CN111798034B - Wind power prediction system and prediction method based on wind flow field space-time image learning - Google Patents

Wind power prediction system and prediction method based on wind flow field space-time image learning Download PDF

Info

Publication number
CN111798034B
CN111798034B CN202010522362.2A CN202010522362A CN111798034B CN 111798034 B CN111798034 B CN 111798034B CN 202010522362 A CN202010522362 A CN 202010522362A CN 111798034 B CN111798034 B CN 111798034B
Authority
CN
China
Prior art keywords
wind power
convolution
prediction
wind
conversion
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
Application number
CN202010522362.2A
Other languages
Chinese (zh)
Other versions
CN111798034A (en
Inventor
程礼临
臧海祥
卫志农
许瑞琦
孙国强
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Hohai University HHU
Original Assignee
Hohai University HHU
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Hohai University HHU filed Critical Hohai University HHU
Priority to CN202010522362.2A priority Critical patent/CN111798034B/en
Publication of CN111798034A publication Critical patent/CN111798034A/en
Application granted granted Critical
Publication of CN111798034B publication Critical patent/CN111798034B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • 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
    • 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
    • G06N3/045—Combinations of 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
    • G06N3/04—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
    • 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

Abstract

The invention discloses a wind power prediction system based on wind flow field space-time image learning, belonging to the technical field of renewable energy development and utilization, and also discloses a prediction method thereof, wherein the method and the system comprise the following steps: firstly, drawing a wind flow field space-time image based on a wind power station site position and wind speed and direction data; secondly, constructing a convolution conversion model by taking the wind flow field image as input, and training the convolution conversion model aiming at the wind power of different wind power stations; then, taking a hidden layer feature graph of a convolution conversion model as a convolution prediction model input, and realizing model combination based on a weight sharing strategy; then, setting the ratio of the conversion error to the prediction error to be 0.1:1, and training a convolution conversion-convolution prediction combined model; and finally, wind power prediction and prediction error verification are realized based on the wind flow field space-time image at any moment. The method can be deployed in a wind power grid-connected power system with multiple densely distributed wind power stations, and meets the requirements of wind power prediction and wind energy resource evaluation of multiple sites.

Description

Wind power prediction system and prediction method based on wind flow field space-time image learning
Technical Field
The invention belongs to the technical field of renewable energy development and utilization, and particularly relates to a wind power prediction system and a prediction method based on wind flow field spatiotemporal image learning.
Background
Over the last two decades, wind energy has become an important renewable resource that alleviates the global energy crisis and climate degradation problems. According to the report of the Global Wind Energy Council (GWEC), by the end of 2019, the global accumulated wind power installed capacity exceeds 651GW, the newly added installed capacity reaches 60.4GW, and the increase is 19%.
Because wind energy has strong randomness and uncertainty and shows obvious peak counter-regulation characteristics, along with the continuous increase of wind power permeability in a power grid, the development of a high-precision wind power prediction method has more and more important significance, and particularly, the method aims to meet the requirements of power system operation scheduling and energy market planning so as to ensure the long-term safe, stable and economic operation of a power system.
For wind power prediction of multiple fans and a multi-site wind power plant, a large number of research results show that the consideration of time-space correlation is favorable for remarkably improving prediction precision.
At present, there are two typical methods for spatio-temporal correlation research, namely a dependency-based method and a probability-based method. The method based on the dependency predicts the power of a certain station by measuring the dependency between stations and taking the wind speed and power information of all related stations as input; in contrast, the probability-based method does not need to directly calculate the time-space correlation between the wind turbine and the site, but considers the historical distribution characteristics of wind and the power generation characteristics of the wind turbine to establish a probability model with implicit correlation. These spatio-temporal prediction methods have two general difficulties: on one hand, the physical modeling calculation of the time-space correlation is complex, and some topographic information is difficult to obtain; on the other hand, using a machine learning model can simplify the modeling process, but how to construct the model structure and input data is a difficult research point of machine learning modeling.
Disclosure of Invention
The purpose of the invention is as follows: the invention aims to provide a wind power prediction system based on wind flow field space-time image learning, which is based on deep learning and convolutional neural network principles and can adaptively extract implicit characteristics of wind flow field images to realize wind power space-time prediction; the invention also discloses a prediction method of the wind power grid-connected power system, and the prediction method can be deployed in a wind power grid-connected power system with a plurality of densely distributed wind power stations and meets the wind power prediction requirements of multiple sites.
The technical scheme is as follows: in order to achieve the purpose, the invention provides the following technical scheme:
a wind power prediction system based on wind flow field space-time image learning comprises a wind power conversion calculation module and a wind power prediction calculation module; the wind power conversion calculation module is based on a convolution conversion model, and utilizes a wind flow field space-time image as input to extract a hidden layer characteristic diagram and calculate a wind power output value of 1-4 hours at present; the wind power prediction calculation module is used for predicting a wind power prediction result which is advanced by 1 hour in the future by combining the current wind power output value of 1-4 hours based on a convolution prediction model; the wind power conversion calculation module and the wind power prediction calculation module are used as input data of the wind power prediction calculation model through a hidden layer characteristic diagram of the wind power conversion calculation module so as to realize the connection of the wind power conversion calculation module and the wind power prediction calculation module; based on the above, the wind power conversion calculation module inputs the hidden layer characteristic diagram extracted in the process of converting the wind power into the wind power prediction calculation module; the wind power prediction calculation module uses the hidden layer characteristic diagram as input to realize the function of predicting wind power 1 hour in advance.
The prediction method of the wind power prediction system based on the wind flow field space-time image learning comprises the following steps:
step 1: recording longitude and latitude position information of a wind power station site by taking a longitude and latitude resolution of 1/500 degrees as a pixel point, acquiring current wind speed and direction data of each wind power station within 1-4 hours, respectively representing wind speed and wind direction by using the length and direction of a wind vector, drawing a spatial image of a wind flow field, and forming a wind flow field space-time image by using the 4-hour spatial image;
step 2: based on a convolutional neural network principle, taking a single-hour wind flow field space image as input and real-time wind power as output, constructing a convolutional conversion model, and training the convolutional conversion model by using a conversion error as a supervised learning index;
and step 3: selecting a hidden layer feature graph of a current 1-4 hour convolution conversion model as input, taking wind power to be predicted 1 hour in advance as output, constructing a convolution prediction model based on a double-layer circulation channel and attention connection mechanism principle, and using a prediction error as a supervised learning index;
and 4, step 4: the method comprises the steps of combining a convolution conversion model and a convolution prediction model based on a weight sharing strategy, setting the weight ratio of a conversion error to a prediction error to be 0.1:1, and training a convolution conversion-convolution prediction combination model;
and 5: based on the time-space image input of the wind flow field at any moment, the wind power prediction of different wind power stations is realized by using a convolution prediction model, and the accuracy of a prediction error check model is calculated.
Further, the step 2 of constructing the convolution conversion model includes the following implementation steps:
step 2.1: the method takes a space-time image of a wind flow field as input and real-time wind power as output, and the calculation formula is as follows:
Figure BDA0002532599430000031
in the formula (I), the compound is shown in the specification,
Figure BDA0002532599430000032
representing the real-time wind power conversion power output, X, of the convolution conversion model at time t0(t) and v (t) representing the wind flow field spatial image and the wind speed value at time t, fCCalculating a function for convolution of a convolution transformation model, PmaxIs the maximum output power, v, of the wind power plantcut-in、vcut-outAnd vratedAnd the cut-in wind speed, the cut-out wind speed and the rated wind speed of the wind turbine are shown.
Step 2.2: the conversion error is used as a supervised learning index of a convolution conversion model to realize model training, and the calculation formula is as follows:
Figure BDA0002532599430000033
in the formula, JCFor conversion errors, Pi(t) and
Figure BDA0002532599430000034
respectively representing the actual wind power and the converted wind power of the ith sample at the moment tPower; n issThe number of samples depends on the total amount of historical data collected during the operation of the wind power plant.
Further, in step 3, a convolution prediction model is constructed based on a double-layer circulation channel and an attention connection mechanism principle, and the method comprises the following implementation steps:
step 3.1: selecting a hidden layer characteristic diagram of a current 1-4 hour convolution conversion model as input, taking wind power to be predicted 1 hour in advance as output, and building a convolution long-short term memory (CLSTM) circulation channel, wherein the calculation formula is as follows:
Figure BDA0002532599430000035
in the formula, Fl(t) is the hidden layer feature map of the convolution conversion model at time t,
Figure BDA0002532599430000036
and Ht–1 CLSTMFor the output of the CLSTM loop path at times t and t-1, It、Gt、OtAnd CtAn input gate unit, a forgetting gate unit, an output gate unit and a memory unit which are respectively a CLSTM circulation channel at the time t, Ct–1Memory cell at time t-1, αhTo activate a function, Wxi、Whi、WciAnd BiIs ItInput weight, hidden layer weight, memory weight and offset value, Wxg、Whg、WcgAnd BgIs GtInput weight, hidden layer weight, memory weight and offset value, Wxo、Who、WcoAnd BoIs OtInput weight, hidden layer weight, memory weight and offset value, Wxc、WhcAnd BcIs CtThe input weight, the hidden layer weight and the bias value,
Figure BDA00025325994300000412
And tanh represents convolution, elemental product and tangent operations;
step 3.2: similarly, a hidden layer characteristic diagram of a current 1-4 hour convolution conversion model is selected as input, wind power to be predicted is selected as output 1 hour ahead of time, a pure convolution circulation channel is built, and the calculation formula is as follows:
Figure BDA0002532599430000041
in the formula, BtAnd VtRepresenting the offset value and convolution kernel at time t,
Figure BDA0002532599430000042
and Ht–1 CThe outputs of the pure convolution cyclic channels at the time t and t-1 respectively, | | | is the channel series operation [ ·]jRepresents the j channel after the series connection;
step 3.3: connecting each output in the CLSTM cyclic channel and the pure convolution cyclic channel based on an attention connection mechanism, wherein the calculation formula is as follows:
Figure BDA0002532599430000043
in the formula, betaiDenotes the ith attention weight, fsmThe function is calculated for the softmax and,
Figure BDA0002532599430000044
and
Figure BDA0002532599430000045
respectively representing the ith output in a pure convolution cycle channel and a CLSTM cycle channel, and e is exponential operation;
step 3.4: based on the attention weight between each output of the double-layer circulation channel, the output value of the pure convolution circulation channel is corrected, and the calculation formula is as follows:
Figure BDA0002532599430000046
in the formula (I), the compound is shown in the specification,
Figure BDA0002532599430000047
representing the output of the modified pure convolution cyclic channel at the time t;
step 3.5: and combining output values of the double-layer circulation channels, and using the prediction error as a supervised learning index, wherein the calculation formula is as follows:
Figure BDA0002532599430000048
Figure BDA0002532599430000049
in the formula (I), the compound is shown in the specification,
Figure BDA00025325994300000410
for wind power prediction values k steps ahead, Pi(t + k) and
Figure BDA00025325994300000411
respectively the actual wind power value and the predicted wind power value f in the ith samplePAnd JPThe convolution prediction model and the prediction error are respectively.
Further, in step 4, the combination of the convolution conversion model and the convolution prediction model is realized based on the weight sharing strategy, which includes the following steps:
step 4.1: aiming at the time t, extracting a hidden layer feature map of a convolution conversion model, wherein the calculation formula is as follows:
Figure BDA0002532599430000051
in the formula, Fl(t) hidden layer feature map of convolution conversion model at time t, { BlAnd { V }lIs the bias and convolution kernel parameter set in the convolution conversion model, ncNumber of layers of convolution, X0(t) is a wind flow field space-time image at time t;
step 4.2: establishing 4 convolution conversion models, respectively converting the current wind flow field space-time image of 1-4 hours into a hidden layer characteristic diagram as the input of a convolution prediction model, wherein the calculation formula is as follows:
Figure BDA0002532599430000052
in the formula (I), the compound is shown in the specification,
Figure BDA0002532599430000053
for the predicted value of wind power k steps ahead, fPFor the convolution prediction model, Fl(t) to Fl(t-3) inputting a hidden layer characteristic diagram in 1-4 hours at present;
step 4.3: sharing bias and convolution kernel parameter sets with 4 convolution conversion models to realize model combination based on weight sharing strategy, wherein the calculation formula is as follows:
Figure BDA0002532599430000054
in the formula, X0(t) to X0(t-3) is a current wind flow field space-time image for 1-4 hours;
step 4.4: setting the weighting ratio of the conversion error to the prediction error to be 0.1:1, training a convolution conversion-convolution prediction combination model, wherein the error formula is as follows:
Figure BDA0002532599430000055
in the formula, Jcombine、JCAnd JPRespectively representing weighted combining error, conversion error and prediction error, Pi(t) and Pi(t + k) represents the actual wind power of the ith sample at times t and t + k,
Figure BDA0002532599430000056
for the conversion of the wind power at the instant t,
Figure BDA0002532599430000057
extracting k steps of wind power prediction power, n, for time tsIs the number of samples.
Further, in step 5, wind power prediction and prediction error calculation of different wind power plant sites are realized by using a convolution prediction model, which relates to all the plant sites in the wind flow field image, and three error evaluation indexes, namely, Root Mean Square Error (RMSE), Mean Absolute Error (MAE) and Mean Absolute Percentage Error (MAPE), are used, and the formula is as follows:
Figure BDA0002532599430000058
Figure BDA0002532599430000061
Figure BDA0002532599430000062
wherein, RMSEk、MAEkAnd MAPEkRespectively the root mean square error, the average absolute error and the average absolute percentage error of the kth station address,
Figure BDA0002532599430000063
and
Figure BDA0002532599430000064
the predicted value and the true value of the wind power of the ith sample representing the kth station site, nsIs the number of samples.
Has the advantages that: compared with the prior art, the wind power prediction system based on the wind flow field space-time image learning is based on module connection, and the wind power conversion calculation module inputs the hidden layer characteristic diagram extracted in the process of converting the wind power into the wind power prediction calculation module; the wind power prediction calculation module takes the hidden layer characteristic diagram as input to realize the function of predicting wind power 1 hour in advance; the invention can directly learn the historical wind flow field image so as to avoid complex space-time correlation physical modeling; according to the method, only the geographical position information of the wind power station, the wind speed and wind direction information of the position of the wind power station are considered, historical wind power generation data and topographic data do not need to be input in a prediction stage, and data requirements and calculation requirements of a system deployment and application stage are simplified; due to the improvement of the prediction precision, the prediction result can more effectively guide the resource evaluation of the wind power station and the operation scheduling of the power system, and the operation stability and reliability of the wind power grid-connected power system are improved; meanwhile, the prediction method and the prediction system can adapt to the prediction application scene of the densely distributed wind power station and meet the wind power prediction and wind energy resource evaluation requirements of multiple sites.
Drawings
FIG. 1 is a schematic diagram of a wind power prediction process for the method and system of the present invention;
FIG. 2 is a schematic diagram of a convolution transformation model according to the method and system of the present invention;
FIG. 3 is a schematic diagram of a convolution prediction model according to the method and system of the present invention;
FIG. 4 is a schematic diagram of a wind power prediction result calculated by the method and system of the present invention.
Detailed Description
The invention will be further described with reference to the following drawings and specific embodiments. The above is only a preferred embodiment of the present invention, and it should be noted that: it will be apparent to those skilled in the art that the location of each facility can be adjusted without departing from the principles of the invention, and such adjustments should be considered within the scope of the invention.
As shown in fig. 1, a wind power prediction system based on wind flow field spatiotemporal image learning includes a wind power conversion calculation module and a wind power prediction calculation module. The wind power conversion calculation module is based on a convolution conversion model shown in FIG. 2, and extracts a hidden layer characteristic diagram and calculates a wind power output value of 1-4 hours at present by taking a wind flow field space-time image as input; the wind power prediction calculation module is used for predicting a wind power prediction result which is advanced by 1 hour in the future by combining the current wind power output value of 1-4 hours based on a convolution prediction model shown in FIG. 3. And the hidden layer characteristic diagram of the wind power conversion calculation module is used as input data of the wind power prediction calculation model so as to realize the connection of the two modules. Based on module connection, the wind power conversion calculation module inputs the hidden layer characteristic diagram extracted in the process of converting wind power into the wind power prediction calculation module; the wind power prediction calculation module uses the hidden layer characteristic diagram as input to realize the function of predicting wind power 1 hour in advance.
A prediction method of a wind power prediction system based on wind flow field space-time image learning comprises the following steps:
step 1: recording longitude and latitude position information of a wind power station site by taking a longitude and latitude resolution of 1/500 degrees as a pixel point, acquiring current wind speed and direction data of each wind power station within 1-4 hours, respectively representing wind speed and wind direction by using the length and direction of a wind vector, drawing a spatial image of a wind flow field, and forming a wind flow field space-time image by using the 4-hour spatial image;
step 2: based on a convolutional neural network principle, taking a single-hour wind flow field space image as input and real-time wind power as output, constructing a convolutional conversion model, and training the convolutional conversion model by using a conversion error as a supervised learning index;
and step 3: selecting a hidden layer feature graph of a current 1-4 hour convolution conversion model as input, taking wind power to be predicted 1 hour in advance as output, constructing a convolution prediction model based on a double-layer circulation channel and attention connection mechanism principle, and using a prediction error as a supervised learning index;
and 4, step 4: the method comprises the steps of combining a convolution conversion model and a convolution prediction model based on a weight sharing strategy, setting the weight ratio of a conversion error to a prediction error to be 0.1:1, and training a convolution conversion-convolution prediction combination model;
and 5: based on the time-space image input of the wind flow field at any moment, the wind power prediction of different wind power stations is realized by using a convolution prediction model, and the accuracy of a prediction error check model is calculated.
Further, the step 2 of constructing a convolution conversion model includes the following implementation steps:
step 2.1: the method takes a space-time image of a wind flow field as input and real-time wind power as output, and the calculation formula is as follows:
Figure BDA0002532599430000071
in the formula (I), the compound is shown in the specification,
Figure BDA0002532599430000072
representing the real-time wind power conversion power output, X, of the convolution conversion model at time t0(t) and v (t) representing the wind flow field spatial image and the wind speed value at time t, fCCalculating a function for convolution of a convolution transformation model, PmaxIs the maximum output power, v, of the wind power plantcut-in、vcut-outAnd vratedAnd the cut-in wind speed, the cut-out wind speed and the rated wind speed of the wind turbine are shown.
Step 2.2: the conversion error is used as a supervised learning index of a convolution conversion model to realize model training, and the calculation formula is as follows:
Figure BDA0002532599430000081
in the formula, JCFor conversion errors, Pi(t) and
Figure BDA0002532599430000082
respectively representing the actual wind power and the converted wind power of the ith sample at the time t; n issThe number of samples depends on the total amount of historical data collected during the operation of the wind power plant.
Further, in step 3, a convolution prediction model is constructed based on a double-layer circulation channel and an attention connection mechanism principle, and the method comprises the following implementation steps:
step 3.1: selecting a hidden layer characteristic diagram of a current 1-4 hour convolution conversion model as input, taking wind power to be predicted 1 hour in advance as output, and building a convolution long-short term memory (CLSTM) circulation channel, wherein the calculation formula is as follows:
Figure BDA0002532599430000083
in the formula, Fl(t) is the hidden layer feature map of the convolution conversion model at time t,
Figure BDA0002532599430000084
and Ht–1 CLSTMFor the output of the CLSTM loop path at times t and t-1, It、Gt、OtAnd CtAn input gate unit, a forgetting gate unit, an output gate unit and a memory unit which are respectively a CLSTM circulation channel at the time t, Ct–1Memory cell at time t-1, αhTo activate a function, Wxi、Whi、WciAnd BiIs ItInput weight, hidden layer weight, memory weight and offset value, Wxg、Whg、WcgAnd BgIs GtInput weight, hidden layer weight, memory weight and offset value, Wxo、Who、WcoAnd BoIs OtInput weight, hidden layer weight, memory weight and offset value, Wxc、WhcAnd BcIs CtThe input weight, the hidden layer weight and the bias value,
Figure BDA0002532599430000085
And tanh represents convolution, elemental product and tangent operations;
step 3.2: similarly, a hidden layer characteristic diagram of a current 1-4 hour convolution conversion model is selected as input, wind power to be predicted is selected as output 1 hour ahead of time, a pure convolution circulation channel is built, and the calculation formula is as follows:
Figure BDA0002532599430000091
in the formula, BtAnd VtRepresenting the offset value and convolution kernel at time t,
Figure BDA0002532599430000092
and Ht–1 CThe outputs of the pure convolution cyclic channels at the time t and t-1 respectively, | | | is the channel series operation [ ·]jRepresents the j channel after the series connection;
step 3.3: connecting each output in the CLSTM cyclic channel and the pure convolution cyclic channel based on an attention connection mechanism, wherein the calculation formula is as follows:
Figure BDA0002532599430000093
in the formula, betaiDenotes the ith attention weight, fsmThe function is calculated for the softmax and,
Figure BDA0002532599430000094
and Hi CLSTMRespectively representing the ith output in a pure convolution cycle channel and a CLSTM cycle channel, and e is exponential operation;
step 3.4: based on the attention weight between each output of the double-layer circulation channel, the output value of the pure convolution circulation channel is corrected, and the calculation formula is as follows:
Figure BDA0002532599430000095
in the formula (I), the compound is shown in the specification,
Figure BDA0002532599430000096
representing the output of the modified pure convolution cyclic channel at the time t;
step 3.5: and combining output values of the double-layer circulation channels, and using the prediction error as a supervised learning index, wherein the calculation formula is as follows:
Figure BDA0002532599430000097
Figure BDA0002532599430000098
in the formula (I), the compound is shown in the specification,
Figure BDA0002532599430000099
for wind power prediction values k steps ahead, Pi(t + k) and
Figure BDA00025325994300000910
respectively the actual wind power value and the predicted wind power value f in the ith samplePAnd JPThe convolution prediction model and the prediction error are respectively.
Further, in step 4, the combination of the convolution conversion model and the convolution prediction model is realized based on the weight sharing strategy, which includes the following steps:
step 4.1: aiming at the time t, extracting a hidden layer feature map of a convolution conversion model, wherein the calculation formula is as follows:
Figure BDA00025325994300000911
in the formula, Fl(t) hidden layer feature map of convolution conversion model at time t, { BlAnd { V }lIs the bias and convolution kernel parameter set in the convolution conversion model, ncNumber of layers of convolution, X0(t) is a wind flow field space-time image at time t;
step 4.2: establishing 4 convolution conversion models, respectively converting the current wind flow field space-time image of 1-4 hours into a hidden layer characteristic diagram as the input of a convolution prediction model, wherein the calculation formula is as follows:
Figure BDA0002532599430000101
in the formula (I), the compound is shown in the specification,
Figure BDA0002532599430000102
for the predicted value of wind power k steps ahead, fPFor the convolution prediction model, Fl(t) to Fl(t-3) inputting a hidden layer characteristic diagram in 1-4 hours at present;
step 4.3: sharing bias and convolution kernel parameter sets with 4 convolution conversion models to realize model combination based on weight sharing strategy, wherein the calculation formula is as follows:
Figure BDA0002532599430000103
in the formula, X0(t) to X0(t-3) is a current wind flow field space-time image for 1-4 hours;
step 4.4: setting the weighting ratio of the conversion error to the prediction error to be 0.1:1, training a convolution conversion-convolution prediction combination model, wherein the error formula is as follows:
Figure BDA0002532599430000104
in the formula, Jcombine、JCAnd JPRespectively representing weighted combining error, conversion error and prediction error, Pi(t) and Pi(t + k) represents the actual wind power of the ith sample at times t and t + k,
Figure BDA0002532599430000105
for the conversion of the wind power at the instant t,
Figure BDA0002532599430000106
extracting k steps of wind power prediction power, n, for time tsIs the number of samples.
Further, in step 5, wind power prediction and prediction error calculation of different wind power plant sites are realized by using a convolution prediction model, which relates to all the plant sites in the wind flow field image, and three error evaluation indexes, namely, Root Mean Square Error (RMSE), Mean Absolute Error (MAE) and Mean Absolute Percentage Error (MAPE), are used, and the formula is as follows:
Figure BDA0002532599430000107
Figure BDA0002532599430000108
Figure BDA0002532599430000109
wherein, RMSEk、MAEkAnd MAPEkRespectively the root mean square error, the average absolute error and the average absolute percentage error of the kth station address,
Figure BDA0002532599430000111
and
Figure BDA0002532599430000112
the predicted value and the true value of the wind power of the ith sample representing the kth station site, nsIs the number of samples.
Recording longitude and latitude position information of a wind power station site by taking a longitude and latitude resolution of 1/500 degrees as a pixel point, acquiring current wind speed and direction data of each wind power station within 1-4 hours, respectively representing wind speed and wind direction by using the length and direction of a wind vector, drawing a spatial image of a wind flow field, and forming a wind flow field space-time image by using the 4-hour spatial image; based on a convolutional neural network principle, taking a single-hour wind flow field space image as input and real-time wind power as output, constructing a convolutional conversion model, and training the convolutional conversion model by using a conversion error as a supervised learning index; selecting a hidden layer feature graph of a current 1-4 hour convolution conversion model as input, taking wind power to be predicted 1 hour in advance as output, constructing a convolution prediction model based on a double-layer circulation channel and attention connection mechanism principle, and using a prediction error as a supervised learning index; the method comprises the steps of combining a convolution conversion model and a convolution prediction model based on a weight sharing strategy, setting the weight ratio of a conversion error to a prediction error to be 0.1:1, and training a convolution conversion-convolution prediction combination model; based on the time-space image input of the wind flow field at any moment, the wind power prediction of different wind power stations is realized by using a convolution prediction model, and the accuracy of a prediction error check model is calculated.
The following describes in detail a specific implementation process of wind power prediction by using the method and system of the present invention with reference to specific embodiments. The method selects wind power data of 2007 to 2012 of 13 wind power stations in Mona of America, spans between 112.24 degrees and 112.08 degrees of west longitude and between 48.63 degrees N and 48.79 degrees N of north latitude, and has a maximum power output range of 12MW to 16 MW. Based on the data set, the method of the invention is implemented by the following steps:
1) according to the geographical range of the wind power station, longitude and latitude resolution 1/500 degrees are used as a pixel point, a range from 112.24 degrees to 112.08 degrees in west longitude and a range from 48.63 degrees N to 48.79 degrees N in north latitude are drawn into 80 x 80 wind flow field space images, and the current space images in 1-4 hours form wind flow field space-time images.
2) A convolution transformation model was constructed as shown in fig. 2. Firstly, a wind flow field space-time image is taken as input, real-time wind power is taken as output, and the calculation formula is as follows:
Figure BDA0002532599430000113
in the formula (I), the compound is shown in the specification,
Figure BDA0002532599430000114
representing the real-time wind power conversion power output (unit MW), X of the convolution conversion model at the time t0(t) and v (t) representing the wind flow field spatial image and the wind speed value at time t, fCCalculating a function for convolution of a convolution transformation model, PmaxIs the maximum output power (in MW), v, of the wind power plantcut-in、vcut-outAnd vratedAnd the cut-in wind speed, the cut-out wind speed and the rated wind speed of the wind turbine are shown. Secondly, the conversion error is used as a supervised learning index of the convolution conversion model to realize model training, and the calculation formula is as follows:
Figure BDA0002532599430000121
in the formula, JCFor conversion errors, Pi(t) and
Figure BDA0002532599430000122
respectively representing the actual wind power and the converted wind power (unit MW) of the ith sample at the moment t, nsIs the number of samples.
3) Based on the principle of a double-layer circulation channel and an attention connection mechanism, a convolution prediction model is constructed, as shown in fig. 3. Firstly, selecting a hidden layer characteristic diagram of a current 1-4 hour convolution conversion model as input, taking wind power to be predicted 1 hour in advance as output, and building a convolution long-short term memory (CLSTM) circulation channel, wherein the calculation formula is as follows:
Figure BDA0002532599430000123
in the formula, Fl(t) is the hidden layer feature map of the convolution conversion model at time t,
Figure BDA0002532599430000124
and Ht–1 CLSTMFor the output of the CLSTM loop path at times t and t-1, It、Gt、OtAnd CtAn input gate unit, a forgetting gate unit, an output gate unit and a memory unit which are respectively a CLSTM circulation channel at the time t, Ct–1Memory cell at time t-1, αhTo activate a function, Wxi、Whi、WciAnd BiIs ItInput weight, hidden layer weight, memory weight and offset value, Wxg、Whg、WcgAnd BgIs GtInput weight, hidden layer weight, memory weight and offset value, Wxo、Who、WcoAnd BoIs OtInput weight, hidden layer weight, memory weight and offset value, Wxc、WhcAnd BcIs CtInput weight, hidden layer weight and biasSetting, setting,
Figure BDA0002532599430000127
And tanh represents convolution, elemental product and tangent operations. Secondly, a hidden layer characteristic diagram of the current 1-4 hour convolution conversion model is selected as input, wind power to be predicted is selected as output 1 hour ahead of time, a pure convolution circulation channel is built, and the calculation formula is as follows:
Figure BDA0002532599430000125
in the formula, BtAnd VtRepresenting the offset value and convolution kernel at time t,
Figure BDA0002532599430000126
and Ht–1 CThe outputs of the pure convolution cyclic channels at the time t and t-1 respectively, | | | is the channel series operation [ ·]jRepresenting the jth channel after the series. Then, each output in the CLSTM cyclic channel and the pure convolution cyclic channel is connected based on an attention connection mechanism, and the calculation formula is as follows:
Figure BDA0002532599430000131
in the formula, betaiDenotes the ith attention weight, fsmThe function is calculated for the softmax and,
Figure BDA0002532599430000132
and Hi CLSTMRespectively representing the ith output in a pure convolution cycle channel and a CLSTM cycle channel, and e is an exponential operation. Then, based on the attention weight between each output of the double-layer circulation channel, the output value of the pure convolution circulation channel is corrected, and the calculation formula is as follows:
Figure BDA0002532599430000133
in the formula，
Figure BDA0002532599430000134
Representing the output of the pure convolution cyclic channel at time t after correction. And finally, combining output values of the double-layer circulation channels, and using the prediction error as a supervised learning index, wherein the calculation formula is as follows:
Figure BDA0002532599430000135
Figure BDA0002532599430000136
in the formula (I), the compound is shown in the specification,
Figure BDA0002532599430000137
for wind power prediction values k steps ahead, Pi(t + k) and
Figure BDA0002532599430000138
respectively the actual wind power value and the predicted wind power value f in the ith samplePAnd JPThe convolution prediction model and the prediction error are respectively.
4) Firstly, aiming at a time t, extracting a hidden layer feature map of the convolution conversion model, wherein a calculation formula is as follows:
Figure BDA0002532599430000139
in the formula, Fl(t) hidden layer feature map of convolution conversion model at time t, { BlAnd { V }lIs the bias and convolution kernel parameter set in the convolution conversion model, ncNumber of layers of convolution, X0And (t) is a space-time image of the wind flow field at the time t. Secondly, establishing 4 convolution conversion models, respectively converting the current 1-4 hour wind flow field space-time image into a hidden layer characteristic diagram serving as the input of a convolution prediction model, and calculatingThe formula is as follows:
Figure BDA00025325994300001310
in the formula (I), the compound is shown in the specification,
Figure BDA00025325994300001311
for the predicted value of wind power k steps ahead, fPFor the convolution prediction model, Fl(t) to FlAnd (t-3) inputting the hidden layer characteristic diagram in 1-4 hours at present. Then, sharing bias and convolution kernel parameter sets with 4 convolution conversion models to realize model combination based on weight sharing strategy, wherein the calculation formula is as follows:
Figure BDA0002532599430000141
in the formula, X0(t) to X0And (t-3) is a current wind flow field space-time image for 1-4 hours. And finally, setting the weighting ratio of the conversion error to the prediction error to be 0.1:1, training a convolution conversion-convolution prediction combination model, wherein the error formula is as follows:
Figure BDA0002532599430000142
in the formula, Jcombine、JCAnd JPRespectively representing weighted combining error, conversion error and prediction error, Pi(t) and Pi(t + k) represents the actual wind power (in MW) for the ith sample at times t and t + k,
Figure BDA0002532599430000143
for the converted wind power at time t (in MW),
Figure BDA0002532599430000144
extracting predicted power (unit MW) of k steps of wind power for time t, nsIs the number of samples.
5) Wind power prediction and prediction error calculation of different wind power station sites are realized by utilizing a convolution prediction model, which relates to all the station sites in a wind flow field image, and three error evaluation indexes, namely Root Mean Square Error (RMSE), Mean Absolute Error (MAE) and Mean Absolute Percentage Error (MAPE), are used, and the formula is as follows:
Figure BDA0002532599430000145
Figure BDA0002532599430000146
Figure BDA0002532599430000147
wherein, RMSEk、MAEkAnd MAPEkRespectively the root mean square error, the average absolute error and the average absolute percentage error of the kth station address,
Figure BDA0002532599430000148
and
Figure BDA0002532599430000149
the predicted value and the true value (unit MW), n of the wind power of the ith sample representing the kth station sitesIs the number of samples. For the tested 13 wind power station sites, the wind power prediction error results calculated by using the system of the invention are shown in table 1. In addition, in order to visually display the prediction error, the station site No. 1 is selected to draw a wind power prediction curve thereof, as shown in fig. 4. As can be seen from the error results of Table 1 and FIG. 4, the method of the present invention can accurately predict the wind power of all 13 tested wind power stations, the total average MAE value is less than 2MW, and the method has high wind power prediction accuracy and reliability.
Wind power prediction error of 13 wind power plant sites tested in table 1
Figure BDA0002532599430000151
In conclusion, the wind power prediction method and the system can directly learn the time-space images of the wind flow field, can avoid complex time-space correlation physical modeling, do not need to input historical wind power generation data and topographic data in the prediction stage, and reduce the calculation requirement in the practical engineering application; the prediction method and the prediction system have higher prediction accuracy, can adapt to the prediction application scene of densely distributed wind power stations, meet the wind power prediction and wind energy resource evaluation requirements of multiple sites, and further improve the safe and stable operation of the wind power grid-connected power system.

Claims (5)

1. A wind power prediction method based on wind flow field space-time image learning is characterized by comprising the following steps: the method is based on a wind flow field space-time image learning wind power prediction system, and the system comprises a wind power conversion calculation module and a wind power prediction calculation module; the wind power conversion calculation module is based on a convolution conversion model, and utilizes a wind flow field space-time image as input to extract a hidden layer characteristic diagram and calculate a wind power output value of 1-4 hours at present; the wind power prediction calculation module is used for predicting a wind power prediction result which is advanced by 1 hour in the future by combining the current wind power output value of 1-4 hours based on a convolution prediction model; the wind power conversion calculation module and the wind power prediction calculation module are used as input data of the wind power prediction calculation model through a hidden layer characteristic diagram of the wind power conversion calculation module so as to realize the connection of the wind power conversion calculation module and the wind power prediction calculation module; based on the above, the wind power conversion calculation module inputs the hidden layer characteristic diagram extracted in the process of converting the wind power into the wind power prediction calculation module; the wind power prediction calculation module takes the hidden layer characteristic diagram as input to realize the function of predicting wind power 1 hour in advance; the method comprises the following steps:
step 1: recording longitude and latitude position information of a wind power station site by taking a longitude and latitude resolution of 1/500 degrees as a pixel point, acquiring current wind speed and direction data of each wind power station within 1-4 hours, respectively representing wind speed and wind direction by using the length and direction of a wind vector, drawing a spatial image of a wind flow field, and forming a wind flow field space-time image by using the 4-hour spatial image;
step 2: based on a convolutional neural network principle, taking a single-hour wind flow field space image as input and real-time wind power as output, constructing a convolutional conversion model, and training the convolutional conversion model by using a conversion error as a supervised learning index;
and step 3: selecting a hidden layer feature graph of a current 1-4 hour convolution conversion model as input, taking wind power to be predicted 1 hour in advance as output, constructing a convolution prediction model based on a double-layer circulation channel and attention connection mechanism principle, and using a prediction error as a supervised learning index;
and 4, step 4: the method comprises the steps of combining a convolution conversion model and a convolution prediction model based on a weight sharing strategy, setting the weight ratio of a conversion error to a prediction error to be 0.1:1, and training a convolution conversion-convolution prediction combination model;
and 5: based on the time-space image input of the wind flow field at any moment, the wind power prediction of different wind power stations is realized by using a convolution prediction model, and the accuracy of a prediction error check model is calculated.
2. The wind power prediction method based on wind flow field spatiotemporal image learning of claim 1, characterized in that, the step 2 of constructing the convolution conversion model comprises the following implementation steps:
step 2.1: the method takes a space-time image of a wind flow field as input and real-time wind power as output, and the calculation formula is as follows:
Figure FDA0003172500530000021
in the formula (I), the compound is shown in the specification,
Figure FDA0003172500530000022
representing the real-time wind power conversion power output, X, of the convolution conversion model at time t0(t) and v (t) representing the wind flow field spatial image and the wind speed value at time t，fCCalculating a function for convolution of a convolution transformation model, PmaxIs the maximum output power, v, of the wind power plantcut-in、vcut-outAnd vratedRepresenting the cut-in wind speed, the cut-out wind speed and the rated wind speed of the wind turbine generator;
step 2.2: the conversion error is used as a supervised learning index of a convolution conversion model to realize model training, and the calculation formula is as follows:
Figure FDA0003172500530000023
in the formula, JCFor conversion errors, Pi(t) and
Figure FDA0003172500530000024
respectively representing the actual wind power and the converted wind power of the ith sample at the time t; n issThe number of samples depends on the total amount of historical data collected during the operation of the wind power plant.
3. The wind power prediction method based on wind flow field spatiotemporal image learning of claim 2 is characterized in that in step 3, a convolution prediction model is constructed based on a double-layer circulation channel and an attention connection mechanism principle, and the method comprises the following implementation steps:
step 3.1: selecting a hidden layer characteristic diagram of a current 1-4 hour convolution conversion model as input, taking wind power to be predicted 1 hour in advance as output, and building a convolution long-term and short-term memory circulation channel, wherein the calculation formula is as follows:
Figure FDA0003172500530000025
in the formula, Fl(t) hidden layer feature map of convolution transformation model at time t, Ht CLSTMAnd Ht–1 CLSTMFor the output of the CLSTM loop path at times t and t-1, It、Gt、OtAnd CtAn input gate unit, a forgetting gate unit, an output gate unit and a memory unit which are respectively a CLSTM circulation channel at the time t, Ct–1Memory cell at time t-1, αhTo activate a function, Wxi、Whi、WciAnd BiIs ItInput weight, hidden layer weight, memory weight and offset value, Wxg、Whg、WcgAnd BgIs GtInput weight, hidden layer weight, memory weight and offset value, Wxo、Who、WcoAnd BoIs OtInput weight, hidden layer weight, memory weight and offset value, Wxc、WhcAnd BcIs CtThe input weight, the hidden layer weight and the bias value,
Figure FDA0003172500530000039
And tanh represents convolution, elemental product and tangent operations;
step 3.2: similarly, a hidden layer characteristic diagram of a current 1-4 hour convolution conversion model is selected as input, wind power to be predicted is selected as output 1 hour ahead of time, a pure convolution circulation channel is built, and the calculation formula is as follows:
Figure FDA0003172500530000031
in the formula, BtAnd VtRepresenting the offset and convolution kernel at time t, Ht CAnd Ht–1 CThe outputs of the pure convolution cyclic channels at the time t and t-1 respectively, | | | is the channel series operation [ ·]jRepresents the j channel after the series connection;
step 3.3: connecting each output in the CLSTM cyclic channel and the pure convolution cyclic channel based on an attention connection mechanism, wherein the calculation formula is as follows:
Figure FDA0003172500530000032
in the formula, betaiDenotes the ith attention weight, fsmCalculating a function for softmax, Hi CAnd Hi CLSTMRespectively representing the ith output in a pure convolution cycle channel and a CLSTM cycle channel, and e is exponential operation;
step 3.4: based on the attention weight between each output of the double-layer circulation channel, the output value of the pure convolution circulation channel is corrected, and the calculation formula is as follows:
Figure FDA0003172500530000033
in the formula (I), the compound is shown in the specification,
Figure FDA0003172500530000034
representing the output of the modified pure convolution cyclic channel at the time t;
step 3.5: and combining output values of the double-layer circulation channels, and using the prediction error as a supervised learning index, wherein the calculation formula is as follows:
Figure FDA0003172500530000035
Figure FDA0003172500530000036
in the formula (I), the compound is shown in the specification,
Figure FDA0003172500530000037
for wind power prediction values k steps ahead, Pi(t + k) and
Figure FDA0003172500530000038
respectively the actual wind power value and the predicted wind power value f in the ith samplePAnd JPThe convolution prediction model and the prediction error are respectively.
4. The wind power prediction method based on wind flow field spatiotemporal image learning of claim 3, characterized in that, in step 4, the combination of the convolution conversion model and the convolution prediction model is realized based on the weight sharing strategy, which comprises the following realization steps:
step 4.1: aiming at the time t, extracting a hidden layer feature map of a convolution conversion model, wherein the calculation formula is as follows:
Figure FDA0003172500530000041
in the formula, Fl(t) hidden layer feature map of convolution conversion model at time t, { BlAnd { V }lIs the bias and convolution kernel parameter set in the convolution conversion model, ncNumber of layers of convolution, X0(t) is a wind flow field space-time image at time t;
step 4.2: establishing 4 convolution conversion models, respectively converting the current wind flow field space-time image of 1-4 hours into a hidden layer characteristic diagram as the input of a convolution prediction model, wherein the calculation formula is as follows:
Figure FDA0003172500530000042
in the formula (I), the compound is shown in the specification,
Figure FDA0003172500530000043
for the predicted value of wind power k steps ahead, fPFor the convolution prediction model, Fl(t) to Fl(t-3) inputting a hidden layer characteristic diagram in 1-4 hours at present;
step 4.3: sharing bias and convolution kernel parameter sets with 4 convolution conversion models to realize model combination based on weight sharing strategy, wherein the calculation formula is as follows:
Figure FDA0003172500530000044
in the formula, X0(t) to X0(t-3) is a current wind flow field space-time image for 1-4 hours;
step 4.4: setting the weighting ratio of the conversion error to the prediction error to be 0.1:1, training a convolution conversion-convolution prediction combination model, wherein the error formula is as follows:
Figure FDA0003172500530000045
in the formula, Jcombine、JCAnd JPRespectively representing weighted combining error, conversion error and prediction error, Pi(t) and Pi(t + k) represents the actual wind power of the ith sample at times t and t + k,
Figure FDA0003172500530000046
for the conversion of the wind power at the instant t,
Figure FDA0003172500530000047
extracting k steps of wind power prediction power, n, for time tsIs the number of samples.
5. The wind power prediction method based on wind flow field spatio-temporal image learning of claim 4, characterized in that, step 5 utilizes a convolution prediction model to realize wind power prediction and calculation prediction errors of different wind power stations, which relate to all stations in the wind flow field image, and uses three error evaluation indexes, namely, root mean square error RMSE, mean absolute error MAE and mean absolute percentage error MAPE, with the formula:
Figure FDA0003172500530000048
Figure FDA0003172500530000051
Figure FDA0003172500530000052
wherein, RMSEk、MAEkAnd MAPEkRespectively the root mean square error, the average absolute error and the average absolute percentage error of the kth station address,
Figure FDA0003172500530000053
and Pi kThe predicted value and the true value of the wind power of the ith sample representing the kth station site, nsIs the number of samples.
CN202010522362.2A 2020-06-10 2020-06-10 Wind power prediction system and prediction method based on wind flow field space-time image learning Active CN111798034B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010522362.2A CN111798034B (en) 2020-06-10 2020-06-10 Wind power prediction system and prediction method based on wind flow field space-time image learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010522362.2A CN111798034B (en) 2020-06-10 2020-06-10 Wind power prediction system and prediction method based on wind flow field space-time image learning

Publications (2)

Publication Number Publication Date
CN111798034A CN111798034A (en) 2020-10-20
CN111798034B true CN111798034B (en) 2021-09-07

Family

ID=72804293

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010522362.2A Active CN111798034B (en) 2020-06-10 2020-06-10 Wind power prediction system and prediction method based on wind flow field space-time image learning

Country Status (1)

Country Link
CN (1) CN111798034B (en)

Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103268572A (en) * 2013-05-06 2013-08-28 国家电网公司 A micro-siting method of wind detecting network of ten-million-kilowatt-class large wind power base
CN104699936A (en) * 2014-08-18 2015-06-10 沈阳工业大学 Sector management method based on CFD short-term wind speed forecasting wind power plant
CN106845863A (en) * 2017-02-23 2017-06-13 沈阳工业大学 A kind of distributed wind-power generator is exerted oneself and heat load sync index Forecasting Methodology
US10133981B2 (en) * 2011-08-18 2018-11-20 Siemens Aktiengesellschaft Method for the computer-assisted modeling of a wind power installation or a photovoltaic installation with a feed forward neural network
CN109214566A (en) * 2018-08-30 2019-01-15 华北水利水电大学 Short-term wind power prediction method based on shot and long term memory network
CN110070226A (en) * 2019-04-24 2019-07-30 河海大学 Photovoltaic power prediction technique and system based on convolutional neural networks and meta learning
CN110443725A (en) * 2019-08-14 2019-11-12 国网甘肃省电力公司信息通信公司 Wind power prediction model based on CNN+LSTM
CN110942194A (en) * 2019-11-27 2020-03-31 徐州上若科技有限公司 Wind power prediction error interval evaluation method based on TCN

Family Cites Families (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109657839B (en) * 2018-11-22 2023-05-26 天津大学 Wind power prediction method based on deep convolutional neural network
CN110458344B (en) * 2019-07-26 2022-05-06 湘潭大学 Ultrashort-term wind power prediction method adaptive to time resolution

Patent Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US10133981B2 (en) * 2011-08-18 2018-11-20 Siemens Aktiengesellschaft Method for the computer-assisted modeling of a wind power installation or a photovoltaic installation with a feed forward neural network
CN103268572A (en) * 2013-05-06 2013-08-28 国家电网公司 A micro-siting method of wind detecting network of ten-million-kilowatt-class large wind power base
CN104699936A (en) * 2014-08-18 2015-06-10 沈阳工业大学 Sector management method based on CFD short-term wind speed forecasting wind power plant
CN106845863A (en) * 2017-02-23 2017-06-13 沈阳工业大学 A kind of distributed wind-power generator is exerted oneself and heat load sync index Forecasting Methodology
CN109214566A (en) * 2018-08-30 2019-01-15 华北水利水电大学 Short-term wind power prediction method based on shot and long term memory network
CN110070226A (en) * 2019-04-24 2019-07-30 河海大学 Photovoltaic power prediction technique and system based on convolutional neural networks and meta learning
CN110443725A (en) * 2019-08-14 2019-11-12 国网甘肃省电力公司信息通信公司 Wind power prediction model based on CNN+LSTM
CN110942194A (en) * 2019-11-27 2020-03-31 徐州上若科技有限公司 Wind power prediction error interval evaluation method based on TCN

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
一种结合CNN和GRU网络的超短期风电预测模型;薛阳;《可再生能源》;20190331;全文 *
深度学习在电力领域的研究现状与展望;曹渝昆;《上海电力学院学报》;20170831;全文 *

Also Published As

Publication number Publication date
CN111798034A (en) 2020-10-20

Similar Documents

Publication Publication Date Title
CN109902863B (en) Wind speed prediction method and device based on multi-factor time-space correlation
CN109524993B (en) Wind power photovoltaic typical cycle output scene generation method for medium-long term optimization scheduling
CN109657839B (en) Wind power prediction method based on deep convolutional neural network
CN106295899B (en) Wind power probability density Forecasting Methodology based on genetic algorithm Yu supporting vector quantile estimate
CN102479347B (en) Method and system for forecasting short-term wind speed of wind farm based on data driving
CN112215428B (en) Photovoltaic power generation power prediction method and system based on error correction and fuzzy logic
CN112101626A (en) Distributed photovoltaic power generation power prediction method and system
CN113984198B (en) Shortwave radiation prediction method and system based on convolutional neural network
CN111814398A (en) Map-based earth surface solar radiance prediction method integrating space-time attention
CN117060407B (en) Wind power cluster power prediction method and system based on similar day division
CN113991711A (en) Capacity configuration method for energy storage system of photovoltaic power station
CN116885691B (en) Wind power climbing event indirect prediction method
CN111798034B (en) Wind power prediction system and prediction method based on wind flow field space-time image learning
El Kounni et al. Solar power output forecasting using Artificial Neural Network
CN116128211A (en) Wind-light-water combined short-term optimization scheduling method based on wind-light uncertainty prediction scene
CN116187540A (en) Wind power station ultra-short-term power prediction method based on space-time deviation correction
CN115765044A (en) Wind, light and water power system combined operation and risk analysis method and system
CN105207255A (en) Electric system peak regulation calculation method suitable for wind power output
Shi et al. Short-term photovoltaic power forecast based on long short-term memory network
Ning et al. Short-term master-slave forecast method for distributed photovoltaic plants based on the spatial correlation
CN116896116B (en) Solar grid-connected regulation and control method and system based on artificial intelligence
Bei et al. Short-term power forecasting model of floating photovoltaic—A case study of Huainan floating photovoltaic plant, China
CN116488158B (en) Method and device for predicting medium-long-term electric quantity of wind power based on transfer learning
CN116468164A (en) Photovoltaic output prediction method based on solar radiance and copula combined distribution function
Chen et al. Short-term power prediction of distributed photovoltaic based on CNN-LSTM and meteorological interpolation

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