CN104268627A - Short-term wind speed forecasting method based on deep neural network transfer model - Google Patents

Abstract

The invention discloses a short-term wind speed forecasting method based on a deep neural network transfer model. The method comprises the following steps that (1) normalization preprocessing and division of sample sets are carried out on data of two or more wind power plants, (2) the deep neural network transfer model is established, (3) layered training is started from bottom to top in an unsupervised learning mode, (4) supervised learning is carried out from top to bottom on the basis of the third step, (5) weight parameters of connection of a top layer and hidden layers are finely adjusted so as to obtain an output layer, corresponding to the wind power plants, in a deep neural network, and (6) inverse normalization is carried out on the result output by a deep neural network so as to obtain the predicted value of wind speed. Transfer learning is introduced to the wind speed forecasting field, knowledge of other wind power plants rich in data is transferred to target wind power plants, and the problem that the newly built wind power plants have few data is solved effectively. By means of the effective transfer scheme based on the deep neural network, the wind speed prediction accuracy of the target wind power plants is greatly improved.

Description

A kind of short-term wind speed forecasting method based on deep neural network migration models

Technical field

The present invention is based on machine Learning Theory and Statistical Learning Theory, by building the regression model based on deep neural network, build the depth migration models of multi output sharing hidden layer on this basis, finally this model is used for minority and forecasts according to the short-term wind speed of wind energy turbine set.

Background technology

In wind speed forecast, a lot of method is suggested.These methods can be divided into four classes: 1) physical model; 2) statistical model; 3) space correlation model; 4) artificial intelligence model and other new models.Physical model uses physical factor, weather data such as landform, pressure and temperature to estimate following wind speed.It compares and has superiority in long-term forecasting, in short-term forecasting, generally cannot provide result accurately.Therefore, the first step that they are often just predicted, as the auxiliary input of other models.Statistical model, based on the correlativity of wind series, is set up forecast model by Model Distinguish and the step such as parameter estimation, model testing, is first described the change of historical wind speed sequence, then predict the change in future.Conventional statistics equivalent model is in Random time sequence model, conventional time series models mainly contain: autoregressive model (Autoregressive, AR), moving average model(MA model) (Moving Average Model, MA), ARMA model (Autoregressive Moving Average Model, and return accumulative formula moving average model(MA model) (Regressive Integrated Moving Average Model, ARIMA) ARMA).Another statistical method is Kalman filtering method.It needs the estimated value of previous wind speed and a nearest observed reading, and adopt the method for state equation and recursion to estimate current air speed value, its solution provides with the form of estimated value.Different with additive method, space correlation model needs to consider many groups air speed data in wind energy turbine set and close several place with it, uses the spatial coherence between the wind speed of several places, carries out forecasting wind speed.The method is very large to source data collection amount, therefore than additive method difficulty, but increases, so prediction effect is better due to the factor considered in forecasting process.Nowadays, along with artificial intelligence and the development of other Forecasting Methodologies, various new model is suggested.Comprising support vector machine (Support Vector Machine, SVM), fuzzy logic method, artificial neural network (Artificial Neural Network, ANN) etc. and some combination forecasting methods.

But, for existing solution, do not relate to the forecasting wind speed problem of the newly-built wind energy turbine set of minority certificate.Have little data, it is not enough to study to a good forecast model.At this moment, we introduce the concept of transfer learning (Transfer learning, TL).Transfer learning is a kind of method of the knowledge learning tasks of helping in new environment will acquired in an environment.It can move knowledge to target domain from other field, from available data migration knowledge to study in the future, efficiently solves the problems such as data are few, updating decision, learning process are consuming time.

In addition, by the comparison of many results, better based on the modelling effect of ANN in short-term wind speed forecasting single method, but neural network is before shallow-layer model mostly.2006, the people such as Hinton proposed the concept of degree of depth study (Deep learning, DL).Obtain after this and pay close attention to widely, and be successfully applied to the fields such as computer vision, speech recognition, natural language processing.Degree of depth study forms more abstract high level by combination low-level feature and represents, to find that the distributed nature of data represents.Ubiquitous local optimum problem in non-convex objective cost function in deep layer network before it solving.Three technology create the success of degree of depth learning method, are a large amount of hidden node respectively, better learning algorithm and better parameter initialization technology.In addition, depth structure is very suitable for transfer learning, and it can extract the feature of high-level abstractions, and some such features are all applicable for many fields.

Transfer learning based on deep neural network achieves good effect in numerous applications.Such as, character recognition, semantic classification, multilingual study, Images Classification etc.The short-term wind speed forecasting problem how this model being applicable to new wind energy turbine set is the key issue that the present invention will solve.

Summary of the invention

In view of the neural network Problems existing of forecasting wind speed facing challenges and classics, the present invention proposes a kind of short-term wind speed forecasting method for wind energy turbine set based on deep neural network, degree of depth study is incorporated into forecasting wind speed field by the present invention, there is very strong feature learning ability, the advanced abstract information hidden in wind speed can be extracted.Transfer learning is incorporated into forecasting wind speed field by the present invention, by the knowledge migration of wind energy turbine set that other data enriched to target wind energy turbine set, efficiently solves the problem that newly-built wind farm data is few.By the effective migration scheme based on deep neural network, improve the precision of prediction of target wind energy turbine set for wind speed greatly.

The technical scheme of a kind of short-term wind speed forecasting method based on deep neural network migration models of the present invention is: this forecasting procedure comprises the following steps:

What step one, setting data integrated record is all with ten minutes air speed datas as interval, first carries out maximin normalization pre-service to the data of two or more wind energy turbine set; By the set of data samples that pretreated Data Placement is in units of sky, and as the input of deep neural network migration models, wherein the proper vector of the sample in the i-th moment is the wind velocity vector of 144 dimensions: x _i=(x _i-143, x _i-142..., x _i-1, x _i), the output of this deep neural network is the forecasting wind speed data of 8 hours after this sample, i.e. the wind velocity vector of 48 dimensions: y _i=(x _i+1, x _i+2..., x _i+47, x _i+48), wherein x _iit is the wind speed in the i-th moment;

Step 2, structure deep neural network migration models, wherein, input layer is the day data sequence after above-mentioned wind energy turbine set process, and hidden layer is two noise reduction automatic coding machines piled up, and output layer is the air speed data sequence of prediction in following 8 hours;

Step 3, the bottom-up beginning order training method of employing unsupervised learning, obtain the initial weight parameter that each hidden layer connects;

The initial weight parameter that step 4, each hidden layer obtained based on step 3 connect carries out retraining further by top-down supervised learning to this deep neural network, thus the weighting parameter that after obtaining finely tuning, each hidden layer connects, in this step use all data sets have label data;

The weighting parameter that step 5, initialization top layer are connected with hidden layer, then the label data that has of each wind energy turbine set is used to finely tune the weighting parameter that top layer is connected with hidden layer respectively, obtain the output layer of corresponding each wind energy turbine set in deep neural network, i.e. the wind series of each wind energy turbine set prediction;

Step 6, the result exported this deep neural network carry out the predicted value of 8 hours wind speed after renormalization obtains, and according to predicted value and actual value computational prediction error.

Compared with prior art, the invention has the beneficial effects as follows:

Show by experiment, deep neural network (the Shared-Hidden-Layer Deep Neural Network of Application share hidden layer in the present invention, SHL-DNN) support vector regression (the Support Vector Regression that model and application is non-migratory, SVR) model, non-migratory deep neural network (Deep Neural Network, DNN) compare, short-term wind speed is carried out to the data of different wind energy turbine set and forecasts its better effects if, reduce predicated error, especially its effect of wind energy turbine set that data volume is fewer is more outstanding, and when target wind farm data is abundant (as training data comprises the wind speed of six months), just do not need from other wind energy turbine set there migration datas, not so, even precision of prediction can be reduced in some cases.

Accompanying drawing explanation

Fig. 1 is the deep neural network migration models structural drawing in the present invention;

Fig. 2 is the result of the different error criterions of SVR, DNN, SHL-DNN model in the wind energy turbine set two weeks data of Ningxia, wherein, a () is the result of mean absolute error criterion, b () is the result of mean absolute percentage error criterion, c () is the result of square error criterion, (d) is the result of root-mean-square error criterion;

Fig. 3 is the result of the different error criterions of SVR, DNN, SHL-DNN model in the wind energy turbine set two weeks data of Jilin, wherein, a () is the result of mean absolute error criterion, b () is the result of mean absolute percentage error criterion, c () is the result of square error criterion, (d) is the result of root-mean-square error criterion;

The predicted data of Fig. 4 (a) raw data and SVR, DNN, SHL-DNN tri-models at data set be two weeks data Ningxia wind energy turbine set on be the results contrast of 10 minutes about predicted time;

The predicted data of Fig. 4 (b) raw data and SVR, DNN, SHL-DNN tri-models at data set be two weeks data Ningxia wind energy turbine set on be the results contrast of 30 minutes about predicted time;

The predicted data of Fig. 4 (c) raw data and SVR, DNN, SHL-DNN tri-models at data set be two weeks data Ningxia wind energy turbine set on be the results contrast of 1 hour about predicted time;

The predicted data of Fig. 4 (d) raw data and SVR, DNN, SHL-DNN tri-models at data set be two weeks data Ningxia wind energy turbine set on be the results contrast of 2 hours about predicted time;

The predicted data of Fig. 5 (a) raw data and SVR, DNN, SHL-DNN tri-models at data set be 3 months data Ningxia wind energy turbine set on be the results contrast of 10 minutes about predicted time;

The predicted data of Fig. 5 (b) raw data and SVR, DNN, SHL-DNN tri-models at data set be 3 months data Ningxia wind energy turbine set on be the results contrast of 30 minutes about predicted time;

The predicted data of Fig. 5 (c) raw data and SVR, DNN, SHL-DNN tri-models at data set be 3 months data Ningxia wind energy turbine set on be the results contrast of 1 hour about predicted time;

The predicted data of Fig. 5 (d) raw data and SVR, DNN, SHL-DNN tri-models at data set be 3 months data Ningxia wind energy turbine set on be the results contrast of 2 hours about predicted time;

The predicted data of Fig. 6 (a) raw data and SVR, DNN, SHL-DNN tri-models at data set be 6 months data Ningxia wind energy turbine set on be the results contrast of 10 minutes about predicted time;

The predicted data of Fig. 6 (b) raw data and SVR, DNN, SHL-DNN tri-models at data set be 6 months data Ningxia wind energy turbine set on be the results contrast of 30 minutes about predicted time; SVR, DNN and SHL-DNN

The predicted data of Fig. 6 (c) raw data and SVR, DNN, SHL-DNN tri-models at data set be 6 months data Ningxia wind energy turbine set on be the results contrast of 1 hour about predicted time;

The predicted data of Fig. 6 (d) raw data and SVR, DNN, SHL-DNN tri-models at data set be 6 months data Ningxia wind energy turbine set on be the results contrast of 2 hours about predicted time.

Embodiment

Below in conjunction with the drawings and specific embodiments, technical solution of the present invention is described in further detail.

The present invention proposes the forecast of short-term wind speed is carried out in a kind of application model based on deep neural network migration models.Degree of depth learning art is utilized to build the deep neural network model of the multiple-input and multiple-output sharing hidden layer, as shown in Figure 1.In this structure, input layer and hidden layer are that all wind energy turbine set are shared, and can regard a common eigentransformation as.Output layer is that each wind energy turbine set is separate, because their Data distribution8 is different.This is a type of knowledge migration, because general feature is migrated to each data set.

As shown in Figure 1, the present invention proposes a kind of short-term wind speed forecasting method based on deep neural network migration models, it is characterized in that, comprises the following steps:

What step one, setting data integrated record is all with ten minutes air speed datas as interval, first carries out maximin normalization pre-service to the data of two or more wind energy turbine set; By the set of data samples that pretreated Data Placement is in units of sky, and as the input of deep neural network migration models, the proper vector of the sample in the i-th moment is the wind velocity vector of 144 dimensions: x _i=(x _i-143, x _i-142..., x _i-1, x _i), the output of this deep neural network is the forecasting wind speed data of 8 hours after this sample, i.e. the wind velocity vector of 48 dimensions: y _i=(x _i+1, x _i+2..., x _i+47, x _i+48), wherein x _iit is the wind speed in i-th moment;

Step 2, structure deep neural network migration models, wherein, input layer is the day data sequence (144 nodes) after above-mentioned wind energy turbine set process, hidden layer is two noise reduction automatic coding machine (Stacked Denoising Auto-encoder piled up, SDA) (being all 100 nodes), output layer is the air speed data sequence (48 nodes) of prediction in following 8 hours.Here, we use noise reduction automatic coding machine (Denoising Autoencoder, DAE) to replace traditional automatic coding machine (Auto-encoder, AE).The idea of DAE is very simple.In order to make hidden layer represent the feature of more robust or make it avoid being equal to input completely, input is replaced with the treated version of part this is by a Random Maps complete.We use the mode at random part input node being become 0.It is identical with AE, be encoded as hiding expression:

$\tilde{y}=f_{\mathrm{\θ}}=\left(\tilde{x}\right)=s\left(W\tilde{x}+b\right)---\left(1\right)$

Wherein s is a Nonlinear Mapping, such as, sigmoid function in our experiment.θ={ W, b} represent its parameter sets, wherein the weight matrix of W to be dimension be d' × d, the bias vector of b to be dimension be d'.Afterwards, the expression of hiding or coding be decoded as by a similar conversion reconstruct:

$\tilde{z}=g_{{\mathrm{\θ}}^{\′}}=\left(\tilde{y}\right)=s\left(W^{\′}\tilde{y}+b^{\′}\right)---\left(2\right)$

Wherein can regard as a prediction, parameter sets is θ '={ W', b'}.The weight matrix W' of inverse mapping can be set to W'=W ^t.Each like this trained values be mapped as one corresponding with a reconstruct the parameter of this model, is respectively θ and θ ', optimisedly makes minimized average reconstructed error:

$\begin{array}{c}{\mathrm{\θ}}^{*},{\mathrm{\θ}}^{\′*}=\underset{\mathrm{\θ},{\mathrm{\θ}}^{\′}}{\arg \min }\frac{1}{n}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}L({\tilde{x}}^{\left(i\right)},{\tilde{z}}^{\left(i\right)})\\ =\underset{\mathrm{\θ},{\mathrm{\θ}}^{\′}}{\arg \min }\frac{1}{n}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}L({\tilde{x}}^{\left(i\right)},g_{{\mathrm{\θ}}^{\′}}\left(f_{\mathrm{\θ}}\left({\tilde{x}}^{\left(i\right)}\right)\right))\end{array}---\left(3\right)$

Wherein, L is loss function, traditional Squared Error Loss of such as our use

Step 3, the excitation function that every layer of DAE is set, learning rate, noise reduction ratio etc.The number of samples of arrange frequency of training, at every turn training together, adopts the bottom-up beginning order training method of SDA unsupervised learning, obtains the initial weight parameter that each hidden layer connects;

The number of samples etc. that step 4, the excitation function arranging deep neural network, learning rate, frequency of training are trained together with each, the initial weight parameter that each hidden layer obtained based on step 3 connects carries out retraining further by top-down supervised learning to this deep neural network, thus the weighting parameter that after obtaining finely tuning, each hidden layer connects, in this step use all data sets have label data;

The weighting parameter that step 5, initialization top layer are connected with hidden layer, then the label data that has of each wind energy turbine set is used to finely tune the weighting parameter that top layer is connected with hidden layer respectively, obtain the output layer of corresponding each wind energy turbine set in deep neural network, i.e. the wind series of each wind energy turbine set prediction;

Step 6, the result exported this deep neural network carry out the predicted value of 8 hours wind speed after renormalization obtains, and according to predicted value and actual value computational prediction error.

Research material of the present invention:

In the forecast of short-term wind speed, the migration models based on deep neural network compares non-migratory SVR, non-migratory DNN model, shows obviously good prediction effect.

The evaluation of wind speed prediction error, the most frequently used index has four, be respectively mean absolute error (Mean absolute error, MAE), mean absolute percentage error (Mean absolute percentage error, MAPE), square error (Mean square error, and root-mean-square error (Root mean square error, RMSE) MSE).Their computing formula is as follows:

$\mathrm{MAE}=\frac{1}{n}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}|y_i-y_i^{\′}|---\left(4\right)$

$\mathrm{MAPE}=\frac{1}{n}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}\frac{|y_i-y_i^{\′}|}{y_i}---\left(5\right)$

$\mathrm{MSE}=\frac{1}{n}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{(y_i-y_i^{\′})}^2---\left(6\right)$

$\mathrm{RMSE}=\sqrt{\frac{1}{n}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{(y_i-y_i^{\′})}^2}---\left(7\right)$

Wherein, n is the number of sample, y _ii-th actual value, y _i' be i-th predicted value.

1, the experiment effect of the few wind energy turbine set of data is migrated to

Present hypothetical target wind energy turbine set has a small amount of data.We have done altogether two experiments.For experiment 1, select Ningxia wind energy turbine set to be aiming field, other three wind energy turbine set are source domain.The data in Ningxia are two weeks wind speed, and other three power plant are the data of a year.For experiment 2, select Jilin wind energy turbine set to be aiming field, other three wind energy turbine set are source domain.The data in Jilin are two weeks wind speed, and the data in Ningxia are a year and a half wind speed, and two other is the data of a year.With above-mentioned four measurement indexs, non-migratory SVR, non-migratory DNN model and our migration models three kinds of models based on deep neural network of design are evaluated.The forecast result of two experiments as shown in Figure 2,3.

To sum up, apply non-migratory SVR model, non-migratory DNN and the migration models based on deep neural network model and respectively the forecast of short-term wind speed is carried out to the data of different wind energy turbine set, experimental result illustrates, the migration models modelling effect more non-migratory than two other is well a lot.Therefore, for the wind energy turbine set that data volume is fewer, transfer learning is a kind of effective method.

2, along with the experiment effect increased of target wind farm data

Design another group experimental research model migration and increase along with aiming field training data the impact predicting the outcome and be subject to.Except the difference of target data amount, other to arrange the same group of experiment identical.Table 1-8 summarizes for different aiming fields, the result of different error measurement index.Table 1-4 is for experiment 1 Ningxia wind energy turbine set, and table 5-8 is for experiment 2 Jilin wind energy turbine set.In addition, Fig. 4 (a)-6 (d) illustrates predicted data for experiment 1 Ningxia wind energy turbine set raw data and three models about the comparison of different prediction step on the training set of different sample size, and this illustrates the prediction effect of different aspect visually.

The different model of table 1 is for different prediction step experimental result about MAE on the wind farm data of experiment 1 Ningxia

The different model of table 2 is for different prediction step experimental result about MAPE on the wind farm data of experiment 1 Ningxia

The different model of table 3 is for different prediction step experimental result about MSE on the wind farm data of experiment 1 Ningxia

The different model of table 4 is for different prediction step experimental result about RMSE on the wind farm data of experiment 1 Ningxia

The different model of table 5 is for different prediction step experimental result about MAE on the wind farm data of experiment 2 Jilin

The different model of table 6 is for different prediction step experimental result about MAPE on the wind farm data of experiment 2 Jilin

The different model of table 7 is for different prediction step experimental result about MSE on the wind farm data of experiment 2 Jilin

The different model of table 8 is for different prediction step experimental result about RMSE on the wind farm data of experiment 2 Jilin

To sum up, apply non-migratory SVR model, non-migratory deep neural network model and carry out the forecast of short-term wind speed based on the data that the migration models of deep neural network model is how many to the difference of different wind energy turbine set respectively, experimental result illustrates, the modelling effect that in most cases migration models is more non-migratory than two other is good.Especially to the forecasting wind speed of 10 minutes afterwards and 30 minutes, when data volume is fewer, migration models is significantly better than other two kinds of methods.But along with the increase of data volume, the importance of migration declines to some extent, especially contrast the prediction after the long period.Therefore, can conclusion be obtained, when target wind farm data is abundant, has not just needed from other wind energy turbine set there migration datas, even can reduce precision of prediction in some cases.

Although invention has been described by reference to the accompanying drawings above; but the present invention is not limited to above-mentioned embodiment; above-mentioned embodiment is only schematic; instead of it is restrictive; those of ordinary skill in the art is under enlightenment of the present invention; when not departing from present inventive concept, can also make a lot of distortion, these all belong within protection of the present invention.

Claims (1)

1., based on a short-term wind speed forecasting method for deep neural network migration models, it is characterized in that, comprise the following steps:

What step one, setting data integrated record is all with ten minutes air speed datas as interval, first carries out maximin normalization pre-service to the data of two or more wind energy turbine set; By the set of data samples that pretreated Data Placement is in units of sky, and as the input of deep neural network migration models, wherein the proper vector of the sample in the i-th moment is the wind velocity vector of 144 dimensions: x _i=(x _i-143, x _i-142..., x _i-1, x _i), the output of this deep neural network is the forecasting wind speed data of 8 hours after this sample, i.e. the wind velocity vector of 48 dimensions: y _i=(x _i+1, x _i+2..., x _i+47, x _i+48), wherein x _iit is the wind speed in the i-th moment;

Step 2, structure deep neural network migration models, wherein, input layer is the day data sequence after above-mentioned wind energy turbine set process, and hidden layer is two noise reduction automatic coding machines piled up, and output layer is the air speed data sequence of prediction in following 8 hours;

Step 3, the bottom-up beginning order training method of employing unsupervised learning, obtain the initial weight parameter that each hidden layer connects;

The initial weight parameter that step 4, each hidden layer obtained based on step 3 connect carries out retraining further by top-down supervised learning to this deep neural network, thus the weighting parameter that after obtaining finely tuning, each hidden layer connects, in this step use all data sets have label data;

The weighting parameter that step 5, initialization top layer are connected with hidden layer, then the label data that has of each wind energy turbine set is used to finely tune the weighting parameter that top layer is connected with hidden layer respectively, obtain the output layer of corresponding each wind energy turbine set in deep neural network, i.e. the wind series of each wind energy turbine set prediction;

Step 6, the result exported this deep neural network carry out the predicted value of 8 hours wind speed after renormalization obtains, and according to predicted value and actual value computational prediction error.