Active power virtual acquisition method of distributed photovoltaic power station based on IDBO-KELM-BiGRU neural network

Abstract

The present invention provides a distributed photovoltaic power station active power virtual collection method based on IDBO‑KELM‑BiGRU neural network, which belongs to the field of distributed photovoltaic operation and maintenance data collection technology. It includes: preliminary grid division of virtual collection area according to geographical location; using random matrix theory and dynamic time warping DTW clustering algorithm correlation analysis method to select similar power stations; establishing IDBO‑KELM‑BiGRU neural network, and performing self-supervised training to obtain current‑power data conversion model; converting the historical current data of the distributed photovoltaic power station to be collected into the historical photovoltaic output power prediction data of the power station to be collected, and further training the IDBO‑KELM‑BiGRU neural network model to obtain the active power data virtual collection model; inputting the real-time current data of the distributed photovoltaic power station to be collected on the collection day, using the active power data virtual collection model to perform data fitting, and obtaining the photovoltaic output power prediction data of the distributed photovoltaic power station to be collected on the collection day in real time.

Description

Technical Field

[0001] The present invention relates to the technical field of distributed photovoltaic operation and maintenance data collection, and in particular to a distributed photovoltaic power station active power virtual collection method based on IDBO-KELM-BiGRU neural network. Background Art

[0002] Distributed photovoltaic systems have developed rapidly due to their flexible installation, outstanding environmental benefits, and coexistence of power generation and consumption. In 2023, the newly installed capacity of distributed photovoltaics will far exceed that of centralized photovoltaics. However, distributed photovoltaic output is highly intermittent and random. The photovoltaic output characteristics will not only be affected by environmental conditions such as light, but also by other related factors such as season, geographical region, time, etc., and will also form power load constraints with the connected adjacent photovoltaics. Large-scale decentralized and disorderly access to the distribution network, simply by adding collection, transmission and control equipment, will not only greatly increase the construction cost of the power grid and users, but also greatly increase the workload of subsequent operation and maintenance management. How to achieve "real-time + virtual" collection of distributed photovoltaic operation data in the entire domain by installing limited distributed photovoltaic collection equipment, and achieve high-precision acquisition of distributed photovoltaic operation data is the key point and difficulty.

[0003] The current distributed photovoltaic power data collection solution is only aimed at predicting future power and repairing historical data. For some distributed photovoltaic power stations that only install current collection terminals but not power collection terminals, the existing technology does not involve virtual collection and prediction of active power based on real-time current data. Summary of the invention

[0004] In view of this, the present invention provides a distributed photovoltaic power station active power virtual collection method based on IDBO-KELM-BiGRU neural network, and provides an active power prediction solution for distributed photovoltaic power stations that only have current collection terminals installed but no power collection terminals installed.

[0005] The technical solution adopted by the embodiment of the present invention to solve the technical problem is:

[0006] A distributed photovoltaic power station active power virtual acquisition method based on IDBO-KELM-BiGRU neural network, comprising:

[0007] Step S1, preliminary grid division of the virtual acquisition area according to the geographical location;

[0008] Step S2, using random matrix theory and the correlation analysis method of the dynamic time warping DTW clustering algorithm to select similar power stations to the distributed photovoltaic power station to be collected;

[0009] Step S3, establishing an IDBO-KELM-BiGRU neural network, and using the historical current data of similar power stations and the historical photovoltaic output power data of similar power stations of each of the similar power stations for self-supervised training to obtain a current-power data conversion model, wherein the IDBO-KELM-BiGRU neural network uses an improved dung beetle optimization algorithm IDBO to optimize the hyperparameters of the KELM neural network;

[0010] Step S4, using the current-power data conversion model, converting the historical current data of the distributed photovoltaic power station to be collected into the historical photovoltaic output power prediction data to be collected, and further training the IDBO-KELM-BiGRU neural network model to obtain an active power data virtual collection model;

[0011] Step S5, input the real-time current data of the distributed photovoltaic power station to be collected on the collection day, use the active power data virtual collection model to perform data fitting, and obtain the photovoltaic output power prediction data of the distributed photovoltaic power station to be collected on the collection day in real time.

[0012] Preferably, the step S1 performs preliminary grid division of the virtual acquisition area according to the geographical location, and the division rule is:

[0013] In the virtual collection area, for the distributed photovoltaic power station area located below 1 km above sea level, it is divided according to the unit area of ​​1km×1km; for the distributed photovoltaic power station area located above 1.5 km above sea level, it is divided according to the unit area of ​​3km×3km; the virtual collection area is preliminarily divided into A sub-areas.

[0014] Preferably, the step S2 uses a correlation analysis method of random matrix theory and a dynamic time warping DTW clustering algorithm to select similar power stations to the distributed photovoltaic power station to be collected, and the specific steps include:

[0015] Step S21, creating a random matrix Z of current data of each distributed photovoltaic power station in the sub-region where the distributed photovoltaic power station to be collected is located:

[0016]

[0017] Among them, the number of historical days is n, m is the total length of the historical current data time series, and the current data at each collection time of each historical day are z 11 , z 12 , …, z 1n ,…,z m1 ,z m2 ,…,z mn , m is 96;

[0018] Step S22, selecting a real-time sliding time window based on the matrix Z, and obtaining a time window matrix Z i , where the length of the real-time sliding time window is N w , width is I w , that is, each sampling is shifted backward by one sampling point, and the real-time sliding time window at time i contains the current data at time i and I w -1 adjacent historical data:

[0019]

[0020] In the formula Indicates the Nth collected at time i w Current data, N w =n,I w =m;

[0021] Step S23, the matrix Z i Perform standardization to obtain a standardized matrix The standardized processing rules are:

[0022]

[0023] In the formula, The standard matrix The element in the i-th row and j-th column of i,j Denote the matrix Z i The actual value of the photovoltaic output power in row i and column j, μ(z j ) and σ(z j ) are z j The mean and standard deviation of and They are The mean and standard deviation of z j Denote the matrix Z i The j-th column matrix of Represents the matrix The j-th column matrix of

[0024] Step S24, according to the matrix Obtain the singular value equivalent matrix and perform standardization operations to obtain n standard matrices Z u ;

[0025]

[0026] In the formula, Represents a standardized matrix The conjugate transposed matrix of , U is the Haar unitary matrix;

[0027] Step S25, calculate n standard matrices Z u The product of is the standard non-Hermitian matrix product

[0028]

[0029] Step S26, calculating the standard non-Hermitian matrix product The standard matrix S of , where:

[0030]

[0031] Where j∈[1,n] represents the matrix product The number of columns, s j represents the column elements of the standard matrix S, Represents the matrix product The elements of each column, Represents the matrix product The standard deviation of the elements in the jth column;

[0032] Step S27, calculating the average spectrum radius MSR statistic R of the standard matrix S at time i MSR,i :

[0033]

[0034] Where q∈[1,N], N represents the number of characteristic roots; λ q represents the characteristic root of the standard matrix S;

[0035] Step S28: construct a real-time average spectrum radius feature data set R of the distributed photovoltaic power station according to the time sequence sliding time window. MSR-c , where R MSR-c = {R MSR-c,1 ,R MSR-c,2 ,…,R MSR-c,m}, c∈[1,r], the number of distributed photovoltaic power stations in the sub-region is r; the average spectrum radius feature data set R' of the sub-region is obtained, R'={R MSR-1 ,R MSR-2 ,…,R MSR-r};

[0036] Step S29, taking each element in the average spectrum radius feature data set R' as a time series, and letting the time series o represent the time series corresponding to the distributed photovoltaic power station to be collected, and the time series e represent the time series corresponding to a distributed photovoltaic power station not to be collected in the average spectrum radius feature data set R':

[0037]

[0038] Step S210, find the regular path W according to the Euclidean distance metric, where the Euclidean distance metric of o and e is expressed as:

[0039] τ(m,m)=(o m -e m ) 2

[0040] W={w1,w2,...,w k}

[0041] Where w k is the shortest distance point pair among all the point pairs in the time series curve of the average spectrum radius characteristic data;

[0042] Step S211, calculating the similarity DTW(o,e) between the time series o and the time series e:

[0043]

[0044] Step S212: sort the calculated similarity values ​​from large to small, and select the non-to-be-collected distributed photovoltaic power stations corresponding to the first ten similarity values ​​as the similar power stations.

[0045] Preferably, the hyperparameters of the KELM neural network in step S3 include internal regularization coefficients and kernel function parameters of KELM.

[0046] Preferably, the step S3 establishes an IDBO-KELM-BiGRU neural network, and uses the historical current data of similar power stations and the historical photovoltaic output power data of similar power stations of each of the similar power stations for self-supervised training to obtain a current-power data conversion model including:

[0047] Step S31, establishing a current-power data conversion model data set and dividing it into a first training set, a first validation set and a first test set, wherein the current-power data conversion model data set includes similar power station historical current data and similar power station historical photovoltaic output power data of all the similar power stations;

[0048] Step S32, using the first training set to perform self-supervised training on the KELM-BiGRU neural network model, wherein the historical current data of similar power stations is used as input data, and the historical photovoltaic output power data of similar power stations is used as a reference object for predicted power data output by the model;

[0049] Step S33, using the root mean square error of the prediction result of the first validation set as the fitness function, and optimizing the hyperparameters through the improved dung beetle optimization algorithm IDBO to minimize the root mean square error;

[0050] Step S34, initializing the size and dimension of the dung beetle population, determining the maximum number of iterations and the search range of the internal regularization coefficient and kernel function parameter;

[0051] Step S35, randomly generating the positions of dung beetle populations, calculating the fitness values ​​corresponding to each dung beetle population, selecting the position of the dung beetle individual with the best fitness, and updating the current dung beetle position;

[0052] Step S36, repeating step S35 until the maximum number of iterations is reached, recording the best position, at which point the best position coordinates of the dung beetle are the best hyperparameters obtained by the improved dung beetle optimization algorithm IDBO;

[0053] Step S37, substituting the KELM neural network hyperparameters optimized by the improved dung beetle optimization algorithm IDBO into the first test set, using the first test set to test the IDBO-KELM-BiGRU neural network model, and when the mean absolute percentage error is less than or equal to 5%, completing the training and obtaining the current-power data conversion model;

[0054] Step S38, output the mean absolute percentage error γ MAPE To evaluate the effect, including:

[0055]

[0056] Where N T Indicates the number of test samples; and i They respectively represent the virtual collection value and the real value of the distributed photovoltaic output power.

[0057] Preferably, the step S32 uses the improved dung beetle optimization algorithm IDBO to optimize the neural parameters of the KELM network, including:

[0058] Step S321, set the initial value of the improved dung beetle algorithm: maximum number of iterations I MAX , population size N, population dimension Dim;

[0059] Step S322: simulate the rolling behavior, and the position update formula is:

[0060] x i (g+1)=x i (g)+α×k×x i (g-1)+b×Δx,

[0061] Δx=|x i (g)-X worst |

[0062] Where g is the current iteration number; xi (g) is the position of the i-th dung beetle in the g-th iteration; k is the deflection coefficient constant, k∈(0,0.2]; α is the natural coefficient, α∈(0,1); b is a constant, b∈(0,1); X worst is the global worst solution; △x simulates the change of light intensity;

[0063] Step S323, simulating dance behavior, the position update formula is:

[0064]

[0065] Where θ is the deflection angle, θ∈[0,π]; the Levy distribution simulation formula is as follows:

[0066]

[0067]

[0068] Among them, u and v both obey normal distribution, β = 1.5;

[0069] Step S324, simulation Dung beetle egg-laying area boundary selection strategy:

[0070]

[0071] Lb * =max(X * ×(1-R),Lb)

[0072] Ub * =min(X * ×(1+R),Ub)

[0073] Where X * is the current optimal position; Ub * and Lb * G is the upper and lower boundaries of the female dung beetle's egg-laying area; max is the maximum number of iterations; Ub and Lb are the upper and lower bounds of the optimization problem;

[0074] Step S325, inspired by the salp swarm algorithm, simulates the reproduction behavior, wherein the Newton's law of motion is used to update the strategy instead of the random number, and the position update formula of the dung ball is:

[0075]

[0076] Where B i (g) is the position of the ith dung ball in the gth iteration; b1 and b2 are two independent random vectors of size ; D is the dimension of the optimization problem; l∈random(-1,1); k=5;

[0077] Step S326: Simulate the dung beetle foraging area boundary selection strategy, which is defined as:

[0078] Lb b =max(X b ×(1-R),Lb)

[0079] Ub b =min(X b ×(1+R),Ub)

[0080] Where X b is the global optimal position; Ub b and Lb b are the upper and lower boundaries of the optimal foraging area;

[0081] Step S327, simulating foraging behavior, introducing an adaptive weight factor, the position update formula of the dung beetle forager is:

[0082]

[0083] x i (g+1)=x i (g)+C1×ω×(x i (g)-Lb b )+C2×ω×(x i (g)-Ub b )

[0084] Where x i (g) is the position of the i-th dung beetle forager in the g-th iteration; C1 is a random number that follows a normal distribution; C2 is a random vector in the range of (0,1);

[0085] Step S328, simulating theft behavior, the location update formula is:

[0086] x i (g+1)=X b +S×m×(|x i (g)-X * |+|x i (g)-X b |)

[0087] Where x i (g) is the position of the i-th dung beetle stealer in the g-th iteration; S is a constant; m is a random vector of size 1×D that follows a normal distribution;

[0088] Step S329, finding the best position of the optimized dung beetle individual according to the fitness value of the dung beetle population;

[0089] Step S3210, determining whether the maximum number of searches has been reached or the search accuracy has been met, if so, proceeding to step S3211, otherwise returning to step S322;

[0090] Step S3211, output the global optimal solution and the algorithm ends.

[0091] Preferably, the step S4 comprises:

[0092] Step S41, using the current-power data conversion model to convert the historical current data of the distributed photovoltaic power station to be collected into the historical power prediction data of the power station to be collected, establishing an active power data virtual collection model data set and dividing it into a second training set, a second verification set and a second test set, the active power data virtual collection model data set includes all the historical current data of similar power stations, all the historical photovoltaic output power data of similar power stations, the historical current data of the power station to be collected, and the historical photovoltaic output power prediction data of the power station to be collected;

[0093] Step S42, performing self-supervised training on the KELM-BiGRU neural network model using the second training set to obtain a prediction error under the second training set;

[0094] Step S43, using the root mean square error of the prediction result of the second verification set as the fitness function, and using the improved dung beetle optimization algorithm IDBO to perform optimization to minimize the root mean square error;

[0095] Step S44, repeating steps S33 to S36 to obtain the optimal hyperparameters after the improved dung beetle optimization algorithm IDBO is optimized;

[0096] Step S45, substitute the KELM neural network hyperparameters optimized by the improved dung beetle optimization algorithm IDBO into the second test set, and use the second test set to test the IDBO-KELM-BiGRU neural network model. When the mean absolute percentage error is less than or equal to 3%, the training is completed, the active power data virtual acquisition model is obtained, and the final prediction result is obtained.

[0097] It can be seen from the above technical scheme that the distributed photovoltaic power station active power virtual collection method based on the IDBO-KELM-BiGRU neural network provided by the embodiment of the present invention first performs preliminary grid division of the virtual collection area according to the geographical location; selects similar power stations of the distributed photovoltaic power station to be collected by using the correlation analysis method of the random matrix theory and the dynamic time warping DTW clustering algorithm; establishes an IDBO-KELM-BiGRU neural network, and uses the historical current data of similar power stations and the historical photovoltaic output power data of similar power stations of each similar power station for self-supervised training to obtain a current-power data conversion model; uses the current-power data conversion model to convert the historical current data of the distributed photovoltaic power station to be collected into the historical photovoltaic output power prediction data of the power station to be collected, and further trains the IDBO-KELM-BiGRU neural network model to obtain the active power data virtual collection model; inputs the real-time current data of the distributed photovoltaic power station to be collected on the collection day, uses the active power data virtual collection model to perform data fitting, and obtains the photovoltaic output power prediction data of the distributed photovoltaic power station to be collected on the collection day in real time. The present invention can realize active power prediction for distributed photovoltaic power stations that only have current collection terminals installed but no power collection terminals installed, so as to achieve full-range data collection by a limited number of intelligent fusion terminals within regions, substations, communities, etc., thereby reducing collection costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0098] Figure 1 This is a flow chart of the virtual collection method of active power of distributed photovoltaic power stations based on IDBO-KELM-BiGRU neural network.

[0099] Figure 2 This is a framework diagram of virtual collection of active power of distributed photovoltaic power stations based on random matrix theory and IDBO-KELM-BiGRU neural network of the present invention.

[0100] Figure 3 This is a flow chart of the improved dung beetle optimization algorithm of the present invention. DETAILED DESCRIPTION

[0101] The technical scheme and technical effects of the present invention are further elaborated in detail below in conjunction with the accompanying drawings of the present invention.

[0102] Aiming at the situation that some distributed photovoltaic power stations are equipped with current collection terminals but not with power collection terminals, that is, the distributed photovoltaic power stations to be collected in this article cannot support the system to directly collect power data. This paper proposes a systematic virtual collection scheme to address the problem that the system cannot directly collect power data. Drawing on the above experience and combining with the needs of actual project engineering, a distributed photovoltaic power station active power virtual collection technology based on random matrix theory and IDBO-KELM-BiGRU neural network is proposed.

[0103] like Figure 1 The flowchart of the distributed photovoltaic power station active power virtual acquisition method based on IDBO-KELM-BiGRU neural network is shown in FIG. Figure 2 The implementation steps of the present invention specifically include:

[0104] Step S1, preliminary grid division of the virtual acquisition area according to the geographical location;

[0105] Step S2, using random matrix theory and the correlation analysis method of the dynamic time warping DTW clustering algorithm to select similar power stations to the distributed photovoltaic power station to be collected;

[0106] Step S3, establishing an IDBO-KELM-BiGRU neural network, and using the historical current data of similar power stations and the historical photovoltaic output power data of similar power stations of each similar power station for self-supervised training to obtain a current-power data conversion model, and using the improved dung beetle optimization algorithm IDBO in the IDBO-KELM-BiGRU neural network to optimize the hyperparameters of the KELM neural network to achieve the fitting of the historical current data of similar power stations and the historical photovoltaic output power data; Here, the IDBO-KELM-BiGRU neural network model refers to the combination of the dung beetle optimization algorithm model and the KELM-BiGRU neural network model, IDBO is used to optimize the hyperparameters, and KELM-BiGRU is used for training;

[0107] Step S4, using the current-power data conversion model, converting the historical current data of the distributed photovoltaic power station to be collected into the historical photovoltaic output power prediction data of the power station to be collected, and further training the IDBO-KELM-BiGRU neural network model to obtain the active power data virtual collection model;

[0108] Step S5, input the real-time current data of the distributed photovoltaic power station to be collected on the collection day, use the active power data virtual collection model to perform data fitting, and obtain the photovoltaic output power prediction data of the distributed photovoltaic power station to be collected on the collection day in real time.

[0109] Step S1 divides the virtual acquisition area based on the geographical location, and the division rules are:

[0110] In the virtual collection area, for the distributed photovoltaic power station area below 1 km above sea level, it is divided according to the unit area of ​​1km×1km; for the distributed photovoltaic power station area above 1.5 km above sea level, it is divided according to the unit area of ​​3km×3km; the virtual collection area is initially divided into A sub-areas. The distributed photovoltaic power station to be collected is located in one of the sub-areas.

[0111] Step S2 uses the random matrix theory and the correlation analysis method of the dynamic time warping DTW clustering algorithm to select similar power stations to the distributed photovoltaic power station to be collected. The specific steps include:

[0112] Step S21, creating a random matrix Z of current data of each distributed photovoltaic power station in the sub-region where the distributed photovoltaic power station to be collected is located:

[0113]

[0114] Among them, the number of historical days is n, m is the total length of the historical current data time series, and the current data at each collection time of each historical day are z 11 , z 12 , …, z 1n ,…,z m1 ,z m2 ,…,z mn , m is 96;

[0115] Step S22, select a real-time sliding time window based on the matrix Z, and obtain the time window matrix Z i , where the length of the real-time sliding time window is N w , width is I w , that is, each sampling is shifted backward by one sampling point, and the real-time sliding time window at time i contains the current data at time i and I w -1 adjacent historical data:

[0116]

[0117] In the formula Indicates the Nth collected at time i w Current data, N w =n,I w =m;

[0118] Step S23: transform the matrix Z i Perform standardization to obtain a standardized matrix The standardized processing rules are:

[0119]

[0120] In the formula, is a standard matrix The element in the i-th row and j-th column of i,j Denotes the matrix Z i The actual value of the photovoltaic output power in row i and column j, μ(z j ) and σ(z j ) are z j The mean and standard deviation of and They are The mean and standard deviation of z j Denotes the matrix Z i The j-th column matrix of Representation Matrix The j-th column matrix of

[0121] Step S24, according to the matrix Obtain the singular value equivalent matrix and perform standardization operations to obtain n standard matrices Z u ;

[0122]

[0123] In the formula, Represents a standardized matrix The conjugate transposed matrix of , U is the Haar unitary matrix;

[0124] Step S25, calculate n standard matrices Z u The product of is the standard non-Hermitian matrix product

[0125]

[0126] Step S26, calculate the standard non-Hermitian matrix product The standard matrix S of , where:

[0127]

[0128] Where j∈[1,n] represents the matrix product The number of columns, s j represents the column elements of the standard matrix S, Represents matrix product The elements of each column, Represents matrix product The standard deviation of the elements in the jth column;

[0129] Step S27, calculate the average spectrum radius MSR statistic R of the standard matrix S at time i MSR,i :

[0130]

[0131] Where q∈[1,N], N represents the number of characteristic roots; λ q represents the characteristic root of the standard matrix S;

[0132] Step S28: construct a real-time average spectrum radius feature data set R of a distributed photovoltaic power station according to the time sequence sliding time window. MSR-c , where R MSR-c = {R MSR-c,1 ,R MSR-c,2 ,…,R MSR-c,m}, c∈[1,r], the number of distributed photovoltaic power stations in the sub-region is r; the average spectrum radius feature data set R' of the sub-region is obtained, R'={R MSR-1 ,R MSR-2 ,…,R MSR-r};

[0133] Step S29, each element in the average spectrum radius feature data set R' is taken as a time series, and the time series o represents the time series corresponding to the distributed photovoltaic power station to be collected, and the time series e represents the time series corresponding to a distributed photovoltaic power station not to be collected in the average spectrum radius feature data set R':

[0134]

[0135] Step S210, find the regular path W according to the Euclidean distance metric, where the Euclidean distance metric of o and e is expressed as:

[0136] τ(m,m)=(o m -e m ) 2 (10)

[0137] W={w1,w2,...,w k} (11)

[0138] Where w k is the shortest distance point pair among all the point pairs in the time series curve of the average spectrum radius characteristic data;

[0139] Step S211, calculate the similarity DTW(o,e) between time series o and time series e:

[0140]

[0141] Step S212: sort the calculated similarity values ​​from large to small, and select the non-to-be-collected distributed photovoltaic power stations corresponding to the first ten similarity values ​​as similar power stations.

[0142] The hyperparameters of the KELM neural network in step S3 include the internal regularization coefficient and kernel function parameters of KELM. Step S3 establishes the IDBO-KELM-BiGRU neural network and uses the historical current data of similar power stations and the historical photovoltaic output power data of similar power stations for self-supervised training to obtain the current-power data conversion model including:

[0143] Step S31, establishing a current-power data conversion model data set and dividing it into a first training set, a first validation set and a first test set, the current-power data conversion model data set including similar power station historical current data and similar power station historical photovoltaic output power data of all similar power stations; since the voltage levels in adjacent areas are almost the same, the current data change trend of the distributed photovoltaic power station is similar to the output power change trend;

[0144] Step S32, using the first training set to perform self-supervised training on the KELM-BiGRU neural network model, wherein historical current data of similar power stations is used as input data, and historical photovoltaic output power data of similar power stations is used as a reference object for predicted power data output by the model;

[0145] Step S33, using the root mean square error of the prediction result of the first validation set as the fitness function, and optimizing the hyperparameters through the improved dung beetle optimization algorithm IDBO to minimize the root mean square error;

[0146] Step S34, initializing the size and dimension of the dung beetle population, determining the maximum number of iterations and the search range of the internal regularization coefficient and kernel function parameter;

[0147] Step S35, randomly generating the positions of dung beetle populations, calculating the fitness values ​​corresponding to each dung beetle population, selecting the position of the dung beetle individual with the best fitness, and updating the current dung beetle position;

[0148] Step S36, repeating step S35 until the maximum number of iterations is reached, recording the best position, at which point the best position coordinates of the dung beetle are the best parameters obtained by the improved dung beetle optimization algorithm IDBO;

[0149] Step S37, substitute the KELM neural network hyperparameters optimized by the improved dung beetle optimization algorithm IDBO into the first test set, use the first test set to test the IDBO-KELM-BiGRU neural network model, and when the mean absolute percentage error is less than or equal to 5%, the training is completed to obtain the current-power data conversion model.

[0150] Step S38, output the mean absolute percentage error γ MAPE To evaluate the effect, including:

[0151]

[0152] Where N T Indicates the number of test samples; and i They respectively represent the virtual collection value and the real value of the distributed photovoltaic output power.

[0153] Please refer to Figure 3 Step S32 uses the improved dung beetle optimization algorithm IDBO to optimize the KELM network neural parameters, including:

[0154] Step S321, set the initial value of the improved dung beetle algorithm: maximum number of iterations I MAX , population size N, population dimension Dim;

[0155] Step S322: simulate the rolling behavior, and the position update formula is:

[0156] x i (g+1)=x i (g)+α×k×x i (g-1)+b×Δx,

[0157] Δx=|x i (g)-X worst | (14)

[0158] Where g is the current iteration number; x i (g) is the position of the i-th dung beetle in the g-th iteration; k is the deflection coefficient constant, k∈(0,0.2]; α is the natural coefficient, α∈(0,1); b is a constant, b∈(0,1); X worst is the global worst solution; △x simulates the change of light intensity;

[0159] Step S323, simulating dance behavior, the position update formula is:

[0160]

[0161] Where θ is the deflection angle, θ∈[0,π]; the Levy distribution simulation formula is as follows:

[0162]

[0163]

[0164] Among them, u and v both obey normal distribution, β = 1.5;

[0165] Step S324, simulation Dung beetle egg-laying area boundary selection strategy:

[0166]

[0167]

[0168] Where X * is the current optimal position; Ub * and Lb * G is the upper and lower boundaries of the female dung beetle's egg-laying area; max is the maximum number of iterations; Ub and Lb are the upper and lower bounds of the optimization problem;

[0169] Step S325, inspired by the salp swarm algorithm, simulates the reproduction behavior, wherein the Newton's law of motion is used to update the strategy instead of the random number, and the position update formula of the dung ball is:

[0170]

[0171] Where B i (g) is the position of the ith dung ball in the gth iteration; b1 and b2 are two independent random vectors of size ; D is the dimension of the optimization problem; l∈random(-1,1); k=5;

[0172] Step S326: Simulate the dung beetle foraging area boundary selection strategy, which is defined as:

[0173]

[0174] Where X b is the global optimal position; Ub b and Lb b are the upper and lower boundaries of the optimal foraging area;

[0175] Step S327, simulating foraging behavior, introducing an adaptive weight factor, the position update formula of the dung beetle forager is:

[0176]

[0177] x i (g+1)=x i (g)+C1×ω×(x i (g)-Lb b )+C2×ω×(x i (g)-Ub b ) (twenty three)

[0178] Where x i (g) is the position of the i-th dung beetle forager in the g-th iteration; C1 is a random number that follows a normal distribution; C2 is a random vector in the range of (0,1).

[0179] Step S328, simulating theft behavior, the location update formula is:

[0180] x i (g+1)=X b +S×m×(|x i (g)-X * |+|x i (g)-X b |) (24)

[0181] Where x i (g) is the position of the i-th dung beetle stealer in the g-th iteration; S is a constant; m is a random vector of size 1×D that follows a normal distribution.

[0182] Step S329, finding the best position of the optimized dung beetle individual according to the fitness value of the dung beetle population;

[0183] Step S3210, determining whether the maximum number of searches has been reached or the search accuracy has been met, if so, proceeding to step S3211, otherwise returning to step S322;

[0184] Step S3211: Output the global optimal solution and the algorithm ends.

[0185] Furthermore, the specific implementation of step S4 includes:

[0186] Step S41, using the current-power data conversion model to convert the historical current data of the distributed photovoltaic power station to be collected into the historical power prediction data of the power station to be collected, establishing a virtual collection model data set for active power data and dividing it into a second training set, a second verification set and a second test set, the virtual collection model data set for active power data includes all similar power station historical current data, all similar power station historical photovoltaic output power data, the historical current data of the power station to be collected, and the historical photovoltaic output power prediction data of the power station to be collected;

[0187] Step S42, using the second training set to perform self-supervised training on the KELM-BiGRU neural network model to obtain a prediction error under the second training set;

[0188] Step S43, using the root mean square error of the prediction result of the second validation set as the fitness function, and optimizing using the improved dung beetle optimization algorithm IDBO to minimize the root mean square error;

[0189] Step S44, repeating steps S33 to S36 to obtain the best hyperparameters after the improved dung beetle optimization algorithm IDBO is optimized;

[0190] Step S45, substitute the KELM neural network hyperparameters optimized by the improved dung beetle optimization algorithm IDBO into the second test set, and use the second test set to test the IDBO-KELM-BiGRU neural network model. When the mean absolute percentage error is less than or equal to 3%, the training is completed, the active power data virtual acquisition model is obtained, and the final prediction result is obtained.

[0191] In summary, the present invention proposes a distributed photovoltaic power station active power virtual acquisition technology based on random matrix theory and IDBO-KELM-BiGRU neural network. First, by dividing the grid area, the influence of the external environment on virtual acquisition is eliminated. Under the premise that only the benchmark power station is equipped with a power acquisition device and the power measurement devices and environmental information of other distributed photovoltaic power stations are incomplete, the power of the benchmark power station is predicted by the IDBO-KELM-BiGRU neural network, and the power of other distributed photovoltaic power stations in the grid area is predicted based on the random matrix theory, which solves the problems of strong intermittent and random distributed photovoltaic output and power load constraints formed by the connected adjacent photovoltaics, reduces the construction cost of the power grid and users, and greatly reduces the subsequent operation and maintenance management workload. For the dung beetle optimization algorithm, an improved dung beetle optimization algorithm is proposed, which introduces levy flight strategy, spiral position update strategy, adaptive weight factor and t distribution variation strategy respectively, effectively solves the problem of premature convergence of particles due to falling into local optimality, and solves the problem of low accuracy of weights and thresholds of the traditional dung beetle optimization algorithm to optimize the KELM-BiGRU neural network.

[0192] The solution of the present invention can achieve low-cost and high-efficiency acquisition of distributed photovoltaic operation data, and achieve full-range data collection by a limited number of intelligent fusion terminals within the scope of regions, substations, communities, etc., thereby reducing collection costs.

[0193] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0194] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0195] These computer program instructions may also be stored in a computer readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0196] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0197] What is disclosed above is only a preferred embodiment of the present invention, which certainly cannot be used to limit the scope of rights of the present invention. A person skilled in the art can understand that all or part of the processes of the above embodiments and equivalent changes made according to the claims of the present invention still fall within the scope of the invention.

Claims

1. A distributed photovoltaic power station active power virtual acquisition method based on IDBO-KELM-BiGRU neural network, characterized in that: include: Step S1, preliminary grid division of the virtual acquisition area according to the geographical location; Step S2, using random matrix theory and the correlation analysis method of the dynamic time warping DTW clustering algorithm to select similar power stations to the distributed photovoltaic power station to be collected; Step S3, establishing an IDBO-KELM-BiGRU neural network, and using the historical current data of similar power stations and the historical photovoltaic output power data of similar power stations of each of the similar power stations for self-supervised training to obtain a current-power data conversion model, wherein the IDBO-KELM-BiGRU neural network uses an improved dung beetle optimization algorithm IDBO to optimize the hyperparameters of the KELM neural network; Step S4, using the current-power data conversion model, converting the historical current data of the distributed photovoltaic power station to be collected into the historical photovoltaic output power prediction data to be collected, and training the IDBO-KELM-BiGRU neural network model to obtain a virtual collection model for active power data; Step S5, inputting the real-time current data of the distributed photovoltaic power station to be collected on the collection day, using the active power data virtual collection model to perform data fitting, and obtaining the photovoltaic output power prediction data of the distributed photovoltaic power station to be collected on the collection day in real time; The step S2 uses the correlation analysis method of random matrix theory and dynamic time warping DTW clustering algorithm to select similar power stations of the distributed photovoltaic power station to be collected. The specific steps include: Step S21, creating a random matrix Z of current data of each distributed photovoltaic power station in the sub-region where the distributed photovoltaic power station to be collected is located: Among them, the number of historical days is n, m is the total length of the historical current data time series, and the current data at each collection time of each historical day are z 11 ,z 12 ,…,z 1n ,…,z m1 ,z m2 ,…,z mn , m is 96; Step S22, selecting a real-time sliding time window based on the matrix Z, and obtaining a time window matrix Z i , where the length of the real-time sliding time window is N w , width is I w , that is, each sampling is shifted backward by one sampling point, and the real-time sliding time window at time i contains the current data at time i and I w -1 adjacent historical data: In the formula Indicates the Nth collected at time i w Current data, N w =n,I w =m; Step S23, the matrix Z i Standardization is performed to obtain the standard matrix The standardized processing rules are: In the formula, The standard matrix The element in the i-th row and j-th column of i,j Denote the matrix Z i The actual value of the photovoltaic output power in row i and column j, μ(z j ) and σ(z j ) are z j The mean and standard deviation of and They are The mean and standard deviation of z j Denote the matrix Z i The j-th column matrix of Represents the matrix The j-th column matrix of Step S24, according to the matrix Obtain the singular value equivalent matrix and perform standardization operations to obtain n standard matrices Z u ; In the formula, Represents the standard matrix The conjugate transposed matrix of , U is the Haar unitary matrix; Step S25, calculate n standard matrices Z u The product of is the standard non-Hermitian matrix product Step S26, calculating the standard non-Hermitian matrix product The standard matrix S of , where: Where j∈[1,n] represents the matrix product The number of columns, s j represents the column elements of the standard matrix S, Represents the matrix product The elements of each column, Represents the matrix product The standard deviation of the elements in the jth column; Step S27, calculating the average spectrum radius MSR statistic R of the standard matrix S at time i MSR,i : Where q∈[1,N], N represents the number of characteristic roots; λ q represents the characteristic root of the standard matrix S; Step S28: construct a real-time average spectrum radius feature data set R of the distributed photovoltaic power station according to the time sequence sliding time window. MSR-c , where R MSR-c = {R MSR-c,1 ,R MSR-c,2 ,…,R MSR-c,m }, c∈[1,r], the number of distributed photovoltaic power stations in the sub-region is r; the average spectrum radius feature data set R' of the sub-region is obtained, R'={R MSR-1 ,R MSR-2 ,…,R MSR-r }; Step S29, taking each element in the average spectrum radius feature data set R' as a time series, and letting the time series o represent the time series corresponding to the distributed photovoltaic power station to be collected, and the time series e represent the time series corresponding to a distributed photovoltaic power station not to be collected in the average spectrum radius feature data set R': Step S210, find the regular path W according to the Euclidean distance metric, where the Euclidean distance metric of o and e is expressed as: τ(m,m)=(o m -e m ) 2 In={in1,in2,...,in k } Where w k is the shortest distance point pair among all the point pairs in the time series curve of the average spectrum radius characteristic data; Step S211, calculating the similarity DTW(o,e) between the time series o and the time series e: Step S212: sort the calculated similarity values ​​from large to small, and select the non-to-be-collected distributed photovoltaic power stations corresponding to the first ten similarity values ​​as the similar power stations.

2. The distributed photovoltaic power station active power virtual acquisition method based on IDBO-KELM-BiGRU neural network as claimed in claim 1 is characterized in that: The step S1 performs preliminary grid division of the virtual acquisition area according to the geographical location, and the division rules are as follows: In the virtual collection area, for the distributed photovoltaic power station area located below 1 km above sea level, it is divided according to the unit area of ​​1km×1km; for the distributed photovoltaic power station area located above 1.5 km above sea level, it is divided according to the unit area of ​​3km×3km; the virtual collection area is preliminarily divided into A sub-areas.

3. The distributed photovoltaic power station active power virtual acquisition method based on IDBO-KELM-BiGRU neural network as claimed in claim 2 is characterized in that: The hyperparameters of the KELM neural network in step S3 include the internal regularization coefficient and kernel function parameters of KELM.

4. The distributed photovoltaic power station active power virtual acquisition method based on IDBO-KELM-BiGRU neural network as claimed in claim 3 is characterized in that: The step S3 establishes an IDBO-KELM-BiGRU neural network, and uses the historical current data of similar power stations and the historical photovoltaic output power data of similar power stations of each of the similar power stations for self-supervised training to obtain a current-power data conversion model including: Step S31, establishing a current-power data conversion model data set and dividing it into a first training set, a first validation set and a first test set, wherein the current-power data conversion model data set includes similar power station historical current data and similar power station historical photovoltaic output power data of all the similar power stations; Step S32, using the first training set to perform self-supervised training on the KELM-BiGRU neural network model, wherein the historical current data of similar power stations is used as input data, and the historical photovoltaic output power data of similar power stations is used as a reference object for predicted power data output by the model; Step S33, using the root mean square error of the prediction result of the first validation set as the fitness function, and optimizing the hyperparameters through the improved dung beetle optimization algorithm IDBO to minimize the root mean square error; Step S34, initializing the size and dimension of the dung beetle population, determining the maximum number of iterations and the search range of the internal regularization coefficient and kernel function parameter; Step S35, randomly generating the positions of dung beetle populations, calculating the fitness values ​​corresponding to each dung beetle population, selecting the position of the dung beetle individual with the best fitness, and updating the current dung beetle position; Step S36, repeating step S35 until the maximum number of iterations is reached, recording the best position, at which point the best position coordinates of the dung beetle are the best hyperparameters obtained by the improved dung beetle optimization algorithm IDBO; Step S37, substituting the KELM neural network hyperparameters optimized by the improved dung beetle optimization algorithm IDBO into the first test set, using the first test set to test the IDBO-KELM-BiGRU neural network model, and when the mean absolute percentage error is less than or equal to 5%, completing the training and obtaining the current-power data conversion model; Step S38, output the mean absolute percentage error γ MAPE To evaluate the effect, including: Where N T Indicates the number of test samples; and i They respectively represent the virtual collection value and the real value of the distributed photovoltaic output power.

5. The distributed photovoltaic power station active power virtual acquisition method based on IDBO-KELM-BiGRU neural network as claimed in claim 4 is characterized in that: The step S32 uses the improved dung beetle optimization algorithm IDBO to optimize the neural parameters of the KELM network, including: Step S321, set the initial value of the improved dung beetle algorithm: maximum number of iterations I MAX , population size N, population dimension Dim; Step S322: simulate the rolling behavior, and the position update formula is: x i (g+1)=x i (g)+α×k×x i (g-1)+b×Δx, Δx=|x i (g)-X worst | Where g is the current iteration number; x i (g) is the position of the i-th dung beetle in the g-th iteration; k is the deflection coefficient constant, k∈(0,0.2]; α is the natural coefficient, α∈(0,1); b is a constant, b∈(0,1); X worst is the global worst solution; △x simulates the change of light intensity; Step S323, simulating dance behavior, the position update formula is: Where θ is the deflection angle, θ∈[0,π]; the Levy distribution simulation formula is as follows: Among them, u and v both obey normal distribution, β = 1.5; Step S324, simulation Dung beetle egg-laying area boundary selection strategy: Lb * =max(X * ×(1-R),Lb) Ub * =min(X * ×(1+R),Ub) Where X * is the current optimal position; Ub * and Lb * G is the upper and lower boundaries of the female dung beetle's egg-laying area; max is the maximum number of iterations; Ub and Lb are the upper and lower bounds of the optimization problem; Step S325, inspired by the salp swarm algorithm, simulates the reproduction behavior, wherein the Newton's law of motion is used to update the strategy instead of the random number, and the position update formula of the dung ball is: Where B i (g) is the position of the ith dung ball in the gth iteration; b1 and b2 are two independent random vectors of size ; D is the dimension of the optimization problem; l∈random(-1,1); k=5; Step S326: Simulate the dung beetle foraging area boundary selection strategy, which is defined as: Lb b =max(X b ×(1-R),Lb) Ub b =min(X b ×(1+R),Ub) Where X b is the global optimal position; Ub b and Lb b are the upper and lower boundaries of the optimal foraging area; Step S327, simulating foraging behavior, introducing an adaptive weight factor, the position update formula of the dung beetle forager is: x i (g+1)=x i (g)+C1×ω×(x i (g)-Lb b )+C2×ω×(x i (g)-Ub b ) Where x i (g) is the position of the i-th dung beetle forager in the g-th iteration; C1 is a random number that follows a normal distribution; C2 is a random vector in the range of (0,1); Step S328, simulating theft behavior, the location update formula is: x i (g+1)=X b +S×m×(|x i (g)-X * |+|x i (g)-X b |) Where x i (g) is the position of the i-th dung beetle stealer in the g-th iteration; S is a constant; m is a random vector of size 1×D that follows a normal distribution; Step S329, finding the best position of the optimized dung beetle individual according to the fitness value of the dung beetle population; Step S3210, determining whether the maximum number of searches has been reached or the search accuracy has been met, if so, proceeding to step S3211, otherwise returning to step S322; Step S3211, output the global optimal solution and the algorithm ends.

6. The distributed photovoltaic power station active power virtual acquisition method based on IDBO-KELM-BiGRU neural network as claimed in claim 5, characterized in that: The step S4 comprises: Step S41, using the current-power data conversion model to convert the historical current data of the distributed photovoltaic power station to be collected into the historical power prediction data of the power station to be collected, establishing an active power data virtual collection model data set and dividing it into a second training set, a second verification set and a second test set, the active power data virtual collection model data set includes all the historical current data of similar power stations, all the historical photovoltaic output power data of similar power stations, the historical current data of the power station to be collected, and the historical photovoltaic output power prediction data of the power station to be collected; Step S42, performing self-supervised training on the KELM-BiGRU neural network model using the second training set to obtain a prediction error under the second training set; Step S43, using the root mean square error of the prediction result of the second verification set as the fitness function, and using the improved dung beetle optimization algorithm IDBO to perform optimization to minimize the root mean square error; Step S44, repeating steps S33 to S36 to obtain the optimal hyperparameters after the improved dung beetle optimization algorithm IDBO is optimized; Step S45, substitute the KELM neural network hyperparameters optimized by the improved dung beetle optimization algorithm IDBO into the second test set, and use the second test set to test the IDBO-KELM-BiGRU neural network model. When the mean absolute percentage error is less than or equal to 3%, the training is completed, the active power data virtual acquisition model is obtained, and the final prediction result is obtained.

Images