CN112529275B

CN112529275B - Wind power climbing event prediction method based on feature extraction and deep learning

Info

Publication number: CN112529275B
Application number: CN202011398774.6A
Authority: CN
Inventors: 韩丽; 乔妍; 王晓静; 李梦洁; 鲁盼盼
Original assignee: China University of Mining and Technology CUMT
Current assignee: China University of Mining and Technology CUMT
Priority date: 2020-12-02
Filing date: 2020-12-02
Publication date: 2023-09-22
Anticipated expiration: 2040-12-02
Also published as: CN112529275A

Abstract

The invention discloses a wind power climbing event prediction method based on feature extraction and deep learning, which comprises the steps of firstly, utilizing an improved revolving door (OpSDA) algorithm to carry out climbing recognition on historical wind power to obtain a historical climbing feature value; then, taking the historical characteristic value and the historical power as prediction models to input, taking the prediction power as output, establishing a CNN-LSTM wind power prediction model, and mining the coupling relation between wind power climbing characteristics and wind power through deep learning, wherein CNN is responsible for secondary extraction of data characteristics, and LSTM is responsible for predicting wind power; and finally, carrying out climbing identification to obtain a climbing event prediction result. The deep learning network is adopted to effectively extract and learn the climbing characteristics in wind power, so that a more accurate climbing prediction result can be obtained.

Description

Wind power climbing event prediction method based on feature extraction and deep learning

Technical Field

The invention relates to a wind power climbing event prediction method based on feature extraction and deep learning, and belongs to the field of wind power uncertainty analysis and new energy prediction.

Background

Wind energy is widely used in China as renewable clean energy, and according to the published statistics of the national energy agency, the new nationwide wind power generation of 2019 increases the installed capacity of the grid by 2574 ten thousand kW, and the accumulated installed capacity of 2.1 hundred million kW. And as the grid-connected scale of wind power is increased, the uncertainty of the wind power is greater in influence on safe and stable operation of the power system. Particularly, a wind power climbing event can cause the problems of unbalanced power generation and power supply of a power grid and the like, and brings great potential safety hazard or serious economic loss to the operation of the power grid. Therefore, it is important for safe operation and economic dispatch of the power system to be able to obtain accurate prediction results of climbing events.

The hill climbing prediction is different from wind power prediction in that a small probability event occurs in a wind power sequence, and takes the event as a prediction object, and the prediction method comprises two methods, namely direct prediction and indirect prediction. The direct prediction needs to train a model through a large number of historical climbing data, a learner models the multi-attribute combined statistical characteristics of climbing, a daily climbing event sequence prediction algorithm is provided, the correlation between the time sequence characteristics specific to wind power and the climbing events is ignored, and the completeness of the historical data can influence the accuracy of prediction for the statistical model. The current research based on wind power climbing event prediction mainly adopts an indirect prediction method, namely, climbing recognition is carried out on the basis of wind power prediction to obtain a climbing prediction result. The climbing event can increase the analysis difficulty of the wind power output characteristics, and the existing research mostly takes relevant characteristic data of wind power as the input of a model to improve the prediction accuracy, for example, the long-term trend characteristics of wind power signals are captured through a physical model before prediction. Based on the consideration of the spatial characteristics of wind power, a learner adopts a method combining deep neural network and multi-task learning, and receives inputs from a plurality of wind power stations simultaneously for prediction based on the spatial correlation of the wind power stations. Still other scholars propose to extract features through Convolutional Neural Network (CNN), but existing researches are not considering the features of climbing event itself in the aspect of climbing prediction, and the climbing features are not strictly defined in the process of extracting the climbing features, but the mutation features in the time sequence of wind power are directly extracted by using CNN as the climbing features, and the characteristics of the wind power signal itself are not combined in the prediction model.

In the improvement process of the current theoretical system for predicting the wind power climbing event, the existing prediction method mainly has the problem that the climbing characteristics of wind power are not fully considered in the input part of the prediction model, and wind power and other influencing factors are directly input into the model. However, the climbing event is used as a small sample event in the wind power sequence, the characteristic information is easy to submerge in the wind power signal, the characteristic information is complex, if the climbing characteristic information is not extracted in advance as input, the model is difficult to learn the complete climbing characteristic condition, and then the wind power climbing event cannot be effectively predicted.

