CN113297791B - Wind power combination prediction method based on improved DBN - Google Patents
Wind power combination prediction method based on improved DBN Download PDFInfo
- Publication number
- CN113297791B CN113297791B CN202110560940.6A CN202110560940A CN113297791B CN 113297791 B CN113297791 B CN 113297791B CN 202110560940 A CN202110560940 A CN 202110560940A CN 113297791 B CN113297791 B CN 113297791B
- Authority
- CN
- China
- Prior art keywords
- wind power
- model
- data
- dbn
- prediction
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 66
- 238000012549 training Methods 0.000 claims abstract description 26
- 238000012847 principal component analysis method Methods 0.000 claims abstract description 17
- 230000006870 function Effects 0.000 claims description 24
- 238000012937 correction Methods 0.000 claims description 12
- 210000002569 neuron Anatomy 0.000 claims description 6
- 230000008569 process Effects 0.000 claims description 6
- 230000008859 change Effects 0.000 claims description 4
- 238000005516 engineering process Methods 0.000 claims description 4
- 230000006872 improvement Effects 0.000 claims description 4
- 230000003044 adaptive effect Effects 0.000 claims description 3
- 230000008901 benefit Effects 0.000 claims description 3
- 230000009467 reduction Effects 0.000 claims description 3
- 238000005096 rolling process Methods 0.000 claims description 3
- 238000012795 verification Methods 0.000 claims description 3
- 230000004913 activation Effects 0.000 claims description 2
- 230000005540 biological transmission Effects 0.000 claims description 2
- 238000005315 distribution function Methods 0.000 claims description 2
- 238000005457 optimization Methods 0.000 claims description 2
- 238000012545 processing Methods 0.000 claims description 2
- 238000010248 power generation Methods 0.000 description 7
- 238000013528 artificial neural network Methods 0.000 description 5
- 238000012706 support-vector machine Methods 0.000 description 4
- 238000013135 deep learning Methods 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000000053 physical method Methods 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 238000007619 statistical method Methods 0.000 description 2
- 238000012876 topography Methods 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 238000010219 correlation analysis Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000005265 energy consumption Methods 0.000 description 1
- 238000003912 environmental pollution Methods 0.000 description 1
- 239000000284 extract Substances 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 238000000513 principal component analysis Methods 0.000 description 1
- 230000000717 retained effect Effects 0.000 description 1
- 230000003068 static effect Effects 0.000 description 1
- 230000009466 transformation Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
- G06F18/2135—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods based on approximation criteria, e.g. principal component analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/047—Probabilistic or stochastic networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/082—Learning methods modifying the architecture, e.g. adding, deleting or silencing nodes or connections
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q50/00—Systems or methods specially adapted for specific business sectors, e.g. utilities or tourism
- G06Q50/06—Electricity, gas or water supply
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y04—INFORMATION OR COMMUNICATION TECHNOLOGIES HAVING AN IMPACT ON OTHER TECHNOLOGY AREAS
- Y04S—SYSTEMS INTEGRATING TECHNOLOGIES RELATED TO POWER NETWORK OPERATION, COMMUNICATION OR INFORMATION TECHNOLOGIES FOR IMPROVING THE ELECTRICAL POWER GENERATION, TRANSMISSION, DISTRIBUTION, MANAGEMENT OR USAGE, i.e. SMART GRIDS
- Y04S10/00—Systems supporting electrical power generation, transmission or distribution
- Y04S10/50—Systems or methods supporting the power network operation or management, involving a certain degree of interaction with the load-side end user applications
Abstract
The invention discloses a wind power combination prediction method based on an improved DBN, which relates to the technical field of wind power prediction and comprises the following steps: training and establishing an improved DBN wind power prediction model based on a Gaussian-Bernoulli limited Boltzmann machine, establishing a combined prediction model according to the improved DBN wind power prediction model, NWP data, a principal component analysis method and a wind power prediction model based on a space correlation method, calculating weight coefficients corresponding to the single methods, and obtaining a wind power prediction result of the combined prediction model, and step 3: judging whether the predicted time is met, if yes, ending the operation, otherwise executing the step 4, and executing the step 4: by using a sliding window, newly generated data are added, the oldest data are deleted, and the DBN model is retrained.
