by continuous rainfall.

[0130] According to the correlation coefficients of different weather factors, we can determine which factors should be considered in the load forecast of the substation, and make relevant adjustments to the joint neural network model according to the specific situation.

[0131] Furthermore, the error analysis in step 4 specifically includes average error analysis and root mean square error analysis;

[0132] The mean absolute error has a higher degree of reflecting the difference between the actual value and the predicted value. The use of the mean absolute error can effectively avoid the problem of positive and negative values ​​of the error. The mean error analysis is specifically shown in the following formula:

[0133]

[0134] Among them, x predicted is the predicted value; x real is the true value;

[0135] The root mean square error is also used as a comprehensive evaluation index. Compared with the mean absolute error and the mean square error, it has a better ability to reflect the precision of the prediction results. The root mean square error analysis is specifically shown in the following formula:

[0136]

[0137] Among them, x predicted is the predicted value; x real is the true value;

[0138] The pass rate analysis is a criterion for judging the quality of the entire prediction result. It calculates the evaluation index of each prediction point by combining the previous evaluation criteria and sets the threshold to 5%. If the pass rate analysis ε of a prediction result is less than 5%, it is determined that the prediction deviation of the point is within the qualified range. Therefore, the pass rate analysis described in step 4 is specifically shown in the following formula:

[0139]

[0140] Where r is the qualified rate of load forecasting; N is the number of relative error points in the forecast results;

[0141] When the relative error point number N in the prediction result is less than 5%, the prediction value is deemed to be the final prediction value.

[0142] The processing method and process of data in the joint neural network model during pre-training are consistent with the specific implementation method of step 3.

[0143] Example 2

[0144] On the basis of Example 1, step 3 in the low voltage area load forecasting method based on neural network of the present invention specifically includes the following steps:

[0145] Step 3.1, preprocess the historical load data to obtain the initial data set;

[0146] The preprocessing specifically includes supplementing missing values, processing outliers and normalizing the historical load data in sequence;

[0147] The missing value supplement is specifically shown in the following formula:

[0148]

[0149] where y n and n+1 are the load values ​​at the corresponding time, y n+k is a missing value, and 0 <k<i;

[0150] Outlier processing is divided into horizontal processing method and vertical processing method. The horizontal processing method is shown as follows:

[0151] like:

[0152] min{|y(n,t)-y(n,t-1)|,|y(n,t)-y(n,t+1)|}>ω(t);

[0153] but:

[0154]

[0155] Among them, y(n,t), y(n,t-1), and y(n,t+1) are the load values ​​at three adjacent moments on the nth day; ω(t) is the preset threshold;

[0156] The vertical processing method is shown as follows:

[0157] like:

[0158] |y(n,t)-m(t)|>ω(t);

[0159] but:

[0160]

[0161] Among them, y(n,t) is the load value at the time of the nth day; m(t) is the average load value at the time t of consecutive days of the same day type; δ(t) is the load fluctuation threshold.

[0162] In order to avoid errors caused by data of different dimensions, it is necessary to normalize different types of data in the input data set. After normalization, all types of data are in the same dimension, which can reduce the training time of the model and avoid the influence of dimension. The normalization process is shown in the following formula:

[0163]

[0164] Among them, X * is the normalized data; X min is the minimum value in the data set; X max is the maximum value in the data set.

[0165] Step 3.2, use Tent mapping to initialize the data set, generate uniform sequences, and calculate their fitness;

[0166] In practical applications, in order to improve the global search capability of the sparrow search algorithm and improve the diversity and uniformity of the population, it is necessary to initialize the population using the Tent chaotic mapping method. The data set initialized using the Tent mapping is shown in the following formula:

[0167]

[0168] Among them, x n+j is the n+1th individual in the population, x n is the nth individual in the population, a is a constant in the interval (0,1);

[0169] The fitness is calculated as shown in the following formula:

[0170] For each sparrow in the D-dimensional solution space, the position of X = (x1, x2, ..., x D ), the fitness value is f i =f(x1,x2,...,x D ).

[0171] For the fitness function f, its calculation formula is as follows:

[0172]

[0173] Among them, x i is the position of different sparrows; α i is the distance from the original position, which will affect the final position of the optimal solution. i When =0, the optimal solution position coincides with the origin, so its specific value can be manually set according to the specific experiment, thereby interfering with the final position of the optimal solution.

[0174] Step 3.3, determine whether the fitness meets the requirements, if so, proceed to step 3.8, if not, proceed to step 3.4;

[0175] Step 3.4, update the locations of the discoverer, joiner, and sentinel based on the sparrow search algorithm;

[0176] Step 3.5, perform Gaussian mutation and Tent perturbation to update the position according to the fitness after the position update. After continuing step 3.4 for the number of iterations, proceed to step 3.6, otherwise continue steps 3.4 to 3.5.

[0177] Step 3.6, input the parameter data after the position change in step 3.5 into the graph convolutional neural network to initialize the parameter data;

[0178] Step 3.7, perform the output of the first layer convolution, the second layer convolution, and the third layer convolution through the graph convolutional neural network, and perform step 3.8 based on the output data;

[0179] Step 3.8, initialize the data, update the weight of the initialized data based on the Grey Wolf Optimization Algorithm, and obtain the preliminary prediction value;

[0180] Step 3.9: Judge the accuracy of the preliminary prediction value. If it meets the requirements, output and perform error and qualification rate analysis. If it does not meet the requirements, return to Step 3.6.

[0181] Embodiment 3

[0182] Based on Embodiment 2, in the low-voltage substation area load prediction method based on neural network of the present invention, the sparrow search algorithm model of the sparrow search algorithm is a heuristic search algorithm inspired by the foraging behavior of sparrows and the behavior of escaping pursuers. Its basic idea is to divide the search space into three populations: discoverers, joiners, and guards, and find the global optimal solution of the objective function according to the different search behaviors of different populations. The sparrow search algorithm first sorts the individuals according to the fitness from excellent to poor. In the present invention, the fitness index is the error, that is, the error is sorted from small to large; the discoverers are a population composed of individuals with better fitness, the joiners are a population composed of individuals with poorer fitness, and the guards are randomly selected individuals.

[0183] Therefore, the position update of the discoverers in Step 3.4 is shown in the following formula:

[0184]

[0185] Where, is the position information matrix of the i-th individual in the j-th dimension at the t-th iteration; α is a random number in the interval [0, 1]; iter max is the maximum number of iterations; Q is a random number obeying the normal distribution; L is a matrix all of whose elements are 1; R2 is the warning value, a random number in the interval [0, 1]; ST is the safety value, a random number in the interval [0.5, 1]; when R2 < ST, the discoverer is in a safe position, and when R2 > ST, the discoverer is in a dangerous position;

[0186] The position update of the joiners is shown in the following formula:

[0187]

[0188] Where, is the position information matrix of the i-th individual in the j-th dimension at the t-th iteration; is the position information matrix of the worst individual in the t-th iteration; is the position information matrix of the best discoverer in the t-th iteration; Q is a random number obeying the normal distribution; A is a matrix randomly assigned 1 or -1, A + = A T (AA T ) -1 ; L is a matrix all of whose elements are 1; n is the number of joiners, is located at the rear of the joiner population;

[0189] ​​The position update of the sentinel is shown as follows:

[0190]

[0191] in, is the position information matrix of the jth dimension of the ith individual in the tth iteration; is the position information matrix of the optimal individual in the tth iteration; β is a random number that obeys a normal distribution with a mean of 0 and a variance of 1; K is a random number in the interval [-1,1]; ε is a very small constant; f i is the current individual fitness, f g is the optimal fitness, f W is the worst fitness.

[0192] In the process of finding the optimal solution of the objective function, the sparrow search algorithm may face the situation where the objective function has a large number of local minima, causing the algorithm to fall into the local optimum. In order to further improve the global optimization accuracy of the algorithm, disturbances are added to the positions of individuals with different fitness: for individuals with better fitness, the position is changed by Gaussian mutation; for individuals with poor fitness, the position is changed by Tent perturbation. The position change by Gaussian mutation is shown in the following formula:

[0193] x i ′=x i (1+N(0,1)),f i <f avg ;

[0194] Among them, x i is the current individual location information; x i ′ is the individual position information after the change; N(0,1) is a random number that conforms to the standard normal distribution; f i is the current individual fitness; f avg is the average fitness of the population;

[0195] The position change through the Tent disturbance is shown as follows:

[0196] x i ′=x i (1+Tent(0,1)),f i >f avg ;

[0197] Among them, x i is the current individual location information; x i ′ is the individual position information after the change; Tent(0,1) is the random number in the Tent chaotic sequence in the interval (0,1); f i is the current individual fitness; f avgis the average fitness of the population.

[0198] Furthermore, for a GCN model, which contains two graph convolutional layers and two dense layers, its input data (X, A) is in graph form, and the intermediate features are generated using the intrinsic topological structure. Its data is fed to the output H of the first layer. (1) Specifically, we get a mixed feature matrix It linearly combines the feature vector of each node with the feature vectors of adjacent nodes using the weights represented by A″. Next, a new set of features is obtained by multiplying by the weight matrix W1 and adding the bias vector b1. Therefore, the first layer convolution output in step 3.7 is as follows:

[0199]

[0200] Among them, W1 is the weight matrix in the second graph convolution layer, and b1 is the bias vector of the second graph convolution layer;

[0201] Similarly, the second layer convolution output is as follows:

[0202] H (2) =ReLu(A″H (1) W2+b2);

[0203] Among them, W2 is the weight matrix in the second graph convolution layer, and b2 is the bias vector of the second graph convolution layer;

[0204] The third layer convolution output is shown as follows:

[0205] H (3) =ReLu(H (2) W3+b3);

[0206] Among them, W3 is the weight matrix in the second graph convolution layer, and b3 is the bias vector of the second graph convolution layer.

[0207] Furthermore, the gray wolf optimization algorithm was proposed by Australian scholar Mirjalili, who was inspired by the hunting behavior of wolves. The algorithm has the characteristics of fewer control parameters, high search efficiency and fast convergence speed.

[0208] The present invention updates the weight of the initialized data based on the gray wolf optimization algorithm, specifically by first updating the nonlinear convergence factor of the initialized data based on the gray wolf optimization algorithm, as shown in the following formula:

[0209]

[0210] Among them, K1 and K2 are adjustment coefficients, and their value range is (0,1); t is the current number of iterations; T is the maximum number of iterations;

[0211] After that, the adaptive position is updated. Since in the actual optimization process, the α wolf is more likely to be the optimal solution. In order to highlight the global leadership of the optimal wolf to the population and strengthen the global search ability of the algorithm, the weight of the α wolf is gradually increased during the position update process, as shown in the following formula:

[0212]

[0213] in, is the current iteration position of different gray wolf individuals; W i is the weight of the corresponding wolf.

[0214] Finally, the dynamic weight is modified. The average value cannot reflect the importance of α, β, and γ. Therefore, the data weight is updated dynamically, as shown in the following formula:

[0215]

[0216] Finally, the wolf pack position is updated to obtain the preliminary prediction value, as shown in the following formula;

[0217]

[0218] in, is the position of the gray wolf at the t+1 iteration.

Claims

1. A low voltage area load forecasting method based on a neural network, characterized in that: The specific steps include: Step 1, collecting historical load data of low voltage areas; Step 2, constructing a joint neural network model and pre-training the joint neural network model; Step 3, input the historical load data into the trained joint neural network model to obtain a preliminary prediction value; Step 4: Analyze the error and pass rate of the preliminary prediction value. If the pass rate meets the requirements, the preliminary prediction value is identified as the final prediction value. If the pass rate does not meet the requirements, return to step 3 to obtain a new preliminary prediction value. Step 3 specifically includes the following steps: Step 3.1, preprocess the historical load data to obtain the initial data set; Step 3.2, use Tent mapping to initialize the data set, generate uniform sequences, and calculate their fitness; Step 3.3, determine whether the fitness meets the requirements, if so, proceed to step 3.8, if not, proceed to step 3.4; Step 3.4, update the locations of the discoverer, joiner, and sentinel based on the sparrow search algorithm; Step 3.5, perform Gaussian mutation and Tent perturbation to change the position according to the fitness after the position update; Step 3.6, input the parameter data after the position change in step 3.5 into the graph convolutional neural network to initialize the parameter data; Step 3.7, perform the output of the first layer convolution, the second layer convolution, and the third layer convolution through the graph convolutional neural network, and perform step 3.8 based on the output data; Step 3.8, initialize the data, update the weight of the initialized data based on the Grey Wolf Optimization Algorithm, and obtain the preliminary prediction value; Step 3.9, judge the accuracy of the preliminary prediction value. If it meets the requirements, output it for error and pass rate analysis. If it does not meet the requirements, return to step 3.

6.

2. The method for low voltage area load forecasting based on neural network according to claim 1 is characterized in that: The preprocessing of historical load data in step 3.1 specifically includes supplementing missing values, processing outliers and normalizing the historical load data in sequence; The missing value supplement is specifically shown in the following formula: ； in and are the load values ​​at the corresponding time, is a missing value, and ; The outlier processing is divided into a horizontal processing method and a vertical processing method; The normalization process is shown in the following formula: ； in, is the data after normalization; is the minimum value in the data set; is the maximum value in the data set.

3. The method for low voltage area load forecasting based on neural network according to claim 1 is characterized in that: In step 3.2, the Tent mapping is used to initialize the data set as shown below: ； in, is the n+1th individual in the population, is the nth individual in the population, is a constant in the interval (0,1); The fitness is calculated as shown in the following formula: ； in, For the location of different sparrows; is the distance from the original position.

4. The method for low voltage area load forecasting based on neural network according to claim 1, characterized in that: The discoverer position update described in step 3.4 is as follows: ； in, is the position information matrix of the jth dimension of the ith individual in the tth iteration; for Random numbers in the interval; is the maximum number of iterations; is a random number that follows a normal distribution; is a matrix of all 1s; is the warning value, A random number in the interval; ST is a security value, Random numbers in the interval; To keep the finder in a safe position, for the discoverer is in a dangerous position; The joiner position update is shown as follows: ； in, is the position information matrix of the jth dimension of the ith individual in the tth iteration; The position information matrix of the body; is the position information matrix of the best finder in the t-th iteration; is a random number that follows a normal distribution; is a matrix randomly assigned with values ​​of 1 or −1, ; is a matrix of all 1s; is the number of participants, To be located at the rear of the joiner population; The position update of the sentinel is shown as follows: ； in, is the position information matrix of the jth dimension of the ith individual in the tth iteration; is the position information matrix of the optimal individual in the tth iteration; is a random number that follows a normal distribution with a mean of 0 and a variance of 1; for Random numbers in the interval; is a very small constant; is the current individual fitness, is the optimal fitness, is the worst fitness.

5. The method for low voltage area load forecasting based on neural network according to claim 4 is characterized in that: The position change by the Gaussian mutation is shown as follows: ； in, is the current individual location information; is the individual position information after the change; is a random number that conforms to the standard normal distribution; is the current individual fitness; is the average fitness of the population; The position change through the Tent disturbance is shown as follows: ； in, is the current individual location information; is the individual position information after the change; For Interval Random numbers in chaotic sequences; is the current individual fitness; is the average fitness of the population.

6. The method for low voltage area load forecasting based on neural network according to claim 4 is characterized in that: Step 3.7 The first layer convolution output is as follows: ； in, is the weight matrix in the second graph convolutional layer, is the bias vector of the second graph convolutional layer; The second layer convolution output is as follows: ； in, is the weight matrix in the second graph convolutional layer, is the bias vector of the second graph convolutional layer; The third layer convolution output is shown as follows: ； in, is the weight matrix in the second graph convolutional layer, is the bias vector of the second graph convolutional layer.

7. The method for low voltage area load forecasting based on neural network according to claim 4 is characterized in that: In step 3.8, the weight of the initialized data is updated by first updating the nonlinear convergence factor of the initialized data based on the gray wolf optimization algorithm, as shown in the following formula: ； in, , is the adjustment coefficient, and its value range is ; is the current iteration number; is the maximum number of iterations; Then update the data dynamic weight, as shown in the following formula: ； in, is the current iteration position of different gray wolf individuals; is the weight corresponding to the wolf Finally, the wolf pack position is updated to obtain the preliminary prediction value, as shown in the following formula; ； in, is the position of the gray wolf at the t+1th iteration.

8. The method for low voltage area load forecasting based on neural network according to claim 1, characterized in that: The error analysis in step 4 specifically includes average error analysis and root mean square error analysis; The average error analysis is specifically shown in the following formula: ； in, is the predicted value; is the true value; The root mean square error analysis is specifically shown in the following formula: ； in, is the predicted value; is the true value.

9. The method for low voltage area load forecasting based on neural network according to claim 8, characterized in that: The qualified rate analysis described in step 4 is specifically shown in the following formula: ； in, is the pass rate of load forecast; is the relative error point in the prediction result; When the relative error point number N in the prediction result is less than 5%, the prediction value is deemed to be the final prediction value.

Images