Disclosure of Invention

The invention aims to: in order to overcome the defects of the prior art, the invention provides a wind power climbing event prediction method based on feature extraction and deep learning, which is characterized in that a climbing is identified and features are extracted through an improved rotation door algorithm (OpSDA), and the precision of the climbing event prediction can be effectively improved by predicting after the climbing features are extracted for the second time by using a CNN-LSTM model.

The technical scheme is as follows: the wind power climbing event prediction method based on feature extraction and deep learning comprises the following steps:

step 1: and carrying out climbing identification on the historical wind power by adopting an OPSDA algorithm, and extracting 4 climbing characteristic values: climbing rate R _R Climbing amplitude R _SW Start time R _ST And duration of R _D ；

Step 2: dividing input data consisting of wind power and climbing characteristic values by utilizing a sliding window, inputting the divided input data into a CNN-LSTM prediction model, extracting climbing characteristics again by CNN, obtaining a plurality of prediction results of 16 steps in advance by LSTM prediction, and obtaining a final rolling multi-step wind power prediction result by extracting a first prediction point of multi-step prediction;

step 3: and (5) carrying out climbing identification on the predicted power to obtain a predicted result of a climbing event.

The beneficial effects are that: according to the wind power climbing event prediction method based on feature extraction and deep learning, firstly, the OpSDA algorithm is utilized to identify wind power climbing to extract climbing features, and the climbing features are input into the prediction model, so that on one hand, the prediction model is beneficial to learning of the relation between wind power and climbing events in the training process, the prediction precision is improved, on the other hand, the safe and stable economic operation of wind power grid connection is facilitated, the standby allowance during power grid dispatching can be reduced, and the impact of wind power climbing on power grid power generation and supply balance is reduced. Second, since wind power hill climbing events are sudden events, hill climbing feature data has discontinuities for continuous wind power sequence data. The LSTM network model can reflect the problem of long-term time sequence association of information, but cannot mine the coupling relation between discontinuous climbing features. The convolution kernel of the CNN network moves on the original feature map, convolution operation is carried out, potential relations of all data in the feature map can be effectively extracted to form feature vectors, and climbing features in the whole data set can be extracted layer by layer from a deep structure through deep learning from a small area. Finally, the LSTM network can learn the connection of data on time sequence, and the wind power and the climbing characteristics extracted through the CNN network are used as the input of the LSTM network, so that the model can learn the relation between the sudden climbing event and the wind power time sequence. The important climbing characteristics are reserved through the memory tuple (cell) special for the LSTM, and the characteristics which are not important are forgotten by the forgetting gate, so that the learning of the climbing characteristics by the network is enhanced.

Drawings

FIG. 1 is a feature diagram of a wind-powered climbing event;

FIG. 2 is a schematic diagram of a revolving door algorithm;

FIG. 3 is an OpSDA identification climbing flow chart;

fig. 4 is a schematic diagram of an op sda hill climbing recognition result;

FIG. 5 is a diagram of the structure of the CNN-LSTM model;

FIG. 6 is a comparison of predictive model performance evaluation indicators under different parameters;

fig. 7 is a schematic diagram of a hill climbing prediction result.

Detailed Description

The invention is further explained below with reference to the drawings.

The wind power climbing event prediction method based on feature extraction and deep learning comprises the following steps:

step 1: and carrying out climbing identification on the historical wind power by adopting an OPSDA algorithm, and extracting 4 climbing characteristic values: climbing rate R _R Climbing amplitude R _SW Start time R _ST And duration of R _D As shown in FIG. 1Shown. The specific steps of the step 1 are as follows:

the wind power climbing event belongs to a small probability event for the whole wind power time sequence, so that when wind power is predicted by using a deterministic model, the influence of climbing characteristics on prediction needs to be fully considered. Four important characteristics of the climbing event are important indexes for measuring whether climbing occurs or not and the severity of climbing, and the larger the climbing rate is, the stronger the nonlinearity of wind power is, and the prediction error is larger compared with a non-climbing section. Therefore, the influence of the climbing characteristic is considered in the wind power prediction process, the OpSDA algorithm is adopted to identify climbing and extract the characteristic as a prediction input, and the prediction accuracy is improved.

Step 1.1: and (3) utilizing a revolving door algorithm (SDA) to construct a parallelogram according to the adjustable parameter door width epsilon to screen sample data, and obtaining SDA segmentation points, as shown in figure 2.

Step 1.2: constructing an objective function P (i, j) on any section of wind power time sequence with a time interval of (i, j) based on the segmentation points obtained in the step 1.1, and identifying the climbing by solving the maximum value of the objective function based on the climbing definition.

Wherein, climbing is defined as:

|P _t+Δt -P _t |＞P _threshold

wherein P is _t+Δt Power at time (t+Δt); p (P) _t Power at time t; p (P) _threshold Is a given threshold.

According to different climbing directions, climbing events can be divided into ascending climbing events and descending climbing events, wherein the ascending climbing events refer to wind power suddenly increasing and descending climbing events in a period of time, and the wind power suddenly decreasing in a period of time.

The objective function and its constraint conditions are:

S(i,j)＞S(i,k)+S(k+1,j)

S(i,j)＝(j-i) ² R(i,j)

wherein S (i, k) is a score value of the corresponding interval (i, k); r (·) is a hill climbing criterion, a hill climbing event occurs when the power in the time interval (i, j) satisfies the hill climbing definition, R (i, j) =1, and conversely R (i, j) =0.

In order to improve the overall climbing recognition accuracy, in the recognition process, firstly, some events (bump events) with small change amplitude and opposite change direction to the adjacent climbing in the wind power time sequence are required to be recognized. Uphill identification determines whether a bump event occurs by the following relationship:

[p _k+1 -p _k ]×[1-B(k,k+1)]≥0

wherein p is _k+1 And p _k Wind power values at (k+1) and k time respectively; b (·) is the bump event criterion, B (k, k+1) =1 when a bump event occurs within a time interval (k, k+1), whereas B (k, k+1) =0. When an uphill event is identified, when p _k+1 -p _k The algorithm continues to execute when > 0, i.e. climbing is identified according to the objective function P (i, j) and its constraint.

Downhill climb identification determines whether a bump event occurs by the following relationship:

[p _k+1 -p _k ]×[1-B(k,k+1)]≤0

when a downhill climbing event is identified, when p _k+1 -p _k The algorithm continues to execute when < 0, i.e., a hill climb is identified based on the objective function P (i, j) and its constraints.

Step 1.3: and (3) extracting the characteristic value of the climbing event after the climbing identification result is obtained according to the step 1.1 and the step 1.2.

The flow of identifying a wind power climbing event and extracting a characteristic value thereof by using the OpSDA algorithm is shown in fig. 3. For indirect climbing prediction, the degree of the quality of early climbing recognition determines the degree of climbing prediction accuracy. Taking wind power data of a wind farm 2019, 4 months, 4 days, 5:30, 4 months, 6 days, 2:00 of the specific period of the company Elia as an example, the identification result obtained by using the OPSDA algorithm is shown in fig. 4, the improved algorithm divides adjacent events with the same climbing direction into the same event according to the SDA segmentation point, and identifies the buffer event and divides the buffer event and the adjacent climbing event into the same event in the 1 st and 3 rd climbing events, so that the defect that the traditional SDA algorithm cannot identify long-period climbing is overcome, the climbing trend of the wind power sequence can be effectively identified by using the OPSDA algorithm, and a data basis is provided for characteristic secondary extraction.

Wind power data with the installed capacity of 3796MW in the year 1 of the wind farm 2020 of Elia belgium is adopted, MATLAB is used for statistics, a wind climbing event is identified through an OpSDA algorithm, characteristic values are extracted, and the identification results are shown in table 1. The start time in table 1 is converted to a digital format representation using the "datenum" function, showing only the eigenvalues of the first 22 wind ramp up events for 1 month due to spread constraints.

TABLE 1 climbing event and eigenvalues thereof

Note that: "≡" means climbing up a slope; "∈" indicates downhill climbing.

Step 2: dividing input data consisting of wind power and climbing characteristic values by utilizing a sliding window, inputting the divided input data into a CNN-LSTM prediction model, extracting climbing characteristics again by CNN, obtaining a plurality of prediction results of 16 steps in advance by LSTM prediction, and obtaining a final rolling multi-step wind power prediction result by extracting a first prediction point of multi-step prediction. The model structure is shown in fig. 5, and the prediction process comprises the following steps:

step 2.1: model input section: and the wind power and 4 climbing characteristic values form a data set X to be used as input. The wind power P adopts wind power data of a certain belgium wind power plant of an Elia company, and each 15 minutes of the wind power data are provided with a data point; the four climbing features are respectively climbing rate R _R Climbing amplitude R _SW Start time R _ST And duration of R _D The method comprises the steps of carrying out a first treatment on the surface of the T is a specific moment in the wind power data; m is the advance prediction step length; n is the width of the sliding window, and is set to be 32, namely the model predicts 32 multiplied by 5 data of the input historical power and the climbing characteristic each time, and the input data set is { X } _T ,X _T-1 ,…,X _T-n+1 ,X _T-n (wherein X is _T ＝{P _T ,(R _R ) _T ,(R _SW ) _T ,(R _ST ) _T ,(R _D ) _T }。

Step 2.2: CNN secondary extraction features: the CNN model part is provided with 3 convolution layers and a maximum pooling layer for deep learning and extracting features from the data set. Wherein the number of convolution kernels is set to be 4,16 and 32 respectively, and the convolution kernel size is set to be 2×2. The CNN moves on the feature graphs to carry out convolution operation to extract data features, the number of the convolution kernels determines the number of the feature graphs, namely the depth of a next layer of convolution network, and as the number of the feature graphs of each layer of network increases, multiple layers of convolution can extract more complex features from low-level features.

The convolution layer performs feature extraction on input data through convolution operation, and a calculation formula is as follows:

in the method, in the process of the invention,outputting a j-th feature map of the first layer; />An ith feature map output representing layer (l-1); "x" means a convolution operation; />A convolution kernel weight matrix for connecting between the jth feature map of the first layer and the ith feature map of the (l-1) layer; />Is a bias matrix; n (N) _j A set of input feature maps; f (·) activate function.

The pooling layer performs secondary feature extraction and information filtering on the output of the convolution layer, thereby preserving the most significant features. In wind power climbing prediction based on climbing characteristic values, the most important information in the climbing characteristic values is extracted in a maximum pooling mode. The calculation formula of the pooling layer is as follows:

where down (·) represents the downsampling function.

The model adopts a nonlinear activation function ReLU to construct a sparse matrix, and adopts an all-zero filling mode to maintain the characteristic dimension.

Step 2.3: LSTM prediction: the LSTM model part takes the output of the CNN model as input, adopts a single-layer LSTM network, sets the number of neurons to be 128, and outputs a wind power prediction result through a full connection layer. The basic unit internal structure of the LSTM model comprises three control gates: input gate, forget gate and output gate, the activation function formula of each gate is:

wherein: sigma is a sigmoid function or a tanh function; x is x _t The input vector is the current t moment; w (W) _xi ,W _hi ,W _ci ,W _xf ,W _xo ,W _ho ,W _co ,W _xc ,W _hc Is a weight parameter matrix; b _i ,b _f ,b _o ,b _c Is a bias vector; c _t A vector that is a state unit and an immediate state; h is a _t Outputting the current t moment of the state unit; f (f) _t Output for forget gate; i.e _t Output as an input gate; o (o) _t Is output for the output gate.

In order to illustrate the influence of feature extraction and deep learning on prediction precision in wind power climbing event prediction, python is adopted as a programming language, the compiling environment is PyCharm Community Edition, the RAM is 16GB, and the processor is AMD Ryzen 74800H. The following 2 cases were set:

1) Case1: the prediction model consists of an OpSDA algorithm and a CNN-LSTM network. And analyzing the performance of the climbing multi-step prediction model under different prediction step sizes, different door widths epsilon and different input data according to the evaluation index of the model.

2) Case2: the prediction model consists of an OpSDA algorithm and a CNN-LSTM network, and is compared with the prediction performance of wind power climbing event by using 2 prediction models of an LSTM neural network and a BP neural network. And comparing the optimal prediction result obtained by the model with the optimal prediction result of the recent climbing prediction research.

Simulation results:

case1 analysis:

considering that in the climbing identification process, the revolving door algorithm firstly compresses the wind power time series data through the door width epsilon to obtain a plurality of SDA segmentation points, and the selection of the door width epsilon can bring a certain degree of influence to the climbing identification and the subsequent climbing prediction. In the multi-step prediction model, different prediction results can be obtained by different prediction step sizes.

To sum up, in order to evaluate different influences of parameters set by a model on the model and show the effect of considering the climbing characteristic value to improve the model precision in the input part of the prediction model, taking data of 1 month in 2020 as an example, a recall ratio R is adopted _C Precision F _A Frequency deviation index B _S Key success index C _SI And analyzing four evaluation indexes, namely analyzing the performance of the climbing multi-step prediction model under different prediction step sizes, different door widths epsilon and different input data.

Recall R _C :

Precision F _A :

Frequency deviation index B _S :

Critical success index C _SI ：

Wherein N is _TP To predict the number of times a hill climbing event occurs and actually occurs; n (N) _FN To predict the number of times a hill climbing event does not occur but actually occurs; n (N) _FP To predict the number of times a hill climb event occurs but does not actually occur.

The prediction step length is respectively advanced by 4 steps (1 h), 16 steps (4 h) and 32 steps (8 h); the door widths epsilon are respectively set as 10,25,50; the input of the prediction model is divided into two cases of wind power data only and wind power data plus a climbing characteristic value. The relationship of the four evaluation indexes is plotted according to the following formula, and an evaluation index chart is shown in fig. 6.

FIG. 6 abscissa represents the precision F _A The ordinate represents the recall R _C The curve represents the key success index C _SI The diagonal lines represent the frequency deviation index B _S . The prediction effect of the model can be intuitively displayed according to the evaluation index points in fig. 6, and the better the prediction effect, the closer the evaluation index is to the upper right corner. The recall ratio after the input power and the eigenvalue is generally better than the condition of only input power, R _C The number of times reaches more than 0.9. The evaluation points are mainly distributed near the diagonal line when epsilon=25, and can meet the recall ratio and the precision ratio more than epsilon=50 and epsilon=10. The estimated index result of the predicted step length of 16 steps (4 h) is tighter, the predicted result is more stable, and the duration of the wind power climbing event is lower than 4.04h by combining the existing statistical result, so that the predicted step length is set to be 16 steps (4 h) to be optimal.

FIG. 6The solid evaluation index point of the middle circle is an optimal parameter point, the predictive step length is 16 steps (4 h) in advance, the gate width is set to epsilon=25, and the model input is added with the climbing characteristic value. Under the condition, the accuracy rate is 0.8587, so that not only the probability of correctly predicting the occurrence of the climbing event to occupy all the predicted climbing results is higher, but also the accuracy of accurately predicting the actual climbing event reaches the approximate F _A Optimal conditions, R _C The numerical value reaches 0.9240, and has higher B _S And C _SI . In summary, the invention selects the situation with optimal parameters for the simulation of the CNN-LSTM prediction model. In practical engineering application, different optimal parameters can be set for prediction according to different wind power data sets.

Case2 analysis:

the wind power climbing event prediction performance is compared with that of the wind power climbing event prediction model using the LSTM neural network and the BP neural network. Selecting data of the 2019 10 months to the 2020 9 months of the Elia website, identifying climbing extraction characteristic values by an OpSDA algorithm in four quarters, inputting the data and wind power data as a model, substituting the model into a CNN-LSTM network model, an LSTM model and a BP neural network model for training, taking 1000 data points in each quarter, taking the first 500 data points as a training set, and taking the last 500 data points as a test set. To evaluate the performance of the prediction model, the average absolute percentage error I of 3 climbing prediction models is respectively obtained _MAPE Accuracy A _CC Rate of missing report M _I Error rate E _R Recall ratio R _C Precision F _A Predicting correct rate S of climbing up slope _NR And predicting the downhill climbing accuracy S _R As shown in table 2.

Table 2 comparison of different predictive model evaluation indices

Average absolute percentage error I of the climbing section of CNN-LSTM model according to Table 2,4 quarters _MAPE Are all lower than 0.1, I _MAPE The smaller the wind power prediction effect is, the better. Wherein the wind power prediction effect of the third quarter is optimal, I _MAPE I of only 0.0603, LSTM model and BP model _MAPE Is obviously larger than the prediction model provided by the invention. The other seven evaluation indexes mainly reflect the prediction accuracy of the prediction model for the climbing event, and the climbing prediction accuracy A of the CNN-LSTM model in the first 3 quarters _CC Are all above 0.82, although quarter A of 4 _CC Slightly insufficient but still higher than the other two prediction models, and has good effect of predicting occurrence of downhill climbing, S _R 0.8041 is reached. Although there are BP model and LSTM model miss rate M _I Or false positive rate E _R Lower than the CNN-LSTM model, as in quarter 3, but A of these 2 methods _CC Are far from the CNN-LSTM model. The CNN-LSTM model can ensure that higher accuracy can be obtained under the condition of relatively low false alarm rate and false alarm rate. And the CNN-LSTM model has higher R in 4 quarters _C And F _A The two indexes in the 2 nd quarter reach more than 0.9, namely the number of the correctly predicted climbing occurrence events accounts for more than nine times of the total number of the predicted climbing events and the total number of the actually generated climbing events. Overall, the accuracy of the prediction of the downslope event by each model is higher than that of the upslope event, which is more difficult to predict than the downslope event.

To further verify the model to predict the wind power climbing performance, the optimal prediction result obtained by the model of the invention is compared with the optimal prediction result of the recent climbing prediction research based on the same evaluation index, as shown in table 3. Although the overall prediction accuracy A of the CNN-LSTM model of the invention _CC Slightly lower than the MLP-BT model and the MLP-MSAR model, but the model of the invention can better balance each evaluation index and ensure the recall ratio R _C Higher and better precision F _A Each index is higher than 0.81, and R _C 、F _A And S is _R The model of the invention predicts the performance of ascending and descending climbing events better when the model reaches more than 0.9.

TABLE 3 comparison of evaluation index of other literature models

The result of predicting the climbing event of month 1 in 2020 obtained by the CNN-LSTM model is shown in FIG. 7. Fig. 7 (a) shows an actual wind power curve. Each rectangle in fig. 7 (b), 7 (c) and 7 (d) represents a climbing event, an ascending climbing event is represented in a first quadrant, a descending climbing event is represented in a fourth quadrant, an abscissa can display duration time of each climbing event and a wind power sample serial number point at the beginning of climbing, and an ordinate can display amplitude value of each climbing event. The invention provides the mode of displaying the predicted climbing event shown in fig. 7, so that the prediction effect of each model can be more intuitively seen, and simultaneously, the climbing amplitude and the climbing time can be predicted to provide a better scheduling basis for the power grid.

The wind turbine generator system can adjust the reserve allowance according to the climbing prediction result during scheduling, so that the prediction result is required to have certain persistence and stability. In contrast, according to fig. 7, the prediction results of both the lstm model and the BP model generate serious fluctuation phenomenon, and in particular, the BP model predicts a long-time climbing event as a plurality of events having short climbing duration but large climbing amplitude. Fig. 7 (b) can obviously show that the CNN-LSTM model has more excellent performance in the prediction of the climbing amplitude and the climbing duration, and after the climbing characteristics are extracted by the CNN for the second time, the LSTM model can learn the continuity of the wind power signal and the climbing event more effectively, so as to obtain the climbing prediction result on the long-time-sequence wind power signal, which is more beneficial to the safe operation and economic dispatching of the power system.

The foregoing is merely a preferred embodiment of the present invention and it should be noted that modifications and adaptations to those skilled in the art may be made without departing from the principles of the present invention, which are intended to be comprehended within the scope of the present invention.

Claims

1. The wind power climbing event prediction method based on feature extraction and deep learning is characterized by comprising the following steps of:

step 3: the prediction result of the climbing event is obtained by carrying out climbing identification on the prediction power;

the step 1 comprises the following specific steps:

step 1.1: constructing a parallelogram according to the adjustable parameter door width epsilon by using a revolving door algorithm to screen sample data, and obtaining a revolving door algorithm segmentation point;

step 1.2: constructing an objective function P (i, j) on any section of wind power time sequence with a time interval of (i, j) based on the segmentation points obtained in the step 1.1, and identifying a climbing by solving the maximum value of the objective function based on the climbing definition;

the climbing is defined as:

|P _t+Δt -P _t |＞P _threshold

wherein P is _t+Δt Power at time (t+Δt), P _t Power at time t, P _threshold Is a given threshold;

the objective function P (i, j) and its constraint conditions are:

S(i,j)＞S(i,k)+S(k+1,j)

S(i,j)＝(j-i) ² R(i,j)

wherein S (i, k) is a score value of the corresponding interval (i, k); r (·) is a hill climbing criterion, a hill climbing event occurs when the power in the time interval (i, j) satisfies the hill climbing definition, R (i, j) =1, otherwise R (i, j) =0;

in order to improve the overall climbing recognition accuracy, in the recognition process, firstly, a lamp event with small change amplitude and opposite change direction to an adjacent climbing in a wind power time sequence needs to be recognized; wherein, the uphill identification determines whether a bump event occurs by the following relation:

[p _k+1 -p _k ]×[1-B(k,k+1)]≥0

wherein p is _k+1 And p _k Wind power values at (k+1) and k time respectively; b (·) is a bump event criterion, B (k, k+1) =1 when a bump event occurs in a time interval (k, k+1), whereas B (k, k+1) =0; when an uphill event is identified, when p _k+1 -p _k When the value is more than 0, the algorithm is continuously executed, namely, climbing is identified according to an objective function P (i, j) and constraint conditions of the objective function P (i, j);

[p _k+1 -p _k ]×[1-B(k,k+1)]≤0

when a downhill climbing event is identified, when p _k+1 -p _k When the algorithm is less than 0, the algorithm is continuously executed, namely climbing is identified according to an objective function P (i, j) and constraint conditions of the objective function P (i, j);

step 1.3: after a climbing recognition result is obtained according to the steps 1.1 and 1.2, extracting a characteristic value of a climbing event;

the step 2 comprises the following specific steps:

step 2.1: model input section: the wind power and 4 climbing characteristic values form a data set X to be used as input; sampling a data point every 15 minutes for wind power data of a wind power plant to obtain wind power P; the input data set is { X } _T ,X _T-1 ,…,X _T-n+1 ,X _T-n (wherein X is _T ＝{P _T ,(R _R ) _T ,(R _SW ) _T ,(R _ST ) _T ,(R _D ) _T T is a specific moment in the wind power data, n is the width of the sliding window, and n is set to 32;

step 2.2: CNN secondary extraction features: the CNN model part is provided with 3 convolution layers and a maximum pooling layer for deep learning of the data set and extracting features; wherein, the number of convolution kernels is set to be 4,16 and 32 respectively, and the size of the convolution kernels is 2 multiplied by 2; the CNN moves on the feature map through the convolution kernel to carry out convolution operation to extract data features;

in the method, in the process of the invention,outputting a j-th feature map of the first layer; />An ith feature map output representing layer (l-1); "x" means a convolution operation; />A convolution kernel weight matrix for connecting between the jth feature map of the first layer and the ith feature map of the (l-1) layer;is a bias matrix; n (N) _j A set of input feature maps; f (·) activate function;

the pooling layer performs secondary feature extraction and information filtering on the output of the convolution layer, so that the most obvious features are reserved; the calculation formula of the pooling layer is as follows:

wherein, down (·) is a downsampling function;

the model adopts a nonlinear activation function ReLU to construct a sparse matrix, and adopts an all-zero filling mode to maintain characteristic dimensions;

step 2.3: LSTM prediction: the LSTM model part takes the output of the CNN model as input, adopts a single-layer LSTM network, sets the number of neurons to be 128, and outputs a wind power prediction result through a full connection layer; the basic unit internal structure of the LSTM model comprises three control gates: input gate, forget gate and output gate, the activation function formula of each gate is: