CN110555515A - Short-term wind speed prediction method based on EEMD and LSTM - Google Patents

Abstract

a short-term wind speed prediction method based on EEMD and LSTM is characterized in that a wind speed prediction model combining EEMD and long-term and short-term memory neural network LSTM is decomposed based on a set empirical mode, a wind speed sequence is decomposed through EEMD to obtain a stable subsequence, the property of original data is well reserved, errors caused by EMD modal aliasing of a traditional empirical mode are optimized, the method is combined with the LSTM prediction model, the hysteresis of the LSTM prediction model is improved, the efficiency is higher compared with that of a traditional method, and meanwhile prediction errors can be effectively reduced.

Description

Short-term wind speed prediction method based on EEMD and LSTM

Technical Field

The invention belongs to the field of atmospheric science, and particularly relates to a short-term wind speed prediction method based on EEMD and LSTM.

Background

wind speed is the velocity of the air stream generated by the movement of air, and wind is the result of the combined action of solar radiation and earth surface long wave radiation on the atmosphere at the earth's surface. The wind speed has the characteristics of uncertainty, volatility and the like, and is one of the most difficult elements to predict in meteorological elements. The wind speed prediction has important significance for improving the accuracy of weather forecast, predicting short-time strong wind disasters, diffusing speed of air pollutants, safe operation of a wind turbine generator and the like. In the aspect of weather forecast, wind speed is used as a basic element in the weather forecast, and the accuracy of prediction of the wind speed influences the accuracy of the weather forecast; the wind speed also influences the migration and diffusion speed of air pollutants and dust, and the accurate prediction of the wind speed is favorable for forecasting the expansion path of the pollutants and taking prevention and treatment measures; in meteorological disasters, strong wind usually appears as disastrous weather such as typhoons, hurricanes, tornadoes and the like, which not only causes huge damage to the ecological natural environment, but also threatens the production, life and property safety of people.

in recent years, new energy industries such as wind power and the like have emerged, so that the wind utilization demand is higher and higher, and how to improve the accuracy of wind speed prediction and the wind utilization rate directly relates to the efficiency and the output value of the industries. At present, wind speed prediction methods mainly comprise a physical method based on a numerical weather forecast prediction result and a statistical method based on historical data modeling. With the development of artificial intelligence, various deep learning algorithms continuously appear, in order to make up for the limitations of physical methods and statistical methods, hybrid algorithm prediction models are generated at the same time, and similar technologies such as a wind power plant short-term wind speed combined prediction method are provided: extracting historical wind speed data from a related data acquisition and monitoring control system; performing sequence analysis on the extracted wind speed data by adopting clustering empirical mode decomposition; respectively establishing a least square support vector machine model for each subsequence obtained by clustering empirical mode decomposition, and comprehensively selecting three parameters influencing the prediction effect of the least square support vector machine by adopting a self-adaptive disturbance particle swarm algorithm and learning effect feedback; selecting optimal parameters for prediction according to the learning effect of the least square support vector machine; superposing the prediction results of the subsequences to obtain a wind speed prediction result; and carrying out error analysis on the wind speed prediction result.

the method for predicting the wind speed has poor support on big data, takes long time and is not accurate enough in prediction. The wind speed prediction model based on the combination of the ensemble empirical mode decomposition EEMD and the long-short term memory neural network LSTM is used for decomposing a wind speed sequence through the EEMD to obtain a stable subsequence, the property of original data is well reserved, errors caused by EMD modal aliasing of the traditional empirical mode are optimized, the wind speed prediction model is combined with the LSTM prediction model, the hysteresis of the LSTM prediction model is improved, the efficiency is higher compared with that of the traditional method, and meanwhile the prediction errors can be effectively reduced.

disclosure of Invention

Because wind has strong paroxysmal and local characteristics and more influence factors, the wind speed is predicted by using machine learning related technologies, the influence is always caused, the prediction accuracy is reduced, and particularly, the accuracy is generally low for the prediction of instantaneous strong wind. Aiming at the problems, the invention provides a short-term wind speed prediction method based on EEMD and LSTM, which adopts EEMD to decompose a wind speed sequence into a plurality of relatively stable frequency domain subsequences, avoids the modal aliasing phenomenon of EMD, and then adopts LSTM to construct a prediction model, thereby improving the accuracy of short-term wind speed prediction.

in order to achieve the purpose, the invention adopts the following technical scheme:

a short-term wind speed prediction method based on EEMD and LSTM is characterized by comprising the following steps:

step 1: preprocessing the actually measured second-by-second wind speed of a meteorological observation station in a certain area into a wind speed sequence with a time interval of tau;

step 2: decomposing the wind speed sequence into a plurality of components using EEMD;

And step 3: determining a time scale, reconstructing each component, and normalizing the unified dimension to obtain a plurality of samples;

and 4, step 4: determining a training set and a testing set from the reconstructed sample;

And 5: respectively establishing an LSTM prediction model for each training set and each test set, wherein the LSTM prediction models are used for predicting each component;

Step 6: and obtaining wind speed prediction components of a plurality of components according to the LSTM prediction model, carrying out reverse normalization on each wind speed prediction component, and then superposing to obtain a wind speed prediction result.

In order to optimize the technical scheme, the specific measures adopted further comprise:

further, in the step 1, tau epsilon [1, 86400] s.

Further, the step 2 is specifically as follows:

Step 2-1: adding white noise which follows normal distribution into the wind speed sequence N (t);

Step 2-2: EMD decomposition is carried out on the wind speed sequence, and s inherent modal function components imf (t) and 1 residual component r (t) are calculated:

Wherein imf _i (t) is the ith imf (t) obtained by EMD decomposition, r (t) is the signal residual component after s imf (t) is decomposed and screened, t is the sequence length, and t is more than 0;

Step 2-3: repeating the step 2-1 and the step 2-2 r times, and adding new white noise each time;

And 2-4, solving the integral average of the component IMF _i (t) after r times of decomposition, and taking the integral average as the IMF component of the wind speed sequence N (t), thereby finally obtaining s inherent modal function components IMF ₁ -IMF _s with different scales and a residual component Res.

further, in step 3, a time scale Ts is determined, the decomposed IMF component and Res component are reconstructed to obtain a reconstruction sequence, and each component is reconstructed into the following form:

in the formula, N _n represents the reconstructed nth sample, and M _n represents the label of the reconstructed nth sample;

total samples are N ═ N ₁, N ₂.., N _n ], normalized by column as follows:

In the formula, N _n,k represents the value of the nth row and k columns of the total sample, N' _n,k represents the normalized value of the nth row and k columns of the total sample, N _k,min represents the minimum value of the kth column, and N _k,max represents the maximum value of the kth column.

further, in step 4, determining a training set from the reconstructed samples as follows:

E＝{(N′₁，M′₁)，(N′₂，M′₂)，…，(N′_m，M′_m)}

The test set is:

Test′＝{(N′_m+1，M′_m+1)，(N′_m+2，M′_m+2)，…，(N′_n，M′_n)}

Where N '_m represents the normalized mth sample, M' _m represents the label of the normalized mth sample, and 1 < M < N.

further, in the step 5, the LSTM prediction model includes an LSTM layer, the input of the network at the time T ' is the historical wind speed sequence N ' _n,t', T ' is an integer between (1, T _s), the output is the predicted value M ' _n,t' at the next time, the output h _t' of the hidden layer is obtained after the hidden layer operation, and the output of the network is:

M′_n，t′＝sigmoid(W_h·h_t′+b)

In the formula, W _h is a weight matrix between a hidden layer and an output layer, b is the offset of the output layer, and Ts historical data before the current time are used as the input of the LSTM network for training and prediction.

further, in step 6, the wind speed prediction components are stacked after being denormalized, and the formula is as follows:

where M ^* _n,t' is the predicted value of the nth sample, M ' _i,n,t' represents the predicted value of wind speed for the ith component, M _max represents the maximum value of the training label, and M _min represents the minimum value of the training label.

The invention has the beneficial effects that:

1. The short-term wind speed prediction method based on the combination of EEMD and LSTM can better fit the actual wind speed sequence, and the effectiveness of the method is verified;

2. The short-term wind speed prediction method based on the combination of EEMD and LSTM improves the hysteresis of the LSTM model, eliminates the adverse effect of EMD algorithm modal aliasing on wind speed prediction, and has the advantages that the prediction result is closer to the actual wind speed value, the error is smaller, and the application value is higher.

drawings

FIG. 1 is a flow chart of a wind speed prediction method of the present invention.

FIG. 2 is a flow chart of the present invention for extracting the ith IMF.

FIG. 3 is a flow chart of EEMD decomposition of the present invention.

FIG. 4 is a graph of the wind speed sequence after EEMD decomposition for the present invention.

FIG. 5 is a graph of wind speed predicted using LSTM versus actual observed values.

FIG. 6 is a plot of wind speed predicted using EMD + LSTM versus actual observed values.

FIG. 7 is a graph of wind speed predicted using EEMD + LSTM versus actual observed values.

Detailed Description

The present invention will now be described in further detail with reference to the accompanying drawings.

As shown in FIG. 1, a short-term wind speed prediction method based on EEMD and LSTM specifically includes the following steps:

1. and preprocessing the wind speed measured by a certain regional meteorological observation station every second into a wind speed sequence with the time interval of tau, wherein tau belongs to [1, 86400] s.

2. The problems of equipment failure and noise interference cannot be avoided in the acquisition process of the wind speed observation data, so that the modal aliasing phenomenon is easy to occur in the EMD decomposition process, and the prediction precision is reduced. The current method for solving the problem is EEMD, the decomposition flow chart is shown in FIG. 3, and the improvement steps are as follows:

1) adding white noise as distributed from normal to the sequence n (t);

2) EMD decomposition of the newly formed sequence, finding s imf (t) components and 1 remaining component r (t):

Wherein imf _i (t) is the ith imf (t) obtained by EMD decomposition, the extraction process is shown in FIG. 2, r (t) is the residual component of the signal after s imf (t) is decomposed and screened, which usually represents the direct current component or the trend of the signal, t is the sequence length, and t is more than 0;

3) repeating the step 1) and the step 2) r times, and adding a new white noise sequence every time;

4) And solving the integral average of the component IMF _i (t) after r times of decomposition to be used as an IMF component of the wind speed sequence N (t), and finally obtaining s stable intrinsic mode function IMF ₁ -IMF _s components with different scales and a residual Res component.

3. Determining a time scale Ts, reconstructing the decomposed IMF components and errors Res, normalizing the unified dimension, and determining a training set and a test set; the method specifically comprises the following steps: firstly, reconstructing the decomposed IMF components and the error Res to obtain a reconstruction sequence, and reconstructing each component into the following form:

in the formula, N _n represents the reconstructed nth sample, and M _n represents the label of the reconstructed nth sample;

And expanding the original sequence into n samples after reconstruction, wherein the selection of the time scale Ts directly influences the prediction result, the information superposition can be caused when the time scale Ts is too small, and the calculated amount can be increased when the time scale Ts is too large. Suitable parameters need to be determined.

The wind speed sequence is decomposed by the EEMD to obtain N stable subsequences and an error sequence, each component is normalized to avoid the influence caused by different dimensions, a total sample is N ═ N ₁, N ₂,.., N _n, and the total sample is normalized as follows:

In the formula, N _n,k represents the value of the nth row and k columns of the total sample, N _n,k represents the normalized value of the nth row and k columns of the total sample, N _k,min represents the minimum value of the kth column, N _k,max represents the maximum value of the kth column, and the normalized range is [0, 1 ].

4. determining a training set by the reconstructed samples as follows:

E＝{(N′₁，M′₁)，(N′₂，M′₂)，…，(N′_m，M′_m)}

the test set is:

Test′＝{(N′_m+1，M′_m+1)，(N′_m+2，M′_m+2)，…，(N′_n，M′_n)}

Where N '_m represents the normalized mth sample, M' _m represents the label of the normalized mth sample, and 1 < M < N.

5. establishing an LSTM prediction model for each training set and each test set respectively to predict a wind speed subsequence, constructing a short-term wind speed prediction model by using the LSTM, wherein the network model comprises an LSTM layer, the input of the network at the time T ' is a historical wind speed sequence N ' _n,t' in the training set, T ' is an integer between (1 and T _s), the output is a predicted value M ' _n,t' at the next time, the output h _t' of the hidden layer is obtained after the hidden layer operation, and the output of the network is as follows:

M′_n，t′＝sigmoid(W_h·h_t′+b)

In the formula, W _h is a weight matrix between a hidden layer and an output layer, b is the offset of the output layer, and Ts historical data before the current time are used as the input of the LSTM network for training and prediction.

6. and after reverse normalization, superposing the wind speed prediction components, and finally calculating the predicted value of the wind speed, wherein the formula is as follows:

wherein M ^* _n,t' is the predicted value of the nth sample, n is more than M, M ' _in,t' represents the predicted value of the wind speed of the ith component, M _max represents the maximum value of the training label, and M _min represents the minimum value of the training label.

Because some missing and redundant data exist in the actual observation process and cannot form a continuous complete sequence every second, the case adopts 29-day wind speed data of 9 months and 10-day wind speed preprocessing time intervals of 2015 of a certain ground meteorological observation station of a Yangtze river basin Nanjing area provided by the Husu province and the maritime administration to preprocess the original data, wherein the preprocessing mode is selected as 10-second average, namely, the data in each 10 seconds are averaged.

In the case, the wind speed sequence is decomposed by the EEMD to obtain more stable 10 subsequences and an error sequence, each component is normalized to avoid the influence caused by different dimensions, and each component sequence is reconstructed, for example, the case obtains a waveform diagram of 10 inherent modal components IMF ₁ -IMF ₁₀ and a residual component Res, which are stable in frequency domain.

As can be seen from FIG. 4, compared with the original wind speed sequence, the decomposed component diagram has relatively stable fluctuation of the subsequence, and the spectral characteristics are sequentially represented by the components from high frequency to low frequency. The method decomposes the original unstable wind speed sequence into a plurality of stable subsequences, and simultaneously preserves the characteristics of the wind speed sequence.

and expanding the original sequence into 10 samples after reconstruction, wherein the selection of the time scale Ts directly influences the prediction result, the information superposition can be caused when the time scale Ts is too small, and the calculated amount can be increased when the time scale Ts is too large. Suitable parameters need to be determined. And only the LSTM is adopted for wind speed forecasting, and different super parameters are set for the LSTM network for experiments. The following table compares the results of the Ts experiments:

TABLE 1 comparison of the results of the Ts experiments

as can be seen by combining the comparison results in Table 1, when Ts is unchanged, EPOCH increases, the prediction error decreases, and the prediction time increases, and as EPOCH increases to a certain extent, the prediction error decreases slowly, but the time does not decrease. When the EPOCH is not changed, Ts is increased, the prediction error is reduced firstly and then is increased, and the prediction time is increased. Compared with the two methods, the method has the advantages that the effect of increasing the EPOCH to a certain extent is better than that of increasing Ts, and the time is short. By contrast, Ts is 10 and EPOCH is 500 as the over-parameter of the network.

the wind speed forecasting method comprises the steps of forecasting the wind speed by adopting three methods, namely short-term wind speed forecasting based on LSTM, short-term wind speed forecasting based on EMD and LSTM and short-term wind speed forecasting based on EEMD and LSTM. And setting the hyper-parameters of the LSTM network by adopting the experimental result. The comparison of the predicted 180s wind speed and the actual observed results is chosen, as shown in fig. 5, 6 and 7. The estimated time was adjusted to 80 minutes and calculated again to obtain the evaluation results of table 2.

TABLE 2 evaluation index of three prediction methods

from table 2, it can be seen that: the prediction by adopting the method combining EMD and LSTM reduces 0.2596m/s, 0.3092m/s and 4.36% respectively compared with the prediction by only using the LSTM method. The prediction by adopting the method combining EEMD and LSTM is respectively reduced by 0.1902m/s, 0.2433m/s and 3.2% compared with the prediction by adopting the method combining EMD and LSTM. Therefore, the method of combining EEMD and LSTM to predict the short-term wind speed can better simulate the wind speed condition of 10s after 80 minutes in the future, and the prediction error is reduced.

The result evaluation indexes in table 2 are specifically calculated as follows:

MAE (mean Absolute error) is the mean Absolute error in units (m/s) and is calculated as:

RMSE (root Mean Squared error) is the root Mean square error in units (m/s) and is calculated as:

MAPE (mean Absolute Percentage error) is the mean Absolute Percentage error in (%), which is calculated as:

t is the running time in units of(s).

it should be noted that the terms "upper", "lower", "left", "right", "front", "back", etc. used in the present invention are for clarity of description only, and are not intended to limit the scope of the present invention, and the relative relationship between the terms and the terms is not limited by the technical contents of the essential changes.

The above is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above-mentioned embodiments, and all technical solutions belonging to the idea of the present invention belong to the protection scope of the present invention. It should be noted that modifications and embellishments within the scope of the invention may be made by those skilled in the art without departing from the principle of the invention.

Claims (7)

1. a short-term wind speed prediction method based on EEMD and LSTM is characterized by comprising the following steps:

step 1: preprocessing the actually measured second-by-second wind speed of a meteorological observation station in a certain area into a wind speed sequence with a time interval of tau;

Step 2: decomposing the wind speed sequence into a plurality of components using EEMD;

And step 3: determining a time scale, reconstructing each component, and normalizing the unified dimension to obtain a plurality of samples;

And 4, step 4: determining a training set and a testing set from the reconstructed sample;

And 5: respectively establishing an LSTM prediction model for each training set and each test set, wherein the LSTM prediction models are used for predicting each component;

Step 6: and obtaining wind speed prediction components of a plurality of components according to the LSTM prediction model, carrying out reverse normalization on each wind speed prediction component, and then superposing to obtain a wind speed prediction result.

2. The method of claim 1 for short term wind speed prediction based on EEMD and LSTM, wherein: in the step 1, tau is equal to [1, 86400] s.

3. The method of claim 1 for short term wind speed prediction based on EEMD and LSTM, wherein: the step 2 is specifically as follows:

step 2-1: adding white noise which follows normal distribution into the wind speed sequence N (t);

Step 2-2: EMD decomposition is carried out on the wind speed sequence, and s inherent modal function components imf (t) and 1 residual component r (t) are calculated:

wherein imf _i (t) is the ith imf (t) obtained by EMD decomposition, r (t) is the signal residual component after s imf (t) is decomposed and screened, t is the sequence length, and t is more than 0;

Step 2-3: repeating the step 2-1 and the step 2-2 r times, and adding new white noise each time;

and 2-4, solving the integral average of the component IMF _i (t) after r times of decomposition, and taking the integral average as the IMF component of the wind speed sequence N (t), thereby finally obtaining s inherent modal function components IMF ₁ -IMF _s with different scales and a residual component Res.

4. the EEMD and LSTM based short term wind speed prediction method as claimed in claim 3, wherein: in the step 3, a time scale Ts is determined, the decomposed IMF component and Res component are reconstructed to obtain a reconstruction sequence, and each component is reconstructed into the following form:

in the formula, N _n represents the reconstructed nth sample, and M _n represents the label of the reconstructed nth sample;

Total samples are N ═ N ₁, N ₂.., N _n ], normalized by column as follows:

In the formula, N _n,k represents the value of the nth row and k columns of the total sample, N' _n,k represents the normalized value of the nth row and k columns of the total sample, N _k,min represents the minimum value of the kth column, and N _k,max represents the maximum value of the kth column.

5. The EEMD and LSTM based short term wind speed prediction method as claimed in claim 4, wherein: in step 4, determining a training set from the reconstructed sample as follows:

E＝{(N′₁，M′₁)，(N′₂，M′₂)，…，(N′_m，M′_m)}

The test set is:

Test′＝{(N′_m+1，M′_m+1)，(N′_m+2，M′_m+2)，…，(N′_n，M′_n)}

Where N '_m represents the normalized mth sample, M' _m represents the label of the normalized mth sample, and 1 < M < N.

6. The method of claim 5, wherein in step 5, the LSTM prediction model comprises an LSTM layer, the input of the network at time T ' is a historical wind speed sequence N ' _n,t', T ' is an integer between (1, T _s), the output is a predicted value M ' _n,t' at the next time, the output h _t' of the hidden layer is obtained after hidden layer operation, and the output of the network is:

M′_n，t′＝sigmoid(W_h·h_t′+b)

in the formula, W _h is a weight matrix between a hidden layer and an output layer, b is the offset of the output layer, and Ts historical data before the current time are used as the input of the LSTM network for training and prediction.

7. the EEMD and LSTM based short term wind speed prediction method as claimed in claim 6, wherein: in the step 6, the wind speed prediction components are stacked after being subjected to inverse normalization, and the formula is as follows:

where M ^* _n,t' is the predicted value of the nth sample, M ' _i,n,t' represents the predicted value of wind speed for the ith component, M _max represents the maximum value of the training label, and M _min represents the minimum value of the training label.