Ocean sea wind power prediction method and system based on adaptive error correction

Abstract

The invention belongs to the technical field of power prediction, and particularly relates to an offshore wind power prediction method and system based on adaptive error correction. The method comprises the following steps: firstly, screening meteorological characteristics with high correlation with power change in a preliminary wind power data set by utilizing a Pearson correlation coefficient, performing modal decomposition on power data by utilizing EEMD (Ensemble Empirical Mode Decomposition) to form a multi-modal power data set, and fusing the meteorological characteristics and the multi-modal power data set to form a wind power amplification data set; and finally, inputting the amplified wind power data set into a GRU model improved based on residual connection and an attention mechanism. By means of the short-term wind power combined prediction model integrating modal decomposition and machine learning, errors generated in the prediction process are reasonably processed, the precision of offshore wind power prediction and the credibility of the prediction result are improved, the stability of operation of a power system is improved, assistance is provided for resource allocation of the power system, and the economic benefits of the power system are improved. And sustainable development of the power system in the future is facilitated.

Description

Technical Field

[0001] The present invention belongs to the technical field of power prediction, and in particular relates to a method and system for predicting offshore wind power based on adaptive error correction. Background Art

[0002] Errors are inevitable in the process of offshore wind power forecasting. In order to improve the accuracy of offshore wind power forecasting and the credibility of its forecast results, it is of great significance to handle errors reasonably. In recent years, with the rapid development of data processing and machine learning technology, some advanced technologies have been widely used in power forecasting, greatly improving the accuracy and efficiency of forecasting. However, there are still certain problems. For example, the invention patent with publication number CN119134284A discloses a short-term wind power forecasting method based on the Spearman-ISSA-BiLSTM combined model, but it has poor adaptability when facing a variety of complex weather conditions, and is prone to insufficient model prediction accuracy under extreme weather conditions.

[0003] For example, the invention patent with publication number CN119005456A discloses a short-term photovoltaic and wind turbine power prediction method and system based on a nuclear extreme learning machine. However, its use of a time series window to extract historical features has certain limitations, because the time series window can only extract local historical data and cannot capture the long-term dependencies between data, resulting in a decrease in the model prediction accuracy.

[0004] To solve the above problems, the present invention uses SOM clustering analysis to divide weather types and improve the adaptability of the prediction model to different weather conditions. At the same time, it combines Pearson correlation analysis and EEMD modal decomposition technology to extract historical weather characteristics and power characteristics, effectively capture the long-term dependency between data, and achieve high-precision power prediction. Summary of the invention

[0005] Purpose of the invention: In order to overcome the shortcomings of the prior art, the present invention provides a method for predicting offshore wind power based on adaptive error correction, which solves the problem of low accuracy of offshore wind power prediction models. The present invention also provides a system for predicting offshore wind power based on adaptive error correction.

[0006] Technical solution: According to a first aspect of the present invention, a method for predicting offshore wind power based on adaptive error correction is provided, the method comprising:

[0007] S1 preprocesses the collected initial offshore wind power data, performs SOM cluster analysis on the preprocessed wind power data, divides different weather types, and forms a preliminary wind power data set;

[0008] S2 performs correlation analysis on the meteorological features in the obtained preliminary wind power data set to screen out meteorological features that are highly correlated with power; processes the power data in the preliminary wind power data set using empirical mode decomposition to obtain multimodal power data, and fuses the screened high-power meteorological features with each component after modal decomposition processing to obtain a wind power augmented data set;

[0009] S3 inputs the obtained wind power amplification data set into the improved GRU model, trains the model with the training set, and predicts the required wind power with the test set, wherein the improved GRU model includes: processing the reset gate in the GRU model with a weight coefficient based on the attention mechanism and adding a residual connection, wherein the residual connection is used to directly add the input and the output, thereby realizing the jump transmission of information;

[0010] S4 uses relevant evaluation indicators to evaluate the effectiveness of the prediction model.

[0011] Further, including:

[0012] The SOM cluster analysis is performed on the pre-processed wind power data to divide different weather types and form a preliminary wind power data set, which specifically includes the following steps:

[0013] S11 initialization: assign small random numbers to each weight vector of the output layer and normalize them to obtain the initial weight W ij (0), determine the initial domain range σ0, the initial learning rate η0 and the number of iterations K;

[0014] S12 Data input: input the preprocessed wind power data X into the SOM clustering model; the wind power data X includes multiple samples, which are expressed as: X = {x1, x2, ..., x n}, where any sample x i Each is a vector containing power data and meteorological data;

[0015] S13 distance calculation: For any input vector x i , calculate the weight vector W ij and the input vector x i The Euclidean distance between them is calculated as:

[0016]

[0017] After the calculation is completed, select the distance input vector x i The closest neuron is called the winning neuron i * , the winning neuron satisfies:

[0018]

[0019] It should be noted that the output layer of the network is a two-dimensional network, and the position of the winning neuron in the network is a two-dimensional coordinate.

[0020] Winning Neuron i * The weight vector can reflect the input data x i features, which helps to cluster the input data later.

[0021] S14 weight update: Assume the winning neuron is i * , the weight vector of the winning neuron and its field neuron needs to be updated. The weight update formula is:

[0022]

[0023] Among them, η(t) is the learning rate, which controls the amplitude of weight adjustment and usually decays gradually as training progresses to ensure the convergence of analysis results. is the neighborhood function, which defines the relationship between neurons i and i * The similarity is calculated as follows:

[0024]

[0025] in, is neurons i and i * , σ(t) is the range of the domain function. As time goes by, the neighborhood range will gradually decrease.

[0026] S15 Learning rate and domain scope update: After completing the above steps, the learning rate and domain scope are updated to provide support for the next training. The learning rate and domain scope update formula is as follows:

[0027]

[0028] Among them, η0 and σ0 are the initial learning rate and domain range, τ η and τ σ is a constant that controls the decay rate.

[0029] S16 repeat iteration: repeat the above steps S11-S15 until the number of iterations reaches K;

[0030] S17 After the above steps, the weight vector of each neuron in the SOM network will gradually adapt to the distribution of the input data, and similar input data will be mapped to adjacent neurons to obtain the clustering result; the output data set is recorded as X1, which is the preliminary wind power data set.

[0031] The winning neuron of the present invention, each time a sample is input, the network will find a winning neuron by calculating the Euclidean distance, and move the weight vector of the neuron and its domain closer to the input sample. After multiple iterations, input samples with similar features will be mapped to the same neuron to achieve clustering analysis. Therefore, the winning neuron does not represent the data in the initial data set, it is a node that exists in the neural network itself, which is used to measure the similarity of features between data and data, so as to achieve clustering.

[0032] Further, including:

[0033] In step S2, correlation analysis is performed on the meteorological features in the obtained preliminary wind power data set to screen out meteorological features that are highly correlated with power, including the following steps:

[0034] S21 uses the Pearson correlation coefficient to analyze the correlation of meteorological characteristics. In the present invention, the meteorological data and power data in the preliminary wind power data set X1 are subjected to correlation analysis to screen out meteorological characteristics that have a high correlation with power changes. The Pearson correlation coefficient method is a statistical method that accurately measures the closeness of the relationship between two variables and is widely used in the field of prediction. The size of the Pearson correlation coefficient can reflect the strength of the linear correlation between two variables. With variable X = [x1, x2, ..., x n ] T and Y = [y1,y2,…,y n ] T For example, their Pearson correlation coefficient can be calculated as follows:

[0035]

[0036] Where: and is the average value of n data. The value range of the correlation coefficient r is (-1, 1), that is, |r|≤1. The closer |r| is to 1, the higher the correlation between x and y. When r=1, it indicates that x and y are completely positively linearly correlated. When r=-1, it indicates that x and y are completely negatively linearly correlated. When r=0, it indicates that there is no linear relationship between x and y.

[0037] Generally speaking, when |r|≥0.8, it indicates that there is a high correlation between x and y; when 0.5≤|r|<0.8, it indicates that there is a moderate correlation between x and y; when 0.3≤|r|<0.5, it indicates that there is a low correlation between x and y; when |r|<0.3, it indicates that there is an extremely weak correlation between x and y, which is considered a nonlinear correlation.

[0038] In the present invention, the input data set includes power data and meteorological features, and the power data is recorded as P, and the meteorological features are recorded as Feature. The input data set is X = [P, Feature]. The Pearson correlation coefficient of several meteorological feature data and power data is calculated to obtain the Pearson correlation coefficient matrix. According to the judgment principle of the principle part, the Pearson correlation coefficient |r||0.8 is recorded as a high correlation feature, namely Feature'.

[0039] In the present invention, the Pearson correlation coefficient can reflect the relationship between power data and meteorological characteristic data, so it is only necessary to calculate the Pearson correlation coefficient between other characteristics and power, and according to the correlation principle, it is possible to measure whether the characteristics and power are highly correlated. The input of the correlation analysis part includes meteorological data and power data, and the output is a coefficient matrix, and the coefficient matrix can be used to screen out high correlation features.

[0040] Further, including:

[0041] In step S2, the power data in the preliminary wind power data set is processed by empirical mode decomposition to obtain multimodal power data, and the screened high-power meteorological features are fused with each component after the mode decomposition process to obtain a wind power augmented data set, which specifically includes:

[0042] S22: The power data set in the preliminary wind power data set X1 is X2. The relevant analysis steps are as follows:

[0043] Perform EMD decomposition on the power data set X2, the total number of times is K, the variable i = 1, and add white noise p to the power data set X2 i (t), the value range is 0.1-0.4 times the ratio of the amplitude to the original signal amplitude, and the original vibration signal after superposition is:

[0044] y i (t) = y(t) + p i (t)i=1,2,…,K (8)

[0045] S23 superimposes y(t) with p(t) of different amplitudes K times, and performs EMD decomposition after each superposition to obtain a set of IMF components c ij and the residual component r i (t), therefore, the result obtained after the i-th decomposition is in the form of:

[0046] y i (t) = ∑c ij (t)+r i (t)i=1,2,…,K (9)

[0047] In the formula, J is the total number of IMF;

[0048] After S24 is decomposed K times, the IMF components can be obtained. The average value of these components will be the IMF components after the white noise is eliminated, that is:

[0049]

[0050] Where: c j (t) is the value of y i (t) The jth component after EEMD decomposition; signal y i (t) is the residual component of the signal y(t) after EEMD decomposition.

[0051] S25 fuses the high-correlation meteorological features obtained by correlation analysis with the IMF components obtained by EEMD decomposition to form a wind power augmentation data set. The fusion process can be expressed as follows:

[0052] Xi nput =[Feature',IMFi] (11)

[0053] Where, X input is the data in the augmented data set, Feature' represents the meteorological features with high correlation with power data, IMF i represents the components obtained after the power data is decomposed by EEMD, integrating all the obtained X input , that is, the wind power augmented data set X3 is obtained.

[0054] In the present invention, EEMD is the abbreviation of Ensemble Empirical Mode Decomposition, which is ensemble empirical mode decomposition in Chinese. It is a noise-assisted data analysis method aimed at the shortcomings of the EMD method. The EEMD decomposition principle is that when the additional white noise is evenly distributed in the entire time-frequency space, the time-frequency space is composed of different scale components divided by the filter bank.

[0055] Further, including:

[0056] In step S3, the reset gate in the GRU model is processed based on the weight coefficient in the attention mechanism, including: replacing the reset gate in the GRU model with the weight coefficient in the attention mechanism, and the weight coefficient in the attention mechanism is used to evaluate the similarity between the Query and each Key, which is expressed as:

[0057]

[0058] Among them, the hidden state h of the previous moment is used t-1 As the query vector (Query), the current input x t As the key vector (Key), Wq and W k is the weight matrix used to transform the hidden state h t-1 and x t Input is transformed into query and keyspace, d k is the dimension of the key vector, which is used to scale the dot product. The Softmax function ensures that the sum of all time step weights is 1. It determines how to weight the hidden state h of the previous moment based on these weights. t-1 .

[0059] Further, including:

[0060] In the step S3, the reset gate in the GRU model is processed using a weight coefficient based on the attention mechanism, and further includes:

[0061] The improved GRU model is recorded as GRU' model, and its corresponding hidden state is expressed by the following formula:

[0062]

[0063] In formula (13), the weight coefficient α is used t Replace the reset gate in the traditional model. After the replacement is completed, the candidate hidden state at the current moment is obtained Then Input into formula (14) to obtain the current hidden state and implement the update of the hidden state.

[0064] Further, including:

[0065] In step S3, a residual connection is added to the GRU' model, and the residual connection is used to directly add the input and the output, thereby realizing the jump transmission of information. The model is recorded as GRU', which is expressed by the following formula:

[0066] h t =GRU(x t ,h t-1 )+x t ; (15)

[0067] Among them, h t is the updated hidden state after the GRU' model is calculated. Since the attention mechanism is introduced to replace the traditional reset gate, x t is the current input data, and the input data x t Added directly to the output of the hidden state.

[0068] In a second aspect, the present invention further provides a system for predicting offshore wind power based on adaptive error correction, the system comprising:

[0069] The data preprocessing module is used to preprocess the collected initial offshore wind power data, perform SOM cluster analysis on the preprocessed wind power data, divide different weather types, and form a preliminary wind power data set;

[0070] The data set expansion module is used to perform correlation analysis on the meteorological features in the obtained preliminary wind power data set, and screen out meteorological features that are highly correlated with power; the power data in the preliminary wind power data set is processed using empirical mode decomposition to obtain multimodal power data, and the screened high-power meteorological features are merged with each component after modal decomposition processing to obtain a wind power expansion data set;

[0071] A model training module, used for inputting the obtained wind power amplification data set into an improved GRU model, training the model with a training set, and predicting the required wind power with a test set, wherein the improved GRU model comprises: processing the reset gate in the GRU model with a weight coefficient based on an attention mechanism and adding a residual connection, wherein the residual connection is used for directly adding the input and the output, thereby realizing the jump transmission of information;

[0072] The effect evaluation module is used to evaluate the effect of the prediction model using relevant evaluation indicators.

[0073] In a third aspect, the present invention further provides a computer-readable storage medium, on which is stored a distant sea wind power prediction program based on adaptive error correction, and when the distant sea wind power prediction program based on adaptive error correction is executed by a processor, the steps of the distant sea wind power prediction method based on adaptive error correction are implemented.

[0074] In a fourth aspect, the present invention further provides an electronic device, comprising a memory and a processor, wherein a computer program is stored in the memory, and the processor is configured to run the computer program to execute the method described above.

[0075] Beneficial effects: Compared with the prior art, the present invention has the following advantages: First, the present invention uses SOM cluster analysis to divide different weather types, and trains the model to perform power prediction under different weather conditions, thereby improving the model's adaptability to different weather conditions; then, the Pearson correlation analysis and EEMD modal decomposition technology are combined to screen power features and meteorological features, effectively capturing the long-term dependencies between data and improving prediction accuracy; secondly, the present invention also introduces an attention mechanism to replace the traditional reset gate to improve the GRU model, thereby realizing the model's automatic review of historical data, and dynamically allocating weights, thereby improving the model's adaptability to long-term data; finally, the residual connection structure is used to improve the GRU, thereby effectively alleviating problems such as gradient explosion and realizing error correction of prediction results. Therefore, the present invention can reasonably handle the errors generated in the prediction process through this short-term wind power combination prediction model that integrates modal decomposition and machine learning, thereby improving the accuracy of offshore wind power prediction and the credibility of its prediction results, which is beneficial to improving the stability of power system operation, providing assistance for its resource allocation, and facilitating the sustainable development of the power system in the future. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] Figure 1 This is a flow chart of a method for predicting offshore wind power based on adaptive error correction according to an embodiment of the present invention;

[0077] Figure 2 This is a prediction result diagram of the combined prediction model in the first scenario described in the embodiment of the present invention;

[0078] Figure 3 This is a prediction result diagram of the combined prediction model in the second scenario described in the embodiment of the present invention;

[0079] Figure 4 This is a diagram of the prediction results of the combined prediction model in the third scenario described in an embodiment of the present invention. DETAILED DESCRIPTION

[0080] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0081] Example 1

[0082] First, the cluster analysis method of this embodiment, such as SOM cluster analysis, can identify the nonlinear relationship between wind power and meteorological data. By training historical data, SOM can find the characteristic patterns of wind power under different weather conditions, thereby automatically dividing different weather types according to the data. Signal processing methods, such as EEMD, are mainly used to decompose nonlinear and non-stationary signals into multiple intrinsic mode functions IMF, and suppress mode aliasing by adding white noise, thereby improving the accuracy of signal analysis. Correlation analysis techniques, such as Pearson correlation analysis methods, can deeply explore the potential relationship between different data and features. In the field of machine learning, models such as LSTM, GRU and MLP are selected for processing different data components because of their advantages in capturing specific data features. The comprehensive application of these models and data processing methods further improves the accuracy of power prediction. In particular, the GRU model can effectively capture long-term dependencies in time series and is suitable for processing the time series characteristics of wind power. At the same time, compared with LSTM, GRU has a simpler structure, fewer parameters, faster calculation speed, is suitable for large-scale data processing, and performs well in the face of complex wind power data.

[0083] Based on the above research status, in order to improve the prediction accuracy of offshore wind power and reasonably handle errors in the prediction process, the present invention specifically uses SOM clustering to divide different types of weather to form an initial wind power feature data set, performs correlation analysis on meteorological data, and performs collective empirical mode decomposition on power data, and the two are merged to form a wind power amplified data set. The amplified data set is then input into the GRU model based on adaptive error correction, and the prediction results are obtained through training. In view of the fact that traditional GRU models are prone to gradient explosion and have limitations in capturing complex time relationships, the present invention introduces a residual connection structure to achieve error correction and alleviate gradient explosion; an attention mechanism is introduced to replace the traditional reset gate to improve the model's ability to represent and learn time series data.

[0084] The goal of this invention is to reasonably handle the errors that may occur in the prediction process and improve the accuracy and credibility of offshore wind power prediction. 2 Indicators such as scalability and cost-efficiency are used to evaluate the prediction accuracy of the model, which is convenient for optimizing the model and providing more reliable and efficient support for the reliable operation and resource allocation of the power system.

[0085] Reference Figure 1 The method for predicting offshore wind power based on adaptive error correction described in the present invention comprises the following steps:

[0086] Step 1: The meteorological data and power data in the corresponding recorders are exported and sorted to form an initial wind power data set. The initial wind power data set is preprocessed, including but not limited to cleaning and dynamic filling, to obtain a processed data set X.

[0087] Step 1.1, preprocessing the original wind power data, mainly including correcting and filling the abnormal values ​​and vacant values ​​contained in the data, and normalizing the data. The processed data set is represented by X;

[0088] Step 2: Reference Figure 1 , perform SOM cluster analysis on the data processed in step 1, classify different types of weather, and form a preliminary wind power data set X1. The process of using SOM cluster analysis includes the following steps:

[0089] Step 2.1, Initialization: Assign small random numbers to each weight vector of the output layer and normalize them to obtain the initial weight W ij (0), determine the initial domain range σ0, the initial learning rate η0 and the number of iterations K;

[0090] Step 2.2, data input. The preprocessed wind power data X is input into the SOM clustering model; the wind power data X includes multiple samples, which are represented as: X = {x1, x2, ..., x n}, where any sample x i It is a vector containing power data and meteorological data.

[0091] Step 2.3, distance calculation. For any input vector x i , calculate the weight vector W ij and the input vector x i The Euclidean distance between them is calculated as:

[0092]

[0093] After the calculation is completed, select the distance input vector x i The closest neuron is called the winning neuron i * , the winning neuron satisfies:

[0094]

[0095] Winning Neuron i * The weight vector can reflect the input data x i features, which helps to cluster the input data later.

[0096] Step 2.4, weight update: Assume that the winning neuron is i * , the weight vector of the winning neuron and its field neuron needs to be updated. The weight update formula is:

[0097]

[0098] Among them, η(t) is the learning rate, which controls the amplitude of weight adjustment and usually decays gradually as training progresses to ensure the convergence of analysis results. is the neighborhood function, which defines the relationship between neurons i and i * The similarity is calculated as follows:

[0099]

[0100] in, is neurons i and i * , σ(t) is the range of the domain function. As time goes by, the neighborhood range will gradually decrease.

[0101] Step 2.5, learning rate and domain scope update: After completing the above steps, update the learning rate and domain scope to provide support for the next training. The learning rate and domain scope update formula is as follows:

[0102]

[0103] Among them, η0 and σ0 are the initial learning rate and domain range, τ η and τ σ is a constant that controls the decay rate.

[0104] Step 2.6, repeat iteration: Repeat the above steps 2.1-2.5 until the number of iterations reaches K;

[0105] Step 2.7 After the above steps, the weight vector of each neuron in the SOM network will gradually adapt to the distribution of the input data, and similar input data will be mapped to adjacent neurons to obtain the clustering result; the output data set is recorded as X1, which is the preliminary wind power data set X1.

[0106] Step 3: Take the preliminary wind power data set X1 obtained in step 2 as the analysis object, use the Pearson correlation coefficient to analyze the meteorological data, and screen out meteorological features with high correlation with power changes; use EEMD to perform modal decomposition on the power data to form a multimodal power data set. Fusion of meteorological features and multimodal power data sets to obtain the augmented wind power data set X3.