Description
Technical Field
The invention relates to the technical field of wind power prediction, in particular to a wind power combination prediction method based on an improved DBN,
background
With the increase of fossil energy consumption and environmental pollution, the effective utilization of renewable energy becomes a necessary trend, wind energy is one of renewable energy sources with the most development prospect, the application of wind energy in a power system is more and more extensive, the global installed capacity of wind power generation reaches 644.5GW by 2018, the accuracy of wind power generation prediction directly influences the safety of a power grid compared with 17.4% of the last year, however, the randomness and intermittence of wind power generation bring difficulty to the prediction of the wind power generation, which challenges a modern power system, so that an accurate wind power generation prediction method needs to be developed to assist the economic dispatch of the power system,
the main categories of wind power generation prediction methods comprise a statistical method, a physical method and spatial correlation, common statistical methods comprise an autoregressive moving average model (ARIMA), a Gaussian Process (GP), an artificial neural network method (ANN), a Support Vector Machine (SVM) and the like, the physical method based on numerical weather forecast (NWP) data and a wind speed-power curve is an indirect prediction method, the problems of insufficient data and the like can be solved, the spatial correlation method (SC) generally analyzes a target wind power field through meteorological information of adjacent wind power fields, and plays an important role in improving prediction precision under the condition of insufficient data of the target wind power field,
however, wind power prediction has a plurality of influence factors, a nonlinear relation is complex, an artificial neural network, a support vector machine and the like belong to a shallow layer model, the expression capability on complex functions is limited, a deep belief network model is formed by stacking a plurality of limited boltzmann machines (RBMs), a deep nonlinear network structure is provided, the complex nonlinear problem can be solved, but the existing research at present often directly inputs data into the underlying structure of a deep learning network, however, when multidimensional meteorological data are input, the input data of each meteorological factor is a continuous input process, if a traditional deep belief network is adopted, the problem of information loss during continuous input is caused, the prediction accuracy is reduced, because the visible layer of the traditional deep belief network is required to be binary distribution,
disclosure of Invention
The invention aims to overcome the defects of the prior art, provides a wind power combination prediction method based on an improved DBN,
the aim of the invention is realized by the following technical scheme:
an improved DBN-based wind power combination prediction method comprises the following steps:
step 1, training and establishing a DBN wind power prediction model based on Gaussian-Bernoulli limited Boltzmann machine improvement,
step 2, establishing a combined prediction model according to the improved DBN wind power prediction model, NWP data, a principal component analysis method and a wind power prediction model based on a space correlation method, calculating weight coefficients corresponding to the single methods to obtain a wind power prediction result of the combined prediction model,
further, the step 1 includes the following substeps:
step 1.1, inputting a proper training set and a verification set,
step 1.2, selecting a rolling dataset of length L,
step 1.3, determining a network structure for improving the DBN model,
step 1.4, selecting parameters of the DBN model,
step 1.5, training and establishing a DBN model,
further, the determining the network structure of the improved DBN model in the step 1.3 includes the following:
adding a Gaussian function into the RBM, and providing a Gaussian-Bernoulli limited Boltzmann machine GBRBM, wherein the energy function of the GBRBM is expressed as:
wherein sigma i 2 Is the variance of the gaussian distribution, m is the number of hidden units in the hidden layer, n is the number of visible units in the visible layer, and the conditional probability between a visible unit and a hidden unit according to the energy function is as follows:
wherein the method comprises the steps ofIs Gao SihanNumber, mean value is μ, variance is σ i 2 The improved DBN model is overlapped according to GBRBM and RBM from bottom to top, the bottom layer is GBRBM, the rest layers are RBM, BP network is arranged at the top of the model,
further, the parameters for selecting the DBN model in the step 1.4 include the following:
an adaptive learning step length ALS technology is adopted to determine the proper learning rate of the DBN model, independent learning rate parameters are adopted to replace the global learning rate for each weight connection, the step length is adjusted according to the change of the sign,
where u represents the increment factor of the learning step, u>1, a step of; d represents the decrement factor of the learning step, d<1，Representing the individual learning rate, if two consecutive updates are in the same direction, the step size will increase, conversely, when the update directions are opposite, the step size will decrease,
further, the step 2 includes the following:
2.1, extracting historical data variables by using a principal component analysis method, inputting the historical data variables into an improved DBN model, carrying out wind power prediction,
2.2, inputting NWP data into the improved DBN model, carrying out wind power prediction,
2.3, extracting the principal component of the original data from the NWP data by using a principal component analysis method, adding wind speed as the input of an improved DBN model to obtain a wind power prediction based on a DBN wind speed correction model,
2.4, wind power prediction is carried out based on a wind power prediction model of a space correlation method,
2.5. weighting each wind power prediction result in 2.1 to 2.4, calculating a weight coefficient corresponding to each single method to obtain a wind power prediction result of the combined prediction model,
further, the method also comprises the following steps:
step 3: judging whether the predicted time is satisfied, if yes, ending the operation, otherwise executing the step 4,
step 4: using the sliding window, adding the newly generated data and deleting the oldest data, returning to execution step 1.2,
the beneficial effects of the invention are as follows:
(1) An improved DBN model is provided, a Gaussian function is added into the RBM, so that the wind power prediction precision is higher than that of the traditional DBN model, the application of the self-adaptive learning step length technology further improves the precision of the system,
(2) The combined prediction method adopting the sliding window strategy effectively combines the advantages of different methods, compared with the traditional DBN and LSTM, GRU, seq seq and other deep neural networks, the prediction accuracy is greatly improved,
(3) By establishing a wind speed correction model based on NWP data, the precision of wind speed data is improved, and further the precision of wind power prediction is improved,
(4) Under the condition that the historical data of the target predicted point is insufficient, the wind power prediction can be completed by utilizing the data of the adjacent wind power plants around the target wind power plant, the influence of the geographic position and the topography on the wind power generation can be simultaneously reflected through the spatial correlation analysis,
drawings
