Wind power cluster output power prediction method and wind power cluster active hierarchical control method based on time-space characteristics

Abstract

The invention discloses a wind power cluster output power prediction method and a wind power cluster active power hierarchical control method based on time-space characteristics, which can enhance the control stability and optimization performance of a system while effectively improving the prediction effect by combining CEEMDAN, GA-BP, an active power hierarchical control method and an improved NSGA-II algorithm.

Description

Technical Field

[0001] The present invention belongs to the technical field of wind farms, and specifically relates to a method for predicting the output power of a wind power cluster based on spatio-temporal characteristics and a method for hierarchical active power control of a wind power cluster. Background Art

[0002] In recent years, with the continuous increase in wind power penetration, its volatility and uncertainty have brought huge challenges to the dispatching operation control of power systems. Currently, most wind power is connected to the power grid in the form of large-scale wind power clusters (WPC). Therefore, on the premise of considering power prediction information, how to perform reasonable and effective active power control on wind power clusters and improve the acceptance capacity of the power grid for large-scale wind power clusters has become one of the urgent problems to be solved.

[0003] As a natural resource with strong volatility and instability, how to improve the accuracy of wind power prediction and process abnormal data during the power generation process is an important task in wind power engineering. At the same time, a large amount of Numerical Weather Prediction (NWP) data can support the prediction model, but too much data input may lead to overfitting of the model or even non-convergence, and the volatility and instability of historical power supply data make it difficult to be fully exploited by the model. Therefore, data preprocessing must be carried out before prediction to minimize the impact of raw data errors and irregularities on the model accuracy.

[0004] Traditional wind power prediction methods have their own applicability and limitations. They usually focus on the power prediction of a single wind farm and mainly rely on the meteorological prediction data, historical wind power, and measured meteorological data of the wind farm. By analyzing the time evolution law of historical and future data or the mapping relationship between meteorological data and power data, the wind speed or power is predicted. Although this method is relatively simple, it only considers limited information sources and ignores the spatial correlation between geographical terrain, meteorology, environment, and wind turbines within the wind farm or cluster area, resulting in limitations in the prediction system and the urgent need to improve and optimize the prediction accuracy. Summary of the Invention

[0005] Objective of the Invention: To solve the problems of how to conduct reasonable and effective active power control for a wind power cluster, improve the acceptance capacity of the power grid for a large-scale wind power cluster, how to improve the accuracy of wind power prediction and process abnormal data during the power generation process, and how to address the limitations of existing wind power prediction methods that ignore the geographical terrain, meteorology, environment, and spatial correlation between wind turbines within a wind farm or cluster area, resulting in limitations in the prediction system and the urgent need to improve and optimize the prediction accuracy. The present invention proposes a method for predicting the output power of a wind power cluster based on spatio-temporal characteristics and a method for hierarchical active power control of a wind power cluster. By combining CEEMDAN, GA-BP, a method for hierarchical active power control, and an improved NSGA-II algorithm, it can effectively improve the prediction effect while enhancing the control stability and optimization performance of the system.

[0006] Technical Solution: A method for predicting the output power of a wind power cluster based on spatio-temporal characteristics, comprising the following steps:

[0007] Step 1: Interpolate and repair abnormal data in the historical power supply data of the wind power cluster and the historical meteorological data of the wind power cluster, and normalize the interpolated and repaired historical power supply data of the wind power cluster and the historical meteorological data of the wind power cluster;

[0008] Analyze the historical power supply data of the wind power cluster using the maximum information coefficient MIC and the cross-correlation function CFF, and screen out wind farm groups with strong correlation from them;

[0009] Step 2: Use the CEEMDAN decomposition algorithm to decompose the historical power supply data of the wind farm groups with strong correlation screened out into intrinsic mode functions, and obtain the decomposed historical power supply data;

[0010] Step 3: Use the normalized historical meteorological data of the wind power cluster and the decomposed historical power supply data to train the GA-BP neural network model, and obtain a trained GA-BP neural network model;

[0011] Step 4: Collect relevant data information required for wind power prediction, input the relevant data information into the trained GA-BP neural network model, and obtain the wind power cluster output power prediction result.

[0012] Further, the specific operation of using the maximum information coefficient MIC and the cross-correlation function CFF to analyze the historical power supply data of the wind power cluster and screen out wind farm group meteorological data with strong correlation includes:

[0013] Analyze the cross-correlation of the historical power supply data of wind farm X and wind farm Y in the wind power cluster using the cross-correlation function CFF, expressed as:

[0014]

[0015] In the formula, X represents the meteorological data of the historical power supply data sequence of wind farm X, and is described by the sequence X = [x 1 , x 2 ,..., x n , where x n is the nth data of the historical power supply data sequence of wind farm X; Y represents the meteorological data of the historical power supply data sequence of wind farm Y, and is described by the sequence Y = [y 1 , y 2 ,..., y n , where y n is the nth data of this sequence; C CFF (τ, X, Y) represents the cross-correlation function between sequence X and sequence Y; τ is the time shift interval. When τ = 0, C CFF (0, X, Y) = R XY , representing the correlation coefficient of the power supply of wind farms X and Y. The correlation relationship between the power supplies of wind farms X and Y increases / decreases with the increase / decrease of R XY .

[0016] The maximum information coefficient MIC is used to measure the correlation between the historical power supply data of wind farms X and Y in the wind power cluster, which is expressed as:

[0017]

[0018] In the formula, p(X, Y) represents the joint probability distribution of sequence X and sequence Y, a and b are the number of data in the two sequences respectively, and B represents the complexity limit parameter;

[0019] Combining the calculated results of C CFF (τ, X, Y) and MIC[X; Y], by setting a threshold to screen out the wind farm groups where both C CFF (τ, X, Y) and MIC[X; Y] are higher than the set threshold as the wind farm groups with strong correlation.

[0020] Furthermore, the GA-BP neural network model is obtained by optimizing the BP neural network using the genetic algorithm;

[0021] In the BP neural network, the output data of the BP neural network is compared with the preset data value, and the error E(n) between the two is expressed as:

[0022]

[0023] In the formula, T k is the output data of the BP neural network, Ok is a preset data value, L is a set of data points, and n is the number of data points; if the error E(n) is greater than the threshold, the output data of the BP neural network is backpropagated in the order of the output layer, hidden layer, and input layer. While this process is ongoing, the weights are updated based on the error E(n), expressed as:

[0024] W ij (n + 1)=ΔW kj (n + 1)+W kj (n)

[0025] Among them, W ij is the weight from the input layer to the hidden layer, and W kj is the weight from the hidden layer to the output layer; ΔW kj is the difference in the weights from the hidden layer to the output layer between data point n + 1 and data point n;

[0026] The optimization of the BP neural network using the genetic algorithm specifically includes the following operations:

[0027] Step ① Encode the initial values of the input features and thresholds in the BP neural network;

[0028] Step ② Determine the fitness function based on the error E(n), expressed as:

[0029]

[0030] Among them, l is the total number of samples, G is the number of data points in the output layer, is the output prediction value, is the true output value;

[0031] Step ③ In each iteration process, sort the individuals in the population according to the fitness function, and generate a new population of wind power data through crossover and mutation operations;

[0032] Step ④ Return to Step ② for loop calculation, and stop the calculation only when the individuals in the final population meet the expected requirements, and output the wind power prediction result.

[0033] Furthermore, it also includes the following steps:

[0034] Use the upper and lower bound estimation model based on the long short-term memory network to obtain the upper and lower bound estimations, and generate a prediction interval according to the upper and lower bound estimations;

[0035] Use the prediction interval to perform confidence evaluation and correction on the trained GA - BP neural network model to obtain a refined wind power prediction model;

[0036] The non-dominated sorting genetic algorithm is used to iteratively solve the improved wind power prediction model to obtain the optimal set P u ;

[0037] Take the average value of the optimal set P u as the final wind power cluster output power prediction result.

[0038] Furthermore, the non-dominated sorting genetic algorithm is used to iteratively solve the improved wind power prediction model to obtain the optimal set P u , which is obtained according to the following steps:

[0039] S1: Divide the meteorological data and the decomposed historical power supply data of the wind power cluster into a training set, a validation set, and a test set. The training set is used to adjust the parameters of the long short-term memory network, the validation set is used to determine the structure of the long short-term memory network, and the test set is used to verify the generalization ability of the upper and lower bound estimation model;

[0040] S2: For each individual i, the parameters in the long short-term memory network are randomly and uniformly sampled for initialization;

[0041] S3: Each individual i can be regarded as a solution to the upper and lower bound estimation model, and the error index and the PINAW index are calculated according to the following formula;

[0042]

[0043]

[0044] Among them, E PIE is the error index, N represents the number of individuals, and W is the value range of the prediction interval; y i is the value of individual i, y i,t and y i,u are the predicted upper and lower bounds; W PINA is the PINAW index;

[0045] ① Suppose IT represents the current iteration time, set it to 1;

[0046] ② For each individual, let t = 1, E PIE = 0, W PINA = 0;

[0047] ③ Suppose the total number of samples in the training set is n, and the input of the long short-term memory network is the previous m observed values. According to the forward propagation process of the long short-term memory network, the upper and lower bounds of y t+m are constructed. If y t+m,u ≤y t+m ≤y t+m,l , let E t+m be 0, otherwise where w represents the penalty factor for adjusting the error metric, and y t+m,l , y t+m,u represent the upper and lower bounds of y t+m , and E t+m is the prediction error that measures the distance of the predicted value y t+m from the nearest boundary of the prediction interval [y t+m,u , y t+m,l ;

[0048] ④ The width calculation formula of the prediction interval is

[0049] ⑤ E PIE ' = E PIE + E t+m , and W PINA ' = W PINA + d t+m ;

[0050] ⑥ If t = n - m, perform S4; otherwise, t = t + 1, and return to ③ of S3;

[0051] S4: Combine the elite strategy to obtain an overall population number P u ;

[0052] S5: Check the dominance relationship between each individual;

[0053] S6: Divide the population into different levels according to the dominance relationship, and then obtain a new population P n ;

[0054] S7: Perform selection, crossover, mutation, and competitive learning on P n to obtain an offspring population P 0 ;

[0055] S8: IT = IT + 1. If IT is greater than the maximum iteration time, the algorithm stops and records the result of P u ; otherwise, return to ② of S3.

[0056] The present invention discloses an active hierarchical control method for a wind power cluster based on spatio-temporal characteristics, including the following steps:

[0057] Divide the control layer into a cluster layer, a field group layer, and a sub-field layer;

[0058] Predict the output power of the wind power cluster to obtain a prediction result;

[0059] Determine the overall scheduling plan of the wind power cluster with the goals of minimizing the overall power output fluctuation of the wind power cluster and maximizing the overall output of the wind power cluster;

[0060] Allocate the overall scheduling plan of the wind power cluster according to the prediction of each individual wind farm to determine the initial individual wind farm scheduling plan;

[0061] Compare the initial individual wind farm scheduling plan with the short-term prediction of each individual wind farm, perform dynamic grouping according to the comparison results, and issue the field group scheduling instruction;

[0062] According to the field group scheduling instruction, with the goal of maximizing the output of the wind farm and minimizing the output fluctuation, determine the optimized output of the wind farm;

[0063] Among them, the prediction of the output power of the wind power cluster to obtain the prediction result is to predict the output power of the wind power cluster by using a method for predicting the output power of a wind power cluster based on spatio-temporal characteristics described in any one of claims 1 to 5 to obtain the prediction result.

[0064] Furthermore, with the goal of minimizing the overall power output fluctuation of the wind power cluster and maximizing the overall output of the wind power cluster, determine the overall scheduling plan of the wind power cluster. The specific operations include:

[0065]

[0066] Among them, δ 1 and δ 2 are the weights for suppressing the output fluctuation of the wind power cluster and maximizing the output respectively, δ 1 ≥0, δ 2 ≥0, δ 1 +δ 2 =1, P Z,t is the existing optimized scheduling power of the wind power cluster at time t on a 15-minute time scale, n is the total number of wind farms in the wind power cluster, R i,t-1 is the output power of wind farm i at time t-1 on a 15-minute time scale, is the power prediction value of wind farm i at time t on a 15-minute time scale;

[0067] The corresponding constraint conditions include:

[0068] The output ramp constraint of the wind power cluster, expressed as:

[0069]

[0070] Among them, D Z is the ramp rate limit of the wind power cluster on a 15-minute time scale, P cluz is the total installed capacity of the wind power cluster;

[0071] The power constraint of the wind power cluster, expressed as:

[0072]

[0073] The wind power cluster tracking plan constraint, expressed as:

[0074] P Z,t ≤P wet,t

[0075] In the formula, P wet,t is the grid planned power;

[0076] The wind power cluster transmission line capacity constraint, expressed as:

[0077] P Z,t ≤P FMAX

[0078] In the formula, P FMAX is the maximum transmission line capacity of the wind power cluster.

[0079] Furthermore, the overall scheduling plan of the wind power cluster is allocated according to the prediction of each single farm, and the initial single-farm scheduling plan is determined, expressed as:

[0080]

[0081] Among them, P i,t is the initial scheduling power of wind farm i at time t on a 15-minute time scale;

[0082] The initial single-farm scheduling plan is compared with the short-term prediction of each single farm, and dynamic grouping is carried out according to the comparison result, expressed as:

[0083]

[0084] Among them, Q i,t is the grouping index. When Q i,t =1, it means that wind farm i is a high-output wind farm at time t. When Q i,t =0, it means that wind farm i is a low-output wind farm at time t. is the power prediction value of the wind farm on a 5-minute time scale.

[0085] Furthermore, the field group scheduling instruction is issued, expressed as:

[0086]

[0087] ΔP D,t,s5 =-ΔP S,t,s5

[0088]

[0089] P i,t,s5 =P i,t +ΔP i,t,s5

[0090] Among them, ΔP S,t,s5 is the additional power that can be generated by multiple wind farm groups in the clustering on a 5 - minute time scale. ΔP D,t,s5 is the additional power that can be generated by under - generating wind farm groups in the clustering on a 5 - minute time scale. m is the number of multiple - generating wind farms, n is the total number of wind farms, =P i,t,s5 is the fine - tuning amount of the dispatching power of wind farm i on a 5 - minute time scale, is the excess amount of the power prediction value of wind farm i over the initial dispatching power on a 5 - minute time scale. P i,t,s5 is the adjusted dispatching power value of wind farm i on a 5 - minute time scale.

[0091] Furthermore, taking the maximum output of wind farms and the minimum output fluctuation as the goal, it is expressed as:

[0092]

[0093] Among them, δ i1 , δ i2 are the weights for suppressing the output fluctuation and maximizing the output of wind farm i respectively. P op i,t,s5 is the optimal dispatching power of wind farm i at time t on a 5 - minute time scale. R i,t-1,s5 is the output power of wind farm i at time t - 1 on a 5 - minute time scale;

[0094] The constraint conditions include:

[0095] 1) The output ramp - up constraint of the wind farm, which is expressed as:

[0096]

[0097] Among them, D iZ,s5 is the ramp - rate limit value of wind farm i on a 5 - minute time scale. P iz is the installed capacity of wind farm i.

[0098] 2) The power constraint of the wind farm, which is expressed as:

[0099]

[0100] 3) The tracking plan constraint of the wind farm, which is expressed as:

[0101]

[0102] 4) The transmission line capacity constraint of the wind farm, which is expressed as:

[0103]

[0104] Among them, P fimaxis the maximum transmission line capacity of wind farm i.

[0105] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0106] (1) The method of the present invention takes into account the errors and losses in the operation of the wind farm and the data acquisition process, avoids the model divergence caused by too high data redundancy, and decomposes the original power supply data using the CEEMDAN decomposition algorithm into intrinsic mode functions, which can facilitate the prediction model to dig deeper into the original data;

[0107] (2) The method of the present invention adds a genetic algorithm to optimize the BP neural network, can eliminate errors by means of the fitness function and cyclic processing of wind power-related data information, and has great superiority in the identification and elimination ability of abnormal data and the accuracy of wind power prediction;

[0108] (3) The method of the present invention considers the coordinated complementarity problem between wind farms at different time scales, establishes a coordinated field group layer between wind farms, and constructs a multi-time-scale hierarchical scheduling model by combining the cluster layer and the sub-field layer, which can effectively balance the curtailment volume, the coordinated output of wind farm energy storage, and the average wind power volatility, and improve the wind power consumption level and the coordinated output of wind farm energy storage;

[0109] (4) The method of the present invention introduces a competitive learning mechanism, and the improved NSGA-II algorithm proposed can effectively achieve multi-objective optimization, provide Pareto optimal solutions with diversity and convergence, and can effectively prove that the proposed method can effectively improve the prediction effect. Description of the Drawings

[0110] Figure 1 is the main flowchart of a method for predicting the output power of a wind power cluster based on spatio-temporal characteristics and a method for hierarchical active power control of a wind power cluster proposed by the present invention;

[0111] Figure 2 is the flowchart of the method for hierarchical active power control of a wind power cluster proposed by the present invention;

[0112] Figure 3 is the flowchart for solving the improved non-dominated sorting genetic algorithm. Detailed Embodiments

[0113] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be further described below in conjunction with the drawings and embodiments.

[0114] Embodiment 1:

[0115] This embodiment proposes a method for predicting the output power of a wind power cluster based on spatio-temporal characteristics, which mainly includes the following steps:

[0116] S1_1: Each meteorological data has a certain reasonable range. By analyzing the historical power supply data of the wind power cluster and the meteorological data (Numerical Weather Prediction, hereinafter referred to as NWP data), the data outside the reasonable range can be considered as overly abnormal values.

[0117] For the types of NWP data of the wind power cluster, the reasonable ranges of each data item are listed in Table 1.

[0118] Table 1 Reasonable ranges of NWP data

[0119] Meteorological data Reasonable range Wind speed 0 - 60 m / s Wind direction 0-360° Air temperature -60℃-60℃ Air pressure 60 - 120 Kpa Wind speed change within 1 h <10m / s Temperature change within 1 h <5℃ Air pressure change within 1 h <5Kpa

[0120] For the historical power supply data of the wind power cluster, its abnormal conditions can be mainly divided into these three characteristics: due to the "wind curtailment" phenomenon, the corresponding power supply value is too small when the wind speed is large; the measured value of the power supply of the wind farm is greater than the installed capacity of the wind farm; the measured value of the power supply at a certain moment is lower than zero.

[0121] For these three abnormal value phenomena, the following treatment solutions are adopted:

[0122] Use the local "wind curtailment" curtailment information to search for the data with too small power supply caused by this reason, and perform interpolation repair on the value at this point through the fan power curve.

[0123] For the power points that are too large and lower than zero, perform interpolation repair on them through the wind power curve.

[0124] Adopt the method of linear interpolation to fill in the missing and incorrect data, and its calculation formula is as shown in Equation (1):

[0125]

[0126] Among them, x t is the missing or abnormal value, x t-1 is the normal data that is not missing at the moment before the abnormal value, and x t+1 is the normal data at the next moment.

[0127] Both the historical power supply data of the wind power cluster and the NWP data can be corrected in this way.

[0128] S1_2: Perform normalization processing on the corrected NWP data and the historical power supply data of the wind power cluster;

[0129] The method of normalizing the historical power supply data of the wind power cluster is: Among them, is the historical power supply data of the wind power cluster after normalization processing, P is the measured value, P Nis the installed capacity of the wind farm.

[0130] The method for normalizing the NWP data is as follows: Among them, is the NWP data after normalization, including but not limited to 10m wind speed, 30m wind speed, temperature, air pressure, etc. X i is the measured value of the NWP data. Since the NWP data is a set of sequences, so X max is the maximum value in the NWP data.

[0131] The wind direction is usually expressed in angles. However, due to its periodic characteristics and discontinuity, it may be difficult for the model to understand the angle data directly. Therefore, the wind direction angle included in the NWP data is decomposed into sine and cosine, and the wind direction is represented as a two-dimensional vector: D c = cos(D), D s = sin(D), where D is the measured wind direction angle in the NWP data, D c and D s are the cosine value and sine value of the wind direction representing the wind speed direction.

[0132] S1_3: Analyze the corrected and normalized historical power supply data of the wind power cluster using the maximum information coefficient MIC and the cross-correlation function CFF; the specific operations include: using two sequences X = [x 1 , x 2 ,..., x n and Y = [y 1 , y 2 ,..., y n to describe the historical power supply data sequences of wind farms X and Y in the wind power cluster. The cross-correlation function C CFF (τ, X, Y) between X and Y describes the linear relationship between the two sequences and can be expressed as:

[0133]

[0134] Among them, n represents the number of data in the sequence, represents the average value of all data in the X sequence, represents the average value of all data in the Y sequence, τ is the order of the time delay in the sequence, here it represents the time shift interval. When τ = 0, C CFF (0, X, Y) = R XY is the correlation coefficient of the power supply of wind farms X and Y. The correlation relationship between X and Y increases / decreases as R XY increases / decreases; this formula represents the correlation between sequence X and sequence Y shifted by τ orders. Its value can be positive or negative. When the value of τ is positive, it means that Y is shifted backward by τ orders, and when it is negative, it means that it is shifted forward.

[0135] There may be a non - linear relationship between the historical power supply data of wind farms X and Y in a wind power cluster. Therefore, using the maximum information coefficient (MIC) to measure the possible non - linear relationship between the two has fairness and universality. Its definition formula is as follows:

[0136]

[0137] In the formula, p(X,Y) represents the joint probability distribution of sequence X and sequence Y. a and b are the number of data in the two sequences respectively. B represents the complexity limit to prevent over - fitting of data, which is taken as the 0.6th power of the data volume.

[0138] Combined with the calculated C CFF (τ,X,Y) and MIC results, identify the wind farm groups with strong correlations, that is, set a threshold to screen out the wind farm groups with both CFF and MIC values higher than the selected level to ensure capturing linear and non - linear correlations.

[0139] S1_4: Use the CEEMDAN decomposition algorithm to decompose the historical power supply data of the wind power cluster into intrinsic mode functions to facilitate the prediction model to dig deeper into the original data.

[0140] The process of the CEEMDAN decomposition algorithm is as follows:

[0141] (1) Let the corrected and normalized historical power supply data sequence of the wind power cluster be X(t), and set the standard deviation of the white noise sequence N(t)=[n 1 (t),n 2 (t),...,n Δt (t)] and the number of sequence additions M, where 0 < i ≤ M and i is an integer, and Δt is the number of time series intervals of X(t).

[0142] (2) Add the Gaussian white noise a i n i (t) sequence obtained by scaling each component to X(t), and perform EMD decomposition on the added new sequence:

[0143] X(t)+a i n i (t)=IMF i1 (t)+r i1 (t) (5)

[0144] Where a i is the scaling factor, representing the amplitude of the noise, which is a constant. IMF i1 (t) is the original sequence after superimposing the Gaussian white noise a i n iThe first IMF component obtained after EMD decomposition at (t), r i1 (t) is the remaining component obtained by subtracting the original sequence, that is, r i1 (t) = X(t) - IMF i1 (t).

[0145] Average the generated IMF i1 (t) component set as the first output component, denoted as:

[0146]

[0147] (3) Add the noise component a i1 obtained by EMD decomposition to the remaining component r 1 E 1 (n i (t)) and perform EMD decomposition on this sequence again:

[0148] r i1 (t) + a 1 E 1 (n i (t)) = IMF i2 (t) + r i2 (t) (7)

[0149] where E 1 (·) represents the first IMF component after EMD decomposition of the sequence in the parentheses, and a 1 E 1 (n i (t)) is the first IMF component after EMD decomposition of the white noise n i (t), repeat (2).

[0150] (4) Obtain the (k + 1)-th remaining component, that is, r ik+1 (t) = X(t) - IMF ik+1 (t), and then repeat the operation in (3), superimpose new Gaussian white noise on the new remaining component for EMD decomposition, generate k + 2 remainders, and calculate

[0151] (5) Repeat the above steps until EMD decomposition cannot be performed and the loop ends.

[0152] S1_5: The relevant data information required for wind power prediction (including key meteorological data and the power supply historical data of the strongly correlated wind farm group screened by CFF and MIC) enters the BP neural network from the input layer. Inside the BP neural network, the input relevant data information is processed and calculated through the hidden layer. When the hidden layer processes the wind power data, the function used is:

[0153]

[0154] Among them, y j is the output data, and is the activation function (log-simoid, LS).

[0155] Compare the output data of the BP neural network with the preset data value, and the error E(n) between the two is expressed as:

[0156]

[0157] In the formula, T k is the output data of the BP neural network, O k is the preset data value, L is the set of data points, and n is the number of data points;

[0158] If the value of the error E(n) of the nth wind power data is greater than the preset threshold, the output value will be BP in the order of the output layer, hidden layer, and input layer. While this process is ongoing, the BP neural network will update the weights according to the error, and the calculation formula for updating the weights is: W ij (n + 1) = ΔW kj (n + 1) + W kj (n); among them, W ij is the data weight from the input layer to the hidden layer, W kj is the data weight from the hidden layer to the output layer, and ΔW kj is the difference in the weights from the hidden layer to the output layer between data point n + 1 and data point n.

[0159] S1_6: Optimize the BP neural network using the genetic algorithm GA to form a GA-BP neural network model. The specific steps include:

[0160] ① Encode the initial values of the wind power weights (i.e., the weights of the input features including the wind direction and wind speed after sine-cosine decomposition) and thresholds in the BP neural network;

[0161] ② Determine the fitness function according to the error between the output value of the BP neural network and the preset value. The fitness function S is expressed as:

[0162]

[0163] Among them, l is the total number of training samples, G is the number of data points in the output layer, is the output value of the individual under the training samples, is the true output value;

[0164] ③ In each iteration process, according to the magnitude of the value finally calculated by the fitness function, the sample individuals in the cluster are sorted, and a new population of wind power data is generated through crossover and mutation operations;

[0165] ④ When the above steps are completed, return to step ② for loop calculation. The calculation will only stop when the individuals in the final population meet the indicators of the expected requirements, so as to output the accurate wind power prediction result.

[0166] Embodiment 2:

[0167] Based on the preliminary prediction model proposed in Embodiment 1, this embodiment further considers the influence of prediction errors. By introducing a competitive learning mechanism, an improved NSGA-II multi-objective optimization algorithm is proposed to obtain the optimized prediction results of effective coordination control of each control layer under the existing scheduling conditions. The specific steps include:

[0168] S2_1: Use the upper and lower bound estimation model based on the Long Short Term Memory (LSTM) network to improve the preliminary wind power prediction model, that is, incorporate the active power layer by layer, generate a prediction interval for the power of each layer by Lower and Upper Bound Estimation (LUBE), and its two outputs are the upper and lower limits. Evaluate the confidence level of the preliminary prediction model to obtain the improved wind power prediction model.

[0169] The PICP index is defined as:

[0170]

[0171] where N is the number of experimental data samples, a i is a binary value parameter, and its calculation formula is as follows:

[0172]

[0173] where y i is the value of test sample i, y i,t and y i,u are the predicted upper and lower bounds. To ensure the accuracy of the prediction interval, it is usually required that PICP is greater than the pre-set confidence level (1-α).

[0174] To improve the prediction efficiency, the PINAW index is introduced for evaluation:

[0175]

[0176] where W is the value range of the prediction interval.

[0177] S2_2: Combine the interval coverage probability (PICP) and the normalized averaged width (PINAW) to obtain a comprehensive index (coverage width-based criterion, C CW ), to evaluate the overall performance of the prediction interval.

[0178] To evaluate the overall performance of the prediction interval, obtain C CW :

[0179]

[0180]

[0181] where γ and μ are two control coefficients of CCW. γ reflects the coverage probability requirement of the prediction interval and can be calculated according to the predetermined confidence level (1 - α). η is the penalty coefficient when the obtained prediction interval does not meet the coverage probability requirement. W PINA is the PINAW parameter, P PIC is the PICP parameter

[0182] The estimation error E PIE is:

[0183]

[0184] S2_3: Introduce a competitive learning mechanism to improve the non-dominated fast sorting genetic algorithm and solve the improved wind power prediction model.

[0185] Step 1) Divide the original NWP dataset and the historical output dataset preprocessed in Step 1 into training, validation, and test subsets. The training set is used to adjust the parameters of the LSTM, the validation set is used to determine the structure of the LSTM, and the test set is used to verify the generalization ability of the model.

[0186] Step 2) Overall initialization: For each individual i, the parameters in the LSTM are initialized by sampling from a random uniform distribution.

[0187] Step 3) Fitness value: Each i can be regarded as a solution to the LUBE model. Calculate the error index and the PINAW index according to Equations (13) and (16).

[0188] ① Assume that IT represents the current iteration time and set it to 1;

[0189] ② For each individual, let t = 1, E PIE = 0, W PINA = 0;

[0190] ③ Suppose the total number of samples in Dtrain is n, and the input of the LSTM is the previous m observations. According to the forward propagation process of the LSTM, the upper and lower bounds of y are constructed. If y t+m ≤ y t+m,u ≤ y t+m ≤ y t+m,l , let E t+m be 0, otherwise

[0191] ④ The width calculation formula of the prediction interval is

[0192] ⑤ E PIE ' = E PIE + E t+m , W PINA ' = W PINA + d t+m ;

[0193] ⑥ If t = n - m, perform step 4), otherwise t = t + 1, and return to ③ in step 3).

[0194] Step 4) Obtain the overall population: The current population combines the elite strategy to obtain an overall population number P u (When IT = 1, the number of elite strategies is set to 0).

[0195] Step 5) Dominance check: Check the dominance relationship between each individual.

[0196] Step 6) Quick sort: Divide the population into different levels according to the dominance relationship. Then, a new population P n is obtained based on density.

[0197] Step 7) Obtain the offspring population: Perform selection, crossover, mutation, and competitive learning on P n to obtain an offspring population P 0 .

[0198] Step 8) Iteration: IT = IT + 1. If IT is greater than the maximum iteration time, the algorithm stops. Otherwise, return to ② in step 3).

[0199] Step 9) Record the result of P n .

[0200] S2_4: Take the average value of P n as the final prediction result.

[0201] Example 3:

[0202] This embodiment proposes a hierarchical active power control method for a wind power cluster, which divides the wind power cluster into three layers: the cluster layer, the farm group layer, and the sub-farm layer for regulation, and generally includes the following steps:

[0203] S3_1: The cluster layer receives the grid dispatching instruction planned value, compares it with the wind power prediction value, and determines the overall dispatching plan of the wind power cluster with the goal of minimizing the overall power output fluctuation of the wind power cluster and maximizing the output. The specific operations include:

[0204] Use the GA-BP neural network model for prediction to obtain the wind power prediction value;

[0205] The objective function of the rolling optimization dispatching model in the cluster layer aims to minimize the fluctuation and maximize the output, which is expressed as:

[0206]

[0207] Among them, δ 1 and δ 2 are the weights for suppressing the output fluctuation of the wind power cluster and maximizing the output respectively, δ 1 ≥0, δ 2 ≥0, δ 1 +δ 2 =1, P Z,t is the optimized dispatching power of the wind power cluster at the 15-minute time scale in the t-th period, n is the total number of wind farms, R i,t-1 is the output power of wind farm i at the 15-minute time scale in the (t - 1)-th period, is the power prediction value of wind farm i at the 15-minute time scale in the t-th period, and n is the number of wind power data.

[0208] The constraint conditions of the cluster layer are:

[0209] 1) The output ramp constraint of the wind power cluster. The output ramp constraint of the wind power cluster reflects the limit of the power change of the wind power cluster, that is, the power change of the wind power cluster in the previous and subsequent periods needs to be within a certain range.

[0210]

[0211] Among them, D Z is the ramp rate limit value of the wind power cluster at the 15-minute time scale, and P cluz is the total installed capacity of the wind power cluster.

[0212] 2) The power constraint of the wind power cluster.

[0213]

[0214] 3) The tracking plan constraint of the wind power cluster.

[0215] P Z,t≤P wet,t (20)

[0216] Among them, P wet,t is the grid planned power. The optimized power of the wind power cluster is less than the planned power to prevent no solution caused by the predicted power being less than the planned power. Therefore, the original equation is relaxed and tightened by maximizing the output through the objective function. For the time periods when the prediction is less than the planned power, only the wind farm energy storage can be used to coordinate the output.

[0217] 4) Capacity constraint of the transmission line of the wind power cluster.

[0218] P Z,t ≤P FMAX (21)

[0219] Among them, P FMAX is the maximum transmission line capacity of the wind power cluster.

[0220] S3_2: Allocate the scheduling plan according to the prediction of each single field to determine the initial single-field scheduling plan, which is expressed as:

[0221]

[0222] Among them, P i,t is the initial scheduling power of wind farm i at time t on a 15-minute time scale.

[0223] S3_3: Compare the initial single-field scheduling plan with the 5-minute ultra-short-term prediction of each single field, perform dynamic grouping according to the comparison results, and issue the field group scheduling instruction, which is expressed as:

[0224] After the optimized power of the cluster is reasonably allocated through the predicted values of each wind farm, it is compared with the values after 3 rounds of prediction with a 5-minute time scale for each wind farm to determine the number of wind farms that generate more and less power than the initial scheduling instruction within this 15 minutes, and form 2 types of field groups based on this. The sum of the reasonably allocated power values is used as the field group scheduling instruction. The grouping comparison model is expressed as:

[0225]

[0226] Among them, Q i,t is the grouping index. When Q i,t =1, it represents that wind farm i is a wind farm with more power generation at time t. When Q i,t =0, it represents that wind farm i is a wind farm with less power generation at time t. P P i,t,s5 is the power prediction value of the wind farm on a 5-minute time scale.

[0227] S3_4: The field group layer dynamically groups through the comparison of the predicted power of different wind farms and the initial scheduling power. According to the output of different field groups, on the premise of ensuring that the total optimized scheduling power of the cluster remains unchanged, the output of each field group is adjusted, and then the output of each wind farm is adjusted, so as to achieve the purpose of reducing wind curtailment.

[0228] The mathematical model of the field group layer is as follows:

[0229]

[0230] ΔP D,t,s5 =-ΔP S,t,s5 (25)

[0231]

[0232] P i,t,s5 =P i,t +ΔP i,t,s5 (27)

[0233] Among them, ΔP S,t,s5 is the power that the over-producing field group can produce more in the grouping on the 5-minute time scale, ΔP D,t,s5 is the power that the under-producing field group can produce more in the grouping on the 5-minute time scale, m is the number of over-producing wind farms, n is the total number of wind farms, ΔP i,t,s5 is the fine-tuning amount of the scheduling power of wind farm i on the 5-minute time scale, ΔP P i,t,s5 is the excess amount of the predicted power of wind farm i compared with the initial scheduling power on the 5-minute time scale, P i,t,s5 is the adjusted scheduling power value of wind farm i on the 5-minute time scale.

[0234] The above model uses the original wind curtailment power to make up for the less-produced wind power of the under-producing wind farms, which can not only reduce the amount of wind curtailment but also reduce the coordinated output of the wind farm energy storage.

[0235] S3_5: The sub-field layer accepts the scheduling instructions of the field group layer and determines the optimal output of the wind farm with the goal of maximizing the output of the wind farm and minimizing the output fluctuation.

[0236] The objective function of the rolling optimization scheduling model of the sub-field layer aims to minimize the fluctuation and maximize the output:

[0237]

[0238] Among them, δ i1 、δ i2 are the weights for suppressing the output fluctuation and maximizing the output of wind farm i respectively, P op i,t,s5 is the optimal scheduling power of wind farm i at time t on the 5-minute time scale, Ri,t-1,s5 is the output power of wind farm i at time period t-1 on a 5-minute time scale.

[0239] Its constraint conditions are as follows:

[0240] 1) Ramp rate constraint of wind farm output.

[0241]

[0242] Among them, D iZ,s5 is the ramp rate limit of wind farm i on a 5-minute time scale, and P iz is the installed capacity of wind farm i.

[0243] 2) Power constraint of wind farm.

[0244]

[0245] 3) Tracking plan constraint of wind farm.

[0246]

[0247] 4) Transmission line capacity constraint of wind farm.

[0248]

[0249] Among them, P fimax is the maximum transmission line capacity of wind farm i.

Claims

1. A method for predicting wind power cluster output power based on spatiotemporal characteristics, characterized in that: The following steps are involved: Step 1: interpolate and repair the abnormal data in the historical power supply data of the wind power cluster and the historical meteorological data of the wind power cluster, and normalize the interpolated and repaired historical power supply data of the wind power cluster and the historical meteorological data of the wind power cluster; The maximum mutual information coefficient MIC and the cross-correlation function CFF are used to analyze the historical power supply data of wind power clusters, and wind farm groups with strong correlation are selected; Step 2: Use the CEEMDAN decomposition algorithm to perform eigenmode function decomposition on the historical power supply data of the screened wind farm group with strong correlation, and obtain the decomposed historical power supply data; Step 3: Use the normalized historical meteorological data of the wind power cluster and the disassembled historical power supply data to train the GA-BP neural network model to obtain the output power prediction results of the wind power cluster.

2. The method for predicting wind power cluster output power based on spatiotemporal characteristics according to claim 1 is characterized by: The use of the maximum mutual information coefficient MIC and the cross-correlation function CFF to analyze the historical power supply data of the wind power cluster and select the meteorological data of the wind farm group with strong correlation, specifically includes: The cross-correlation function CFF is used to analyze the cross-correlation of the historical power supply data of wind farm X and wind farm Y in the wind power cluster, which is expressed as: Where X represents the historical data sequence of power supply of wind farm X, and the sequence X = [x1, x2, ..., x n ] to describe, x n is the nth data of the historical data sequence of power supply of wind farm X; Y represents the meteorological data of the historical data sequence of power supply of wind farm Y, and the sequence Y = [y1, y2, ..., y n ] to describe, y n is the nth data of the sequence; C CFF (τ,X,Y) represents the cross-correlation function between sequence X and sequence Y; τ is the time shift interval, when τ=0, C CFF (0,X,Y)=R XY , which is the correlation coefficient of the power supply of wind farm X and Y. The correlation between the power supply of wind farm X and Y varies with R XY increases / decreases with the increase / decrease of The maximum mutual information coefficient MIC is used to measure the correlation between the historical power supply data of wind farm X and wind farm Y in the wind power cluster, which is expressed as: In the formula, p(X,Y) represents the joint probability distribution of sequence X and sequence Y, a and b are the number of data in the two sequences respectively, and B represents the complexity limit parameter; Combined with the calculated C CFF (τ, X, Y) and MIC[X; Y] results, and filter out C by setting a threshold CFF The wind farm groups whose (τ, X, Y) and MIC[X; Y] are both higher than the set threshold are regarded as the wind farm groups with strong correlation.

3. The method for predicting wind power cluster output power based on spatiotemporal characteristics according to claim 1 is characterized by: The GA-BP neural network model is obtained by optimizing the BP neural network using a genetic algorithm; In the BP neural network, the output data of the BP neural network is compared with the preset data value, and the error E(n) between the two is expressed as: Where, T k is the output data of BP neural network, O k is the preset data value, L is the set of data points, and n is the number of data points; if the error E(n) is greater than the threshold, the output data of the BP neural network is BP-ed in the order of output layer, hidden layer, and input layer. During this process, the weights are updated according to the error E(n), which is expressed as: IN ij (n+1)=ΔW kj (n+1)+W kj (n) Among them, W ij is the weight from the input layer to the hidden layer, W kj is the weight from the hidden layer to the output layer; ΔW kj is the difference between the weights from the hidden layer to the output layer of data point n+1 and data point n; The genetic algorithm is used to optimize the BP neural network, and the specific operations include: Step ① Encode the input features and initial values ​​of the threshold in the BP neural network; Step ② determines the fitness function according to the error E(n), expressed as: Among them, l is the total number of samples, G is the number of data points in the output layer, is the output prediction value, is the real output value; Step ③ In each iteration, the individuals in the cluster are sorted according to the fitness function, and a new population of wind power data is generated through crossover and mutation operations; Step ④ returns to step ② for loop calculation. The calculation is stopped only when the last population individual meets the expected required indicators, and the wind power prediction result is output.

4. The method for predicting wind power cluster output power based on spatiotemporal characteristics according to claim 1 is characterized in that: The following steps are also included: Using the upper and lower bound estimation model based on long short-term memory network, the upper and lower bound estimates are obtained, and the prediction interval is generated according to the upper and lower bound estimates; The trained GA-BP neural network model is corrected by confidence assessment using the prediction interval to obtain a perfect wind power prediction model. The non-dominated quick sort genetic algorithm is used to iteratively solve the improved wind power prediction model and obtain the optimal set P u ; Take the optimal set P u The average value result is taken as the final wind power cluster output power prediction result.

5. The method for predicting wind power cluster output power based on spatiotemporal characteristics according to claim 4 is characterized by: The non-dominated quick sort genetic algorithm is used to iteratively solve the improved wind power prediction model to obtain the optimal set P u , follow these steps to get: S1: The meteorological data and the historical power supply data of the disassembled wind power cluster are divided into a training set, a validation set and a test set. The training set is used to adjust the parameters of the long short-term memory network, the validation set is used to determine the structure of the long short-term memory network, and the test set is used to verify the generalization ability of the upper and lower bound estimation model. S2: For each individual i, the parameters in the long short-term memory network are initialized by sampling from a random uniform distribution; S3: Each individual i can be regarded as a solution to the upper and lower bound estimation model, and the error index and PINAW index are calculated according to the following formula; Among them, E PIE is the error index, N represents the number of individuals, and W is the range of the prediction interval; y i is the value of individual i, y i,t and i,u W is the upper and lower bounds of the prediction; PINA is the PINAW index; ① Assume that IT represents the current iteration time and is set to 1; ②For each individual, let t = 1, E PIE =0,W PINA =0; ③ Assuming that the total number of samples in the training set is n, the input of the long short-term memory network is the previous m observations. According to the forward propagation process of the long short-term memory network, y is constructed. t+m The upper and lower bounds of y t+m,u ≤y t+m ≤y t+m,l , let E t+m is 0, otherwise Where w represents the penalty factor for adjusting the error metric, y t+m,l ,y t+m,u Represents y t+m The upper and lower bounds of E t+m is a measure of the predicted value y t+m Relative to the prediction interval [y t+m,u ,y t+m,l ]Prediction error of the nearest boundary distance; ④The calculation formula for the width of the prediction interval is ⑤E PIE '=E PIE +E t+m ,W PINA '=W PINA +d t+m ; ⑥If t=nm, proceed to S4, otherwise t=t+1, return to S3③; S4: Combined with the elite strategy, we obtain a total population number P u ; S5: Check the dominance relationship between each individual; S6: Divide the population into different levels according to the dominance relationship, and then obtain a new group P based on the density n ; S7: P n Perform selection, crossover, mutation and competitive learning to obtain an offspring population P0; S8: IT = IT + 1. If IT is greater than the maximum iteration time, the algorithm stops and records P. u otherwise, return to ② of S3.

6. A wind power cluster active power hierarchical control method based on spatiotemporal characteristics, characterized by: The following steps are involved: The control layer is divided into cluster layer, field group layer and sub-field layer; Predict the output power of the wind power cluster and obtain the prediction result; Determine the overall dispatching plan of the wind power cluster with the goal of minimizing the fluctuation of the overall power output of the wind power cluster and maximizing the overall output of the wind power cluster; Allocate the overall dispatching plan of the wind power cluster according to the forecast situation of each single field and determine the initial dispatching plan of each single field; Compare the initial single-field dispatch plan with the short-term forecast of each single field, dynamically group them according to the comparison results, and issue dispatch instructions for the field groups; According to the dispatching instructions of the wind farm group, the optimal output of the wind farm is determined with the goal of maximizing the output of the wind farm and minimizing the output fluctuation; The wind power cluster output power prediction to obtain the prediction result is to predict the wind power cluster output power by using a wind power cluster output power prediction method based on spatiotemporal characteristics as described in any one of claims 1 to 5 to obtain the prediction result.

7. The method for hierarchical control of wind power cluster active power based on spatiotemporal characteristics according to claim 6 is characterized by: The overall dispatching plan of the wind power cluster is determined with the goal of minimizing the fluctuation of the overall power output of the wind power cluster and maximizing the overall output of the wind power cluster. The specific operations include: Among them, δ1 and δ2 are the weights of suppressing the fluctuation of wind power cluster output and maximizing output, δ1≥0, δ2≥0, δ1+δ2=1, P Z,t is the existing optimized dispatching power of the wind power cluster in the 15-min time scale in period t, n is the total number of wind farms in the wind power cluster, R i,t-1 is the output power of wind farm i in period t-1 on a 15-min time scale, is the power forecast value of wind farm i in period t on a 15-minute time scale; The corresponding constraints include: The output ramp constraint of wind power cluster is expressed as: Among them, D Z is the climbing rate limit of the wind power cluster on a 15-minute time scale, P cluz is the total installed capacity of the wind power cluster; The wind power cluster power constraint is expressed as: The wind power cluster tracking plan constraint is expressed as: P Z,t ≤P wet,t Where P wet,t Plan power for the grid; The capacity constraint of wind power cluster transmission line is expressed as: P Z,t ≤P FMAX Where P FMAX is the maximum transmission line capacity of the wind power cluster.

8. The method for hierarchical control of active power of wind power cluster based on spatiotemporal characteristics according to claim 7 is characterized by: The overall dispatching plan of the wind power cluster is allocated according to the forecast situation of each single field, and the initial single field dispatching plan is determined, which is expressed as: Among them, P i,t is the initial dispatch power of wind farm i in period t on a 15-min time scale; The initial single-field scheduling plan is compared with the short-term forecast of each single field, and dynamic grouping is performed according to the comparison result, which is expressed as: Among them, Q i,t is the clustering index, Q i,t =1 means that wind farm i is a multi-generation wind farm in period t, Q i,t =0 means that wind farm i is a low-generation wind farm in period t, It is the power prediction value of the wind farm at a time scale of 5 minutes.

9. The method for hierarchical control of wind power cluster active power based on spatiotemporal characteristics according to claim 8 is characterized by: The issuing of the field group dispatching instruction is expressed as: ΔP D,t,s5 =-ΔP S,t,s5 P i,t,s5 =P i,t +ΔP i,t,s5 Among them, ΔP S,t,s5 is the additional power that can be generated by the multi-generation wind farm group in the 5-minute time scale, ΔP D,t,s5 is the power that can be generated by the less-generating wind farm group in the 5-minute time scale, m is the number of more-generating wind farms, n is the total number of wind farms, ΔP i,t,s5 is the fine-tuning amount of wind farm i dispatching power under the 5-minute time scale, is the excess of the predicted power of wind farm i over the initial dispatched power at a time scale of 5 minutes, P i,t,s5 is the adjusted dispatch power value of wind farm i on a 5-minute time scale.

10. The method for hierarchical control of active power of wind power cluster based on spatiotemporal characteristics according to claim 9, characterized in that: The goal of maximizing the output of the wind farm and minimizing the output fluctuation is expressed as: Among them, δ i1 , δ i2 are the weights of suppressing the output fluctuation of wind farm i and maximizing the output, P op i,t,s5 is the optimal dispatching power of wind farm i in period t at a time scale of 5 minutes, R i,t-1,s5 is the output power of wind farm i in period t-1 at a time scale of 5 minutes; The constraints include: 1) Wind farm output ramp constraint, expressed as: Among them, D iZ,s5 is the ramp rate limit of wind farm i in the 5-min time scale, P iz is the installed capacity of wind farm i. 2) Wind farm power constraint, expressed as: 3) Wind farm tracking plan constraints, expressed as: 4) Wind farm transmission line capacity constraint, expressed as: Among them, P fimax is the maximum transmission line capacity of wind farm i.

Images