[0107] Step 3.1, using the Pearson correlation coefficient to analyze the correlation of meteorological characteristics. In the present invention, the meteorological data and power data in the preliminary wind power data set X1 are subjected to correlation analysis to screen out meteorological characteristics with high correlation with power changes. The Pearson correlation coefficient method is a statistical method that accurately measures the closeness of the relationship between two variables and is widely used in the field of prediction. The size of the Pearson correlation coefficient can reflect the strength of the linear correlation between two variables. With variable X = [x1, x2, ..., x n ]T and Y = [y1,y2,…,y n ] T For example, their Pearson correlation coefficient can be calculated as follows:

[0108]

[0109] Where: and is the average value of n data. The value range of the correlation coefficient r is (-1, 1), that is, |r|≤1. The closer |r| is to 1, the higher the correlation between x and y. When r=1, it indicates that x and y are completely positively linearly correlated. When r=-1, it indicates that x and y are completely negatively linearly correlated. When r=0, it indicates that there is no linear relationship between x and y.

[0110] Generally speaking, when |r|≥0.8, it indicates that there is a high correlation between x and y; when 0.5≤|r|<0.8, it indicates that there is a moderate correlation between x and y; when 0.3≤|r|<0.5, it indicates that there is a low correlation between x and y; when |r|<0.3, it indicates that there is an extremely weak correlation between x and y, which is considered a nonlinear correlation.

[0111] In the embodiment of the present invention, the input data set includes power data and meteorological features, where the power data is denoted as P and the meteorological features are denoted as Feature. The input data set is X=[P, Feature]. The Pearson correlation coefficient of several meteorological feature data and power data is calculated to obtain a Pearson correlation coefficient matrix. According to the judgment principle of the principle part, the Pearson correlation coefficient |r|≥0.8 is recorded as a high correlation feature, i.e., Feature'.

[0112] Step 3.2: Use EEMD to perform modal decomposition on the power data set to form a multi-modal power data set. The power data set in the preliminary wind power data set X1 is called X2, and the relevant analysis steps are as follows:

[0113] Step 3.2.1, perform EMD decomposition on the power data set X2, the total number of times is K, variable i = 1,

[0114] Step 3.2.2, add white noise p to X2 i (t), the value range is 0.1-0.4 times the ratio of the amplitude to the original signal amplitude, and the original vibration signal after superposition is:

[0115] y i (t) = y(t) + p i (t)i=1,2,…,K (8)

[0116] Step 3.2.3, superimpose p(t) with different amplitudes K times on y(t), and perform EMD decomposition after each superposition to obtain a set of IMF components c ij and the residual component r i (t). The result obtained after the i-th decomposition is in the form of:

[0117] y i (t) = ∑c ij (t)+r i (t)i=1,2,…,K (9)

[0118] Step 3.2.4, after K decompositions, the IMF components can be obtained. Taking the average of these components will obtain the IMF components after white noise is eliminated, that is:

[0119]

[0120] Where: c j (t) is the value of y i (t) is the jth component after EEMD decomposition, and J is the total number of IMFs.

[0121] In step 3.2.5, the signal y(t) is the residual component of the signal y(t) after EEMD decomposition.

[0122] Step 3.3, the high correlation meteorological features obtained by correlation analysis are fused with the IMF components obtained by EEMD decomposition to form a wind power augmentation data set. The fusion process can be expressed as follows:

[0123] Xi nput =[Feature',IMFi] (11)

[0124] Where, X input is the data in the augmented data set, Feature' represents the meteorological features with high correlation with power data, IMF i represents the components obtained after the power data is decomposed by EEMD, integrating all the obtained X input , we can get the wind power augmented data set X3.

[0125] Step 4: Input the augmented data set X3 obtained in step 3 into the GRU model transformed based on residual connection, and perform power prediction after training.

[0126] Step 4.1, GRU model principle and parameter setting. GRU stands for gated recurrent unit, which is a neural network model used to process sequence data. Compared with the traditional LSTM network, GRU has a simpler structure, fewer parameters, and faster training speed. The basic formula of the GRU network is shown in (8)-(11):

[0127] z t =σ(W z [h t-1 ,x t ]+b z ) (12)

[0128] r t =σ(W t [h t-1 ,x t ]+b r ) (13)

[0129] h t =tanh(W h [r t ×h t-1 ,x t ]+b h ) (14)

[0130] h t =(1-z t )×h t-1 +z t ×h t (15)

[0131] Where: z t represents the update gate; r t Represents the reset gate; h t represents the candidate hidden state at time t; h t represents the hidden state at time t; x t represents the input data at the current time t; h t-1 represents the hidden state at time t-1; σ(.) represents the sigmoid function; tanh(.) represents the hyperbolic tangent activation function; W z , W t , W h Represents the weight matrices of the update gate, reset gate and candidate hidden state respectively; b z 、b r 、b h Represent the update gate, reset gate and bias vector of the candidate hidden state respectively.

[0132] Step 4.2: Introduce the attention mechanism to replace the traditional reset gate r t The traditional reset gate depends on the input data and the hidden state of the previous moment to determine the update degree of the hidden state at the current moment. It cannot make flexible adjustments according to the actual needs of different time steps, resulting in limitations in the model in capturing long-term dependencies. The attention mechanism can dynamically adjust the attention weight according to the similarity between the current moment input and historical information, allowing the model to selectively review the historical data most relevant to the current moment input and dynamically adjust the attention weight αt , effectively capturing the long-term dependencies between data. In the attention mechanism, given a certain element Query, the source module is used to calculate the similarity or correlation between the Query and each Key, obtain the weight coefficient of the Value corresponding to each Key, and then perform weighted summation on the Value to obtain the final Attention value. Therefore, in essence, the Attention mechanism is a weighted summation of the Value values ​​of the elements in the Source, and the Query and Key are used to calculate the weight coefficient of the corresponding Value.

[0133] In this embodiment, the reset gate r t is replaced by the attention-based weight coefficient α t The calculation formula of the weight coefficient is shown in (16).

[0134]

[0135] In the above formula, the hidden state h at the previous moment is used t-1 As the query vector Query, the current input x t As the key vector Key. W q and W k is the weight matrix used to transform the hidden state h t-1 and x t Input is transformed into query and keyspace. k is the dimension of the key vector, which is used to scale the dot product. The Softmax function ensures that the sum of all time step weights is 1, and the model can decide how to weight the hidden state h at the previous moment based on these weights. t-1 .

[0136] Replace the reset gate in the traditional model with the weight coefficient α t , after the replacement is completed, the hidden state of GRU can be expressed by equations (17) and (18):

[0137]

[0138] In formula (17), the weight coefficient α is used t Replace the reset gate in the traditional model. After the replacement is completed, the candidate hidden state at the current moment is obtained Then Input into formula (18) to obtain the current hidden state and update the hidden state.

[0139] Step 4.3, use the residual connection structure to transform the GRU model so that it can perform adaptive error correction. The residual connection directly adds the input and output to achieve information jump transmission. For the residual connection, the output formula of the network is:

[0140] y=F(x)+x (19)

[0141] Among them, F(x) is the nonlinear transformation function of the network and x is the input.

[0142] The GRU model modified by the residual connection structure can perform adaptive error correction, making it more robust to small disturbances or errors. The GRU model based on residual connection can be expressed as follows:

[0143] h t =GRU(x t ,h t-1 )+x t (20)

[0144] Among them, h t is the updated hidden state after GRU calculation. In the embodiment of the present invention, since the attention mechanism is introduced to replace the traditional reset gate, h t As shown in formula (20), x t is the current input. By converting the input x t By adding it directly to the output of the hidden state, the model can retain the original input information to a certain extent, thereby achieving adaptive error correction.

[0145] Step 4.4: Model training and power prediction. In the present invention, the wind power amplified data set X3 is used as the input of the GRU algorithm model, and the test set X3 is divided into test and training set X train , of which the test set accounts for 20% and the training set accounts for 80%. The test set and training set can be expressed by the following formula:

[0146] X test =(x i ,y i ) N (twenty one)

[0147]

[0148] Among them, x i Represents the input data set, y i Represents the output samples corresponding to the input data.

[0149] The model is trained using the data from the training set and validated using the data from the test set to obtain the prediction results of the input data on the test set.

[0150] Step 5: Use the mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R-Square, R2 ) is used to measure the accuracy of the power prediction of the combined model in step 4. The specific calculation formulas for each indicator are as follows:

[0151]

[0152] Where: y i is the actual value of power, x i is the predicted value of power, n represents the number of samples, is the average value of power.

[0153] Example 2

[0154] This embodiment is based on the Python platform, and performs simulation and optimization analysis in the WIN11 operating system, 13th Gen Inter(R) Core i7-13650HX, 2.6GHz processor environment. The specific solution flow chart is as follows Figure 1 shown.

[0155] In this embodiment, the power and meteorological data of a region in East China in 2022 and 2023 are used. The first 80% of the data is used as the training sample set, and the last 20% of the data is used as the test set data. The predicted results of the training set are compared with the test set data to verify the effectiveness of the model. The data sampling frequency is 15 minutes per time.

[0156] In order to verify the effectiveness of the method proposed in Example 1 of the present invention, this embodiment sets up the following four models for comparative analysis: Model 1: using the SOM-EEMD-GRU model for power prediction; Model 2: using the SOM-EEMD-LSTM model for power prediction; Model 3: using the SOM-EEMD-MLP model for power prediction; Model 4: using the SOM-EEMD-ResidualGRU model proposed in Example 1 of the present invention for power prediction.

[0157] In three different scenarios, the comparison curves of the prediction curves of each model and the actual values ​​are as follows: Figure 2 , 3 As shown in Figure 4, where the horizontal axis is time and the vertical axis is power data, it can be seen that the power data predicted by the GRU model based on the residual connection structure and the improved attention mechanism is closer to the actual value. This is because the ResidualGRU model can perform adaptive error correction and reasonably handle the errors that may occur during the prediction process, thereby improving the prediction accuracy. At the same time, compared with other models, the ResidualGRU model has better adaptability to complex data and can effectively prevent the occurrence of phenomena such as gradient explosion. Therefore, the prediction curve obtained by the SOM-EEMD-ResidualGRU model is better than that of other models.

[0158] The prediction accuracy evaluation indicators of the four models in different scenarios are shown in Tables 1, 2, and 3.

[0159] Table 1 Prediction indicators of each model in scenario 1

[0160]

[0161] Table 2 Prediction indicators of each model in scenario 2

[0162]

[0163] Table 3 Prediction indicators of each model in scenario 3

[0164]

[0165]

[0166] It can be seen that in different scenarios, the method proposed in the present invention has achieved better prediction results compared with each comparison model. Taking scenario 2 as an example, under the RMSE index, the proposed method is reduced by 2.79%, 5.22%, and 19.21% respectively compared with each comparison model; under the MAE index, it is reduced by 1.79%, 3.02%, and 21.54% respectively compared with each comparison model; under the R2 index, it is increased by 0.43%, 0.83%, and 4.21% respectively compared with each comparison model. In summary, the model proposed in the present invention has good applicability in short-term offshore wind power forecasting.

[0167] Example 3

[0168] This embodiment provides a SOM-EEMD-GRU combined model based on adaptive error correction for the offshore wind power short-term prediction method based on adaptive error correction described in Embodiment 1, including:

[0169] The preprocessing module is used to correct and dynamically fill in abnormal values ​​and missing values ​​in historical wind power data.

[0170] The SOM cluster analysis module is used to perform cluster analysis on historical wind power data, classify different weather types, and form a preliminary wind power data set.

[0171] EEMD and correlation analysis module. The correlation analysis module is used to analyze the meteorological data in the preliminary wind power data set and screen out meteorological features with high correlation with power changes. EEMD is used to analyze the power data in the preliminary wind power data set, extract the IMF component, and form a multimodal power data set. The meteorological features are fused with the multimodal power data set to form a wind power augmented data set.

[0172] The model prediction module inputs the wind power amplification data set into the GRU model based on the residual connection structure and attention mechanism. 80% of the data is used as the training set and 20% of the data is used as the test set. The prediction results are obtained after the model is trained. The residual connection structure realizes adaptive error correction in the prediction process and ensures the prediction accuracy.

[0173] The evaluation module uses the mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R-Square, R 2 ) to evaluate the accuracy of the prediction model.

[0174] Example 4

[0175] This embodiment provides a storage medium, in which a computer program is stored. When the computer program is executed by a processor, the device where the storage medium is located controls the method for short-term prediction of offshore wind power based on adaptive error correction described in Embodiment 1.

[0176] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0177] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0178] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0179] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0180] The embodiments are preferred implementations of the present invention, but the present invention is not limited to the above-mentioned implementations. Any obvious improvements, substitutions or modifications that can be made by those skilled in the art without departing from the essential content of the present invention belong to the protection scope of the present invention.

Claims

1. A method for predicting offshore wind power based on adaptive error correction, characterized in that: The method includes: S1 preprocesses the collected initial offshore wind power data, performs SOM cluster analysis on the preprocessed wind power data, divides different weather types, and forms a preliminary wind power data set; S2 performs correlation analysis on the meteorological features in the obtained preliminary wind power data set to screen out meteorological features that are highly correlated with power; processes the power data in the preliminary wind power data set using empirical mode decomposition to obtain multimodal power data, and fuses the screened high-power meteorological features with each component after modal decomposition processing to obtain a wind power augmented data set; S3 inputs the obtained wind power amplification data set into the improved GRU model, trains the model with the training set, and predicts the required wind power with the test set, wherein the improved GRU model includes: processing the reset gate in the GRU model with a weight coefficient based on the attention mechanism and adding a residual connection, wherein the residual connection is used to directly add the input and the output, thereby realizing the jump transmission of information; S4 uses relevant evaluation indicators to evaluate the effectiveness of the prediction model.

2. The offshore wind power prediction method based on adaptive error correction according to claim 1 is characterized in that: The SOM cluster analysis is performed on the pre-processed wind power data to divide different weather types and form a preliminary wind power data set, which specifically includes the following steps: S11 initialization: assign small random numbers to each weight vector of the output layer and normalize them to obtain the initial weight W ij (0), determine the initial domain range σ0, the initial learning rate η0 and the number of iterations K; S12 Data input: input the preprocessed wind power data X into the SOM clustering model; the wind power data X includes multiple samples, which are expressed as: X = {x1, x2, ..., x n }, where any sample x i Each is a vector containing power data and meteorological data; S13 distance calculation: For any input vector x i , calculate the weight vector W ij and the input vector x i The Euclidean distance between them is calculated as: After the calculation is completed, select the distance input vector x i The closest neuron is called the winning neuron i * , the winning neuron satisfies: Winning Neuron i * The weight vector can reflect the input data x i The features of the classification are helpful for subsequent clustering of input data; S14 weight update: Assume the winning neuron is i * , the weight vector of the winning neuron and its field neuron needs to be updated, and the weight update formula is: Where η(t) is the learning rate, which controls the amplitude of weight adjustment and usually decays gradually as training progresses to ensure the convergence of analysis results. is the neighborhood function, which defines the relationship between neurons i and i * The similarity is calculated as follows: in, is neurons i and i * The distance between them, σ(t) is the range of the domain function. As time goes by, the neighborhood range will gradually decrease; S15 Learning rate and domain scope update: After completing the above steps, the learning rate and domain scope are updated to provide support for the next training. The learning rate and domain scope update formulas are as follows: Among them, η0 and σ0 are the initial learning rate and domain range, τ η and τ σ is a constant that controls the decay rate; S16 repeat iteration: repeat the above steps S11-S15 until the number of iterations reaches K; S17 After the above steps, the weight vector of each neuron in the SOM network will gradually adapt to the distribution of the input data, and similar input data will be mapped to adjacent neurons to obtain the clustering result; the output data set is recorded as X1, which is the preliminary wind power data set.

3. The offshore wind power prediction method based on adaptive error correction according to claim 1 is characterized in that: In step S2, correlation analysis is performed on the meteorological features in the obtained preliminary wind power data set to screen out meteorological features that are highly correlated with power, including: The input preliminary wind power data set contains power data and meteorological characteristics. The power data is denoted as P and the meteorological characteristics are denoted as Feature. The input data set is X=[P,Feature]. The Pearson correlation coefficients of several meteorological characteristic data and power data are calculated to obtain the Pearson correlation coefficient matrix. According to the judgment principle in the principle part, the Pearson correlation coefficient |r|≥0.8 is recorded as a high correlation feature, namely Feature'.

4. The offshore wind power prediction method based on adaptive error correction according to claim 3 is characterized in that: In step S2, the power data in the preliminary wind power data set is processed by empirical mode decomposition to obtain multimodal power data, and the screened high-power meteorological features are fused with each component after the mode decomposition process to obtain a wind power augmented data set, which specifically includes: S22: The power data set in the preliminary wind power data set X1 is X2. The relevant analysis steps are as follows: Perform empirical mode decomposition on the power data set X2, with a total number of K times and variable i = 1. Add white noise p to the power data set X2. i (t), the value range is 0.1-0.4 times the ratio of the amplitude to the original signal amplitude, and the original vibration signal after superposition is: y i (t)=y(t)+p i (t)i=1,2,…,K (7) S23 superimposes p(t) with different amplitudes K times on y(t), and performs empirical mode decomposition after each superposition to obtain a set of IMF components c ij and the residual component r i (t), therefore, the result obtained after the i-th decomposition is in the form of: y i (t)=∑c ij (t)+r i (t)i=1,2,…,K (8) After S24 is decomposed K times, the IMF components can be obtained. The average value of these components will be the IMF components after the white noise is eliminated, that is: Where: c j (t) is the value of y i (t) is the jth component after the ensemble empirical mode decomposition, J is the total number of IMFs, signal y i (t) is the residual component of the signal y(t) after ensemble empirical mode decomposition; S25 fuses the high-correlation meteorological features obtained by correlation analysis with the IMF components obtained by ensemble empirical mode decomposition to form a wind power augmented data set. The fusion process can be expressed as follows: X input =[Feature',IMF i ] (10) Where, X input is the data in the augmented data set, Feature' represents the meteorological features with high correlation with power data, IMF i represents the component obtained after the power data is subjected to the empirical mode decomposition of the set, and integrates all the obtained X input , that is, the wind power augmented data set X3 is obtained.

5. The offshore wind power prediction method based on adaptive error correction according to claim 1 is characterized in that: In step S3, the reset gate in the GRU model is processed based on the weight coefficient in the attention mechanism, including: replacing the reset gate in the GRU model with the weight coefficient in the attention mechanism, and the weight coefficient in the attention mechanism is used to evaluate the similarity between the Query and each Key, which is expressed as: Among them, the hidden state h of the previous moment is used t-1 As the query vector Query, the current input x t As the key vector Key, W q and W k is the weight matrix used to transform the hidden state h t-1 and x t Input is transformed into query and keyspace, d k is the dimension of the key vector, which is used to scale the dot product. The Softmax function ensures that the sum of all time step weights is 1. It decides how to weight the hidden state h of the previous moment based on these weights. t-1 .

6. The offshore wind power prediction method based on adaptive error correction according to claim 5 is characterized in that: In the step S3, the reset gate in the GRU model is processed using a weight coefficient based on the attention mechanism, and further includes: The improved GRU model is recorded as GRU' model, and its corresponding hidden state is expressed by the following formula: In formula (13), the weight coefficient α is used t Replace the reset gate in the traditional model. After the replacement is completed, the candidate hidden state at the current moment is obtained Then Input into formula (14) to obtain the current hidden state and implement the update of the hidden state.

7. The offshore wind power prediction method based on adaptive error correction according to claim 6 is characterized in that: In step S3, a residual connection is added to the GRU' model, and the residual connection is used to directly add the input and the output, thereby realizing the jump transmission of information. The model is recorded as GRU', which is expressed by the following formula: h t =GRU(x t ,h t-1 )+x t ; (15) Among them, h t is the updated hidden state after the GRU' model is calculated. Since the attention mechanism is introduced to replace the traditional reset gate, x t is the current input data, and the input data x t Added directly to the output of the hidden state.

8. A offshore wind power prediction system based on adaptive error correction, characterized in that: The system includes: The data preprocessing module is used to preprocess the collected initial offshore wind power data, perform SOM cluster analysis on the preprocessed wind power data, divide different weather types, and form a preliminary wind power data set; The data set expansion module is used to perform correlation analysis on the meteorological features in the obtained preliminary wind power data set, and screen out meteorological features that are highly correlated with power; the power data in the preliminary wind power data set is processed using empirical mode decomposition to obtain multimodal power data, and the screened high-power meteorological features are merged with each component after modal decomposition processing to obtain a wind power expansion data set; A model training module, used for inputting the obtained wind power amplification data set into an improved GRU model, training the model with a training set, and predicting the required wind power with a test set, wherein the improved GRU model comprises: processing the reset gate in the GRU model with a weight coefficient based on an attention mechanism and adding a residual connection, wherein the residual connection is used for directly adding the input and the output, thereby realizing the jump transmission of information; The effect evaluation module is used to evaluate the effect of the prediction model using relevant evaluation indicators.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a program for predicting offshore wind power based on adaptive error correction, and when the program for predicting offshore wind power based on adaptive error correction is executed by a processor, the steps of the method for predicting offshore wind power based on adaptive error correction as described in any one of claims 1 to 7 are implemented.

10. An electronic device comprising a memory and a processor, characterized in that: A computer program is stored in the memory, and the processor is configured to run the computer program to perform the method according to any one of claims 1 to 7.

Images