Figure 1 is a diagram of the network architecture of the DBN of the present invention,
figure 2 is a block diagram of the GBRBM-DBN of the invention,
figure 3 is a flow chart of a spatial correlation method,
figure 4 is a flow chart of an NWP wind speed correction model,
figure 5 is a flow chart of the combined weighted wind power prediction of the present invention,
Detailed Description
The technical solution of the present invention will be described in further detail with reference to the accompanying drawings, but the scope of the present invention is not limited to the following,
wind power prediction has a plurality of influence factors, a nonlinear relation is complex, an artificial neural network, a support vector machine and the like belong to a shallow layer model, the expression capability of complex functions is limited, a deep belief network model is formed by stacking a plurality of limited boltzmann machines (RBMs), a deep nonlinear network structure is provided, the complex nonlinear problem can be solved, but the existing research at present always directly inputs data into the underlying structure of the deep learning network, when multidimensional meteorological data are input, the input data of each meteorological factor is a continuous input process, if the traditional deep belief network is adopted, the problem of information loss during continuous input is caused, the prediction accuracy is reduced, because the visible layer of the traditional deep belief network is required to be binary distribution,
the visible layer based on the traditional deep belief network must be binary distribution, which can cause the problem of information loss during continuous input, and lead to the problem of prediction accuracy reduction, and the technical scheme provided by the application is as follows:
step 1, training and establishing a DBN wind power prediction model based on Gaussian-Bernoulli limited Boltzmann machine improvement,
step 2, establishing a combined prediction model according to the improved DBN wind power prediction model, NWP data, a principal component analysis method and a wind power prediction model based on a space correlation method, calculating weight coefficients corresponding to the single methods to obtain a wind power prediction result of the combined prediction model,
wherein the DBN is a probability generation model composed of RBM and BP network, the topology structure is shown in figure 1, the input data is received by the visible layer of the first RBM and transmitted to the hidden layer, the hidden layer of the last RBM is visible to the next RBM, after each hidden layer is trained, the data characteristics are extracted from the current hidden layer, the information is transmitted to the next hidden layer, the last RBM transmits the trained data to the output layer through back transmission,
wherein the RBM is a typical two-layer neural network comprising a visible layer for receiving input data and a hidden layer for detecting characteristics of the data from the visible layer based on connection weights, adjacent layer RBM neurons are connected to each other, while neurons of the same layer are not connected, the visible layer is assumed to be a binary value, and for a given visible layer and hidden layer, the energy function of the RBM is expressed as:
wherein m and n represent the number of neurons, w ij For the connection weight between hidden and visible layers, a i And b j Deviation of visible layer and hidden layer, v i And h j The parameters of the RBM model, which are neurons of the visible and hidden layers, respectively, can be expressed as:
θ＝[W,a,b] (2)
the joint probability distribution function of the visible layer and the hidden layer is as follows:
in RBM model, when v i =1 or h j When=1, the visible node and the hidden node are independent of each other, the activation probability distribution is obtained by the formulas (4) and (5),
where sigmoid () is a sigmoid function, expressed as:
the conventional RBM directly adopts Bernoulli-Bernoulli, the visible layer of which must be binary values, however, the input meteorological data of the wind power prediction model is continuously distributed, and the direct use of the conventional RBM can cause information loss, so by adding a Gaussian function, the Gaussian-Bernoulli limited Boltzmann machine is proposed, the input data is not limited to binary values, and the energy function of the GBRBM is expressed as:
wherein sigma i 2 Is the variance of the gaussian distribution, m is the number of hidden units in the hidden layer, n is the number of visible units in the visible layer, and the conditional probability between a visible unit and a hidden unit according to the energy function is as follows:
wherein the method comprises the steps ofIs a Gaussian function with a mean of μ and a variance of σ i 2 The built GBRBM-DBN model architecture is shown in figure 2, the bottom layer is GBRBM, the rest layers are RBM, the improved DBN model is overlapped according to GBRBM and RBM from bottom to top, BP network is arranged at the top of the model, parameters of the DBN model are further adjusted,
the training of the DBN model usually involves the training of a plurality of RBMs, each iteration needs a long time, and therefore, a proper learning rate needs to be set, which is very important for using the DBN model, the training process is unstable due to the overlarge learning rate, the convergence rate is slow and the training time is long due to the overlarge learning rate, an adaptive learning step length (ALS) technology is provided for determining the proper learning rate, an independent learning rate parameter is adopted for replacing the global learning rate for each weight connection to obtain the satisfactory training rate, the step length is adjusted according to the change of the sign,
where u represents the increment factor of the learning step, u>1, a step of; d represents the decrement factor of the learning step, d<1，Representing the individual learning rate, if two successive updates are in the same direction, the step size will increase, conversely, when the update directions are opposite, the step size will decrease, avoiding the problem caused by improper step size, improving the convergence rate of the DBN model,
in some embodiments, where redundancy problems exist between different variables in a multi-variable dataset, resulting in increased computational complexity and reduced prediction accuracy, PCA is a method of reducing the dimensionality of the dataset by building a new set of variables that reflect linear combinations between the original variables, thereby capturing the relationship between them, n-dimensional data can be mapped to k-dimensional data (k < n) by orthogonal transformation, these new variables being uncorrelated, called principal components, in contrast to the original information in the original variables being retained to a maximum,
assume that there is a sample set q= { Q 1 ,q 2 ,...,q k The mean value is 0, and the new coordinate obtained by projection is w= { W 1 ,w 2 ,...,w k }，w i Representing a characteristic value χ i Is converted into a orthonormal basis W i || 2 =1, the projection of the sample point onto the plane is denoted W T q i The variance after projection isThe optimization problem is defined as follows:
maxtr(W T QQ T W)s.t.W T W＝1 (13)
processing the formula (13) by using Lagrangian multiplier method to obtain formula (14),
QQ T W＝χ W (14)
the eigenvalue derived from equation (14) is χ 1 ≥χ 2 ≥...≥χ k Ranking, variance contribution and total variance formula are as follows:
new coordinate V obtained by dimension reduction p ＝(v 1 ,v 2 ,...,v p ) The first p eigenvectors containing eigenvalues, which are solutions to principal component analysis, will typically be considered as the primary information containing the original reference data if ηΣ (p) > 85%.
In some embodiments, if only historical data is used, the prediction horizon is typically less than 6 hours, and the dependence on the historical data can be reduced using numerical weather forecast data, which is based on calculation of specific starting and edge values to solve equations of hydrodynamic and thermodynamic systems, including wind speed, wind direction, temperature, humidity, air pressure, air density and rainfall, to predict weather data using a physical model, to provide basis for wind power prediction,
the NWP data has a great influence on the accuracy of wind power prediction, while the meteorological system is an unstable dynamic system, the NWP data may not be consistent with the actual data, and due to the cubic relationship between wind power and wind speed, a small NWP error may cause a huge error of wind power prediction, based on this, a wind speed correction model based on a DBN model is established according to the NWP wind speed and the actual wind speed, and principal component variables are extracted by using a principal component analysis method, and the correction model has the advantage of improving the accuracy of wind speed data prediction, and fig. 4 is a flowchart of the correction model.
In some embodiments, it is reasonable in prediction to analyze and consider the spatial correlation of the target wind farm with the neighboring wind farms when predicting the wind power of the target wind farm, furthermore, when the historical data of the target wind farm is missing, the data of the neighboring wind farms can be used to assist in the prediction of the target wind farm, and the method can reflect the effect of the geographic location and topography by using the data of the surrounding observation points, assuming that the wind speed sequence of the target wind farm is expressed asRepresenting the wind speed sequence, y, of adjacent wind farms it Represents the wind direction sequence, theta, of adjacent wind farms i Indicating the wind direction y it And the angle i of the observation point relative to the target point th The difference between the two is shown in fig. 3, a flow chart of a spatial correlation method is shown, the time difference between the adjacent wind power plant and the target wind power plant needs to be obtained, the selected adjacent observation point is b, the target wind power plant is a, the pearson correlation coefficient is used as a standard for determining the time difference, and the method can be expressed as follows:
wherein N is the data number, v a And v bi Wind speed, eta of the target wind power plant and the adjacent wind power plant respectively v The closer to 1, the higher the correlation between wind speeds, Σv a Sum sigma v bi Can be expressed as:
wherein deltat represents the time difference between the adjacent wind power plant and the target wind power plant, the pearson correlation coefficient changes along with the time delay, when the wind speed sequence similarity of the adjacent wind power plant and the target wind power plant is highest, the corresponding pearson correlation coefficient is the largest,
as a possible implementation manner, based on the implementation manner, the method extracts the historical data variable by using the principal component analysis method and inputs the historical data variable into the improved DBN model to perform wind power prediction, the NWP data is input into the improved DBN model to perform wind power prediction, the NWP data is used for extracting the principal component of the original data by using the principal component analysis method and adding wind speed as the input of the improved DBN model to obtain a wind power correction model based on the DBN to perform wind power prediction, the wind power prediction model based on the spatial correlation method to perform wind power prediction, a combined prediction method based on GBRBM-DBN is provided, each wind power prediction result is weighted, weight coefficients corresponding to each single method are calculated to obtain a wind power prediction result of the combined prediction model,
assume that the wind power predicted by the prediction model is P 1 、P 2 、P 3 ，…P M The corresponding weighting factor is beta 1 、β 2 、β 3 ，…β M The wind power predicted by the combined prediction method is expressed as:
the prediction error of each prediction model is epsilon 1 、ε 2 、…、ε M Its corresponding variance is ζ 1 、ζ 2 、…、ζ i I=1, 2, …, M, variance formula of the combined weighting method is:
in the case where the formula (22) is satisfied, the determination weight factor is optimized by minimizing the variance in the formula (21),
and solving the minimum value of the variance by using a Lagrange multiplier method to obtain a weighting factor of the combined model, wherein the Lagrange function is expressed as follows:
beta derived in (23) i And lambda is zero, the optimal weighting factor beta can be obtained 1 、β 2 、…、β M ，
In some embodiments, for a traditional static window, based on the proposed wind power prediction model, a mapping relationship between input variables and wind power is established, the relationship is kept unchanged with time, however, in a wind power sequence, the value at a certain moment has a great correlation with the value at an adjacent moment, the correlation with the data far from the moment is not great, the data for establishing the model should be continuously updated and changed, and under the condition that the window length is certain, the latest data is utilized for establishing the model to ensure the accuracy of a prediction result,
assume that there is a continuous data set (x 1 ,y 1 ),…,(x t ,y t ),…,(x L ,y L )∈R n The modeling information for time t is obtained from a set of data before time t, then a training data set is created with the length of the training data set being constant, the window will move forward over time until the last time is reached, the oldest data of the same length is removed when newly generated data is added, the predictive model is continuously retrained with updated training data, and the variable s is input i (t) and output data r i The data samples (each time using a sliding window) of (t) are as follows:
s i (t)＝[x i (t-L),x i (t-L+2),…,x i (t-1)] (24)
r i (t)＝[x i (t),x i (t+1),…,x i (t+d-1)] (25)
wherein x is i (t) is the value of the ith variable at time t, d is the prediction range, assuming that the current moment is t, training set data is from time t-L to t-1, the length is L, firstly, a wind power prediction model is established by using training data, the wind power of the time t is predicted, a newly generated data sample is added until the next time t+1, the data at the time t-L is deleted, modeling data is changed from t-L+1 to t, the length is still L, the prediction adopting a sliding window strategy is a dynamic process taking window length as a unit,
as a possible implementation manner, based on the foregoing implementation, as shown in fig. 5, the wind power prediction method of the present application may be described as follows:
step 1, training and establishing a DBN wind power prediction model based on Gaussian-Bernoulli limited Boltzmann machine improvement,
step 1.1, inputting a proper training set and a verification set,
step 1.2, selecting a rolling dataset of length L,
step 1.3, determining a network structure for improving the DBN model,
step 1.4, selecting parameters of the DBN model,
step 1.5, training and establishing a DBN model,
step 2, establishing a combined prediction model according to the improved DBN wind power prediction model, NWP data, a principal component analysis method and a wind power prediction model based on a space correlation method, calculating weight coefficients corresponding to the single methods to obtain a wind power prediction result of the combined prediction model,
2.1, extracting historical data variables by using a principal component analysis method, inputting the historical data variables into an improved DBN model, carrying out wind power prediction,
2.2, inputting NWP data into the improved DBN model, carrying out wind power prediction,
2.3, extracting the principal component of the original data from the NWP data by using a principal component analysis method, adding wind speed as the input of an improved DBN model to obtain a wind power prediction based on a DBN wind speed correction model,
2.4, wind power prediction is carried out based on a wind power prediction model of a space correlation method,
2.5. weighting each wind power prediction result in 2.1 to 2.4, calculating a weight coefficient corresponding to each single method to obtain a wind power prediction result of the combined prediction model,
step 3: judging whether the predicted time is satisfied, if yes, ending the operation, otherwise executing the step 4,
step 4: using the sliding window, adding the newly generated data and deleting the oldest data, returning to execution step 1.2,
it should be noted that types 1,2, 3, and 4 respectively represent spring, summer, autumn, and winter, the method is applied according to the actual conditions of different wind speeds in different seasons,
the foregoing is merely a preferred embodiment of the present invention, and it should be understood that the described embodiments are some, but not all, embodiments of the present invention, and that all other embodiments obtained by persons of ordinary skill in the art without making any inventive effort based on the embodiments of the present invention are within the scope of protection of the present invention, which is not to be limited to the form disclosed herein, but is to be used in various other combinations, modifications and environments, and can be modified within the scope of the idea described herein by the above teachings or the skill or knowledge of the relevant art, and all modifications and changes made by those skilled in the art without departing from the spirit and scope of the invention are to be considered as within the scope of the appended claims.
Claims (1)
1. The wind power combination prediction method based on the improved DBN is characterized by comprising the following steps of:
step 1, training and establishing a DBN wind power prediction model based on the improvement of a Gaussian-Bernoulli limited Boltzmann machine,
step 2, establishing a combined prediction model according to the improved DBN wind power prediction model, NWP data, a principal component analysis method and a wind power prediction model based on a space correlation method, and calculating weight coefficients corresponding to the single methods to obtain a wind power prediction result of the combined prediction model;
said step 1 comprises the sub-steps of:
step 1.1, inputting a proper training set and a verification set,
step 1.2, selecting a rolling dataset of length L,
step 1.3, determining a network structure for improving the DBN model,
step 1.4, selecting parameters of the DBN model,
step 1.5, training and establishing a DBN model;
the network structure for improving the DBN model determined in the step 1.3 comprises the following contents:
adding a Gaussian function into the RBM, and providing a Gaussian-Bernoulli limited Boltzmann machine GBRBM, wherein the energy function of the GBRBM is expressed as:
wherein the method comprises the steps ofIs the variance of the gaussian distribution, m is the number of hidden units in the hidden layer, n is the number of visible units in the visible layer, and the conditional probability between a visible unit and a hidden unit according to the energy function is as follows:
wherein the method comprises the steps ofIs a Gaussian function with a mean of μ and a variance of +.>The improved DBN model is overlapped according to GBRBM and RBM from bottom to top, the bottom layer is GBRBM, the rest layers are RBM, and BP network is arranged at the top of the model;
the parameters of the DBN model selected in the step 1.4 comprise the following contents:
an adaptive learning step length ALS technology is adopted to determine the proper learning rate of the DBN model, independent learning rate parameters are adopted to replace the global learning rate for each weight connection, the step length is adjusted according to the change of the sign,
where u > 1, u represents an increment factor of the learning step, d < 1, d represents a decrement factor of the learning step,representing the individual learning rate, if two consecutive updates are in the same direction, the step size will increase; conversely, when the update direction is reversed, the step size will decrease;
the step 2 comprises the following steps:
2.1, extracting historical data variables by using a principal component analysis method, inputting the historical data variables into an improved DBN model, carrying out wind power prediction,
2.2, inputting NWP data into the improved DBN model, carrying out wind power prediction,
2.3, extracting the principal component of the original data from the NWP data by using a principal component analysis method, adding wind speed as the input of an improved DBN model to obtain a wind power prediction based on a DBN wind speed correction model,
2.4, wind power prediction is carried out based on a wind power prediction model of a space correlation method,
2.5. weighting all wind power prediction results in 2.1 to 2.4, and calculating weight coefficients corresponding to all the single methods to obtain wind power prediction results of the combined prediction model;
the method also comprises the following steps:
step 3: judging whether the predicted time is satisfied, if yes, ending the operation, otherwise executing the step 4,
step 4: using the sliding window, adding the newly generated data and deleting the oldest data, and returning to execute the step 1.2;
the DBN is a probability generation model composed of RBM and BP networks, input data are received by a visible layer of a first RBM and transmitted to a hidden layer, the hidden layer of the first RBM is visible to a next RBM, each hidden layer is trained, after data features are extracted from the current hidden layer, information is transmitted to the next hidden layer, and the last RBM transmits trained data to an output layer through back transmission;
wherein, the energy function of RBM is expressed as:
wherein m and n represent the number of neurons, w ij For the connection weight between hidden and visible layers, a i And b j Deviation of visible layer and hidden layer, v i And h j The parameters of the RBM model, which are neurons of the visible and hidden layers, respectively, can be expressed as:
θ＝[W，a，b]
the joint probability distribution function of the visible layer and the hidden layer is as follows:
in RBM model, when v i =1 or h j When=1, the visible node and the hidden node are independent of each other, and the activation probability distribution is obtained by the following formula:
where sigmoid () is a sigmoid function, expressed as:
by adding a Gaussian function, a Gaussian-Bernoulli limited Boltzmann machine is proposed, and the energy function of the GBRBM is expressed as:
wherein the method comprises the steps ofIs the variance of the gaussian distribution, m is the number of hidden units in the hidden layer, n is the number of visible units in the visible layer, and the conditional probability between a visible unit and a hidden unit according to the energy function is as follows:
wherein the method comprises the steps ofIs a Gaussian function with a mean of μ and a variance of σ i 2 Establishing a GBRBM-DBN model, wherein the bottom layer of the GBRBM-DBN model is the GBRBM, the rest layers are the RBMs, the improved DBN model is overlapped according to the GBRBM and the RBM from bottom to top, and a BP network is arranged at the top of the model to further adjust parameters of the DBN model;
each weight connection is replaced by independent learning rate parameter to replace global learning rate, step length is adjusted according to the change of the sign,
wherein u represents an increment factor of the learning step, u > 1; d represents a decrement factor of the learning step, d < 1;representing the individual learning rate, if two continuous updates are in the same direction, the step length is increased, and conversely, when the update directions are opposite, the step length is reduced, so that the problem caused by improper step length is avoided, and the convergence speed of the DBN model is improved;
with sample set q= { Q 1 ，q 2 ，...，q k The mean value is 0, and the new coordinate obtained by projection is w= { W 1 ，w 2 ，...，w k }，w i Representing a characteristic value χ i Is converted into a orthonormal basis W i || 2 =1, the projection of the sample point onto the plane is denoted W T q i The variance after projection isThe optimization problem is defined as follows:
maxtr(W T QQ T W)s.t.W T W＝1
processing by Lagrangian multiplier method to obtain formula:
QQ T W＝χW
characteristic value according to χ 1 ≥χ 2 ≥...≥χ k Ranking, variance contribution and total variance formula are as follows:
new coordinate V obtained by dimension reduction p ＝(v 1 ，v 2 ，...，v p ) The first p eigenvectors containing eigenvalues, if η Σ (p) > 85%, the new variable will be regarded as the main information containing the original reference data;
according to the NWP wind speed and the actual wind speed, a wind speed correction model based on a DBN model is established, a principal component variable is extracted by using a principal component analysis method, and the correction model has the advantage of improving the wind speed data prediction accuracy;
when the historical data of the target wind power plant is missing, the data of the adjacent wind power plants are used for assisting the prediction of the target wind power plant, and the wind speed sequence of the target wind power plant is assumed to be expressed asx it Representing the wind speed sequence, y, of adjacent wind farms it Represents the wind direction sequence, theta, of adjacent wind farms i Indicating the wind direction y it And the angle i of the observation point relative to the target point th The difference between the adjacent wind power plant and the target wind power plant needs to be obtained, the selected adjacent observation point is b, the target wind power plant is a, the pearson correlation coefficient is used as a standard for determining the time difference, and the time difference is expressed as follows:
wherein N is the data number, v a And v bi Wind speeds of the target wind power plant and the adjacent wind power plants respectively; η (eta) v The closer to 1, the higher the correlation between wind speeds; sigma v a Sum sigma v bi Can be expressed as:
wherein Δt represents the time difference between the adjacent wind farm and the target wind farm; the pearson correlation coefficient varies with time delay; when the wind speed sequence similarity of the adjacent wind power plant and the target wind power plant is highest, the corresponding pearson correlation coefficient is the largest;
extracting historical data variables by a principal component analysis method, inputting the historical data variables into an improved DBN model, carrying out wind power prediction, inputting NWP data into the improved DBN model, carrying out wind power prediction, extracting principal components of original data by the NWP data by the principal component analysis method, adding wind speed as input of the improved DBN model, obtaining a wind power correction model based on the DBN, carrying out wind power prediction by a wind power prediction model based on a spatial correlation method, providing a GBM-DBN-based combined prediction method, weighting all wind power prediction results, calculating weight coefficients corresponding to all single methods, and obtaining a wind power prediction result of the combined prediction model;
the wind power predicted by the prediction model is P 1 、P 2 、P 3 ，...P M The corresponding weighting factor is beta 1 、β 2 、β 3 ，...β M The wind power predicted by the combined prediction method is expressed as:
the prediction error of each prediction model is epsilon 1 、ε 2 、...、ε M Its corresponding variance is ζ 1 、ζ 2 、...、ζ i I=1, 2, …, M, variance formula of the combined weighting method is:
optimizing the determination of the weight factor by minimizing the variance;
solving the minimum value of the variance by using a Lagrange multiplier method to obtain a weighting factor of the combined model; the Lagrangian function is expressed as:
when beta is i And lambda is zero, the optimal weighting factor beta can be obtained 1 、β 2 、...、β M ；
With successive data sets (x 1 ，y 1 )，…，(x t ，y t )，…，(x L ，y L )∈R n The modeling information for time t is obtained from a set of data before time t, then a training data set is created with the length of the training data set being constant, the window will move forward over time until the last time is reached, the oldest data of the same length is removed when newly generated data is added, the predictive model is continuously retrained with updated training data, and the variable s is input i (t) and output data r i The data sample of (t) is as follows:
s i (t)＝[x i (t-L)，x i (t-L+2)，…，x i (t-1)]
r i (t)＝[x i (t)，x i (t+1)，…，x i (t+d-1)]
wherein x is i And (t) is the value of the ith variable at time t, d is a prediction range, the current moment is t, the training set data is from time t-L to t-1, the length is L, firstly, a wind power prediction model is established by using the training data, the wind power of the time t is predicted, a newly generated data sample is added until the next time t+1, the data at the time t-L is deleted, the modeling data is changed from t-L+1 to t, the length is still L, and the prediction adopting the sliding window strategy is a dynamic process taking the window length as a unit.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110560940.6A CN113297791B (en) | 2021-05-18 | 2021-05-18 | Wind power combination prediction method based on improved DBN |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110560940.6A CN113297791B (en) | 2021-05-18 | 2021-05-18 | Wind power combination prediction method based on improved DBN |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113297791A CN113297791A (en) | 2021-08-24 |
| CN113297791B true CN113297791B (en) | 2024-02-06 |
Family
ID=77323908
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110560940.6A Active CN113297791B (en) | 2021-05-18 | 2021-05-18 | Wind power combination prediction method based on improved DBN |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113297791B (en) |
Citations (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107725283A (en) * | 2017-09-19 | 2018-02-23 | 江苏方天电力技术有限公司 | A kind of fan trouble detection method based on depth belief network model |
| CN108776843A (en) * | 2018-04-12 | 2018-11-09 | 天津大学 | Photovoltaic ultra-short term power forecasting method based on Boltzmann machine |
| WO2019141040A1 (en) * | 2018-01-22 | 2019-07-25 | 佛山科学技术学院 | Short term electrical load predication method |
| CN110288136A (en) * | 2019-06-11 | 2019-09-27 | 上海电力学院 | Wind power multi-step Predictive Model method for building up |
| CN110580543A (en) * | 2019-08-06 | 2019-12-17 | 天津大学 | Power load prediction method and system based on deep belief network |
| CN110929958A (en) * | 2019-12-10 | 2020-03-27 | 西安邮电大学 | Short-term traffic flow prediction method based on deep learning parameter optimization |
| CN110991737A (en) * | 2019-11-29 | 2020-04-10 | 河海大学 | Ultra-short-term wind power prediction method based on deep belief network |
| CN111091233A (en) * | 2019-11-26 | 2020-05-01 | 江苏科技大学 | Wind power plant short-term wind power prediction modeling method based on wavelet analysis and multi-model AdaBoost depth network |
| CN112001066A (en) * | 2020-07-30 | 2020-11-27 | 四川大学 | Deep learning-based method for calculating limit transmission capacity |
| CN112288164A (en) * | 2020-10-29 | 2021-01-29 | 四川大学 | Wind power combined prediction method considering spatial correlation and correcting numerical weather forecast |
| CN112330051A (en) * | 2020-11-20 | 2021-02-05 | 国网辽宁省电力有限公司营口供电公司 | Short-term load prediction method based on Kmeans and FR-DBN |
| CN112329979A (en) * | 2020-09-23 | 2021-02-05 | 燕山大学 | Ultra-short-term wind power prediction method based on self-adaptive depth residual error network |
| CN112418526A (en) * | 2020-11-24 | 2021-02-26 | 国网天津市电力公司 | Comprehensive energy load control method and device based on improved deep belief network |
-
2021
- 2021-05-18 CN CN202110560940.6A patent/CN113297791B/en active Active
Patent Citations (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107725283A (en) * | 2017-09-19 | 2018-02-23 | 江苏方天电力技术有限公司 | A kind of fan trouble detection method based on depth belief network model |
| WO2019141040A1 (en) * | 2018-01-22 | 2019-07-25 | 佛山科学技术学院 | Short term electrical load predication method |
| CN108776843A (en) * | 2018-04-12 | 2018-11-09 | 天津大学 | Photovoltaic ultra-short term power forecasting method based on Boltzmann machine |
| CN110288136A (en) * | 2019-06-11 | 2019-09-27 | 上海电力学院 | Wind power multi-step Predictive Model method for building up |
| CN110580543A (en) * | 2019-08-06 | 2019-12-17 | 天津大学 | Power load prediction method and system based on deep belief network |
| CN111091233A (en) * | 2019-11-26 | 2020-05-01 | 江苏科技大学 | Wind power plant short-term wind power prediction modeling method based on wavelet analysis and multi-model AdaBoost depth network |
| CN110991737A (en) * | 2019-11-29 | 2020-04-10 | 河海大学 | Ultra-short-term wind power prediction method based on deep belief network |
| CN110929958A (en) * | 2019-12-10 | 2020-03-27 | 西安邮电大学 | Short-term traffic flow prediction method based on deep learning parameter optimization |
| CN112001066A (en) * | 2020-07-30 | 2020-11-27 | 四川大学 | Deep learning-based method for calculating limit transmission capacity |
| CN112329979A (en) * | 2020-09-23 | 2021-02-05 | 燕山大学 | Ultra-short-term wind power prediction method based on self-adaptive depth residual error network |
| CN112288164A (en) * | 2020-10-29 | 2021-01-29 | 四川大学 | Wind power combined prediction method considering spatial correlation and correcting numerical weather forecast |
| CN112330051A (en) * | 2020-11-20 | 2021-02-05 | 国网辽宁省电力有限公司营口供电公司 | Short-term load prediction method based on Kmeans and FR-DBN |
| CN112418526A (en) * | 2020-11-24 | 2021-02-26 | 国网天津市电力公司 | Comprehensive energy load control method and device based on improved deep belief network |
Non-Patent Citations (9)
| Title |
|---|
| Deep belief network based k-means cluster approach for short-term wind power forecasting;Kejun Wang等;Energy;840-852 * |
| Ultra-Short-Term Wind Power Forecasting Based on Deep Belief Network;Sen Wang等;2019 Chinese Control Conference (CCC);1-5 * |
| 基于KMO-Bartlett典型风速选取的PCA-WNN短期风速预测;解坤;张俊芳;;发电设备(第02期);20-25 * |
| 基于主成分分析与遗传优化BP神经网络的风电场短期功率预测研究;张泽龙;钱勇;刘华兵;;宁夏电力(第06期);5-10+38 * |
| 基于主成分分析法的风电功率短期组合预测;吴金浩;杨秀媛;孙骏;;电气技术(第07期);56-62 * |
| 基于深度学习网络的风电场功率短期预测研究;潘志刚;《中国优秀硕士学位论文全文数据库 工程科技Ⅱ辑》;正文第12页-第58页 * |
| 基于深度学习网络的风电场功率短期预测研究;潘志刚;刘三明;李莹;朱晓伟;杨阳;;科技与创新(19);4-6 * |
| 深度学习驱动的电力系统暂态稳定预防控制进化算法;苏童;刘友波;沈晓东;刘挺坚;邱高;刘俊勇;;中国电机工程学报(第12期);103-114 * |
| 组合预测技术在风电中的研究分析;钟宏宇;王超;张旭光;赵家志;赵立峰;任俊明;;电器与能效管理技术(第16期);65-71 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113297791A (en) | 2021-08-24 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108564192B (en) | Short-term photovoltaic power prediction method based on meteorological factor weight similarity day | |
| CN107766990B (en) | Method for predicting power generation power of photovoltaic power station | |
| CN112288164B (en) | Wind power combined prediction method considering spatial correlation and correcting numerical weather forecast | |
| CN110942194A (en) | Wind power prediction error interval evaluation method based on TCN | |
| CN113705877A (en) | Real-time monthly runoff forecasting method based on deep learning model | |
| CN116937579B (en) | Wind power interval prediction considering space-time correlation and interpretable method thereof | |
| CN114298140A (en) | Wind power short-term power prediction correction method considering unit classification | |
| Li et al. | Deep spatio-temporal wind power forecasting | |
| CN115099296A (en) | Sea wave height prediction method based on deep learning algorithm | |
| CN114757427A (en) | Autoregressive corrected LSTM intelligent wind power plant ultra-short term power prediction method | |
| CN114117852B (en) | Regional heat load rolling prediction method based on finite difference working domain division | |
| CN114897129A (en) | Photovoltaic power station short-term power prediction method based on similar daily clustering and Kmeans-GRA-LSTM | |
| Lyu et al. | Dynamic feature selection for solar irradiance forecasting based on deep reinforcement learning | |
| Zhang et al. | A local semi-supervised ensemble learning strategy for the data‐driven soft sensor of the power prediction in wind power generation | |
| Liu et al. | An online transfer learning model for wind turbine power prediction based on spatial feature construction and system-wide update | |
| CN108832623B (en) | Physical-statistical hybrid two-stage wind power prediction method | |
| CN110163437B (en) | Day-ahead photovoltaic power generation power prediction method based on DPK-means | |
| CN113297791B (en) | Wind power combination prediction method based on improved DBN | |
| CN114444763A (en) | Wind power prediction method based on AFSA-GNN | |
| CN115423091A (en) | Conditional antagonistic neural network training method, scene generation method and system | |
| CN114943277A (en) | Method for identifying operation condition and predicting operation performance of comprehensive energy system | |
| Li et al. | GA-SVR Traffic Flow Prediction Based on Phase Space Reconstruction with Improved KNN Method | |
| Huang et al. | Short-term load forecasting based on similar day approach and intelligent algorithm using analytic hierarchy process | |
| Li et al. | Short-term Photovoltaic Power Forecasting Using SOM-based Regional Modelling Methods | |
| Bashar et al. | Optimized Hybrid Neural Network for Wind Speed Forecasting |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |