CN116957849A - Wind power prediction method based on double-attention mechanism - Google Patents
Wind power prediction method based on double-attention mechanism Download PDFInfo
- Publication number
- CN116957849A CN116957849A CN202310923863.5A CN202310923863A CN116957849A CN 116957849 A CN116957849 A CN 116957849A CN 202310923863 A CN202310923863 A CN 202310923863A CN 116957849 A CN116957849 A CN 116957849A
- Authority
- CN
- China
- Prior art keywords
- wind power
- neural network
- input
- time sequence
- wind
- 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.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 72
- 230000007246 mechanism Effects 0.000 title claims abstract description 26
- 238000013528 artificial neural network Methods 0.000 claims abstract description 76
- 238000012545 processing Methods 0.000 claims abstract description 8
- 238000010183 spectrum analysis Methods 0.000 claims abstract description 7
- 230000000694 effects Effects 0.000 claims abstract description 4
- 238000011156 evaluation Methods 0.000 claims abstract description 4
- 239000011159 matrix material Substances 0.000 claims description 62
- 238000012549 training Methods 0.000 claims description 31
- 238000010606 normalization Methods 0.000 claims description 19
- 102100024549 Tenascin-X Human genes 0.000 claims description 18
- 238000005070 sampling Methods 0.000 claims description 17
- 238000012795 verification Methods 0.000 claims description 16
- 239000013598 vector Substances 0.000 claims description 11
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 claims description 10
- 230000006870 function Effects 0.000 claims description 6
- 238000012886 linear function Methods 0.000 claims description 6
- 230000009467 reduction Effects 0.000 claims description 6
- 238000000354 decomposition reaction Methods 0.000 claims description 5
- 230000004927 fusion Effects 0.000 claims description 4
- 238000004422 calculation algorithm Methods 0.000 claims description 3
- 238000012544 monitoring process Methods 0.000 claims description 3
- 238000012935 Averaging Methods 0.000 claims description 2
- 238000011176 pooling Methods 0.000 claims description 2
- 238000007781 pre-processing Methods 0.000 abstract 1
- 238000013527 convolutional neural network Methods 0.000 description 3
- 238000007405 data analysis Methods 0.000 description 3
- 238000005516 engineering process Methods 0.000 description 3
- 238000007726 management method Methods 0.000 description 3
- 238000013473 artificial intelligence Methods 0.000 description 2
- 238000007637 random forest analysis Methods 0.000 description 2
- 238000012706 support-vector machine Methods 0.000 description 2
- 230000002411 adverse Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000007796 conventional method Methods 0.000 description 1
- 238000013480 data collection Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 238000013136 deep learning model Methods 0.000 description 1
- 238000009795 derivation Methods 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000003993 interaction Effects 0.000 description 1
- 238000012417 linear regression Methods 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000001537 neural effect Effects 0.000 description 1
- 230000000306 recurrent effect Effects 0.000 description 1
- 238000000611 regression analysis Methods 0.000 description 1
- 238000013179 statistical model Methods 0.000 description 1
- 238000012731 temporal analysis Methods 0.000 description 1
- 230000002123 temporal effect Effects 0.000 description 1
- 238000000700 time series analysis Methods 0.000 description 1
- 238000012876 topography Methods 0.000 description 1
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
- G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
- G06Q50/06—Energy or water supply
-
- 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
- G06N3/0464—Convolutional networks [CNN, ConvNet]
-
- 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
- G06Q10/00—Administration; Management
- G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
-
- 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
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
- Y02A30/00—Adapting or protecting infrastructure or their operation
Landscapes
- Engineering & Computer Science (AREA)
- Business, Economics & Management (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Health & Medical Sciences (AREA)
- Economics (AREA)
- General Physics & Mathematics (AREA)
- Strategic Management (AREA)
- General Health & Medical Sciences (AREA)
- Human Resources & Organizations (AREA)
- Biophysics (AREA)
- Data Mining & Analysis (AREA)
- Software Systems (AREA)
- General Business, Economics & Management (AREA)
- Marketing (AREA)
- Mathematical Physics (AREA)
- General Engineering & Computer Science (AREA)
- Computing Systems (AREA)
- Molecular Biology (AREA)
- Evolutionary Computation (AREA)
- Tourism & Hospitality (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Computational Linguistics (AREA)
- Public Health (AREA)
- Quality & Reliability (AREA)
- Operations Research (AREA)
- Entrepreneurship & Innovation (AREA)
- Game Theory and Decision Science (AREA)
- Development Economics (AREA)
- Water Supply & Treatment (AREA)
- Primary Health Care (AREA)
- Wind Motors (AREA)
Abstract
A wind power prediction method based on a dual-attention mechanism includes the steps of firstly, collecting historical wind speed, wind direction, wind power, temperature, humidity and atmospheric pressure of a wind power plant; then, preprocessing the historical wind speed, wind direction and wind power by adopting a method of combining singular spectrum analysis and a UNet neural network to remove noise; finally, the data of the historical wind speed, wind direction, temperature, humidity and atmospheric pressure are passed through a Seq2Seq neural network based on a dual-attention mechanism to obtain predicted wind power, and a prediction effect evaluation scheme is designed to enhance the reliability of a prediction result. Compared with the traditional method, the method is more excellent in processing nonlinearity, time characteristic correlation and the like, and solves the problem of data noise. The method can better excavate potential modes and correlations in wind energy data, improves the accuracy of power prediction, and provides important support for operation management and power grid dispatching of the wind power plant.
Description
Technical Field
The invention belongs to the field of energy data analysis, and particularly relates to a wind power prediction method based on a dual-attention mechanism.
Background
Wind energy is an important form of renewable energy and is widely used and developed worldwide. However, due to wind instability and difficulty in predictability, the power fluctuation of wind farms is large, which presents challenges to the stability and reliability of the power system. Thus, accurately predicting the power output of a wind farm is critical to power system operation and planning.
Traditional wind power prediction methods are generally based on statistical models or physical models, such as regression analysis, time series analysis, and the like. These methods often require large amounts of historical data and perform poorly in processing non-linear, non-stationary wind energy data. Moreover, these methods often fail to adequately account for correlations and dynamic changes between different features, resulting in limited accuracy of the predicted results. Specifically, the conventional wind power prediction has some technical problems, including but not limited to the following aspects:
(1) Instability of wind power: wind energy is affected by various factors such as geographical location, topography, seasons, climate, etc., resulting in high temporal and spatial variability of wind speed and wind direction. This makes wind power prediction subject to wind instability, and conventional methods have difficulty in accurately predicting the power output of a wind farm.
(2) Nonlinearity and non-stationarity: wind energy data is often characterized by non-linearities and non-stationarity. Traditional linear regression and time series methods often fail to capture nonlinear relationships and dynamic changes in wind energy data, resulting in limited accuracy of the predicted results.
(3) Utilization of feature association and timing information: wind power is affected by various meteorological factors such as wind speed, wind direction, temperature, etc. There is a complex association between these features and there is a timing dependency. The efficient use of the correlation and timing information of these features to accurately predict wind power is a challenging problem.
(4) Expression ability of predictive model: traditional predictive models have certain limitations in capturing complex nonlinear relationships and interactions between multiple variables. In order to improve the accuracy of the predictions, it is necessary to design more expressive models to fully exploit the correlation between potential patterns and features in the wind energy data.
In order to overcome the limitations of the traditional method, in recent years, data analysis and artificial intelligence technology are widely applied in the field of wind power prediction. For example, machine learning algorithms such as Support Vector Machines (SVMs), random Forest (Random Forest), and deep learning models such as Recurrent Neural Networks (RNNs) and Convolutional Neural Networks (CNNs) have been applied to improve accuracy and reliability of wind power predictions.
In addition, the attentional mechanism is an important mechanism, and has been widely studied and applied in the field of deep learning in recent years. By introducing an attention mechanism, the model can automatically learn and pay attention to information particularly important to tasks, so that the expression capacity and the prediction performance of the model are improved. In wind power prediction, attention mechanisms are introduced to help the model to pay attention to correlation among characteristics such as wind speed, wind direction and temperature, so that accuracy of power prediction is improved.
Disclosure of Invention
Based on the background technology, the invention provides a wind power prediction method based on a double-attention mechanism, and the accuracy and reliability of wind power plant power prediction are improved by combining data analysis and artificial intelligence technology. The method can provide effective operation management and power grid dispatching decision support for the wind power industry, and further promote sustainable development of renewable energy sources.
A wind power prediction method based on a dual-attention mechanism comprises the following steps:
step 1, acquiring original wind power plant data from a data acquisition and monitoring control SCADA system, and extracting a wind speed time sequence X from the original wind power plant data ws Time series X of wind direction wd Wind power time series X wp Temperature time series X t Humidity time series X h Atmospheric pressure time series X p ；
Step 2, using singular spectrum analysis method to respectively obtain wind speed time series X ws Wind direction time series X wd Wind power time series X wp Denoising;
step 3, respectively carrying out normalization processing on each time sequence by using a linear function normalization method to obtain normalized time sequences;
step 4, constructing a UNet neural network, wherein the UNet neural network consists of a double convolution neural network layer DoubluConv, an encoder, a decoder and a convolution layer, the encoder consists of 4 downsampled neural network layers Down, and the decoder consists of 4 upsampled neural network layers Up;
step 5, obtaining a UNet input matrix XU based on the normalized wind speed time sequence, wind direction time sequence and wind power time sequence, and obtaining a UNet label matrix ZU based on the normalized denoising wind speed time sequence, denoising wind direction time sequence and denoising wind power time sequence; taking a part of the front columns from the UNet input matrix XU and the UNet label matrix ZU respectively to obtain UNet training input matrices XU respectively 1 UNet training tag matrix ZU 1 ；
Step 6, training the UNet neural network by using an Adam optimizer and a BCE loss function; training data input is XU 1 Training data tag ZU 1 The trained model is exported and stored;
step 7, taking a part of columns from the UNet input matrix XU to obtain a UNet verification input matrix XU 2 Inputting the trained UNet neural network to obtain an UNet verification output matrix YU;
step 8, taking the latter part of the columns from the normalized temperature time sequence, humidity time sequence and atmospheric pressure time sequence respectively to obtain input temperature time sequence TNX respectively t Inputting humidity time sequence TNX h Inputting the atmospheric pressure time series TNX p ；
Step 9, performing row vector decomposition on the UNet verification output matrix YU to obtain a UNet output wind speed time sequence NYU ws UNet output wind direction time series NYU wa UNet output wind power time sequence NYU wp Combining the input time sequences obtained in the step 8 to obtain a Seq2Seq input matrix XS, and dividing the input matrix XS into a Seq2Seq training matrix XS 1 And a Seq2Seq verification matrix XS 2 ；
Step 10, defining Attention layer Attention, taking input parameters of the Attention layer Attention as an input time sequence, and converting the input parameters into three vectors of q, k and v through a full connection layer;
step 11, constructing a Seq2Seq neural network, wherein the Seq2Seq neural network consists of an Attention layer Attention, an encoder, an Attention layer Attention and a decoder;
step 12, training the Seq2Seq matrix XS 1 Input into the Seq2Seq neural network for training: outputting the UNet to wind power time series NYU wp Training the trained label by using an Adam optimizer and an MSE loss function, and deriving and storing the trained model;
step 13, verifying the sequence 2 sequence verification matrix XS 2 Inputting the trained Seq2Seq neural network to obtain a predicted normalized wind power time sequence NYS wp The method comprises the steps of carrying out a first treatment on the surface of the Pair NYS wp Carrying out normalization reduction to obtain a predicted wind power time sequence YS wp The method comprises the steps of carrying out a first treatment on the surface of the From wind power time series X wp The sequence of the latter part is taken out to obtain a time sequence TX of the post wind power wp The method comprises the steps of carrying out a first treatment on the surface of the According to the wind power time sequence YS wp Post wind power time series TX wp Calculating average absolute error and root mean square error as evaluation indexes of the final wind power prediction effect;
step 14, obtaining a predicted wind speed time sequence P from weather forecast information ws Time series P of predicted wind direction wd Predicted temperature time series P t Predicted humidity time series P h Predicted barometric pressure time series P p The method comprises the steps of carrying out a first treatment on the surface of the Respectively carrying out linear function normalization to obtain a prediction input matrix NP, and inputting NP into the trained Seq2Seq neural network to obtain a weather prediction normalized wind power time sequence NP wp For NP wp Carrying out normalization reduction to obtain weather forecast windTime series of electric power P wp I.e. the result of wind power prediction.
The wind power prediction method based on the double-attention mechanism has the following beneficial effects:
(1) The problem of data noise is solved: noise interference is often present in wind energy data, and these factors can adversely affect the predicted outcome. The method corrects noise by adopting a method combining singular spectrum analysis and a UNet neural network, reduces noise influence, and improves reliability and robustness of a prediction result.
(2) The prediction accuracy is improved: by introducing a double-attention mechanism, a Seq2Seq neural network is constructed, and the method can better mine potential modes and correlations in wind energy data, so that the accuracy and reliability of wind power plant power prediction are improved.
(3) Facilitating wind farm management and grid scheduling: accurate wind power prediction is critical to operation management and grid scheduling of wind farms. The prediction method can provide accurate power prediction results, help optimize the operation plan of the wind power plant, adjust the power grid load, effectively cope with wind energy fluctuation and improve the economy and reliability of the wind power plant.
Drawings
FIG. 1 is a flowchart of steps of a wind power prediction method based on a dual-attention mechanism in an embodiment of the present invention.
FIG. 2 is a flowchart of a wind power prediction method based on a dual-attention mechanism according to an embodiment of the present invention.
Detailed Description
The technical scheme of the invention is further described in detail below with reference to the attached drawings.
A wind power prediction method based on a double-attention mechanism, referring to fig. 1 and 2, comprises the following main steps:
step 1: collecting original wind power plant data from a data collection and monitoring control SCADA system, and extracting a wind speed time sequence X from the original wind power plant data ws Time series X of wind direction wd Time series X of wind power wp Temperature time series X t Time series of humidity X h Atmospheric pressure time series X p 。
Step 2: method of singular spectrum analysis for wind speed time series X ws Time series X of wind direction wd Time series X of wind power wp And (5) processing. And for the time sequence X, decomposing and embedding the time sequence X into a track matrix Y, then carrying out singular value decomposition on the track matrix Y, and carrying out three-feature grouping and anti-diagonal averaging to obtain a denoising time sequence Z. Wind speed time series X ws Time series X of wind direction wd Time series X of wind power wp After singular spectrum analysis processing, respectively obtaining a denoising wind speed time sequence Z ws Time series Z of denoising direction wd Denoising wind power time sequence Z wp 。
Step 3: wind speed time series X using a linear function normalization (Min-Max Scaling) method ws Time series X of wind direction wd Time series X of wind power wp Denoised wind speed time series Z ws Time series Z of denoising direction wd Denoising wind power time sequence Z wp Temperature time series X t Time series of humidity X h Atmospheric pressure time series X p Carrying out normalization processing to obtain normalized wind speed time series NX ws Normalized wind direction time series NX wd Normalized wind power time series NX wp Normalized denoised wind speed time series NZ ws Normalized de-noised wind direction time series NZ wd Normalized denoising wind power time series NZ wp Normalized temperature time series NX t Normalized humidity time series NX h Normalized barometric pressure time series NX p 。
Step 4: defining a double convolution neural network layer DoubluConv, wherein input parameters are an input dimension in, an output dimension out and an intermediate dimension mid, and the neural network contained in the neural network layer is described as follows:
(1) A 2-dimensional convolution layer, wherein the input dimension is in, the output dimension is mid, the convolution kernel size is 3 multiplied by 3, and the padding is 1; (2) a 2-dimensional batch normalization layer; (3) a ReLu layer; (4) A 2-dimensional convolution layer, wherein the input dimension is mid, the output dimension is out, the convolution kernel size is 3×3, and the padding is 1; (5) a 2-dimensional batch normalization layer; (6) ReLu layer.
The Down sampling neural network module Down is defined, and input parameters are an input dimension in and an output dimension out. The neural network contained in the module is described as follows:
(1) 2-dimensional maximum pooling layer, window size is 2; (2) The double convolutional neural network layer doubiliConv, the input parameters in=in, out=out, mid=out.
The Up-sampling neural network module Up is defined, and input parameters are an input dimension in, an output dimension out and a down-sampling skip layer skip. The neural network contained in the module is described as follows:
(1) The Upsampled layer adopts bilinear interpolation up-sampling algorithm, and the multiple is 2; (2) The double convolution neural network layer DoubluConv has input parameters of in=in, out=out, mid=out/2; (3) And the feature fusion layer is used for carrying out feature fusion with the downsampling skip layer skip.
And constructing a UNet neural network, wherein the UNet neural network consists of an encoder and a decoder. The encoder consists of 4 downsampled neural network layers and the decoder consists of 4 upsampled neural network layers. The neural network contained by UNet is described as follows:
(1) Double convolution neural network layer DoubluConv, input parameters are in=1, out=32, mid=32, and output is stored as variable x 1 The method comprises the steps of carrying out a first treatment on the surface of the (2) Downsampling neural network module Down, input parameters are in=32, out=64, output is stored as variable x 2 The method comprises the steps of carrying out a first treatment on the surface of the (3) Downsampling neural network module Down, input parameters are in=64, out=128, output is stored as variable x 3 The method comprises the steps of carrying out a first treatment on the surface of the (4) Downsampling neural network module Down, input parameters are in=128, out=256, output is stored as variable x 4 The method comprises the steps of carrying out a first treatment on the surface of the (5) The Down-sampling neural network module Down has input parameters of in=256 and out=256; (6) The Up-sampling neural network module Up has input parameters of in=512, out=128, skip=x 4 The method comprises the steps of carrying out a first treatment on the surface of the (7) The Up-sampling neural network module Up has input parameters of in=256, out=64, skip=x 3 The method comprises the steps of carrying out a first treatment on the surface of the (8) Upward miningThe neural network module Up has input parameters of in=128, out=32 and skip=x 2 The method comprises the steps of carrying out a first treatment on the surface of the (9) The Up-sampling neural network module Up has input parameters of in=64, out=32, skip=x 1 The method comprises the steps of carrying out a first treatment on the surface of the (10) A 2-dimensional convolution layer, with an input dimension of 32, an output dimension of 1, and a convolution kernel size of 1 x 1.
Step 5: definition of UNet input matrix xu= [ NX ] ws ，NX wd ，NX wp ] T UNet tag matrix zu= [ NZ ws ，NZ wd ，NZ wp ] T Where T represents the matrix transpose operator, NX ws For normalized wind speed time series, NX wd For normalized wind direction time series, NX wp NZ is normalized wind power time series ws NZ is a normalized denoised wind speed time series wd NZ is a normalized de-noised wind direction time series wp And (5) a normalized denoising wind power time sequence. The first 20% of columns are taken from the UNet input matrix XU and UNet label matrix ZU respectively, and UNet training input matrix XU is obtained respectively 1 UNet training tag matrix ZU 1 。
Step 6: training the UNet neural network: training uses Adam optimizer and BCE loss function, the Adam optimizer learning rate is 2×10 -5 Training data input is XU 1 Training data tag ZU 1 Wherein XU is 1 Training the input matrix for UNet, ZU 1 The tag matrix is trained for UNet. The trained model export is stored as file unet.
Step 7: 80% of columns are taken from the UNet input matrix XU to obtain a UNet verification input matrix XU 2 . Loading a UNet neural network from a file UNet. Pkt, and inputting a UNet verification input matrix XU 2 Inputting into the UNet neural network to obtain the UNet verification output matrix YU.
Step 8: from normalized temperature time series NX, respectively t Normalized humidity time series NX h Normalized barometric pressure time series NX p The 80% columns are taken out to obtain input temperature time series TNX respectively t Input humidity time series TNX h Time sequence of input of barometric pressureColumn TNX p 。
Step 9: performing row vector decomposition on the UNet verification output matrix YU to obtain a UNet output wind speed time sequence NYU ws Time series NYU of UNet output wind direction wd Time sequence NYU of UNet output wind power wp . Define the Seq2Seq input matrix xs= [ NYU ] ws ，NYU wd ，TNX t ，TNX h ，TNX p ] T Wherein T represents a matrix transpose operator, TNX t To input a temperature time series, TNX h To input humidity time series, TNX p For the input of the barometric pressure time series. Input matrix XS for Seq2Seq is 8:2 column scale division Seq2Seq training matrix XS 1 And a Seq2Seq verification matrix XS 2 。
Step 10: defining Attention layer Attention, converting input time sequence X into three vectors of q, k and v through full connection layer, defining L as length of input time sequence X. The correlation vector Sim is initialized to a length L. Traversing vector k such that Let α=softmax (Sim), then attention +.>The input time series X refers to input parameters of Attention layer Attention. Specifically, after Attention layer Attention is referenced in step 11, the actual incoming parameter is XS 1 Or XS 2 Etc.
Step 11: the method comprises the steps of constructing a Seq2Seq neural network, wherein the Seq2Seq neural network consists of Attention layer Attention, an encoder, attention layer Attention and a decoder. The neural network comprised by the Seq2Seq neural network is described as follows: (1) Attention layer Attention; (2) an Encoder layer; (3) Attention layer Attention; (4) a Decoder layer.
Step 12: training the Seq2Seq matrix XS 1 Input into the Seq2Seq neural networkTraining: outputting the UNet to wind power time series NYU wp As a training label, training used Adam optimizer and MSE loss function, the Adam optimizer learning rate was 1×10 -4 . The trained model derivation is stored as file dseq2seq.
Step 13: loading the Seq2Seq neural network from the file dseq2Seq. Pkt, validating the matrix XS with the Seq2Seq 2 Inputting into a Seq2Seq neural network to obtain a predicted normalized wind power time sequence NYS wp . Pair NYS wp Carrying out normalization reduction to obtain a predicted wind power time sequence YS wp . From wind power time series X wp The sequence of the last 16% is obtained to obtain a time sequence TX of the last wind power wp . Calculating the average absolute error And root mean square errorAs an evaluation index of the final wind power prediction effect. The two formulas calculate a value in power, W, where a smaller value indicates a smaller error between the predicted value and the actual value, i.e., a smaller value.
Step 14: obtaining a predicted wind speed time sequence P from weather forecast information ws Time series P of predicted wind direction wd Predicted temperature time series P t Predicted humidity time series P h Predicted barometric pressure time series P p . Will P ws 、P wd ，P t ，P h ，P p Respectively carrying out linear function normalization to obtain a normalized predicted wind speed time sequence NP ws Normalized predicted wind direction time series NP wd Normalized predicted temperature time series NP t Normalized predicted humidity time series NP h Normalized predicted barometric pressure time series NP p . Let the prediction input matrix np= [ NP ] ws ，NP wd ，NP t ，NP h ，NP p ] T Inputting NP into the trained Seq2Seq neural network to obtain weather forecast normalized wind power time sequence NP wp . For NP wp Carrying out normalization reduction to obtain a weather forecast wind power time sequence P wp 。P wp I.e. the result of wind power prediction.
The above description is merely of preferred embodiments of the present invention, and the scope of the present invention is not limited to the above embodiments, but all equivalent modifications or variations according to the present disclosure will be within the scope of the claims.
Claims (10)
1. A wind power prediction method based on a double-attention mechanism is characterized by comprising the following steps of: the method comprises the following steps:
step 1, acquiring original wind power plant data from a data acquisition and monitoring control SCADA system, and extracting a wind speed time sequence X from the original wind power plant data ws Time series X of wind direction wd Time series X of wind power wp Temperature time series X t Time series of humidity X h Atmospheric pressure time series X p ；
Step 2, using singular spectrum analysis method to respectively obtain wind speed time series X ws Time series X of wind direction wd Time series X of wind power wp Denoising;
step 3, respectively carrying out normalization processing on each time sequence by using a linear function normalization method to obtain normalized time sequences;
step 4, constructing a UNet neural network, wherein the UNet neural network consists of a double convolution neural network layer DoubluConv, an encoder, a decoder and a convolution layer, the encoder consists of 4 downsampled neural network layers Down, and the decoder consists of 4 upsampled neural network layers Up;
step 5, obtaining a UNet input matrix X based on the normalized wind speed time sequence, wind direction time sequence and wind power time sequenceU, obtaining a UNet tag matrix ZU based on the normalized denoising wind speed time sequence, the denoising wind direction time sequence and the denoising wind power time sequence; taking a part of the front columns from the UNet input matrix XU and the UNet label matrix ZU respectively to obtain UNet training input matrices XU respectively 1 UNet training tag matrix ZU 1 ；
Step 6, training the UNet neural network by using an Adam optimizer and a BCE loss function; training data input is XU 1 Training data tag ZU 1 The trained model is exported and stored;
step 7, taking a part of columns from the UNet input matrix XU to obtain a UNet verification input matrix XU 2 Inputting the trained UNet neural network to obtain an UNet verification output matrix YU;
step 8, taking the latter part of the columns from the normalized temperature time sequence, humidity time sequence and atmospheric pressure time sequence respectively to obtain input temperature time sequence TNX respectively t Inputting humidity time sequence TNX h Inputting the atmospheric pressure time series TNX p ；
Step 9, performing row vector decomposition on the UNet verification output matrix YU to obtain a UNet output wind speed time sequence NYU ws Time series NYU of UNet output wind direction wd Time sequence NYU of UNet output wind power wp Combining the input time sequences obtained in the step 8 to obtain a Seq2Seq input matrix XS, and dividing the input matrix XS into a Seq2Seq training matrix XS 1 And a Seq2Seq verification matrix XS 2 ；
Step 10, defining Attention layer Attention, taking input parameters of the Attention layer Attention as an input time sequence, and converting the input parameters into three vectors of q, k and v through a full connection layer;
step 11, constructing a Seq2Seq neural network, wherein the Seq2Seq neural network consists of an Attention layer Attention, an encoder, an Attention layer Attention and a decoder;
step 12, training the Seq2Seq matrix XS 1 Input into the Seq2Seq neural network for training: outputting the UNet to wind power time series NYU wp As a training label, training is usedThe Adam optimizer and the MSE loss function derive and store the trained model;
step 13, verifying the sequence 2 sequence verification matrix XS 2 Inputting the trained Seq2Seq neural network to obtain a predicted normalized wind power time sequence NYS wp The method comprises the steps of carrying out a first treatment on the surface of the Pair NYS wp Carrying out normalization reduction to obtain a predicted wind power time sequence YS wp The method comprises the steps of carrying out a first treatment on the surface of the From wind power time series X wp The sequence of the latter part is taken out to obtain a time sequence TX of the post wind power wp The method comprises the steps of carrying out a first treatment on the surface of the According to the wind power time sequence YS wp Post wind power time series TX wp Calculating average absolute error and root mean square error as evaluation indexes of the final wind power prediction effect;
step 14, obtaining a predicted wind speed time sequence P from weather forecast information ws Time series P of predicted wind direction wd Predicted temperature time series P t Predicted humidity time series P h Predicted barometric pressure time series P p The method comprises the steps of carrying out a first treatment on the surface of the Respectively carrying out linear function normalization to obtain a prediction input matrix NP, and inputting NP into the trained Seq2Seq neural network to obtain a weather prediction normalized wind power time sequence NP wp For NP wp Carrying out normalization reduction to obtain a weather forecast wind power time sequence P wp I.e. the result of wind power prediction.
2. The wind power prediction method based on the dual-attention mechanism according to claim 1, wherein the method comprises the following steps: in the step 2, for the time sequence X, decomposing and embedding the time sequence X to form a track matrix Y, then carrying out singular value decomposition on the track matrix Y, and obtaining a denoising time sequence Z after three-feature grouping and anti-diagonal line averaging; wind speed time series X ws Wind direction time series X wd Wind power time series X wp After singular spectrum analysis processing, respectively obtaining a denoising wind speed time sequence Z ws Denoise wind time series Z wd Denoising wind power time sequence Z wp 。
3. The wind power prediction method based on the dual-attention mechanism according to claim 1, wherein the method comprises the following steps: in step 4, the input parameters are an input dimension in, an output dimension out, and an intermediate dimension mi, which include the following parts: a 2-dimensional convolution layer, wherein the input dimension is in, the output dimension is mid, the convolution kernel size is 3 multiplied by 3, and the padding is 1; 2-dimensional batch normalization layer; a ReLu layer; a 2-dimensional convolution layer, wherein the input dimension is mid, the output dimension is out, the convolution kernel size is 3×3, and the padding is 1; 2-dimensional batch normalization layer; a ReLu layer;
the Down sampling neural network module Down, the input parameters are input dimension in and output dimension out, and the Down sampling neural network module comprises the following parts: 2-dimensional maximum pooling layer, window size is 2; the double convolution neural network layer doubiliConv, the input parameters in=in, out=out, mid=out;
the Up-sampling neural network module Up, the input parameters are input dimension in and output dimension out, and the down-sampling skip layer skip comprises the following parts: the Upsampled layer adopts bilinear interpolation up-sampling algorithm, and the multiple is 2; the double convolution neural network layer DoubluConv has input parameters of in=in, out=out, mid=out/2; and the feature fusion layer is used for carrying out feature fusion with the downsampling skip layer skip.
4. A wind power prediction method based on a dual-attention mechanism according to claim 3, wherein: in step 4, the UNet neural network comprises the following parts:
double convolution neural network layer DoubluConv, input parameters are in=1, out=32, mid=32, and output is stored as variable x 1 The method comprises the steps of carrying out a first treatment on the surface of the Downsampling neural network module Down, input parameters are in=32, out=64, output is stored as variable x 2 The method comprises the steps of carrying out a first treatment on the surface of the Downsampling neural network module Down, input parameters are in=64, out=128, output is stored as variable x 3 The method comprises the steps of carrying out a first treatment on the surface of the Downsampling neural network module Down, input parameters are in=128, out=256, output is stored as variable x 4 The method comprises the steps of carrying out a first treatment on the surface of the The Down-sampling neural network module Down has input parameters of in=256 and out=256; the Up-sampling neural network module Up has input parameters of in=512, out=128, skip=x 4 The method comprises the steps of carrying out a first treatment on the surface of the UpsamplingThe neural network module Up has input parameters of in=256, out=64, skip=x 3 The method comprises the steps of carrying out a first treatment on the surface of the The Up-sampling neural network module Up has input parameters of in=128, out=32, skip=x 2 The method comprises the steps of carrying out a first treatment on the surface of the The Up-sampling neural network module Up has input parameters of in=64, out=32, skip=x 1 The method comprises the steps of carrying out a first treatment on the surface of the A 2-dimensional convolution layer, with an input dimension of 32, an output dimension of 1, and a convolution kernel size of 1 x 1.
5. The wind power prediction method based on the dual-attention mechanism according to claim 1, wherein the method comprises the following steps: in step 6, the learning rate of the Adam optimizer is 2×10 -5 。
6. The wind power prediction method based on the dual-attention mechanism according to claim 1, wherein the method comprises the following steps: in step 9, the Seq2Seq input matrix xs= [ NYU ] ws ,NYU wd ,TNX t ,TNX h ,TNX p ] T 。
7. The wind power prediction method based on the dual-attention mechanism according to claim 1, wherein the method comprises the following steps: in step 9, the Seq2Seq input matrix XS divides the Seq2Seq training matrix XS at a column ratio of 8:2 1 And a Seq2Seq verification matrix XS 2 。
8. The wind power prediction method based on the dual-attention mechanism according to claim 1, wherein the method comprises the following steps: in step 10, defining L as the length of the input time sequence; initializing a correlation vector Sim, wherein the length of the correlation vector Sim is L; traversing vector k such that Let α=softmax (Sim), then attention +.>
9. The wind power prediction method based on the dual-attention mechanism according to claim 1, wherein the method comprises the following steps: in step 12, the learning rate of Adam optimizer is 1×10 -4 。
10. The wind power prediction method based on the dual-attention mechanism according to claim 1, wherein the method comprises the following steps: in step 13, absolute errorRoot mean square error
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310923863.5A CN116957849A (en) | 2023-07-26 | 2023-07-26 | Wind power prediction method based on double-attention mechanism |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310923863.5A CN116957849A (en) | 2023-07-26 | 2023-07-26 | Wind power prediction method based on double-attention mechanism |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN116957849A true CN116957849A (en) | 2023-10-27 |
Family
ID=88448655
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310923863.5A Pending CN116957849A (en) | 2023-07-26 | 2023-07-26 | Wind power prediction method based on double-attention mechanism |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116957849A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117808175A (en) * | 2024-03-01 | 2024-04-02 | 南京信息工程大学 | Short-term multi-energy load prediction method based on DTformer |
-
2023
- 2023-07-26 CN CN202310923863.5A patent/CN116957849A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117808175A (en) * | 2024-03-01 | 2024-04-02 | 南京信息工程大学 | Short-term multi-energy load prediction method based on DTformer |
| CN117808175B (en) * | 2024-03-01 | 2024-05-17 | 南京信息工程大学 | DTformer-based short-term multi-energy load prediction method |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Ma et al. | A hybrid attention-based deep learning approach for wind power prediction | |
| Yin et al. | Multi-temporal-spatial-scale temporal convolution network for short-term load forecasting of power systems | |
| Liu et al. | Wind speed forecasting using deep neural network with feature selection | |
| CN116070799B (en) | Photovoltaic power generation amount prediction system and method based on attention and deep learning | |
| CN102562469B (en) | Short-term wind driven generator output power predicting method based on correction algorithm | |
| CN117290800B (en) | Timing sequence anomaly detection method and system based on hypergraph attention network | |
| JP7441553B2 (en) | Short-term wind speed prediction method and system based on rolling time series and support vector machine | |
| CN113222289B (en) | Prediction method of energy power based on data processing | |
| CN116957849A (en) | Wind power prediction method based on double-attention mechanism | |
| Wang et al. | A wind speed forecasting system for the construction of a smart grid with two-stage data processing based on improved ELM and deep learning strategies | |
| CN114298134A (en) | Wind power prediction method and device and electronic equipment | |
| Liu et al. | A novel hybrid model based on GA-VMD, sample entropy reconstruction and BiLSTM for wind speed prediction | |
| CN118227979B (en) | ENSO (open ended sensing algorithm) method for predicting sea temperature abnormality of tropical Pacific subsurface layer based on improved convolutional neural network | |
| CN114330934A (en) | Model parameter self-adaptive GRU new energy short-term power generation power prediction method | |
| CN116702831A (en) | Hybrid short-term wind power prediction method considering massive loss of data | |
| CN116881665A (en) | CMOA optimization-based TimesNet-BiLSTM photovoltaic output prediction method | |
| CN116151464A (en) | Photovoltaic power generation power prediction method, system and storable medium | |
| CN114118401A (en) | Neural network-based power distribution network flow prediction method, system, device and storage medium | |
| CN114676622A (en) | Short-term photovoltaic power prediction method based on self-encoder deep learning model | |
| CN114037143A (en) | Short-term wind power combination prediction method | |
| Hou et al. | A hybrid machine learning forecasting model for photovoltaic power | |
| CN116090635A (en) | Meteorological-driven new energy generation power prediction method | |
| Wang et al. | Efficient shrinkage temporal convolutional network model for photovoltaic power prediction | |
| CN114545066A (en) | Non-invasive load monitoring model polymerization method and system | |
| CN112036672A (en) | New energy power generation ultra-short term power prediction method and system based on iterative correction |
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 |