New energy station power interval prediction method

Abstract

The present application belongs to the technical field of new energy station power prediction, and particularly relates to a new energy station power interval prediction method. In view of the deficiency that the prediction accuracy of the existing new energy station power interval prediction method has room for improvement, the present application adopts the following technical scheme: the new energy station power interval prediction method comprises: constructing a VMD-SA-TCN-BiLSTM-KDE prediction model; obtaining historical output power data of a wind power station, and inputting the preprocessed data into SA-TCN-BiLSTM for training, wherein the preprocessing comprises decomposing the data by using VMD; inputting the preprocessed part of the historical output power data into the trained SA-TCN-BiLSTM to obtain a prediction result; performing KDE kernel density estimation on the prediction error, and obtaining a historical output power prediction interval of the wind power station according to the prediction result and the prediction error. The present application has the beneficial effect of higher prediction accuracy.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of new energy station power prediction, and particularly relates to a new energy station power interval prediction method. BACKGROUND

[0002] New energy stations have become a core component of power supply. However, new energy output is affected by weather conditions such as wind speed and temperature, and presents the characteristics of high intermittency and randomness. This uncertainty poses a serious challenge to power grid optimization scheduling.

[0003] New energy station power interval prediction is to predict the upper and lower boundaries of new energy output at the to-be-predicted time under a given confidence level, and is an important method to improve the economic operation efficiency of the power system and reduce the safety risk coefficient of the power grid. Because new energy output is affected by many factors and has high uncertainty, a single model cannot effectively and fully capture the complex characteristics of new energy power output, and noise interference in the original data will also lead to an excessively wide prediction interval or insufficient confidence.

[0004] To solve this problem, signal decomposition and deep neural network prediction structures are widely used, but there are still the following shortcomings: 1. Empirical Mode Decomposition (EMD) has a modal aliasing phenomenon, which requires manual selection of effective components, increasing the computational burden; 2. Deep neural networks do not fully combine the advantages of each neural network, rely on a single structure, and are difficult to model local details and global trends at the same time, with low time series modeling efficiency; 3. Weak representation ability for irregular time series characteristics (such as intermittency and suddenness of wind and light power); 4. There is still room for further improvement in the accuracy of new energy station power interval prediction.

[0005] Related literature:

[0006] Patent application 1, CN108717579A - a short-term wind power interval prediction method;

[0007] Patent application 2, CN118503724A - a wind and light power interval prediction method based on modal decomposition and hybrid neural network combination;

[0008] Patent application 3, CN118508430A - a short-term wind power prediction method based on deep learning.

[0009] Related terms:

[0010] Variational mode decomposition (VMD);

[0011] Self-Attention (SA);

[0012] Temporal Convolutional Network (TCN);

[0013] Bidirectional Long Short-Term Memory (BiLSTM);

[0014] Kernel Density Estimation (KDE). SUMMARY

[0015] The present application aims at the deficiency that the prediction accuracy of the existing new energy station power interval prediction method has room for improvement, and provides a new energy station power interval prediction method based on VMD-SA-TCN-BILSTM-KDE. First, the original output power data of the wind power station is processed by using the variational mode decomposition method WMD. Then, the deterministic prediction of each frequency term is performed by using the bidirectional long short-term memory time convolutional neural network TCN-BiLSTM based on the self-attention mechanism SA. The prediction error of the deterministic prediction result is estimated by using the kernel density estimation method KDE. The prediction interval is superimposed on the predicted value to obtain the overall interval prediction result. Compared with other classical interval prediction models, the present application has higher prediction accuracy.

[0016] To achieve the above-mentioned purpose, the present application adopts the following technical solution: a new energy station power interval prediction method, which comprises the following steps:

[0017] S1, constructing a VMD-SA-TCN-BiLSTM-KDE prediction model;

[0018] S2, obtaining the historical output power data of the wind power station, preprocessing and inputting into the SA-TCN-BiLSTM for training to obtain the trained SA-TCN-BiLSTM. The preprocessing includes decomposing the historical output power data of the wind power station by using the variational mode decomposition method VMD.

[0019] S3, inputting the preprocessed part of the historical output power data into the trained SA-TCN-BiLSTM model to obtain the prediction result;

[0020] S4, performing KDE kernel density estimation on the prediction error, and obtaining the historical output power prediction interval of the wind power station according to the prediction result and the prediction error.

[0021] As an improvement, in S2, the time scale of the historical output power data is one year; the output power data is collected every 10 minutes.

[0022] As an improvement, in S2, the historical output power data of the wind farm station is decomposed into three intrinsic modal functions by using the variational modal decomposition method VMD; in order to ensure that the decomposition sequence is a modal function with a central frequency and a minimum sum of bandwidths of each modal function, the constraint condition is that the sum of all modal functions is equal to the power time sequence;

[0023] By continuously iterating and updating to find the optimal solution, the frequency of the output power data of the wind farm station is extracted and separated, and the high-frequency term, the medium-frequency term, the low-frequency term and the residual term corresponding to the output power data are obtained.

[0024] As an improvement, in S2, the preprocessing also includes standardizing the high-frequency term, the medium-frequency term, the low-frequency term and the residual term of the output power data; after standardization, the data set is scaled to a distribution interval with a mean of 0 and a standard deviation of 1, so that the numerical values of each modal component are in a similar order of magnitude, and the scale of the multi-modal component is unified and the model convergence speed is accelerated.

[0025] As an improvement, in S2, each frequency term of the standardized wind farm station output power data is used as input data of SA-TCN-BiLSTM, and the output data of SA-TCN-BiLSTM after inverse standardization is the prediction result of the output power of the wind farm station.

[0026] As an improvement S2, the standardized wind farm station output power data is divided into training set, validation set and test set according to a certain proportion;

[0027] In S2, the training set data is input into SA-TCN-BiLSTM for training, and the weight parameters in the training model are minimized to minimize the mean square error between the predicted value and the true value;

[0028] In S3, the validation set data is input into SA-TCN-BiLSTM-KDE to statistically analyze the prediction error and find the confidence interval range.

[0029] The new energy station power interval prediction method further includes: S5, inputting the test set data into SA-TCN-BiLSTM-KDE to evaluate the model.

[0030] As an improvement, in S2, the self-attention mechanism SA dynamically allocates weights to the input time series output power data, captures the dependency between different positions in the time series, and enhances the attention to key parts;

[0031] The time convolution network (TCN) adopts a 4-layer dilated convolution structure, and the dilated factor is multiplied layer by layer to extract long-term trend features in time series.

[0032] The bidirectional long short-term memory network (BiLSTM) is composed of two LSTMs, which process time series data from the forward and reverse directions respectively, and the number of hidden units is 32, and the total number of units in the bidirectional structure is 64, which fully learns the forward and backward dependencies of time series.

[0033] As an improvement, in S4, the probability density function used is expressed as:

[0034]

[0035]

[0036] wherein, n is the number of samples for one prediction, h is a bandwidth function used to control the smoothing degree, K is a Gaussian kernel function, is the mean of the prediction error; is the true value; is the predicted value;

[0037] Through fitting of the prediction error distribution, the probability density function of the prediction error is obtained.

[0038] As an improvement, in S4, the probability density function is further integrated to obtain the prediction error probability cumulative distribution function, which is expressed as follows:

[0039] ;

[0040] Based on the obtained prediction error probability cumulative distribution function, the statistical characteristics of the error are quantified and the confidence interval is generated, which is expressed as follows:

[0041]

[0042]

[0043] wherein, is the probability of the confidence interval, is the upper limit of the corresponding confidence interval, is the lower limit of the corresponding confidence interval, is the predicted value of the deterministic prediction;

[0044] Through statistical modeling of the prediction error and combining the prediction results of the training set, the output power prediction interval of the wind farm station is obtained.

[0045] As an improvement, the evaluation indexes include PICP, PINAW, CWC and MPICD.

[0046] The PICP index is used for evaluating coverage of the probability interval prediction model;

[0047] The PINAW index is used for evaluating average width of the prediction interval of the probability interval prediction model, and the prediction accuracy and generalization ability of the model are judged by comparison with the range of the observed value;

[0048] The CWC index comprehensively considers the PICP and the PINAW, that is, the coverage degree and the narrowness rate of the prediction interval are comprehensively considered;

[0049] The MPICD index is used for evaluating the deviation degree of the midpoint of the prediction interval from the true value.

[0050] The new energy station power interval prediction method of the embodiment has the following beneficial effects:

[0051] (1) The variational mode decomposition VMD adopted can adaptively decompose the power sequence of the new energy station, and by constraining the center frequency and bandwidth of the mode component, the mode aliasing problem existing in the traditional empirical mode decomposition is overcome, and clearer high-frequency and low-frequency sub-sequences are obtained.

[0052] (2) The self-attention mechanism SA is used in time series prediction to enhance the dependency relationship between different positions in the time series, and by assigning different weights to each position, the attention of the model to the key part is enhanced.

[0053] (3) The time series convolution network TCN adopted expands the receptive field through dilated causal convolution, realizes efficient capture of long-distance dependence under the premise of ensuring time series causality, and utilizes the parallel computing advantage to improve the learning efficiency, and has stronger time series modeling efficiency compared with the local feature extraction of the traditional CNN network.

[0054] (4) The bidirectional long short-term memory network BiLSTM adopted synchronously captures the forward and backward time series dynamic characteristics through the bidirectional structure, and the representation ability of the intermittent and sudden characteristics (such as the intermittence and suddenness of the wind and light power) of the time series is significantly enhanced compared with the unidirectional LSTM network, and by fusing the historical and future hidden state information, the periodic fluctuation and long-term trend change in the power sequence can be more accurately modeled.

[0055] (5) The adaptive bandwidth optimization algorithm is introduced in the probability prediction of the kernel density estimation KDE, and by dynamically adjusting the bandwidth of the kernel function, the uncertainty of the new energy power interval can be better and more finely represented. BRIEF DESCRIPTION OF DRAWINGS

[0056] Figure 1 is a flowchart of the new energy station power interval prediction method of the embodiment.

[0057] Figure 2 is a process schematic diagram of the new energy station power interval prediction method of the embodiment of the present application.

[0058] Figure 3 is a process schematic diagram of the new energy station power interval prediction method of the embodiment of the present application (also showing the internal structure of the model).

[0059] Figure 4 is a comparison diagram of the predicted interval and the true value obtained by using the method of the embodiment one of the present application.

[0060] Figure 5 is a comparison diagram of the predicted interval and the true value obtained by using the SA-CNN-LSTM-KDE model.

[0061] Figure 6 is a comparison diagram of the predicted interval and the true value obtained by using the VMD-SA-CNN-LSTM-KDE model. DETAILED DESCRIPTION

[0062] The technical solutions of the embodiments of the present application are explained and described below, but the following embodiments are only preferred embodiments of the present application, not all. Based on the embodiments in the embodiments, other embodiments obtained by those skilled in the art without creative labor are within the protection scope of the present application.

[0063] Referring to Figure 1 , the new energy station power interval prediction method of the embodiment one of the present application comprises the following steps:

[0064] S1, constructing a VMD-SA-TCN-BiLSTM-KDE prediction model;

[0065] S2, obtaining historical output power data of a wind power station, pre-processing and inputting SA-TCN-BiLSTM for training to obtain the trained SA-TCN-BiLSTM, the pre-processing including decomposing the historical output power data of the wind power station by using a variational mode decomposition method VMD;

[0066] S3, inputting the pre-processed part of the historical output power data into the trained SA-TCN-BiLSTM model to obtain a prediction result;

[0067] S4, performing KDE kernel density estimation on the prediction error, and obtaining a historical output power prediction interval of the wind power station according to the prediction result and the prediction error.

[0068] Referring to Figure 2 , the main idea of the prediction method of the embodiment of the present application is:

[0069] Firstly, the historical output power data of a single wind turbine is obtained, and the historical output power data of the wind farm station is summarized;

[0070] Then, the historical output power data of the wind farm station is processed by VMD to obtain high-frequency, medium-frequency, low-frequency and residual terms.

[0071] Secondly, the data processed by VMD is input into the SA-TCN-BiLSTM model for prediction to obtain deterministic results (point values).

[0072] Finally, the interval prediction (interval value) is obtained according to the KDE kernel density interval prediction and the deterministic results and output.

[0073] Referring to Figure 3 , the historical output power data of the wind farm station is obtained by collecting single wind turbines. Specifically, the output power data of the wind farm single wind turbine can be collected every 10 minutes.

[0074] Then, the historical output power data of the wind farm station is divided into training set, validation set and test set according to the ratio of 5:2:3. The training set data is used to train the weight parameters in the neural network, and the mean square error between the predicted value and the true value is minimized. The validation set is used for statistical analysis of the prediction error to find the confidence interval range; the test set is used to evaluate the generalization performance of the model based on PICP, PINAW, CWC and MPICD indicators.

[0075] At the same time, based on the output power data of the wind farm station, VMD is used to decompose the historical output power data of the wind farm station into three intrinsic mode functions. In order to ensure that the decomposition sequence is a modal function with a central frequency and a minimum bandwidth, the constraint condition is that the sum of all modal functions is equal to the power time sequence, and the VMD constrained variation model is as follows:

[0076]

[0077] Wherein, is the modal function; is the modal function center frequency; is the Dirac function; is the historical output power data set of the wind farm station.

[0078] Through continuous iteration and update to find the optimal solution , , the frequency of the output power data of the wind farm station is extracted and separated, and the high-frequency term (IMF1), the medium-frequency term (IMF2), the low-frequency term (IMF3) and the residual term (res) corresponding to the output power data are obtained.

[0079] Furthermore, the high-frequency, mid-frequency, low-frequency, and residual terms of the output power data are standardized using the following formula:

[0080]

[0081]

[0082]

[0083] in, The input time series dataset; The mean of the dataset; The standard deviation of the dataset; This is a standardized time series dataset.

[0084] After standardization, the dataset is scaled to a distribution range with a mean of 0 and a standard deviation of 1, so that the values ​​of each modal component are at similar magnitudes, unifying the scale of multimodal components and accelerating the convergence speed of the model.

[0085] Then, the high-frequency, mid-frequency, low-frequency, and residual data of the standardized wind farm power output are input into the SA-TCN-BILSTM for training. The SA-TCN-BILSTM model is trained based on the training set data. The time series input to the SA-TCN-BILSTM consists of the frequency terms of the standardized wind farm power output data (training set), and the output data, after destandardization, becomes the predicted wind farm power output for one day later.

[0086] Temporal Convolutional Neural Networks (TCNs) capture the long-term trend and local fluctuation features of the output power sequence through convolution. Dilated convolution is used to increase the coverage of the receptive field. In addition, residual connections are used to alleviate the gradient vanishing problem in deep networks. The expression for dilated convolution is shown below:

[0087]

[0088] in, For the first in the output sequence One element, For the first in the input sequence One element, As the expansion factor, The location of the convolution kernel. N The kernel size is [size]. These are the kernel weights.

[0089] The residual block is composed of dilated convolution, weight normalization, activation function and Dropout layer, and is directly added to the input through a skip connection, where weight normalization accelerates convergence and improves model stability, and Dropout layer randomly discards part of the neurons with a probability of 0.2 to prevent overfitting.

[0090] The bidirectional long short-term memory network BiLSTM is used to capture the forward and backward time sequence relationship of the power sequence, and simultaneously processes the time sequence data from two directions of forward and reverse. The internal structure of the LSTM unit includes an input gate, a forget gate and an output gate, which enables the LSTM unit to save useful information learned for a long time, better learn the long-term dependence in the time sequence, and avoid the problems of gradient disappearance and gradient explosion. Its expression is as follows:

[0091]

[0092]

[0093]

[0094]

[0095]

[0096] wherein, , , are the weight matrices of the input gate, the forget gate and the output gate, respectively; , , are the output sequences of the input gate, the forget gate and the output gate, respectively; , , are the corresponding biases; is the candidate cell state, representing new information that can be added to the cell state; is the current time step cell state, containing information processed by the forget gate and the input gate; is the current time step input sequence; is the previous time step hidden state; is a sigmoid activation function; tanh is a hyperbolic tangent activation function.

[0097] The self-attention mechanism SA is used in time series prediction to enhance the dependence between different positions in the time series. By assigning different weights to each position, the model's attention to key parts is enhanced. In SA, the input sequence is mapped to three different space vectors. Among them, the key is used to measure the relevance of the query to other elements in the sequence; the query is used to determine the part of the sequence that is focused on; and the value is the output feature obtained after weighted summation by the attention weight. The calculation expression is as shown below:

[0098]

[0099]

[0100]

[0101]

[0102] wherein, , , are the query matrix, the key matrix and the value matrix respectively; , , are the corresponding weight matrix; is the input time series; is the scaling factor; is the attention score; is the weight coefficient; is the final attention output. SA, TCN and BILSTM can all use existing models.

[0103] After training, the validation set data is input into the trained SA-TCN-BILSTM to obtain the prediction result, and then the prediction error is subjected to KDE kernel density estimation. The probability density function used is as shown below:

[0104]

[0105]

[0106] wherein, n is the number of samples for one prediction, h is the bandwidth function used to control the smoothing degree, K is the Gaussian kernel function, is the average value of the prediction error; is the true value; is the predicted value.

[0107] The probability density function of the prediction error can be obtained by fitting the prediction error distribution, and the prediction error probability cumulative distribution function can be obtained by further integrating the probability density function, as shown in the following expression:

[0108]

[0109] Based on the obtained prediction error probability cumulative distribution function, the statistical characteristics of the error are quantified and the confidence interval is generated, as shown in the following expression:

[0110]

[0111]

[0112] wherein, the probability of the confidence interval, the upper limit of the corresponding confidence interval, the lower limit of the corresponding confidence interval, the predicted value of the deterministic prediction.

[0113] S3, after adjusting the parameters of SA-CNN-SLTM through the validation set, inputting the test set into SA-TCN-BiLSTM-KDE to obtain the prediction result, and using PICP, PINAW, CWC and MPICD indexes to measure and evaluate the prediction result.

[0114]

[0115] wherein I is an indicator function, which takes the value 1 when the condition in the parentheses is true, and 0 otherwise. This formula represents the probability that the actual value falls within the prediction interval. When PICP is higher, it indicates that the prediction interval of the model is more reliable, and the prediction ability is better.

[0116]

[0117] By calculating PINAW, the average width of the probability interval prediction model relative to the true value range can be evaluated, and the prediction accuracy and generalization ability of the model can be judged. The smaller the value, the smaller the proportion of the prediction interval relative to the true value range, and the higher the prediction accuracy.

[0118]

[0119] Since there is a certain contradiction between PICP and PINAW indexes, the larger PINAW is, the value of PICP tends to be larger, so CWC considers the coverage degree and narrowness rate of the prediction interval. Wherein is the set confidence level, and the 90% confidence level is 0.1; if PICP ≥ , then is 0, otherwise 1; when PICP is less than the confidence interval, is 1, there is a penalty coefficient for the prediction error. Therefore, the smaller the CWC, the better the performance of the prediction model.

[0120]

[0121] The formula indicates that when the midpoint of the prediction interval is closer to the true value, the MPICD index is better, and the time series prediction is more accurate.

[0122] Referring to Figures 4 to 6 , the prediction interval obtained by the VMD-SA-TCN-BiLSTM-KDE model of the embodiment shown in Figure 4 is compared with existing models (such as the SA-CNN-SLTM-KDE of Figure 5 , the VMD-SA-CNN-LSTM-KDE of Figure 6 ). Figures 4 to 6 It can be seen from that the power prediction interval obtained by the VMD-SA-TCN-BiLSTM-KDE of the embodiment is closer to the actual value, verifying the effectiveness of the proposed power interval prediction model VMD-SA-TCN-BILSTM-KDE, which can provide more reliable and narrower confidence intervals for power interval prediction compared with existing models. Specifically, the evaluation indexes (PICP, PINAW and CWC) are used to compare the model of the embodiment and the existing model, and the comparison results are as follows.

[0123] Table 1 Comparison of evaluation indexes of 30% confidence interval of prediction model

[0124]

[0125] Table 2 Comparison of evaluation indexes of 60% confidence interval of prediction model

[0126]

[0127] Table 3 Comparison of evaluation indexes of 90% confidence interval of prediction model

[0128]

[0129] As can be seen from Tables 1 to 3, the VMD-SA-TCN-BiLSTM-KDE model of the embodiment has a higher PICP value than the existing VMD-SA-CNN-LSTM-KDE or SA-CNN-SLTM-KDE, indicating that the prediction interval of the model of the embodiment is more reliable and the prediction ability is better; the PINAW is not much different from the existing model; the CWC is significantly smaller than the existing model, and the performance of the model is better.

[0130] The new energy station power interval prediction method of the embodiment of the application is a prediction model based on VMD-SA-TCN-BiLSTM-KDE. First, VMD decomposes the new energy station power sequence into multiple sub-sequences with different frequency characteristics, effectively decomposes the nonlinear components in the data, and provides information reference in the time domain and the frequency domain for fine modeling of the new energy station power. In view of the long-term characteristics of the new energy station power driven by seasonal light and wind speed changes, a TCN is used to stack multiple layers of convolution layers with increasing expansion coefficients, so as to efficiently capture the multi-scale global features of the input time sequence. In order to further depict the short-term intermittent fluctuations in the new energy station power, a BiLSTM is introduced on the basis of the high-dimensional global features extracted by the TCN, which combines the forward and backward time dimension information through the bidirectional gating mechanism, and enhances the learning ability of the model for local features. In addition, an attention mechanism SA is introduced in the neural network to capture the dependency relationship between different positions in the new energy station power sequence, and by assigning different weights to each position, the attention of the model to the key part is enhanced, and the model has better learning ability for the complementary features of the new energy station power sequence in space and time.

[0131] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto. Those skilled in the art should understand that the present application includes but is not limited to the above specific embodiments. Any modification that does not deviate from the functional and structural principles of the present application will be included in the scope of the claims.

Claims

1. A method for power interval prediction of a new energy station, characterized in that: The new energy station power interval prediction method comprises the following steps: S1, constructing a VMD-SA-TCN-BiLSTM-KDE prediction model; S2, obtaining historical output power data of the wind farm station, preprocessing and inputting SA-TCN-BiLSTM for training to obtain the trained SA-TCN-BiLSTM, the preprocessing including decomposing the historical output power data of the wind farm station by using the variational mode decomposition method VMD; S3, inputting the preprocessed part of the historical output power data into the trained SA-TCN-BiLSTM model to obtain a prediction result; S4, performing KDE kernel density estimation on the prediction error, and obtaining the historical output power prediction interval of the wind farm station according to the prediction result and the prediction error; In S2, the historical output power data of the wind farm station is decomposed into three intrinsic mode functions by using the variational mode decomposition method VMD, in order to ensure that the decomposed sequence is a mode function with a central frequency and a minimum sum of bandwidths of each mode function, and the constraint condition is that the sum of all mode functions is equal to the power time sequence; In S2, the optimal solution is found by continuously iterating and updating, so as to extract and separate the frequency of the output power data of the wind farm station, and obtain the high-frequency term, the medium-frequency term, the low-frequency term and the residual term of the corresponding output power data.

2. The method of claim 1, wherein: In S2, the time scale of the historical output power data is one year; the output power data is collected every 10 minutes.

3. The method of claim 1, wherein: In S2, the preprocessing further includes standardizing the high-frequency term, the medium-frequency term, the low-frequency term and the residual term of the output power data, and after standardization, the data set is scaled to a distribution interval with a mean value of 0 and a standard deviation of 1, so that the numerical values of each mode component are in a similar order of magnitude, the scales of multiple mode components are unified, and the model convergence speed is accelerated.

4. The method of claim 3, wherein: In S2, each frequency term of the standardized wind farm station output power data is used as the input data of SA-TCN-BiLSTM, and the output data of SA-TCN-BiLSTM is the prediction result of the wind farm station output power after being de-standardized.

5. The method of claim 4, wherein: In S2, the standardized wind farm station output power data is divided into a training set, a validation set and a test set according to a certain proportion; In S2, the training set data is input into SA-TCN-BiLSTM for training, the weight parameters in the training model are trained, and the mean square error between the predicted value and the true value is minimized; In S3, the validation set data is input into SA-TCN-BiLSTM-KDE, the prediction error is statistically analyzed, and the confidence interval range is found; The new energy station power interval prediction method further comprises: S5, inputting the test set data into SA-TCN-BiLSTM-KDE to evaluate the model.

6. The method of claim 5, wherein: In S2, the self-attention mechanism SA dynamically allocates weights to the input time series output power data, captures the dependency relationship between different positions in the time series, and enhances the attention to the key part; The time series convolution network TCN adopts a 4-layer dilated convolution structure, and the dilation factor is multiplied by a layer, so as to extract the long-term trend characteristics in the time series; The bidirectional long short-term memory network BiLSTM is composed of two LSTMs, which respectively process time series data from the forward and reverse directions, and has 32 hidden units, and a total of 64 units in the bidirectional structure, which sufficiently learns the forward and backward dependencies of the time series.

7. The method of claim 5, wherein: In S4, the probability density function used is expressed as: wherein, n is the number of samples for prediction, h is a bandwidth function for controlling the degree of smoothing, K is a Gaussian kernel function, is the mean of the prediction errors; is the true value; is the predicted value; Through fitting of the prediction error distribution, the probability density function of the prediction error is obtained.

8. The method of claim 7, wherein: In S4, the probability density function is further integrated to obtain the prediction error probability cumulative distribution function, which is expressed as follows: ； Based on the obtained prediction error probability cumulative distribution function, the statistical characteristics of the error are further quantified and the confidence interval is generated, which is expressed as follows: wherein, is a probability of the confidence interval, is an upper bound of the corresponding confidence interval, is a lower bound of the corresponding confidence interval, is a prediction value of the deterministic prediction; Through statistical modeling of the prediction error, combined with the prediction results of the training set, the output power prediction interval of the wind farm station is obtained.

9. The method of claim 5, wherein: The evaluation indexes include PICP, PINAW, CWC and MPICD; The PICP index is used to evaluate the coverage rate of the probability interval prediction model; The PINAW index is used to evaluate the average width of the prediction interval of the probability interval prediction model, and by comparing with the observed value range, the prediction accuracy and generalization ability of the model are further judged; The CWC index comprehensively considers PICP and PINAW, that is, it comprehensively considers the coverage degree and narrowness rate of the prediction interval; The MPICD index is used to evaluate the deviation degree of the midpoint of the prediction interval from the true value.

Images