Photovoltaic household transformer relation prediction method based on EEMD decomposition and NRBO optimization

Abstract

The invention discloses a photovoltaic user-transformer relation prediction method based on EEMD decomposition and NRBO optimization. EEMD decomposition is carried out on a user-transformer relation original data set; a DRSN-BIGRU-ATTENTION network model is established, and the DRSN-BIGRU-ATTENTION network model is utilized to carry out multi-user data prediction on the user change relation data obtained after EEMD decomposition is carried out on the user change relation data; the optimal parameters of the DRSN-BIGRU-ATTENTION network model are searched for by means of a Newton-Raphson optimization algorithm NRBO (Newton Raphson Optimization); and outputting a user-transformer relation data prediction value. According to the method, a plurality of independent prediction models are combined, and more accurate and reliable overall prediction is obtained by synthesizing the prediction results of the prediction models.

Description

Technical Field

[0001] The present invention relates to the technical field of household-to-household transformer relationship prediction, and in particular to a photovoltaic household-to-household transformer relationship prediction method based on EEMD decomposition and NRBO optimization. Background Art

[0002] The prediction of household-to-transformer relationship of household photovoltaic in low-voltage distribution network is crucial to the reliability, safety and economy of energy system. In capacity planning, according to the power generation of household photovoltaic, the power is reversed to match the distribution equipment with the demand of energy resources. In terms of power grid security, the correct prediction of household-to-transformer relationship can reduce potential power grid security risks. In terms of operating costs, the accurate prediction of household-to-transformer demand of photovoltaic users can help power grid operators to reasonably arrange resources and reduce operating costs.

[0003] Predicting the household-to-transformer relationship of household photovoltaic users in low-voltage distribution networks usually requires a large amount of historical data and a suitable prediction model. Common prediction methods include time series analysis, regression analysis, neural networks, support vector machines, etc. At the same time, considering the complexity and real-time nature of the power system, it is also necessary to combine real-time data acquisition and distributed control strategies to ensure that the prediction model can respond to changes in the system state in a timely and accurate manner. To achieve the prediction of household-to-transformer relationship of household photovoltaics in the distribution network, it is necessary to collect historical data, including the power consumption and power generation information of household photovoltaic users in the distribution network, the operating status of distribution transformers, environmental conditions, etc. These data may come from smart meters, sensors, SCADA systems, etc. Then, the data is preprocessed, including missing value processing, outlier detection and denoising, to ensure the quality and accuracy of the data. It is necessary to perform feature engineering on the data, that is, to extract useful features from the original data. These features may include statistical characteristics of power load (such as mean, standard deviation), environmental conditions (such as temperature, humidity, sunshine intensity), date and time features (such as hours, etc.). Good feature selection can improve the performance of the prediction model.

[0004] According to the nature of the problem and the characteristics of the data, select a suitable prediction model. Common prediction models include time series models (such as ARIMA, Prophet), machine learning models (such as linear regression, decision tree, random forest), deep learning models (such as recurrent neural network, long short-term memory network), etc. For the prediction of household photovoltaic household-to-household relationship, it is often necessary to take into account the characteristics of time series and the complex correlation between data. Use historical data to train the selected model, and use the validation set to evaluate and tune the model to ensure the generalization ability and prediction accuracy of the model. During the evaluation process, some indicators such as root mean square error (RMSE), mean absolute error (MAE), etc. can be used to evaluate the performance of the model. After the model training and evaluation are completed, the model can be deployed to the actual system for application. In actual applications, real-time data can be used to update and adjust the model to adapt to changes in system status. Summary of the invention

[0005] In view of the above problems, the present invention proposes a combined prediction method for household-to-transformer relationship prediction of household photovoltaic users in low-voltage distribution networks based on EEMD decomposition and Newton-Raphson optimization algorithm (NRBO) optimization of DRSN-BIGRU-ATTENTION network model. The method is suitable for household-to-transformer relationship prediction tasks of household photovoltaic users in low-voltage distribution networks. The method fully considers the characteristics of low-voltage distribution networks with complex topological structures and dynamically changing load characteristics, and can achieve good results in practical applications. The method solves the problems of basis function selection and mode overlap in EMD decomposition, and provides better signal-to-noise ratio and decomposition accuracy. The deep residual shrinkage network DRSN is combined with BIGRU-ATTENTION to form a DRSN-BIGRU-ATTENTION network model. The combined DRSN-BIGRU-ATTENTION network model can improve the accuracy and reliability of household photovoltaic user household-to-transformer relationship prediction. Finally, the Newton-Raphson optimization algorithm NRBO is used to find the optimal parameters of the DRSN-BIGRU-ATTENTION network model. This method combines multiple independent forecasting models to obtain a more accurate and reliable overall forecast by integrating their forecast results.

[0006] The technical solution adopted by the present invention is:

[0007] The photovoltaic household-to-substation relationship prediction method based on EEMD decomposition and NRBO optimization includes the following steps:

[0008] Step 1: Perform EEMD decomposition on the original dataset of household change relations;

[0009] Step 2: Build a DRSN-BIGRU-ATTENTION network model and use it to perform multi-user data prediction on the household change relationship data after EEMD decomposition;

[0010] Step 3: Use the Newton-Raphson optimization algorithm NRBO to find the optimal parameters of the DRSN-BIGRU-ATTENTION network model;

[0011] Step 4: Output the predicted value of household change relationship data.

[0012] The step 1 comprises the following steps:

[0013] Step 1.1: Clean the data and remove or repair outliers;

[0014] Step 1.2: Perform EEMD decomposition on the household-to-household relationship of household photovoltaic power generation in low-voltage distribution network. The EEMD decomposition steps are as follows: S1.2.1: Add white noise of equal length to the original sequence X(t) to generate a new signal X i (t)

[0015] X i (t) = X (t) + n i (t) (1)

[0016] In formula (1), X i (t) is the new signal after adding white noise of equal length in the ith trial; n i (t) represents the white noise sequence added for the i-th time;

[0017] S1.2.2: Use EMD to analyze the new signal X i (t) is decomposed to obtain a new sequence C ij (t) and R i (t);

[0018]

[0019] In formula (2), C ij (t) and R i (t) are respectively X i (t) The jth modal component and residual component after the i-th noise decomposition, which is the average trend of the signal; m is the number of IMFs after signal decomposition.

[0020] S1.2.3: Repeat steps S1.2.1 and S1.2.2 to obtain the N group weights and residuals, and calculate their arithmetic mean, which is the decomposition result based on EEMD. The final result obtained by EEMD decomposition is as follows:

[0021]

[0022] In the above formula, C j (t) is the jth IMF component after EEMD decomposition; R(t) is the final residual; N is the number of repetitions of S1.2.1 and S1.2.2.

[0023] In the step 2, the DRSN-BIGRU-ATTENTION network model includes: a deep residual shrinkage network DRSN, a bidirectional gated recurrent unit BIGRU, and an attention mechanism ATTENTION;

[0024] S2.1: The Deep Residual Shrinkage Network (DRSN) is built with multiple layers of one-dimensional convolution and fully connected layers, including the following:

[0025] S2.1.1: One-dimensional convolution,In one-dimensional convolution in RSN, the convolution kernel slides along the time dimension of the time series data; Figure 5 As shown:

[0026] S2.1.2: DRSN-CW sets a threshold for each channel, adds the absolute value to the input feature map x, and then performs global average pooling (GAP) on the obtained |x| to reduce the dimension into a one-dimensional vector;

[0027]

[0028] Where: H is the feature map height, W is the feature map width, and k represents the kth channel;

[0029] S2.1.3: Input the vector obtained in S2.1.2 into the fully connected layer, and then input it into the second fully connected layer for batch normalization and ReLU activation; the number of neuron sources output is the number of channels in the feature map x, and then normalized using the sigmoid function. The normalized number is between 0 and 1, which is recorded as the coefficient α;

[0030] ReLU(x)=max(0,x) (6)

[0031]

[0032] S2.1.4: The weight of a specific channel is multiplied by the average value of the specific channel of the input feature map to obtain the threshold of each channel. The final threshold is recorded as α×|x GAP ;

[0033] S2.1.5: DRSN uses a small sub-network to learn a set of thresholds, soft-thresholds each channel of the feature map, sets redundant features to zero, and makes important features have non-zero outputs; then performs identity mapping to obtain a feature sequence;

[0034] S2.2: Input the feature sequence obtained in S2.1 into the bidirectional gated recurrent unit BIGRU, and divide the feature sequence into positive and negative direction inputs, respectively. and reverse sequence The output of the bidirectional gated recurrent unit BIGRU is represented as h t , including the following steps:

[0035] S2.2.1: After the input data enters the GRU network unit, it is processed by the update gate and the reset gate;

[0036] When the input data x t When entering the GRU network unit, they pass through their own weights W z Multiply them together for processing; at the same time, the hidden state h t-1 It also goes through the same process, it retains the information of the previous t-1 units and combines it with its own weight U z Multiply.

[0037] Add the results obtained above, and then apply the sigmoid activation function to compress the result to between 0 and 1. The formula is as follows:

[0038] z t =sigmoid(W z x t +U z h t-1 ) (8)

[0039] In formula (8), z t For updating the door.

[0040] Reset the gate input and pass the input data x t The weight W corresponding to the hidden state r and U r Multiply, add the results and apply the sigmoid activation function; the formula is as follows:

[0041] r t =sigmoid(W r x t +U r h t-1 ) (9)

[0042] In formula (9), r t is the reset gate output; W r is the weight matrix corresponding to the input data; U r is the weight matrix corresponding to the hidden state.

[0043] S2.2.2: The new candidate state is obtained by the following formula:

[0044] Input data x t Multiply by weight W, hidden state h t-1 Multiply by weight U; calculate reset gate r t With Uh t-1 The product of , from which we can determine what information needs to be retained in the previous state.

[0045] Add the above results and apply the tanh activation function to get

[0046]

[0047] In formula (10), is the candidate state obtained.

[0048] S2.2.3: Update the gate z t Data and h′ t Perform element-by-element multiplication on (1-z t ) and obtained Perform element-by-element multiplication and add the results.

[0049]

[0050] S2.2.4: The data is passed through the forward GRU and the reverse GRU to obtain the combination of the hidden states in the two directions, and the final output h of the bidirectional gated recurrent unit BIGRU is obtained. t .

[0051] S2.3: Due to the large number of temporal features, the attention mechanism ATTENTION selects key features from many features by calculating the probability distribution of attention. The specific calculation includes the following steps:

[0052] S2.3.1: First, the input feature h t With the corresponding weight W h Multiply them together and then add the bias b of the attention mechanism h , calculate the correlation vector s between the output and input features through the tanh function t , which is shown in formula (12):

[0053] s t =tanh(h t ·W h +b h ) (12)

[0054] S2.3.2: The output of formula (12) is normalized and the number of attention points is obtained through the Softmax function. The expression is shown in formula (13):

[0055] a t =Softmax(exp(s t )|∑exp(s t )) (13)

[0056] In formula (13), a t is the normalized result; exp(s t ) is the exponential form of the correlation vector obtained in formula (12); ∑exp(s t ) is the pair exp(s t ) and sum.

[0057] S2.3.3: Final a t With h t Multiply and weighted add to get the final output attention value vector, as shown in formula (14):

[0058] s=∑a t h t (14)

[0059] In formula (14), s is the final attention vector value.

[0060] In the step 3, the Newton-Raphson optimization algorithm NRBO is used to find the optimal parameters of the DRSN-BIGRU-ATTENTION network model;

[0061] S3.1: Population initialization, NRBO starts the search for the optimal solution by generating an initial random population within the boundary of candidate solutions. Given that there are N populations, each population covers fuzzy decision variables or vectors. These variables or vectors constitute the initial set of search space, laying the foundation for further optimization exploration. The following formula is used to generate a random population:

[0062]

[0063] in, represents the position of the jth dimension of the n population; ub represents x j Upper bound; lb represents x j lower bound; rand represents a random number (0-1); n is the population; N is the total population size; dim is the dimension.

[0064] The following formula represents the population matrix of all dimensions:

[0065]

[0066] In formula (16), X n is the population matrix; is the position of the first dimension of the first population; is the position of the second dimension of the first population; is the position of the dim-th dimension of the first population; is the position of the first dimension of the second population; is the position of the second dimension of the second population; is the position of the dim-th dimension of the second population; is the position of the first dimension of the Nth population; is the position of the dim-th dimension of the N-th population; is the position of the dim-th dimension of the N-th population.

[0067] S3.2: Use NRSR (Newton-Raphson) search rule to promote exploration trend and enable convergence. NRSR is based on the concept of NRM, which starts from an assumed initial solution and moves to the next position in a certain direction. Adjacent positions X are represented by x n +Δx and x n -Δx. In this case, NRBO will be at position X b Replace with position x n -Δx, because X b In its neighborhood, it has a larger position than x n A better position, while the position x n +Δx is replaced by position X w , because X w In its neighborhood, it has a larger position than x n Another advantage of this approach is that it uses the position x n Instead of its fitness f(x n ). The formula of NRSR is as follows:

[0068]

[0069] In formula (17), randn is a normally distributed random number with a mean of 0 and a variance of 1, Δx is a relative position variable; x n The location of population n; X w represents the worst position; X b represents the optimal position; the equation can be used to enhance the current solution by using the current solution to assist in updating its position. The expression of Δx is as follows:

[0070]

[0071] In formula (18), For X n The matrix obtained after T iterations; X b represents the optimal solution, and rand(1, dim) is a random number of decision variables with dim dimension. And an adaptive coefficient can be used to enhance the algorithm. The formula of this adaptive coefficient is as follows:

[0072]

[0073] In formula (19), T represents the current number of iterations, and MaxT represents the maximum number of iterations. This coefficient will be automatically adjusted during the iteration process, so that the exploration phase and the development phase can be balanced. The value range of this coefficient is (-1, 1). The updated position is obtained through NRSR, and the updated position is expressed as follows:

[0074] x n+1 =x n -NRSR (20)

[0075] And, in order to further improve the proposed NRBO method, another parameter is introduced, called ρ, which serves to guide the population in the right direction. The expression of parameter ρ is as follows:

[0076]

[0077] In formula (21), a and b are random numbers between (0-1); r1 and r2 are different integers randomly selected from the population. for The matrix obtained after T iterations etc.; for The matrix obtained after T iterations etc.

[0078] The NRSR formula obtained by improving NRM is as follows:

[0079]

[0080] In formula (22), randn is a normally distributed random number with a mean of 0 and a variance of 1; Δx is the relative position variable; x n The location of population n.

[0081] y w and b Z n+1 and x n Get the positions of the two vectors, the formula is as follows:

[0082]

[0083] In formula (23), Mean(Z n+1+x n ) to obtain Z n+1 +x n The mean of ; r1 is a random number between (0-1).

[0084] Z n+1 The expression is as follows:

[0085]

[0086] Vector current position It can be expressed by the following formula:

[0087]

[0088] In formula (25), To update The new vector obtained.

[0089] Substitute X by the result obtained in the above formula b Construct the latest new vector

[0090]

[0091] For local search limitations, NRBO uses the above formula to improve both the diversification and reinforcement phases. The new position vector during the next iteration is expressed as:

[0092]

[0093] In formula (27), is the new position vector; To update The new vector obtained is: To update The new vector obtained is: To update and The new vector is obtained; δ is the adaptive coefficient of the enhancement algorithm; The matrix for the nth population iteration T times.

[0094] Trap Avoidance Operation (TAO) can improve the effectiveness of NRBO in dealing with practical problems and avoid local optimality. b and the current vector position To be enhanced If the value of rand is less than DF, the following formula is used to generate

[0095]

[0096] In formula (28), is the new position vector; is the matrix of the nth population iteration T times; x b The best position for the population; Seek The mean of is the enhanced vector position.

[0097] Where: θ1 and θ2 are consistent random numbers between (-1, 1) and (-0.5, 0.5), respectively. DF represents the determining factor that controls the performance of NRBO. μ1 and μ2 are random numbers generated by the following formula:

[0098]

[0099] In formula (29), β represents 1 or 0. If the value of △ is greater than or equal to 0.5, β is equal to 0, otherwise it is 1.

[0100] The optimal parameters of the neural network in step 3 are found by the Newton-Raphson optimization algorithm, including the following steps:

[0101] S4.1: Population initialization, setting the parameters of this optimization algorithm, population size N, maximum number of iterations MaxT and DF.

[0102] S4.2: Specify the worst solution X w and the optimal solution X b , set the adaptive coefficient, and when n=1:N, randomly select r1 and r2 in (1, N).

[0103] S4.3: X obtained in step 4.2 w and X b Substitute the above formula to calculate NRSR and ρ.

[0104] S4.4: Determine and , and then determine the The position is updated by formula (27).

[0105] S4.5: By judging whether the value of rand is less than DF, the formula (27) is used to determine Update the worst solution X through the position of TAO w and the optimal solution X b location.

[0106] S4.6: Update the position through formula (28) to obtain the optimal solution X b .

[0107] The present invention provides a photovoltaic household transformer relationship prediction method based on EEMD decomposition and NRBO optimization, and the technical effects are as follows:

[0108] 1) In order to solve the problem of insufficient signal feature extraction in the original method, the present invention performs EEMD decomposition on the original data set of household photovoltaic household transformer relationship in the low-voltage distribution gateway. EEMD decomposition can more effectively decompose the signal into multiple intrinsic mode functions IMFs, and at the same time solve the problem of insufficient feature extraction in the EMD method.

[0109] 2) The present invention uses the obtained IMF components as input and passes them to the combined model DRSN-BIGRU-ATTENTION network model to achieve prediction of multi-user data.

[0110] 3) The present invention can better capture the temporal characteristics and correlations between multi-user data by combining models such as the deep residual shrinkage network DRSN, the bidirectional gated recurrent unit BIGRU and the attention mechanism ATTENTION. The BIGRU structure can avoid gradient explosion and gradient disappearance, and has the advantage of strong long-term memory. The attention mechanism ATTENTION selects key features from the many features of household photovoltaics by calculating the probability distribution of attention, thereby improving the computational efficiency of the model.

[0111] 4) The DRSN-BIGRU-ATTENTION network model of the present invention, this combined model can more comprehensively consider the spatiotemporal relationship of multi-user data, and achieve accurate prediction of multi-user data through the deep learning ability of the deep residual shrinkage network DRSN and the sequence modeling ability of the bidirectional gated recurrent unit BIGRU, as well as the importance weighting ability of the attention mechanism ATTENTION.

[0112] 5) The present invention adopts the NRBO optimization algorithm to find the optimal parameters in the DRSN-BIGRU-ATTENTION network model. Through the NRBO optimization algorithm, the network parameters can be adjusted more effectively to improve the prediction accuracy and reliability of the model for multi-user data. Finally, according to the set error threshold, the multi-user prediction value of the household-to-household relationship of household photovoltaics is output. BRIEF DESCRIPTION OF THE DRAWINGS

[0113] The present invention will be further described below in conjunction with the accompanying drawings and embodiments:

[0114] Figure 1 It is a schematic diagram of the overall process framework of the present invention.

[0115] Figure 2 This is a flow chart of the EEMD decomposition steps in the present invention.

[0116] Figure 3 Flow chart of the BIGRU network of the present invention.

[0117] Figure 4 Flowchart of the optimization algorithm for NRBO.

[0118] Figure 5 Schematic diagram of the convolution kernel sliding along the time dimension of time series data in the one-dimensional convolution in RSN.

[0119] Figure 6 This is the flow chart of S2.1.3.

[0120] Figure 7 This is the flow chart of S2.1.4.

[0121] Figure 8 This is the flow chart of S2.1.5.

[0122] Fig. 9 Schematic diagram of the bidirectional gated recurrent unit BIGRU structure.

[0123] Fig.10 This is a schematic diagram of the predicted user area.

[0124] Fig.11 This is the original photovoltaic voltage data diagram.

[0125] Fig.12 EEMD decomposition diagram of the original photovoltaic voltage data.

[0126] Fig.13 is the characteristic kurtosis plot.

[0127] Fig.14 is the feature mean graph.

[0128] Fig.15 It is the characteristic standard deviation graph.

[0129] Fig.16 Optimization parameters map for NRBO. DETAILED DESCRIPTION

[0130] The photovoltaic household-to-substation relationship prediction method based on EEMD decomposition and NRBO optimization includes the following steps:

[0131] Step 1: Perform EEMD decomposition on the original dataset of household change relations;

[0132] Step 2: Build a DRSN-BIGRU-ATTENTION network model and use it to perform multi-user data prediction on the household change relationship data after EEMD decomposition;

[0133] Step 3: Use the Newton-Raphson optimization algorithm NRBO to find the optimal parameters of the DRSN-BIGRU-ATTENTION network model;

[0134] Step 4: Output the predicted value of household change relationship data.

[0135] The step 1 comprises the following steps:

[0136] Step 1.1: Clean the data to ensure that there are no outliers in the data. If there are outliers, delete them or use interpolation methods to repair them.

[0137] Step 1.2: Perform EEMD decomposition on the household-to-household relationship of household photovoltaics in the low-voltage distribution network. The essence of the EEMD method is a multiple empirical mode decomposition of superimposed Gaussian white noise. Then, the corresponding IMF obtained by multiple EMDs is averaged to offset the added white noise, thereby effectively suppressing the generation of modal aliasing.

[0138] The empirical mode decomposition (EMD) method was proposed by Huang. It is a new time-frequency analysis method and an adaptive time-frequency localization analysis method. The EEMD decomposition steps are as follows:

[0139] S1.2.1: Add white noise of equal length to the original sequence X(t) to generate a new signal X i (t)

[0140] X i (t) = X (t) + n i (t) (1)

[0141] In formula (1), X i (t) is the new signal after adding white noise of equal length in the ith trial; n i (t) represents the white noise sequence added for the i-th time;

[0142] S1.2.2: Use EMD to analyze the new signal X i (t) is decomposed to obtain a new sequence C ij (t) and R i (t);

[0143]

[0144] In formula (2), C ij (t) and R i (t) are respectively X i (t) The jth modal component and residual component after the i-th noise decomposition, which is the average trend of the signal; m is the number of IMFs after signal decomposition.

[0145] S1.2.3: Repeat steps S1.2.1 and S1.2.2 to obtain the N group weights and residuals, and calculate their arithmetic mean, which is the decomposition result based on EEMD. The final result obtained by EEMD decomposition is as follows:

[0146]

[0147] In the above formula, C j (t) is the jth IMF component after EEMD decomposition; R(t) is the final residual; N is the number of repetitions of S1.2.1 and S1.2.2.

[0148] In the step 2, the DRSN-BIGRU-ATTENTION network model includes: a deep residual shrinkage network DRSN, a bidirectional gated recurrent unit BIGRU, and an attention mechanism ATTENTION;

[0149] S2.1: The deep residual shrinkage network DRSN is built by multiple layers of one-dimensional convolution and fully connected layers, including the following steps:

[0150] S2.1.1: One-dimensional convolution,In one-dimensional convolution in RSN, the convolution kernel slides along the time dimension of the time series data; Figure 5 As shown:

[0151] S2.1.2: DRSN-CW sets a threshold for each channel, adds the absolute value to the input feature map x, and then performs global average pooling (GAP) on the obtained |x| to reduce the dimension into a one-dimensional vector;

[0152]

[0153] Where: H is the feature map height, W is the feature map width, and k represents the kth channel;

[0154] S2.1.3: Input the vector obtained in S2.1.2 into the fully connected layer, and then input it into the second fully connected layer for batch normalization and ReLU activation. The number of output neuron sources is the number of channels in the feature map x, and then normalized using the sigmoid function. The normalized number is between 0 and 1, which is recorded as the coefficient α; Figure 6 As shown:

[0155] ReLU(x)=max(0,x) (6)

[0156]

[0157] S2.1.4: The weight of a specific channel is multiplied by the average value of the specific channel of the input feature map to obtain the threshold of each channel. The final threshold is recorded as α×|x GAP ;like Figure 7shown.

[0158] S2.1.5: DRSN uses a small subnetwork to learn a set of thresholds, soft-thresholds each channel of the feature map, sets redundant features to zero, and makes important features have non-zero outputs. Then, an identity mapping is performed to obtain a feature sequence, such as Figure 8 shown.

[0159] The bidirectional gated recurrent unit BIGRU is composed of two layers of GRU network units stacked together. GRU is composed of multiple gated recurrent units connected together. Each unit includes a reset gate and an update gate. The reset gate is used to control whether to ignore past memories. It determines which parts of past memories will be discarded in order to better adapt to current inputs. The update gate is used to control whether new information is integrated into the current memory. Its value is between 0 and 1, which determines the weight of old and new information in the memory unit. The two gated units work together to enable GRU to effectively process sequence data while alleviating the problem of gradient disappearance. Its structural diagram is shown below: Fig. 9 As shown:

[0160] x t : Input information at the current moment;

[0161] h t-1 : The hidden state at the previous moment. The hidden state acts as the neural network memory, which contains information about the data that the node has seen before.

[0162] h t : The hidden state passed to the next moment.

[0163] Candidate hidden states.

[0164] r t : Reset the door.

[0165] z t : Update gate.

[0166] σ: sigmoid function, through this function the data can be converted into a value in the range of 0-1.

[0167] tanh: tanh function, through this function, the data can be converted into a value in the range of [-1,1]

[0168] S2.2: Input the feature sequence obtained in S2.1 into the bidirectional gated recurrent unit BIGRU, and divide the feature sequence into positive and negative direction inputs, respectively. and reverse sequence The output of the bidirectional gated recurrent unit BIGRU can be expressed as h t , including the following steps:

[0169] S2.2.1: After the input data enters the GRU network unit, it is processed by the update gate and the reset gate;

[0170] When the input data x t When entering the GRU network unit, they pass through their own weights W z Multiply them together for processing; at the same time, the hidden state h t-1 It also goes through the same process, it retains the information of the previous t-1 units and combines it with its own weight U z Multiply.

[0171] Add the results obtained above, and then apply the sigmoid activation function to compress the result to between 0 and 1. The formula is as follows:

[0172] z t =sigmoid(W z x t +U z h t-1 ) (8)

[0173] In formula (8), z t For updating the door.

[0174] Reset the gate input and pass the input data x t The weight W corresponding to the hidden state r and U r Multiply, add the results and apply the sigmoid activation function; the formula is as follows:

[0175] r t =sigmoid(W r x t +U r h t-1 ) (9)

[0176] In formula (9), r t is the reset gate output; W r is the weight matrix corresponding to the input data; U r is the weight matrix corresponding to the hidden state.

[0177] S2.2.2: The new candidate state is obtained by the following formula:

[0178] Input data x t Multiply by weight W, hidden state h t-1 Multiply by weight U; calculate reset gate r t With Uh t-1 The product of , from which we can determine what information needs to be retained in the previous state.

[0179] Add the above results and apply the tanh activation function to get

[0180]

[0181] In formula (10), is the candidate state obtained.

[0182] S2.2.3: Update the gate z t Data and h′ t Perform element-by-element multiplication on (1-z t ) and obtained Perform element-by-element multiplication and add the results.

[0183]

[0184] S2.2.4: The data is passed through the forward GRU and the reverse GRU to obtain the combination of the hidden states in the two directions, and the final output h of the bidirectional gated recurrent unit BIGRU is obtained. t .

[0185] S2.3: Due to the large number of temporal features, the attention mechanism ATTENTION selects key features from many features by calculating the probability distribution of attention, thereby improving the computational efficiency of the model. The specific calculation includes the following steps:

[0186] S2.3.1: First, the input feature h t With the corresponding weight W h Multiply them together and then add the bias b of the attention mechanism h , calculate the correlation vector s between the output and input features through the tanh function t , which is shown in formula (12):

[0187] s t =tanh(h t ·W h +b h ) (12)

[0188] S2.3.2: The output of formula (12) is normalized and the number of attention points is obtained through the Softmax function. The expression is shown in formula (13):

[0189] a t =Softmax(exp(s t )|∑exp(s t )) (13)

[0190] In formula (13), a t is the normalized result; exp(s t ) is the exponential form of the correlation vector obtained in formula (12); ∑exp(s t ) is the pair exp(s t ) and sum.

[0191] S2.3.3: Final a t With h t Multiply and weighted add to get the final output attention value vector, as shown in formula (14):

[0192] s=∑a t h t (14)

[0193] In formula (14), s is the final attention vector value.

[0194] In the step 3, the Newton-Raphson optimization algorithm NRBO is used to find the optimal parameters of the DRSN-BIGRU-ATTENTION network model;

[0195] The search space can be explored using several sets of vectors and two operators: NRSR and TAO. NRSR is based on the concept of NRM, and uses NRM (Newton-Raphson Method) to identify the search area and define the search path. NRM can guide the search by analyzing the gradient and Hessian matrix of the function to quickly and accurately determine the optimization direction. The NRSR operator can improve the accuracy and convergence speed of the search; while the TAO operator helps avoid local optimal points. The combined application of these vector sets and operators can guide the optimizer to move towards the global optimal solution during the search process, thereby exploring the search space more effectively.

[0196] S3.1: Population initialization, NRBO starts the search for the optimal solution by generating an initial random population within the boundary of candidate solutions. Given that there are N populations, each population covers fuzzy decision variables or vectors. These variables or vectors constitute the initial set of search space, laying the foundation for further optimization exploration. The following formula is used to generate a random population:

[0197]

[0198] in, represents the position of the jth dimension of the n population; ub represents x j Upper bound; lb represents x j lower bound; rand represents a random number (0-1); n is the population; N is the total population size; dim is the dimension.

[0199] The following formula represents the population matrix of all dimensions:

[0200]

[0201] In formula (16), X n is the population matrix; is the position of the first dimension of the first population; is the position of the second dimension of the first population; is the position of the dim-th dimension of the first population; is the position of the first dimension of the second population; is the position of the second dimension of the second population; is the position of the dim-th dimension of the second population; is the position of the first dimension of the Nth population; is the position of the dim-th dimension of the N-th population; is the position of the dim-th dimension of the N-th population.

[0202] S3.2: Use NRSR (Newton-Raphson) search rule to promote exploration trend and enable convergence. NRSR is based on the concept of NRM, which starts from an assumed initial solution and moves to the next position in a certain direction. Adjacent positions X are represented by x n +Δx and x n -Δx. In this case, NRBO will be at position X b Replace with position x n -Δx, because X b In its neighborhood, it has a larger position than x n A better position, while the position x n +Δx is replaced by position X w , because X w In its neighborhood, it has a larger position than x n Another advantage of this approach is that it uses the position x n Instead of its fitness f(x n ). The formula of NRSR is as follows:

[0203]

[0204] In formula (17), randn is a normally distributed random number with a mean of 0 and a variance of 1, Δx is a relative position variable; x n The location of population n; X w represents the worst position; X brepresents the optimal position; the equation can be used to enhance the current solution by using the current solution to assist in updating its position. The expression of Δx is as follows:

[0205]

[0206] In formula (18), For X n The matrix obtained after T iterations; X b represents the optimal solution, and rand(1, dim) is a random number of decision variables with dim dimension. And an adaptive coefficient can be used to enhance the algorithm. The formula of this adaptive coefficient is as follows:

[0207]

[0208] In formula (19), T represents the current number of iterations, and MaxT represents the maximum number of iterations. This coefficient will be automatically adjusted during the iteration process, so that the exploration phase and the development phase can be balanced. The value range of this coefficient is (-1, 1). The updated position is obtained through NRSR, and the updated position is expressed as follows:

[0209] x n+1 =x n -NRSR (20)

[0210] And, in order to further improve the proposed NRBO method, another parameter is introduced, called ρ, which serves to guide the population in the right direction. The expression of parameter ρ is as follows:

[0211]

[0212] In formula (21), a and b are random numbers between (0-1); r1 and r2 are different integers randomly selected from the population. for The matrix obtained after T iterations etc.; For X r2 The matrix obtained after T iterations etc.

[0213] The NRSR formula obtained by improving NRM is as follows:

[0214]

[0215] In formula (22), randn is a normally distributed random number with a mean of 0 and a variance of 1; Δx is the relative position variable; x n The location of population n.

[0216] y w and b Zn+1 and x n Get the positions of the two vectors, the formula is as follows:

[0217]

[0218] In formula (23), Mean(Z n+1 +x n ) to obtain Z n+1 +x n The mean of ; r1 is a random number between (0-1).

[0219] Z n+1 The expression is as follows:

[0220]

[0221] Vector current position It can be expressed by the following formula:

[0222]

[0223] In formula (25), To update The new vector obtained.

[0224] Substitute X by the result obtained in the above formula b Construct the latest new vector

[0225]

[0226] For local search limitations, NRBO uses the above formula to improve both the diversification and reinforcement phases. The new position vector during the next iteration is expressed as:

[0227]

[0228] In formula (27), is the new position vector; To update The new vector obtained is: To update The new vector obtained is: To update and The new vector is obtained; δ is the adaptive coefficient of the enhancement algorithm; The matrix for the nth population iteration T times.

[0229] Trap Avoidance Operation (TAO) improves the effectiveness of NRBO in dealing with practical problems and avoids local optimality. b and the current vector position To be enhanced If the value of rand is less than DF, the following formula is used to generate

[0230]

[0231] In formula (28), X n ′T+1 is the new position vector; X n ′T is the matrix of the nth population iteration T times; x b The best position for the population; Seek The mean of is the enhanced vector position.

[0232] Where: θ1 and θ2 are consistent random numbers between (-1, 1) and (-0.5, 0.5), respectively. DF represents the determining factor that controls the performance of NRBO. μ1 and μ2 are random numbers generated by the following formula:

[0233]

[0234] In formula (29), β represents 1 or 0. If the value of △ is greater than or equal to 0.5, β is equal to 0, otherwise it is 1.

[0235] The optimal parameters of the neural network in step 3 are found by the Newton-Raphson optimization algorithm, including the following steps:

[0236] S4.1: Population initialization, setting the parameters of this optimization algorithm, population size N, maximum number of iterations MaxT and DF.

[0237] S4.2: Specify the worst solution X w and the optimal solution X b , set the adaptive coefficient, and when n=1:N, randomly select r1 and r2 in (1, N).

[0238] S4.3: X obtained in step 4.2 w and X b Substitute the above formula to calculate NRSR and ρ.

[0239] S4.4: Determine and , and then determine the The position is updated by formula (27).

[0240] S4.5: By judging whether the value of rand is less than DF, the formula (27) is used to determine Update the worst solution X through the position of TAOw and the optimal solution X b location.

[0241] S4.6: Update the position through formula (28) to obtain the optimal solution X b .

[0242] Verification example:

[0243] 1. Data preprocessing and EEMD decomposition:

[0244] Data preprocessing: Clean the data and remove outliers and missing values ​​from the improved 24-hour voltage data of 150 household photovoltaic users.

[0245] EEMD decomposition: The ensemble empirical mode decomposition (EEMD) technique is used to decompose the non-stationary PV voltage data into a series of relatively stable intrinsic mode functions (IMFs) and a residual term. This helps to reduce the non-stationarity of the data and improve the prediction accuracy of the subsequent model. Each data is decomposed into five IMFs. The original data is as follows Fig.11 The decomposition results are shown in Fig.12 shown.

[0246] Loop through each column of data and perform EEMD decomposition on it. Generate IMF results by calling the custom eemd function. Set the number of decomposed IMFs to 5. IMF is the signal component of different frequency bands obtained by EEMD decomposition.

[0247] 2. Feature extraction and NRBO optimization:

[0248] Feature construction: Extract the characteristic voltage mean, standard deviation, kurtosis and other statistical features from the IMF and residual terms after EEMD decomposition. Here, the size of the data is first determined, and the feature matrix is ​​constructed to store the features of each sample, including the mean, standard deviation and kurtosis. The matrix size is 750 rows and 72 columns (24×3). In the loop, the mean, standard deviation and kurtosis of each data are calculated in turn and stored in the feature matrix. Fig.13 is the characteristic kurtosis plot, Fig.14 is the feature mean graph. Fig.15 It is the characteristic standard deviation graph.

[0249] NRBO optimization: The Newton-Raphson optimization algorithm (NRBO) is used to further optimize the voltage feature set after feature selection, and the optimal feature combination or weight is found through iterative search to maximize the recognition performance of the model. Fig.16 Optimization parameters map for NRBO.

[0250] 3. Model construction:

[0251] A composite model is constructed by combining the Deep Residual Shrinkage Network (DRSN), Bidirectional Gated Recurrent Unit (BIGRU) and Attention Mechanism (ATTENTION). DRSN is used to process deep features and suppress noise, BIGRU is used to capture bidirectional dependencies in time series data, and ATTENTION mechanism is used to highlight important features.

[0252] Input the obtained feature data into the DRSN-BIGRU-ATTENTION code, classify it, and output the result.

[0253] The input layer of DRSN receives the input feature matrix, multiple convolutional layers extract feature data, and the residual layer adds residuals between convolutional layers to prevent gradient disappearance.

[0254] The BIGRU layer processes data in the time dimension. Each feature data is decomposed from 24 hours of voltage data. The bidirectional GRU layer captures both forward and backward sequence information to enhance the information capture capability.

[0255] Apply the attention mechanism (ATTENTION), calculate the importance weight of each feature, and combine it with the output of GRU to increase the model's attention to important features. The output of GRU is weighted summed by the attention weight to obtain the final representation vector. The output of the attention layer is processed through one or more fully connected layers. Classify the data into specified categories.

[0256] Those with similar characteristics are grouped together. Based on the similar voltage characteristics of the unified substation, the original multiple data are allocated to each group to realize the subordinate relationship between users and substations, that is, the household-to-substation relationship.

[0257] 4. Model verification and evaluation:

[0258] The data set is divided into training set, validation set and test set. The training set is the first 100 users, the validation set is the next 30 users, and the last 20 users are the test set.

[0259] 5. Use the model to predict which stations the user belongs to based on the data. Fig.10 The same color means the same area.

Claims

1. The photovoltaic household-to-substation relationship prediction method based on EEMD decomposition and NRBO optimization is characterized by The following steps are involved: Step 1: Perform EEMD decomposition on the original dataset of household change relations; Step 2: Build a DRSN-BIGRU-ATTENTION network model and use it to perform multi-user data prediction on the household change relationship data after EEMD decomposition; Step 3: Use the Newton-Raphson optimization algorithm NRBO to find the optimal parameters of the DRSN-BIGRU-ATTENTION network model; Step 4: Output the predicted value of household change relationship data.

2. The photovoltaic household transformer relationship prediction method based on EEMD decomposition and NRBO optimization according to claim 1 is characterized by: The step 1 comprises the following steps: Step 1.1: Clean the data and remove or repair outliers; Step 1.2: Perform EEMD decomposition on the household-to-household relationship of household photovoltaic power generation in the low-voltage distribution network. The EEMD decomposition steps are as follows: S1.2.1: Add white noise of equal length to the original sequence X(t) to generate a new signal X i (t) X i (t)=X(t)+n i (t) (1) In formula (1), X i (t) is the new signal after adding white noise of equal length in the ith trial; n i (t) represents the white noise sequence added for the i-th time; S1.2.2: Use EMD to analyze the new signal X i (t) is decomposed to obtain a new sequence C ij (t) and R i (t); In formula (2), C ij (t) and R i (t) are respectively X i (t) The jth modal component and residual component after the i-th noise decomposition, which is the average trend of the signal; m is the number of IMFs after signal decomposition; S1.2.3: Repeat steps S1.2.1 and S1.2.2 to obtain the N group weights and residuals, and calculate their arithmetic mean, which is the decomposition result based on EEMD. The final result obtained by EEMD decomposition is as follows: In the above formula, C j (t) is the jth IMF component after EEMD decomposition; R(t) is the final residual; N is the number of repetitions of S1.2.1 and S1.2.

2.

3. The photovoltaic household transformer relationship prediction method based on EEMD decomposition and NRBO optimization according to claim 1 is characterized by: In the step 2, the DRSN-BIGRU-ATTENTION network model includes: a deep residual shrinkage network DRSN, a bidirectional gated recurrent unit BIGRU, and an attention mechanism ATTENTION; S2.1: The Deep Residual Shrinkage Network (DRSN) is built with multiple layers of one-dimensional convolution and fully connected layers, including the following: S2.1.1: One-dimensional convolution. In the one-dimensional convolution in RSN, the convolution kernel slides along the time dimension of the time series data; as shown in Figure 5: S2.1.2: DRSN-CW sets a threshold for each channel, adds the absolute value to the input feature map x, and then performs global average pooling (GAP) on the obtained |x| to reduce the dimension into a one-dimensional vector; Where: H is the feature map height, W is the feature map width, and k represents the kth channel; S2.1.3: Input the vector obtained in S2.1.2 into the fully connected layer, and then input it into the second fully connected layer for batch normalization and ReLU activation; the number of neuron sources output is the number of channels in the feature map x, and then normalized using the sigmoid function. The normalized number is between 0 and 1, which is recorded as the coefficient α; ReLU(x)=max(0,x) (6) S2.1.4: The weight of a specific channel is multiplied by the average value of the specific channel of the input feature map to obtain the threshold of each channel. The final threshold is recorded as α×|x GAP ; S2.1.5: DRSN uses a small sub-network to learn a set of thresholds, soft-thresholds each channel of the feature map, sets redundant features to zero, and makes important features have non-zero outputs; then performs identity mapping to obtain a feature sequence; S2.2: Input the feature sequence obtained in S2.1 into the bidirectional gated recurrent unit BIGRU, and divide the feature sequence into positive and negative direction inputs, respectively. and reverse sequence The output of the bidirectional gated recurrent unit BIGRU is represented as h t , including the following steps: S2.2.1: After the input data enters the GRU network unit, it is processed by the update gate and the reset gate; When the input data x t When entering the GRU network unit, they pass through their own weights W z Multiply them together for processing; at the same time, the hidden state h t-1 It also goes through the same process, it retains the information of the previous t-1 units and combines it with its own weight U z multiply; Add the results obtained above, and then apply the sigmoid activation function to compress the result to between 0 and 1; the formula is as follows: z t =sigmoid(W z x t +U z h t-1 ) (8) In formula (8), z t To update the door; Reset the gate input and pass the input data x t The weight W corresponding to the hidden state r and U r Multiply, add the results and apply the sigmoid activation function; the formula is as follows: r t =sigmoid(W r x t +U r h t-1 ) (9) In formula (9), r t is the reset gate output; W r is the weight matrix corresponding to the input data; U r is the weight matrix corresponding to the hidden state; S2.2.2: The new candidate state is obtained by the following formula: Input data x t Multiply by weight W, hidden state h t-1 Multiply by weight U; calculate reset gate r t With Uh t-1 The product of , from which we can determine what information needs to be retained in the previous state; Add the above results and apply the tanh activation function to get In formula (10), is the candidate state obtained; S2.2.3: Update the gate z t Data and h′ t Perform element-by-element multiplication on (1-z t ) and obtained Perform element-by-element multiplication and add the results; S2.2.4: The data is passed through the forward GRU and the reverse GRU to obtain the combination of the hidden states in the two directions, and the final output h of the bidirectional gated recurrent unit BIGRU is obtained. t ; S2.3: Due to the large number of temporal features, the attention mechanism ATTENTION selects key features from many features by calculating the probability distribution of attention. The specific calculation includes the following steps: S2.3.1: First, the input feature h t With the corresponding weight W h Multiply them together and then add the bias b of the attention mechanism h , calculate the correlation vector s between the output and input features through the tanh function t , which is shown in formula (12): S t =tanh(h t ·W h +b h ) (12) S2.3.2: The output of formula (12) is normalized and the number of attention points is obtained through the Softmax function. The expression is shown in formula (13): a t =Softmax(exp(s t )|∑exp(s t )) (13) In formula (13), a t is the normalized result; exp(s t ) is the exponential form of the correlation vector obtained in formula (12); ∑exp(s t ) is the pair exp(s t )Sum; S2.3.3: Final a t With h t Multiply and weighted add to get the final output attention value vector, as shown in formula (14): s=∑a t h t (14) In formula (14), s is the final attention vector value.

4. The photovoltaic household transformer relationship prediction method based on EEMD decomposition and NRBO optimization according to claim 1 is characterized by: In the step 3, the Newton-Raphson optimization algorithm NRBO is used to find the optimal parameters of the DRSN-BIGRU-ATTENTION network model; S3.1: Population initialization, NRBO starts the search for the optimal solution by generating an initial random population within the boundary of candidate solutions; given that there are N populations, each population covers fuzzy decision variables or vectors; these variables or vectors constitute the initial set of search space, laying the foundation for further optimization exploration; the random population is generated using the following formula: in, represents the position of the jth dimension of the n population; ub represents x j Upper bound; lb represents x j The lower bound; rand represents a random number (0-1); n is the population; N is the total population size; dim is the dimension; The following formula represents the population matrix of all dimensions: In formula (16), X n is the population matrix; is the position of the first dimension of the first population; is the position of the second dimension of the first population; is the position of the dim-th dimension of the first population; is the position of the first dimension of the second population; is the position of the second dimension of the second population; is the position of the dim-th dimension of the second population; is the position of the first dimension of the Nth population; is the position of the dim-th dimension of the N-th population; is the position of the dim-th dimension of the N-th population; S3.2: Use NRSR (Newton-Raphson) search rule to promote exploration trend and enable convergence. NRSR is based on the concept of NRM. NRM starts from an assumed initial solution and moves to the next position along a certain direction. Adjacent positions X are represented by x n +Δx and x n -Δx means; in this case, NRBO will be at position X b Replace with position x n -Δx, because X b In its neighborhood, it has a larger position than x n A better position, while the position x n +Δx is replaced by position X w , because X w In its neighborhood, it has a larger position than x n worse position; another advantage of this method is that it uses the position x n Instead of its fitness f(x n );The formula of NRSR is as follows: In formula (17), randn is a normally distributed random number with a mean of 0 and a variance of 1, Δx is a relative position variable; x n The location of population n; X w represents the worst position; X b represents the optimal position; the equation can be used to enhance the current solution by using the current solution to assist in updating its position; the expression of Δx is as follows: In formula (18), For X n The matrix obtained after T iterations; X b represents the optimal solution, rand(1, dim) is a random number of a decision variable with dim dimension; and an adaptive coefficient can be used to enhance the algorithm; the formula of this adaptive coefficient is as follows: In formula (19), T represents the current number of iterations, and MaxT represents the maximum number of iterations. This coefficient will be automatically adjusted during the iteration process to ensure that the exploration phase and the development phase are balanced. The value range of this coefficient is (-1, 1). The updated position is obtained through NRSR, and the updated position is expressed as follows: x n+1 =x n -NRSR (20) And, in order to further improve the proposed NRBO method, another parameter is introduced, called ρ, which is used to guide the population in the right direction; the expression of parameter ρ is as follows: In formula (21), a and b are random numbers between (0-1); r1 and r2 are different integers randomly selected from the population; for The matrix obtained after T iterations etc.; For X r2 The matrix obtained after T iterations etc.; The NRSR formula obtained by improving NRM is as follows: In formula (22), randn is a normally distributed random number with a mean of 0 and a variance of 1; Δx is the relative position variable; x n The location of population n; y w and b Z n+1 and x n Get the positions of the two vectors, the formula is as follows: In formula (23), Mean(Z n+1 +x n ) to obtain Z n+1 +x n The mean of; r1 is a random number (0-1); Z n+1 The expression is as follows: Vector current position It can be expressed by the following formula: In formula (25), To update The new vector obtained is: Substitute X by the result obtained in the above formula b Construct the latest new vector As local search is limited, NRBO uses the above formula to improve both the diversification and reinforcement phases; the new position vector during the next iteration is expressed as: In formula (27), is the new position vector; To update The new vector obtained is: To update The new vector obtained is: To update and The new vector is obtained; δ is the adaptive coefficient of the enhancement algorithm; is the matrix of the nth population iteration T times; Trap avoidance operation (TAO) can improve the effectiveness of NRBO in dealing with practical problems and avoid local optimality; by combining the best position X b and the current vector position To be enhanced If the value of rand is less than DF, the following formula is used to generate In formula (28), is the new position vector; is the matrix of the nth population iteration T times; x b The best position for the population; Seek The mean of is the enhanced vector position; Where: θ1 and θ2 are consistent random numbers between (-1, 1) and (-0.5, 0.5), respectively; DF represents the determining factor that controls the performance of NRBO; μ1 and μ2 are random numbers generated by the following formula: In formula (29), β represents 1 or 0. If the value of △ is greater than or equal to 0.5, β is equal to 0, otherwise it is 1.

5. The photovoltaic household transformer relationship prediction method based on EEMD decomposition and NRBO optimization according to claim 4 is characterized by: The optimal parameters of the neural network in step 3 are found by the Newton-Raphson optimization algorithm, including the following steps: S4.1: Population initialization, setting the parameters of this optimization algorithm, population size N, maximum number of iterations MaxT and DF; S4.2: Specify the worst solution X w and the optimal solution X b , set the adaptive coefficient, when n = 1:N, randomly select r1 and r2 in (1, N); S4.3: X obtained in step 4.2 w and X b Substitute the above formula to calculate NRSR and ρ; S4.4: Determine and , and then determine the The position is updated by formula (27); S4.5: By judging whether the value of rand is less than DF, the formula (27) is used to determine Update the worst solution X through the position of TAO w and the optimal solution X b location; S4.6: Update the position through formula (28) to obtain the optimal solution X b .

Images