CN115330085A - Wind speed prediction method based on deep neural network without future information leakage - Google Patents

Abstract

The invention discloses a wind speed prediction method based on a deep neural network and without future information leakage, which comprises the following steps: step 1: respectively screening and preprocessing the wind speed sequence data; screening the effective component Tc according to the sub-modal energy proportion and taking the effective component Tc as an output value of a prediction model; preprocessing the wind speed sequence data by adopting a real-time rolling decomposition strategy to obtain an input value of a prediction model without information leakage; constructing a data set with input values and output values; and 2, step: determining a prediction step size: determining the future time steps needing to be predicted according to the actual prediction demand; and step 3: constructing a prediction model: constructing a bidirectional long-short term memory network combined with an attention mechanism as a prediction model, and training and testing the prediction model by using a data set; and 4, step 4: and (3) wind speed prediction: and predicting the wind speed by using the prediction model obtained by training to obtain a wind speed prediction result. The method avoids the risk of data leakage and improves the practicability.

Description

Wind speed prediction method based on deep neural network without future information leakage

technical field

The invention belongs to the technical field of wind speed prediction, in particular to a wind speed prediction method based on a deep neural network and without future information leakage.

Background technique

Accurate wind speed prediction is crucial in the improvement of weather forecast accuracy, early warning of meteorological disasters, and the development of clean wind power. In the field of wind power, accurate short-term wind speed prediction can allow operators to reasonably regulate the overall operation of wind farms, maintain stable energy supply and reduce the impact on the grid. With the gradual development of domestic onshore wind farms to low wind speed areas, more and more wind farms are located in mountainous areas. Due to the influence of terrain changes and other factors, mountainous areas are prone to wake effects and complex flow phenomena such as flow separation and reattachment, and shear flow. Therefore, in order to ensure the safe and stable operation of wind farms in mountainous areas, it is urgent to develop methods for accurately predicting short-term wind speed.

Currently commonly used forecasting methods are divided into the following categories: physical methods, statistical methods, artificial intelligence methods and hybrid methods. Physical methods consume a lot of computing resources and time, and perform well in mid- and long-term forecasts, but it is difficult to meet the timeliness requirements of short-term forecasts. Statistical methods are difficult to model and predict nonlinear wind speed sequences reasonably due to their inherent linear assumptions. The artificial intelligence method only uses a shallow neural network, which lacks the ability to extract deep nonlinear features of the wind speed sequence.

With the rapid development of computer capabilities, it has become feasible to use deep neural networks for wind speed prediction. Although the short-term wind speed prediction model based on deep learning has achieved good prediction results, it is undeniable that a single neural network model may be difficult to apply to different wind speed sequences. Moreover, there is a large amount of noise information in the measured wind speed sequence, which seriously affects the accuracy of wind speed prediction. Therefore, hybrid forecasting models combined with data preprocessing methods have been widely developed, among which the mode decomposition method is the most widely used. The data preprocessing operation can indeed significantly reduce the randomness and high-frequency noise of wind speed, improve the predictability of wind speed and the performance of the prediction model, and greatly improve the accuracy of wind speed prediction. However, this kind of hybrid forecasting model decomposes all the wind speed sequence data during data preprocessing, divides the training set and test set on the decomposed sub-modes and establishes the forecasting model. This means that the test set data that should be unknown is regarded as known, which inevitably leads to the leakage of future information. When new data is added to the original wind speed sequence data, it needs to be decomposed again, and then substituted into the sub-mode prediction model for prediction. But even if only part of the new data is added at the end of the original data, the decomposed submodality will produce obvious changes at the end of the sequence. This part of terminal data is the most critical information in time series prediction. This means that predictive models built on submodalities of the original data may be difficult to apply to new data, resulting in the lack of practicality of such hybrid predictive models.

Contents of the invention

In view of this, the purpose of the present invention is to provide a wind speed prediction method based on a deep neural network and without future information leakage, which adopts an after-the-fact evaluation method, which avoids the risk of data leakage and improves practicability.

To achieve the above object, the present invention provides the following technical solutions:

A method for predicting wind speed based on a deep neural network without future information leakage, comprising the following steps:

Step 1: Perform data screening and data preprocessing on the wind speed series data respectively; screen the effective component Tc according to the proportion of sub-modal energy and take it as the output value of the prediction model; use the real-time rolling decomposition strategy to preprocess the wind speed series data to obtain infinite The input value of the prediction model of information leakage; construct the data set with the input value and output value;

Step 2: Determine the forecast step size: according to the actual forecast demand, determine the number of future time steps that need to be forecasted;

Step 3: Build a prediction model: build a bidirectional long-short-term memory network combined with an attention mechanism as a prediction model, and use the data set to train and test the prediction model;

Step 4: Wind speed prediction: use the prediction model obtained from training to predict the wind speed, and obtain the wind speed prediction result.

Further, in the step 1, the method for screening out the effective component Tc from the wind speed sequence data is:

S1: Divide the wind speed sequence data into training set, verification set and test set, and form three data groups with the divided data. The first data group includes the training set, the second data group includes the verification set, and the third data group Include the test set;

S2: Perform modal decomposition on the wind speed sequence data in the three data groups;

S3: Judging the index based on the energy principle, selecting the appropriate sub-module to obtain the effective division Tc _i that can represent most of the energy of the wind speed sequence data, i=1, 2 or 3, respectively representing the selection from the wind speed sequence data of the three data groups the effective share obtained;

S4: Splicing the effective components Tc _i in order to obtain the target output value Tc of the prediction model.

Further, in the step S3, when the sum of the energy of the first m sub-modes is greater than or equal to 99% of the energy of the original wind speed sequence, _Tci is the sum of the first m sub-modes, and the calculation method of the ER of the sub-modes is:

表示第k个子模态序列；X表示风速序列数据；N表示风速序列数据X的长度。Among them, ER(k) represents the energy ratio of the kth sub-mode;

Represents the kth sub-mode sequence; X represents the wind speed sequence data; N represents the length of the wind speed sequence data X.

Further, in the first step, the method for preprocessing the wind speed sequence data by using the real-time rolling decomposition strategy is:

(1) Determine the window length L: analyze the spectral characteristics of the wind speed sequence data X, find the frequency peak point, obtain the corresponding cycle, and select the cycle number as the window length required by RTRD;

(2) Determine the sliding stride s;

(3) Reconstruct the wind speed signal: according to the window length L and the sliding step s, the wind speed sequence data X is reconstructed into a two-dimensional matrix X _R , the matrix shape is (N-L+s, L), where N is the wind speed sequence length of data X;

(4) Execute RTRD: decompose X _R line by line, and get the decomposition result without information leakage

Further, the predictive model includes a layer of input layer, two layers of Bi-LSTM network, a layer of attention layer and a layer of fully connected layer arranged in sequence.

Further, the principle of the LSTM network is:

f _t ＝σ(W _xf x _t +W _hf h _t-1 +b _f )

g _t ＝tanh(W _xg x _t +W _hg h _t-1 +b _g )

i _t = σ(W _xi x _t +W _hi h _t-1 +b _i )

c _t =f _t ⊙c _t-1 +i _t ⊙g _t

o _t ＝σ(W _xo x _t +W _ho h _t-1 +b _o )

y _t ＝h _t ＝o _t ⊙tanh(c _t )

Among them, x _t is the input vector, c _t is the long-term state, h _t is the short-term state, y _t is the output vector; W _xf , W _xg , W _xi , W _xo are the weight matrices connected with x _t respectively, W _hf , W _hg , W _hi , W _ho are the weight matrices connected to h _t-1 respectively, b _f , b _g , _bi , b _o are the four-layer bias items respectively; f _t is the controller of the forget gate, and f _t determines which parts of c _t-1 should be deleted; g _t is the intermediate output of the LSTM network, which is used to analyze x _t and h _t-1 ; it _is the controller of the input gate, which determines which important parts of g _t should be added to c _t ; o _t is the controller of the output gate, which part of c _t should be read and applied to h _t and y _t ; σ represents the activation function.

Further, the attention layer adopts the Luong-Attention model, and its calculation method is:

The calculation method of score is:

The calculation method of attention weight α _ts is:

和h_t分别为编码器输出的全部隐藏状态以及最后一个隐藏状态；W表示权重矩阵；in,

and h _t are all hidden states output by the encoder and the last hidden state; W represents the weight matrix;

计算方法为：context vector

The calculation method is:

是全部隐藏状态；where α _ts is the attention weight,

is all hidden state;

The calculation method of the attention vector a _t is:

为上下文向量；h_t为最后一个隐藏状态；W_c表示将要学习的模型参数。in,

is the context vector; h _t is the last hidden state; W _c represents the model parameters to be learned.

Further, the attention _vector at is input to the fully connected layer to obtain the wind speed prediction result.

The beneficial effects of the present invention are:

Existing single-point wind speed combination forecasting models generally adopt a fully decomposed data preprocessing strategy. This data preprocessing strategy has the risk of data leakage, which may lead to the lack of engineering practicability of such combined forecasting models. Therefore, accurate prediction The effective components of wind speed appear to be more critical and feasible. Considering that there is a large amount of high-frequency noise in the measured wind speed, this part of high-frequency noise is an important source of wind speed randomness and volatility, and the low-frequency effective components of wind speed occupy most of the energy of the wind speed. The data preprocessing method extracts the effective components of the wind speed sequence and establishes an expression method without leakage of future information, which is the premise of obtaining reliable and practical wind field prediction results.

Considering that the traditional combination prediction model generally has the problem of data leakage, and the measured wind speed data also has a certain degree of noise components, the present invention is based on a deep neural network and no future information leakage wind speed prediction method, and adopts a data preprocessing strategy without information leakage The short-term wind speed forecasting framework aims to provide a method for accurately predicting the effective components of wind speed series. The prediction method includes three parts: data screening, data preprocessing based on real-time rolling decomposition, and Bi-LSTM neural network (Bi-LSTM-Attention) integrated with the Attention mechanism, which avoids the risk of data leakage and effectively analyzes noise components. processing, and the method of the present invention uses post-assessment, so the truth value obtained through modal decomposition is obtained with a delay.

Description of drawings

In order to make the purpose, technical scheme and beneficial effect of the present invention clearer, the present invention provides the following drawings for illustration:

Fig. 1 is the schematic diagram of the wind speed prediction method based on deep neural network and no future information leakage in the present invention;

Figure 2 is an illustration of the flow of data in the model training phase, verification phase, and testing phase (i.e., the use phase);

Figure 3 shows the single-step forecast curves for all models.

Detailed ways

The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments, so that those skilled in the art can better understand the present invention and implement it, but the examples given are not intended to limit the present invention.

As shown in FIG. 1 , the method for predicting wind speed based on a deep neural network and without future information leakage in the present invention includes the following steps.

Step 1: Carry out data screening and data preprocessing for wind speed sequence data respectively; screen the effective component Tc according to the proportion of sub-modal energy and use it as the output value of the prediction model; adopt the real-time rolling decomposition (RTRD) strategy to analyze the wind speed The sequence data is preprocessed to obtain the input values of the predictive model without information leakage; the data set is constructed with the input values and output values.

Specifically, the method of screening out the effective component Tc from the wind speed sequence data is:

S1: Divide the wind speed sequence data into training set, verification set and test set. In this embodiment, the wind speed sequence data is divided into training set (1-0.6N), verification set (0.6N-0.8N) and For the test set (0.8N-N), three data groups are formed with the divided data, the first data group includes the training set, the second data group includes the training set, and the third data group includes the test set.

S2: Perform modal decomposition on the wind speed sequence data in the three data groups. The modal decomposition method is used to decompose the wind speed data within the range of the first data group (1-0.6N), the second data group (0.6N-0.8N) and the third data group (0.8N-N), Among them, VMD and SSA algorithms need to specify the number of decomposition layers, while EMD and ICEEMDAN algorithms can adaptively determine the number of decomposition layers.

S3: Judging the index based on the energy principle, selecting the appropriate sub-module to obtain the effective division Tc _i that can represent most of the energy of the wind speed sequence data, i=1, 2 or 3, respectively representing the selection from the wind speed sequence data of the three data groups The effective share obtained. Specifically, the number of sub-modes required to screen the effective component Tc according to the energy ratio (Energy ratio, ER) of the sub-modes (the sub-modes are arranged in the order of low frequency to high frequency) is: the sum of the energy of the current m sub-modes When it is greater than or equal to 99% of the energy of the original wind speed sequence, it can be considered that Tc _i is the sum of the first m sub-modes. The ER calculation method of the sub-mode is:

表示第k个子模态序列；X表示风速序列数据；N表示风速序列数据X的长度。Among them, ER(k) represents the energy ratio of the kth sub-mode;

Represents the kth sub-mode sequence; X represents the wind speed sequence data; N represents the length of the wind speed sequence data X.

S4: Splicing the effective components Tc _i in order to obtain the target output value Tc of the prediction model, Tc is the data source of the subsequent model target output, that is, the true value.

Specifically, the method of preprocessing the wind speed sequence data using the real-time rolling decomposition strategy is as follows:

(1) Determine the window length L: analyze the spectral characteristics of the wind speed sequence data X, find the frequency peak point, obtain the corresponding period, and select the appropriate period number as the window length required by RTRD;

(2) Determine the sliding step s. When the short-term wind speed prediction is carried out in a step-by-step manner, the sliding step should be set to 1;

(3) Reconstruct the wind speed signal: according to the window length L and the sliding step s, the wind speed sequence data X is reconstructed into a two-dimensional matrix X _R , the matrix shape is (N-L+s, L), where N is the wind speed sequence length of data X;

将

每一行数据的前m个子模态相加，可以构成输入矩阵

计算

与Tc的相关系数；利用相关系数准则判断不同分解算法下构成

的最优子模态数量。(4) Execute RTRD: decompose X _R line by line, and get the decomposition result without information leakage

Will

The first m sub-modalities of each row of data are added to form an input matrix

calculate

Correlation coefficient with Tc; use the correlation coefficient criterion to judge the composition under different decomposition algorithms

The optimal number of submodals for .

Step 2: Determine the forecast step size: According to the actual forecast demand, determine the number of future time steps that need to be forecasted.

Step 3: Build a prediction model: Build a bidirectional long-term short-term memory network (Bi-LSTM-Attention) combined with an attention mechanism as a prediction model, and use the data set to train and test the prediction model. Specifically, in this embodiment, the prediction model includes an input layer, two Bi-LSTM networks, an attention layer, and a fully connected layer arranged in sequence.

The LSTM network of this embodiment uses the following method to update the hidden state at time t:

f _t ＝σ(W _xf x _t +W _hf h _t-1 +b _f )

g _t ＝tanh(W _xg x _t +W _hg h _t-1 +b _g )

i _t = σ(W _xi x _t +W _hi h _t-1 +b _i )

c _t =f _t ⊙c _t-1 +i _t ⊙g _t

o _t ＝σ(W _xo x _t +W _ho h _t-1 +b _o )

y _t ＝h _t ＝o _t ⊙tanh(c _t )

Among them, x _t is the input vector, c _t is the long-term state, h _t is the short-term state, y _t is the output vector; W _xf , W _xg , W _xi , W _xo are the weight matrices connected with x _t respectively, W _hf , W _hg , W _hi , W _ho are the weight matrices connected to h _t-1 respectively, b _f , b _g , _bi , b _o are the four-layer bias items respectively; f _t is the controller of the forget gate, and f _t determines which parts of c _t-1 should be deleted; g _t is the intermediate output of the LSTM network, which is used to analyze x _t and h _t-1 ; it _is the controller of the input gate, which determines which important parts of g _t should be added to c _t ; o _t is the controller of the output gate, which part of c _t should be read and applied to h _t and y _t ; σ represents the activation function.

The attention layer of this embodiment adopts the Luong-Attention model, and its calculation method is:

The calculation method of score is:

The calculation method of attention weight α _ts is:

和h_t分别为编码器输出的全部隐藏状态以及最后一个隐藏状态；W表示权重矩阵；in,

and h _t are all hidden states output by the encoder and the last hidden state; W represents the weight matrix;

计算方法为：context vector

The calculation method is:

是全部隐藏状态；where α _ts is the attention weight,

is all hidden state;

The calculation method of the attention vector a _t is:

为上下文向量；h_t为最后一个隐藏状态；W_c表示将要学习到的模型参数。in,

is the context vector; h _t is the last hidden state; W _c represents the model parameters to be learned.

Finally, the attention _vector at is input to the fully connected layer to obtain the wind speed prediction result.

Step 4: Wind speed prediction: use the prediction model obtained from training to predict the wind speed, and obtain the wind speed prediction result.

Figure 2 illustrates how the data is run through the training, validation, and testing phases of the neural network. In Figure 2(a), when constructing the input to the model, the modal decomposition follows a real-time rolling decomposition strategy, which is performed on the blue data pane. This strategy ensures that the input of the model always uses the information before the predicted value, that is, the confidentiality of future information is guaranteed, which means that there is no information leakage. The predicted value mainly refers to the effective component of the predicted wind speed, represented by Tc, which follows the meaning of "true value". In Fig. 2(b), during the model training process, since the data has been collected on the complete time axis, the true value Tc ₁ can be obtained by modality decomposition when training the neural network. Therefore, the error can be directly calculated and used to update the parameters of the neural network. In Fig. 2(b), during the verification and testing process, the parameters of the neural network have been determined, and the error Tc ₂ or Tc ₃ between the predicted result and the real value is calculated in a delayed manner instead of real-time calculation. This means that errors can be calculated after a period of operation, and they are only used to prove the accuracy of the forecast. In conclusion, no matter in the training process or in the verification and testing process, the real-time rolling decomposition strategy is always adhered to, and avoiding information leakage is the first principle of our algorithm.

Experimental verification

The wind speed prediction method based on the deep neural network and without future information leakage of the present invention will be verified below in conjunction with specific examples.

The 10-min interval wind speed data X of a wind measuring tower in Yunnan, China, at a height of 70m for 91 consecutive days was selected as the input signal, and the selected dates were from 2010.4.1 to 2010.6.30.

1. Evaluation standard of prediction result error

In this embodiment, three commonly used performance indicators are selected as judgment standards for measuring prediction model errors, including MAPE, MAE and RMSE, and their definitions are as follows.

代表预测值。MAPE、MAE和RMSE的值越小，预测精度越高。Among them: x _n represents the target output value (also known as the true value),

represents the predicted value. The smaller the value of MAPE, MAE and RMSE, the higher the prediction accuracy.

In addition, the relative improvement ratio P _metric of the three error indicators RMSE, MAPE, and MAE is defined to further quantitatively evaluate the performance of different models. P _metric is defined as follows

Among them, metric represents the three indicators of MAPE, MAE and RMSE, and E _a and E _b represent the prediction errors of model a and model b.

2. Comparison of prediction model with different models

The deep learning platform selected in this embodiment is GPU-based TensorFlow 2.3, Python 3.7 version, and builds the Bi-LSTM-Attention model. The training process uses the "Adam" optimizer, the learning rate lr is set to 0.001, the loss function is selected as "Mean Squared Error", and the number of iterations is fixed at 100 epochs. The parameters of each layer of the model are jointly determined by manual selection and random search.

Table 1 shows the comparison between the persistence model (Persistence, PR), BP model, GRU, LSTM and Bi-LSTM-Attention proposed by the present invention. Table 1 and Table 2 show the RMSE, P _MSE , MAPE, P _MAPE , MAE and P _MAE error results of the single-step forecast and 1-6-step forecast of different forecasting models.

Table 1 Prediction performance of each model based on VMD method—one-step prediction

Table 2 Prediction performance of each model based on VMD method—1-6 step prediction

As can be seen from the comparison of all the above neural network models, the method of the present invention can accurately predict the effective component Tc that occupies most of the energy of the wind speed sequence, and the accuracy of the single-step and 1-6-step prediction of the model is better than that of the PR model. And the Bi-LSTM-Attention prediction model proposed in the present invention has achieved the best prediction accuracy, which fully proves the effectiveness of the RTRD-Bi-LSTM-Attention prediction method proposed in the present invention. Specifically, under single-step prediction, the accuracy improvement ratio P _metric of the Bi-LSTM-Attention prediction module relative to the PR model is between 23.79% and 32.77%; under 1-6 step prediction, the Bi-LSTM-Attention prediction module is relatively The accuracy improvement ratio P _metric of the PR model is between 9.90% and 18.09%. When the prediction step size increases, the P _metric of all neural network models decreases, indicating that the performance of neural network models in multi-step wind speed prediction decreases.

The above-mentioned embodiments are only preferred embodiments for fully illustrating the present invention, and the protection scope of the present invention is not limited thereto. Equivalent substitutions or transformations made by those skilled in the art on the basis of the present invention are all within the protection scope of the present invention. The protection scope of the present invention shall be determined by the claims.

Claims (8)

1. A wind speed prediction method based on a deep neural network and without future information leakage is characterized by comprising the following steps: the method comprises the following steps:

step 1: respectively screening and preprocessing the wind speed sequence data; screening the effective component Tc according to the sub-mode energy ratio and taking the effective component Tc as an output value of a prediction model; preprocessing the wind speed sequence data by adopting a real-time rolling decomposition strategy to obtain an input value of a prediction model without information leakage; constructing a data set with input values and output values;

step 2: determining a prediction step size: determining the future time steps needing to be predicted according to the actual prediction demand;

and step 3: constructing a prediction model: constructing a bidirectional long-short term memory network combined with an attention mechanism as a prediction model, and training and testing the prediction model by using a data set;

and 4, step 4: and (3) wind speed prediction: and predicting the wind speed by using the prediction model obtained by training to obtain a wind speed prediction result.

2. The wind speed prediction method based on the deep neural network and without the leakage of future information as claimed in claim 1, wherein: in the first step, the method for screening out the effective component Tc from the wind speed sequence data comprises the following steps:

s1: dividing the wind speed sequence data into a training set, a verification set and a test set, and establishing three data groups by using the divided data groups, wherein the first data group comprises the training set, the second data group comprises the verification set, and the third data group comprises the test set;

s2: modal decomposition is carried out on the wind speed sequence data in the three data groups respectively;

s3: based on the energy principle judgment index, a proper submodule is selected to obtain the effective fraction Tc which can represent most of the energy of the wind speed sequence data _i I =1,2 or 3, respectively, representing the effective fractions screened from the wind speed sequence data of the three data sets;

s4: sequentially adding effective components Tc _i And splicing to obtain a target output value Tc of the prediction model.

3. The wind speed prediction method based on the deep neural network and without the future information leakage as claimed in claim 2, characterized in that: in step S3, when the sum of the current m sub-modal energies is greater than or equal to 99% of the energy of the original wind speed sequence, tc is _i The ER calculation mode of the submode is the sum of the first m submodes:

wherein ER (k) represents the energy fraction of the kth sub-mode;

representing a kth sub-modality sequence; x represents wind speed sequence data; n represents the length of the wind speed sequence data X.

4. The wind speed prediction method based on the deep neural network and without the leakage of future information as claimed in claim 1, wherein: in the first step, the method for preprocessing the wind speed sequence data by adopting the real-time rolling decomposition strategy comprises the following steps:

(1) Determining the window length L: analyzing the frequency spectrum characteristics of the wind speed sequence data X, searching a frequency peak point to obtain a corresponding period, and selecting the period number as the window length required by RTRD;

(2) Determining a sliding step s;

(3) Reconstructing a wind speed signal: reconstructing the wind speed sequence data X into a two-dimensional matrix X according to the window length L and the sliding step s _R The matrix shape is (N-L + s, L), wherein N is the length of the wind speed sequence data X;

(4) And executing RTRD: progressive decomposition of X _R Obtaining the decomposition result without information leakage

5. The wind speed prediction method based on the deep neural network and without the leakage of future information as claimed in any one of claims 1 to 4, wherein: the prediction model comprises a layer of input layer, two layers of Bi-LSTM networks, a layer of attention layer and a layer of full connection layer which are sequentially arranged.

6. The wind speed prediction method based on the deep neural network and without the leakage of future information as claimed in claim 5, wherein: the principle of the LSTM network is as follows:

f _t ＝σ(W _xf x _t +W _hf h _t-1 +b _f )

g _t ＝tanh(W _xg x _t +W _hg h _t-1 +b _g )

i _t ＝σ(W _xi x _t +W _hi h _t-1 +b _i )

c _t ＝f _t ⊙c _t-1 +i _t ⊙g _t

o _t ＝σ(W _xo x _t +W _ho h _t-1 +b _o )

y _t ＝h _t ＝o _t ⊙tanh(c _t )

wherein x is _t As an input vector, c _t In a long-term state, h _t In a short-term state, y _t Is an output vector; w is a group of _xf ,W _xg ,W _xi ,W _xo Are respectively and x _t Weight matrix of connections, W _hf ,W _hg ,W _hi ,W _ho Are respectively h _t-1 Weight matrix of connections, b _f ,b _g ,b _i ,b _o Four layers of bias terms respectively; f. of _t For forgetting the controller of the door, by f _t Determination c _t-1 Which parts of (a) should be deleted; g _t Is the intermediate output of the LSTM network, which acts to analyze x _t And h _t-1 ；i _t For the controller of the input gate, judge g _t Which important parts of c should be added to _t ；o _t For the controller of the output gate, determining that c should be read _t And applied to h _t And y _t Performing the following steps; σ denotes laserA live function.

7. The wind speed prediction method based on the deep neural network and without the future information leakage as claimed in claim 5, characterized in that: the Attention layer adopts a Luong-Attention model, and the calculation method comprises the following steps:

the score calculation method is as follows:

attention weight α _ts The calculation method comprises the following steps:

wherein,

and h _t All hidden states and the last hidden state output by the encoder are respectively; w represents a weight matrix;

context vector

The calculation method comprises the following steps:

wherein alpha is _ts It is the weight of attention that is being weighted,

is a fully hidden state;

attention vector a _t The calculation method comprises the following steps:

wherein,

is a context vector; h is _t Is the last hidden state; w _c Representing the model parameters to be learned.

8. The wind speed prediction method based on the deep neural network and without the future information leakage as claimed in claim 6, characterized in that: attention vector a _t And inputting the wind speed prediction result into a full connection layer to obtain a wind speed prediction result.

Images