Ultra-short-term prediction method of small hydropower station power generation based on grid meteorological data

Abstract

The present invention relates to an ultra-short-term prediction method for power generation of a small hydropower station based on grid meteorological data, comprising: dividing the area where the hydropower station is located into grids, obtaining the rainfall data of the grids; calculating the correlation coefficient between the historical power generation of the hydropower station and the historical rainfall of the grids, and screening the grids with large correlation of power generation; using a local linear embedding method to reduce the dimension of the grid rainfall data to obtain reduced-dimensional grid rainfall data; establishing a power generation prediction model; using the reduced-dimensional historical grid rainfall data and the historical power generation of the hydropower station as data sets of the power generation prediction model, training and testing the prediction model; and using the trained power generation prediction model to perform ultra-short-term prediction of the power generation of the hydropower station. The present invention establishes an ultra-short-term prediction model for power generation of a small hydropower station that takes into account the influence of the temporal and spatial distribution of rainfall. By reducing the dimension of the input data, the computational complexity of the prediction model is reduced, and the computational efficiency and accuracy of the prediction model are improved.

Description

Technical Field

[0001] The invention belongs to the field of hydropower station power generation prediction, and in particular relates to an ultra-short-term prediction method for small hydropower station power generation based on grid meteorological data. Background Art

[0002] As an important component of clean energy, small hydropower has positive significance for optimizing energy structure and energy conservation and emission reduction. With the advancement of my country's energy transformation, the installed capacity of small hydropower will gradually increase. However, small hydropower has shortcomings such as poor regulation ability and being far away from the load center. Its power generation is affected by rainfall, and frequent output fluctuations will have an impact on the power grid. In severe cases, it will affect the grid dispatch and squeeze the transmission channel. The power generation of small hydropower is affected by the meteorological conditions of the basin, and the spatial distribution range of the basin is large. The water supply of the basin is affected by the recent historical rainfall in the region. How to extract effective feature data from the massive meteorological data in the area related to the power generation of small hydropower for the power generation prediction of small hydropower is difficult.

[0003] The use of scientific forecasting methods to improve the accuracy of small hydropower generation forecasting has positive significance for power generation planning and promoting the consumption of new energy. Therefore, the ultra-short-term forecasting method of small hydropower generation based on grid meteorological data is studied. Summary of the invention

[0004] The technical problem of the present invention is that small hydropower is generally far away from the load center of the power grid and has poor regulation ability. Its power generation is easily affected by meteorological conditions such as rainfall and has large uncertainty. The regional meteorological data of small hydropower is complex, and the existing technology lacks a method to effectively use meteorological data to perform high-precision hydropower power forecasting.

[0005] The purpose of the present invention is to provide a method for ultra-short-term prediction of small hydropower station power generation based on grid meteorological data in response to the above problems. The method uses grid meteorological data as a basis, utilizes correlation analysis to screen grids spatially related to small hydropower generation power, and then introduces a local linear embedding algorithm to further reduce the dimension of the relevant grid meteorological data. Finally, the grid meteorological data after dimension reduction and the historical power generation of small hydropower are substituted as input into the long short-term memory neural network training to construct a high-precision ultra-short-term prediction model for small hydropower generation power that takes into account the influence of the temporal and spatial distribution of rainfall.

[0006] The technical solution of the present invention is an ultra-short-term prediction method for power generation of a small hydropower station based on grid meteorological data, comprising the following steps:

[0007] Step 1: Divide the area where the hydropower station is located into grids and obtain the rainfall data of the grids;

[0008] Step 2: Obtain the historical power generation data of the hydropower station, calculate the correlation coefficient between the historical power generation of the hydropower station and the historical rainfall of each grid, and select the grids with high correlation of the power generation of the hydropower station;

[0009] Step 3: Use the local linear embedding method to reduce the dimension of the grid rainfall data to obtain the reduced-dimensional grid rainfall data;

[0010] Step 4: Use the Long Short-Term Memory (LSTM) artificial neural network to establish a power generation prediction model. The input of the power generation prediction model is the grid rainfall data with reduced dimension, and the output of the power generation prediction model is the power generation of the hydropower station.

[0011] Step 5: Use the dimension-reduced historical grid rainfall data and the historical power generation of the hydropower station as the data set of the power generation prediction model, set the parameters of the power generation prediction model, and use the data set to train and test the power generation prediction model to meet the prediction accuracy requirements;

[0012] Step 6: Use the trained power generation prediction model to perform ultra-short-term prediction of the power generation of the hydropower station.

[0013] Furthermore, step 2 specifically includes:

[0014] Step 2.1: Use the Pearson correlation coefficient to measure the correlation between the power generation of the hydropower station and the rainfall in each grid;

[0015] The area where the hydropower station is located is divided into grids, and the grid rainfall data in the area where the hydropower station is located is represented by matrix X.

[0016]

[0017] E i =[e i1 ,e i2 ,...,e it ] (2)

[0018] Where e ij , i=1,2…n,j=1,2…t represents the rainfall of the i-th grid in the j-th time period, n represents the total number of divided grids, and t represents the total number of time periods; E i , i=1,2…n represents the rainfall data vector of the i-th grid in the t-time period; the calculation formula of the correlation coefficient between the grid rainfall and the power generation of the hydropower station is as follows:

[0019]

[0020] Where r i represents the correlation coefficient between the power generation of the hydropower station and the rainfall of the i-th grid; represents the average rainfall of the ith grid in the t period, p j represents the average power generation of the hydropower station in the jth period, represents the average power generation of the hydropower station in t time periods; the correlation coefficient r i ∈[-1,1], when |r i When | is larger, it indicates that the correlation between the rainfall in the i-th grid and the power generation of the hydropower station is higher;

[0021] Step 2.2: Select the correlation coefficient threshold R t As the screening threshold of the grid with large correlation, |r i |≥R t The grids are selected and the rainfall data X′ of the grids with high correlation with the power generation of the hydropower station is obtained.

[0022] X′=[E f ,E g ,…,E h ] T (4)

[0023] Where E f 、E g 、E h They are respectively the rainfall data vectors of the f-th grid, the g-th grid, and the h-th grid in the t time period represented in the grid rainfall data matrix X before screening.

[0024] Preferably, the area where the hydropower station is located is divided into grids, and the unit grids are square areas with a side length of 3 kilometers.

[0025] Preferably, the screening threshold of the grid is R t =0.6.

[0026] Furthermore, step 3 includes the following sub-steps:

[0027] Step 3.1: Calculate the similarity measure of regional rainfall sections;

[0028] Column vector [e 1j ,…e mj ] T , j = 1, 2…t represents the rainfall in the jth time period of the selected grid with a large correlation with the power generation of the hydropower station, that is, the rainfall section in the jth time period, denoted as F j , m represents the number of screened grids with high correlation with the power generation of the hydropower station;

[0029] The Euler distance is used as the similarity measure of regional rainfall sections. The calculation formula of the similarity measure is as follows:

[0030] Dij =|F i -F j | (5)

[0031] Where D ij ,i=1,2…t,j=1,2…t represents the rainfall section F i With F j The similarity measurement result of ij The smaller the value, the higher the similarity;

[0032] Step 3.2: Solve the linearized reconstruction coefficient matrix;

[0033] The preset neighbor parameter k will be i , i=1,2…t The k nearest samples are regarded as a neighboring set, denoted as H (i) , solve the coefficients in the reconstruction coefficient matrix M with the goal of minimizing the linearization error function ε. The reconstruction error function ε and the constraints are as follows:

[0034]

[0035]

[0036] Where a ij ,i=1,2…t,j=1,2…t represents the reconstruction coefficient, representing the rainfall section F i and rainfall section F j If the linear relationship between Then let a ij =0;

[0037] Step 3.3: Perform dimensionality reduction processing according to the reconstruction coefficient matrix M and the preset target dimension d;

[0038] Using the reconstruction coefficient matrix M, the optimization problem is constructed with the goal of minimizing the loss function σ, and the grid rainfall dataset Y is obtained after dimension reduction and normalization.

[0039] The mathematical model corresponding to the construction optimization problem is as follows:

[0040]

[0041] Objective function: min[σ] (9)

[0042] Constraints:

[0043]

[0044] Where U i =[e 1i ,…e di ]T Indicates the i-th rainfall section F in the region i =[e 1i ,…e mi ] T The result after dimensionality reduction, U j Indicates the jth rainfall section F in the region j The result after dimensionality reduction, i = 1, 2…t, j = 1, 2…t, d represents the dimension after dimensionality reduction of the number of screened grids m, d <m;I d represents the d-order unit matrix, 0 represents the d-dimensional zero vector; Y=[U1,...,U t ] is the grid rainfall dataset Y after dimension reduction;

[0045] Step 3.4: Determine the optimal target dimension d and nearest neighbor parameter k with the goal of minimizing the residual variance;

[0046] The residual variance μ is used to evaluate the grid rainfall data after dimensionality reduction in step 3.3. The smaller the residual variance μ, the greater the correlation between the data before and after dimensionality reduction. The calculation formula of the residual variance μ is as follows:

[0047] μ=1-ρ 2 (12)

[0048]

[0049] Where μ represents the residual variance, ρ represents the linear correlation coefficient, and D ij represents the rainfall section F before dimensionality reduction i With rainfall section F j The similarity measure of ij Represents the rainfall section U after dimensionality reduction i and rainfall section U j The similarity measure of Represents a set {D ij The average value of all data in}, Denote the set {D′ ij The average value of all data in};

[0050] D′ ij The calculation formula is as follows:

[0051] D′ ij =|U i -U j | (14)

[0052] Step 3.5: Repeat steps 3.2 to 3.4 above, and use formula (12) to evaluate the dimensionality reduction results until the target dimension d and the nearest neighbor parameter k corresponding to the minimum residual variance μ are obtained, which are the optimal parameters; according to the optimal parameters k and d, repeat steps 3.2 to 3.3 to solve and obtain the optimal dimensionality reduction grid rainfall data Y.

[0053] Compared with the prior art, the beneficial effects of the present invention include:

[0054] 1) The method of the present invention establishes an ultra-short-term prediction model for small hydropower power generation that takes into account the influence of the temporal and spatial distribution of rainfall. By reducing the dimension of the model input data, the computational complexity of the power prediction model is reduced, and the computational efficiency and accuracy of the power prediction model are improved;

[0055] 2) The present invention screens out the grid areas with high correlation with the power generation of the hydropower station through correlation analysis, and realizes the ultra-short-term prediction of the power generation of the hydropower station by using the rainfall data of the grids with high correlation;

[0056] 3) The present invention uses a local linear embedding algorithm to reduce the dimension of the screened grid meteorological data, finds rainfall sections with similar rainfall distribution characteristics through similarity calculation, and realizes dimensionality reduction by linearizing the relative position of samples in a small area, and represents grids with similar rainfall characteristics on a time scale as virtual grids, retaining the main characteristics of the original grid rainfall. The dimensionality reduction effect on the grid rainfall data at a specific moment is significant, which greatly reduces the data volume of the grid meteorological data at a single moment used for power generation prediction, so that the training of the power generation prediction model focuses more on key feature data, prevents overfitting in the training of the prediction model, and improves the training efficiency of the prediction model and the accuracy of the power generation prediction result;

[0057] 4) The present invention adopts a long short-term memory artificial neural network to establish a power generation prediction model, which effectively realizes the hydropower power prediction in a long time window, fully considers the influence of the power generation and rainfall data before the small hydropower prediction time on the power generation prediction value at the current time, and solves the influence of the time lag effect and cumulative effect on the small hydropower power generation to a large extent, especially the prediction effect of ultra-short-term power generation is good. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0059] Figure 1 The figure is a flow chart of a method for ultra-short-term prediction of power generation of a small hydropower station according to an embodiment of the present invention.

[0060] Figure 2 It is a schematic diagram of a process of selecting grids with a high correlation with the power generation capacity of a small hydropower station according to an embodiment of the present invention.

[0061] Figure 3 Schematic diagram of the flow of the local linear embedding algorithm according to an embodiment of the present invention.

[0062] Figure 4 The figure is a flow chart of the LSTM power generation prediction model training process according to an embodiment of the present invention.

[0063] Figure 5 Schematic diagram of an LSTM unit according to an embodiment of the present invention. DETAILED DESCRIPTION

[0064] like Figure 1 As shown, the ultra-short-term prediction method of power generation of a small hydropower station based on grid meteorological data includes the following steps: Step 1: Divide the area where the hydropower station is located into grids, and obtain the rainfall data of the grids;

[0065] In the embodiment, grid rainfall data of small hydropower areas and small hydropower historical power generation time series from May to October every year in the past three years are collected as model training data, and the time interval between grid rainfall data and small hydropower historical power generation time series is 15 minutes.

[0066] Step 2: Obtain the historical power generation data of the hydropower station, calculate the correlation coefficient between the historical power generation of the hydropower station and the historical rainfall of each grid, and select the grids with high correlation of the power generation of the hydropower station, such as Figure 2 shown.

[0067] Step 2.1: Use the Pearson correlation coefficient to measure the correlation between the power generation of the hydropower station and the rainfall in each grid;

[0068] The area where the hydropower station is located is divided into square grids with a side length of 3 kilometers. The grid rainfall data in the area where the hydropower station is located is represented by matrix X.

[0069]

[0070] E i =[e i1 ,e i2 ,...,e it ] (2)

[0071] Where e ij , i=1,2…n,j=1,2…t represents the rainfall of the i-th grid in the j-th time period, n represents the total number of divided grids, and t represents the total number of time periods; E i , i=1,2…n represents the rainfall data vector of the i-th grid in the t time period.

[0072] The calculation formula of the correlation coefficient between grid rainfall and hydropower station power generation is as follows:

[0073]

[0074] Where r i represents the correlation coefficient between the power generation of the hydropower station and the rainfall of the i-th grid; represents the average rainfall of the ith grid in the t period, p j represents the average power generation of the hydropower station in the jth period, represents the average power generation of the hydropower station in t time periods; the correlation coefficient r i ∈[-1,1], when |r i The larger | is, the higher the correlation between the rainfall in the i-th grid and the power generation of the hydropower station is.

[0075] Step 2.2: Select the correlation coefficient threshold R t As the screening threshold of the grid with large correlation, |r i |≥R t The grid, R t =0.6, and the precipitation data X′ of the grids with high correlation with the power generation of the hydropower station is obtained.

[0076]

[0077] Where E f 、E g 、E h They are respectively the rainfall data vectors of the f-th grid, the g-th grid, and the h-th grid in the t time period represented in the grid rainfall data matrix X before screening.

[0078] Step 3: Use the local linear embedding algorithm to reduce the dimension of the grid rainfall data to obtain the reduced dimension grid rainfall data, such as Figure 3 shown.

[0079] Step 3.1: Calculate the similarity measure of regional rainfall sections;

[0080] Column vector [e 1j ,…e mj ] T , j = 1, 2…t represents the rainfall in the jth time period of the selected grid with a large correlation with the power generation of the hydropower station, that is, the rainfall section in the jth time period, denoted as F j , m represents the number of screened grids with high correlation with the power generation of the hydropower station.

[0081] The Euler distance is used as the similarity measure of regional rainfall sections. The calculation formula of the similarity measure is as follows:

[0082] D ij =|F i -F j| (5)

[0083] Where D ij ,i=1,2…t,j=1,2…t represents the rainfall section F i With F j The similarity measurement result of ij The smaller the value, the higher the similarity.

[0084] Step 3.2: Solve the linearized reconstruction coefficient matrix;

[0085] The preset neighbor parameter k will be i , i=1,2…t The k nearest samples are regarded as a neighboring set, denoted as H (i) , solve the coefficients in the reconstruction coefficient matrix M with the goal of minimizing the linearization error function ε. The reconstruction error function ε and the constraints are as follows:

[0086]

[0087]

[0088] a ij ,i=1,2…t,j=1,2…t represents the reconstruction coefficient, representing the rainfall section F i and rainfall section F j If the linear relationship between Then let a ij =0.

[0089] Step 3.3: Perform dimensionality reduction processing according to the reconstruction coefficient matrix M and the preset target dimension d;

[0090] Using the reconstruction coefficient matrix M, the optimization problem is constructed with the goal of minimizing the loss function σ, and the grid rainfall data set Y normalized after dimension reduction is obtained. The mathematical model of the corresponding optimization problem is as follows:

[0091]

[0092] Objective function: min[σ] (9)

[0093] Constraints:

[0094]

[0095] Where U i =[e 1i ,…e di ] T Indicates the i-th rainfall section F in the region i =[e 1i ,…e mi ]T The result after dimensionality reduction, U j Indicates the jth rainfall section F in the region j The result after dimensionality reduction, i = 1, 2…t, j = 1, 2…t, d represents the dimension after dimensionality reduction of the number of screened grids m, d <m;I d represents the d-order unit matrix, 0 represents the d-dimensional zero vector; Y=[U1,...,U t ] is the grid rainfall dataset Y after dimension reduction.

[0096] Step 3.4: Determine the optimal target dimension d and nearest neighbor parameter k with the goal of minimizing the residual variance;

[0097] The residual variance μ is used to evaluate the grid rainfall data after dimensionality reduction. The smaller the residual variance μ, the greater the correlation between the data before and after dimensionality reduction. The calculation formula of the residual variance μ is as follows:

[0098] μ=1-ρ 2 (12)

[0099]

[0100] Where μ represents the residual variance, ρ represents the linear correlation coefficient, and D ij represents the rainfall section F before dimensionality reduction i With rainfall section F j The similarity measure of ij Represents the rainfall section U after dimensionality reduction i and rainfall section U j The similarity measure of Represents a set {D ij The average value of all data in}, Denote the set {D′ ij The average value of all data in};

[0101] D′ ij The calculation formula is as follows:

[0102] D′ ij =|U i -U j | (14)

[0103] Step 3.5: Repeat the above steps 3.2 to 3.4, and use formula (12) to evaluate the dimensionality reduction results until the target dimension d and the nearest neighbor parameter k corresponding to the minimum residual variance μ are obtained, which are the optimal parameters; according to the optimal parameters k and d, the optimal dimensionality reduction virtual grid rainfall data Y is obtained.

[0104] Step 4: Use the long short-term memory artificial neural network to establish an LSTM power generation prediction model. The input of the LSTM power generation prediction model is the reduced-dimensional grid rainfall data and the historical power generation data of small hydropower stations. The output of the LSTM power generation prediction model is the power generation of the hydropower station.

[0105] Step 5: Use the reduced-dimensional historical grid rainfall data and the historical power generation of the hydropower station as the data set of the LSTM power generation prediction model, set the LSTM power generation prediction model parameters, and use the data set to train and test the LSTM power generation prediction model to meet the prediction accuracy requirements, such as Figure 4 shown.

[0106] Step 5.1: Set the hyperparameters of the LSTM ANN.

[0107] The set hyperparameters include: the number of input nodes N1, the number of hidden nodes N2, the number of output nodes N3, the learning rate α, the error threshold cost, the maximum number of iterations s, and the number of LSTM nuclei z; in the embodiment, N1=d+1, z<3.

[0108] Step 5.2: Initialize weight bias;

[0109] Example LSTM Figure 5 As shown, the mathematical model of LSTM is as follows:

[0110]

[0111]

[0112]

[0113]

[0114]

[0115] h t =o t ×tanh(C t ) (20)

[0116] Where x t represents the input of the neural network at time t; h t-1 represents the output of the neural network at time t-1; h t represents the output of the neural network at time t; C t-1 , C t represents the unit state output of the neural network at time t-1 and t; f t 、i t , o t Represent the output values ​​of the forget gate, input gate, and output gate respectively. Represents the intermediate quantity of the state of the input unit; Respectively represent x t The weight matrix of the forget gate, input gate, output gate, and unit state; h t-1 The weight matrix of the forget gate, input gate, output gate, and unit state; b f , b i , b o , b C Represent the bias vectors of the forget gate, input gate, output gate, and unit state respectively.

[0117] σ() and tanh() are both activation functions, where σ() represents the Sigmoid function and tanh() represents the hyperbolic tangent function. The calculation formula is as follows:

[0118]

[0119]

[0120] Step 5.3: Forward calculation;

[0121] The forward calculation formula of the long short-term memory artificial neural network is shown in equations (15) to (20). The forward calculation obtains the output value of each neuron, that is, f t 、i t , o t , h t The values ​​of the five vectors.

[0122] Through forward calculation, LSTM can more effectively determine which information is forgotten and which is retained, and process long time series more efficiently.

[0123] Step 5.4: Update the network weights and biases according to the loss function η;

[0124] The loss function η represents the error size of the network. The smaller its value is, the more accurate the model is, and vice versa. The calculation formula of the loss function is as follows:

[0125]

[0126] W′=W-α·2[h t -h′ t ] (twenty four)

[0127] b′=b-α·2[h t -h′ t ] (25)

[0128] Where L represents the total number of training samples, h trepresents the predicted value of network power generation of the tth training sample, h′ t represents the actual power generation corresponding to the t-th training sample, α is the learning rate, W and b represent the weight and bias before updating, respectively, and W′ and b′ represent the weight and bias after updating, respectively.

[0129] Changing the weight bias value can change the size of η, and using the gradient descent algorithm to quickly find the appropriate weights and biases to minimize η to reach the error threshold cost and speed up the network convergence. The gradient descent algorithm of the embodiment refers to the gradient descent algorithm disclosed in the paper "A Review of Gradient Descent Algorithm Research" by Li Xingyi and Yue Yang published in "Software Engineering" No. 2, 2020. Step 5.5: Repeat steps 5.3 and 5.4 until the number of iterations is equal to s or the error is less than the error threshold cost, and stop training. Step 6: Use the trained LSTM power generation prediction model to perform ultra-short-term prediction of hydropower station power generation.

[0130] The implementation results show that the method of the present invention significantly improves the accuracy of small hydropower generation power forecasting, and has positive significance for the formulation of power grid generation plans and the promotion of new energy consumption.

Claims

1. An ultra-short-term prediction method for power generation of small hydropower stations based on grid meteorological data, characterized in that: The following steps are involved: Step 1: Divide the area where the hydropower station is located into grids and obtain the rainfall data of the grids; Step 2: Obtain the historical power generation data of the hydropower station, calculate the correlation coefficient between the historical power generation of the hydropower station and the historical rainfall of each grid, and select the grids with high correlation of the power generation of the hydropower station; Step 3: Use the local linear embedding method to reduce the dimension of the grid rainfall data to obtain the reduced-dimensional grid rainfall data; Step 4: Use long short-term memory artificial neural network to establish power generation prediction model. The input of power generation prediction model is grid rainfall data with reduced dimension and historical power generation of small hydropower. The output of power generation prediction model is power generation of hydropower station. Step 5: Use the dimension-reduced historical grid rainfall data and the historical power generation of the hydropower station as the data set of the power generation prediction model, set the parameters of the power generation prediction model, and use the data set to train and test the power generation prediction model to meet the prediction accuracy requirements; Step 6: Use the trained power generation prediction model to perform ultra-short-term prediction of the power generation of the hydropower station.

2. The method for ultra-short-term prediction of power generation of a small hydropower station according to claim 1, characterized in that: Step 2 specifically includes: Step 2.1: Use the Pearson correlation coefficient to measure the correlation between the power generation of the hydropower station and the rainfall in each grid; The area where the hydropower station is located is divided into grids, and the grid rainfall data in the area where the hydropower station is located is represented by matrix X. AND i =[and i1 ,And i2 ,...,And it ] (2) Where e ij , i=1,2…n,j=1,2…t represents the rainfall of the i-th grid in the j-th time period, n represents the total number of divided grids, and t represents the total number of time periods; E i , i=1,2…n represents the rainfall data vector of the i-th grid in the t-time period; the calculation formula of the correlation coefficient between the grid rainfall and the power generation of the hydropower station is as follows: Where r i represents the correlation coefficient between the power generation of the hydropower station and the rainfall of the i-th grid; represents the average rainfall of the ith grid in the t period, p j represents the average power generation of the hydropower station in the jth period, represents the average power generation of the hydropower station in t time periods; Correlation coefficient r i ∈[-1,1], when |r i When | is larger, it indicates that the correlation between the rainfall in the i-th grid and the power generation of the hydropower station is higher; Step 2.2: Select the correlation coefficient threshold R t As the screening threshold of the grid with large correlation, |r i |≥R t The grids are selected and the rainfall data X′ of the grids with high correlation with the power generation of the hydropower station is obtained. X′=[E f ,AND g ,…,AND h ] T (4) Where E f 、E g 、E h They are respectively the rainfall data vectors of the f-th grid, the g-th grid, and the h-th grid in the t time period represented in the grid rainfall data matrix X before screening.

3. The method for ultra-short-term prediction of power generation of a small hydropower station according to claim 2 is characterized in that: The area where the hydropower station is located is divided into grids, and the unit grid is a square area with a side length of 3 kilometers.

4. The method for ultra-short-term prediction of power generation of a small hydropower station according to claim 2 is characterized in that: The screening threshold of the grid is R t =0.

6.

5. The method for ultra-short-term prediction of power generation of a small hydropower station according to claim 2, characterized in that: Step 3 includes the following sub-steps: Step 3.1: Calculate the similarity measure of regional rainfall sections; Column vector [e 1j ,…e mj ] T , j = 1, 2…t represents the rainfall in the jth time period of the selected grid with a large correlation with the power generation of the hydropower station, that is, the rainfall section in the jth time period, denoted as F j , m represents the number of screened grids with high correlation with the power generation of the hydropower station; The Euler distance is used as the similarity measure of regional rainfall sections. The calculation formula of the similarity measure is as follows: D ij =|F i -F j | (5) Where D ij ,i=1,2…t,j=1,2…t represents the rainfall section F i With F j The similarity measurement result of ij The smaller the value, the higher the similarity; Step 3.2: Solve the linearized reconstruction coefficient matrix; The preset neighbor parameter k will be i , i=1,2…t The k nearest samples are regarded as a neighboring set, denoted as H (i) , solve the coefficients in the reconstruction coefficient matrix M with the goal of minimizing the linearization error function ε. The reconstruction error function ε and the constraints are as follows: Where a ij ,i=1,2…t,j=1,2…t represents the reconstruction coefficient, representing the rainfall section F i and rainfall section F j If the linear relationship between Then let a ij =0; Step 3.3: Perform dimensionality reduction processing according to the reconstruction coefficient matrix M and the preset target dimension d; Using the reconstruction coefficient matrix M, the optimization problem is constructed with the goal of minimizing the loss function σ, and the grid rainfall dataset Y is obtained after dimension reduction and normalization. The mathematical model corresponding to the construction optimization problem is as follows: Objective function: min[σ] (9) Constraints: Where U i =[e 1i ,…e di ] T Indicates the i-th rainfall section F in the region i =[e 1i ,…e mi ] T The result after dimensionality reduction, U j Indicates the jth rainfall section F in the region j The result after dimensionality reduction, i = 1, 2…t, j = 1, 2…t, d represents the dimension after dimensionality reduction of the number of screened grids m, d <m;I d represents the d-order unit matrix, 0 represents the d-dimensional zero vector; Y=[U1,...,U t ] represents the grid rainfall dataset Y after dimension reduction; Step 3.4: Determine the optimal target dimension d and nearest neighbor parameter k with the goal of minimizing the residual variance; The residual variance μ is used to evaluate the grid rainfall data after dimensionality reduction in step 3.

3. The smaller the residual variance μ, the greater the correlation between the data before and after dimensionality reduction. The calculation formula of the residual variance μ is as follows: μ=1-ρ 2 (12) Where μ represents the residual variance, ρ represents the linear correlation coefficient, and D ij represents the rainfall section F before dimensionality reduction i With rainfall section F j The similarity measure of ij Represents the rainfall section U after dimensionality reduction i and rainfall section U j The similarity measure of Represents a set {D ij The average value of all data in}, Denotes the set {D′ ij The average value of all data in}; D′ ij The calculation formula is as follows: D′ ij =|U i -U j | (14) Step 3.5: Repeat steps 3.2 to 3.4 above, and use formula (12) to evaluate the dimensionality reduction results until the target dimension d and the nearest neighbor parameter k corresponding to the minimum residual variance μ are obtained, which are the optimal parameters; according to the optimal parameters k and d, repeat steps 3.2 to 3.3 to solve and obtain the optimal dimensionality reduction grid rainfall data Y.

6. The method for ultra-short-term prediction of power generation of a small hydropower station according to claim 5, characterized in that: Step 5 includes the following sub-steps: Step 5.1: Set the hyperparameters of the LSTM ANN. The hyperparameters set include: number of input nodes N1, number of hidden nodes N2, number of output nodes N3, learning rate α, error threshold cost, number of nuclei z, and maximum number of iterations s; Step 5.2: Initialize weight bias; The mathematical model of the memory unit of the long short-term memory artificial neural network is as follows: h t =o t ×tanh(C t ) (20) Where x t represents the input of the neural network at time t; h t-1 represents the output of the neural network at time t-1; h t represents the output of the neural network at time t; C t-1 , C t represents the unit state output of the neural network at time t-1 and t; f t 、i t , o t Represent the output values ​​of the forget gate, input gate, and output gate respectively. Represents the intermediate quantity of the state of the input unit; Respectively represent x t The weight matrix of the forget gate, input gate, output gate, and unit state; h t-1 The weight matrix of the forget gate, input gate, output gate, and unit state; b f 、b i 、b o 、b C Represent the bias vectors of forget gate, input gate, output gate, and unit state respectively; σ() and tanh() are both activation functions, where σ() represents the Sigmoid function and tanh() represents the hyperbolic tangent function. The calculation formula is as follows: Step 5.3: Forward calculation; The forward calculation formula of the long short-term memory artificial neural network is shown in equations (15) to (20). The forward calculation obtains the output value of each neuron, that is, f t 、i t , o t , h t The values ​​of the five vectors; Through forward calculation, long short-term memory artificial neural networks can more effectively determine which information is forgotten and which information is retained, and process long time series more efficiently; Step 5.4: Update the network weights and biases according to the loss function η; The loss function η represents the error size of the network. The smaller its value is, the more accurate the model is, and vice versa. The loss function is calculated as follows: W′=W-α…2[h t -h′ t ] (24) b′=b-α·2[h t -h′ t ] (25) Where L represents the total number of training samples, h t represents the predicted value of network power generation of the tth training sample, h′ t represents the actual power generation corresponding to the t-th training sample, α is the learning rate, W and b represent the weight and bias before updating, respectively, and W′ and b′ represent the weight and bias after updating, respectively; Changing the weight bias value can change the size of η and find the appropriate weight and bias to minimize η to reach the error threshold; Step 5.5: Repeat steps 5.3 and 5.4 until the number of iterations is equal to s or the error is less than the error threshold cost, then stop training.

Images