CN115994325B - Fan icing power generation data enhancement method based on TimeGAN deep learning method - Google Patents
Fan icing power generation data enhancement method based on TimeGAN deep learning method Download PDFInfo
- Publication number
- CN115994325B CN115994325B CN202310292758.6A CN202310292758A CN115994325B CN 115994325 B CN115994325 B CN 115994325B CN 202310292758 A CN202310292758 A CN 202310292758A CN 115994325 B CN115994325 B CN 115994325B
- Authority
- CN
- China
- Prior art keywords
- data
- power
- icing
- timegan
- network
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- 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
- Y02E—REDUCTION OF GREENHOUSE GAS [GHG] EMISSIONS, RELATED TO ENERGY GENERATION, TRANSMISSION OR DISTRIBUTION
- Y02E10/00—Energy generation through renewable energy sources
- Y02E10/70—Wind energy
- Y02E10/72—Wind turbines with rotation axis in wind direction
Abstract
A fan icing power generation data enhancement method based on a TimeGAN deep learning method collects original data from a wind farm and sorts the data into a format of time stamp-wind speed-temperature-power; manually marking a fan icing time period for the data, and arranging and generating a data set under the icing condition; for the marked data, fitting and classifying the data according to a wind speed-power corresponding relation curve by using a function, and further expanding the data by matching similar wind power curves of a plurality of power stations; the extracted data is imported into a TimeGAN network for training; generating a new data set by using a generating network Generator, and verifying the correlation of the generated data set on a verification set; splicing the sorted data sets; and importing the generated new data set into a power prediction model for retraining, extracting icing data in the verification set, and performing a prediction test on the result. By adopting the method and the device, the power forecasting accuracy under the condition of ice coating of the fan is improved.
Description
Technical Field
The invention belongs to the technical field of wind power generation power prediction, and particularly relates to a fan icing power generation power data enhancement method based on a TimeGAN deep learning method.
Background
Wind power generation is regarded as clean and pollution-free energy, has better economic benefit and social benefit and is valued by countries around the world. The rapid development of wind power also brings a series of problems, the wind power plant in the high-altitude area is affected by low temperature to cause icing of the fan blade, the icing can change the shape and aerodynamic layout of the blade, and the flow field distribution on the surface of the blade is caused to influence the wind energy utilization rate of the blade, so that the output power of the fan is affected. Wind farm icing generally occurs in low temperature periods with a low probability of occurrence, so that sufficient icing data cannot be obtained through normal data collection and integration. And most of wind power plant related information is imperfect in record, so that available icing power generation data and limited data are caused. Therefore, the high accuracy rate of the traditional power method cannot be achieved in the icing period of the fan, and the method creates a barrier for large-scale grid-connected application of wind power. Therefore, how to improve the early warning accuracy of the fan ice coating and the power forecasting accuracy of the fan in the ice coating condition become a new challenge puzzling the industry.
Related patent literature: CN115163430a discloses a fan icing prediction method based on data driving, which comprises the following steps: step 1, determining atmospheric parameters to be measured according to prediction requirements and measuring in real time; step 2, establishing a data fitting model to obtain an atmospheric parameter predicted value; step 3, converting the three-dimensional model of the wind driven generator blade into a two-dimensional model; step 4, constructing a wind driven generator blade icing quality prediction function based on the atmospheric parameter prediction value; step 5, reconstructing and synthesizing a three-dimensional icing shape of the wind driven generator blade; and 6, reconstructing an atmospheric parameter predicted value according to the real-time parameters, and updating an icing quality prediction function of the wind driven generator blade. CN115143057a discloses a fan icing monitoring method based on frozen environment atmospheric parameters, which comprises the following steps: step 1, determining atmospheric parameters to be detected according to monitoring requirements; step 2, measuring real-time atmospheric parameters; step 3, converting the three-dimensional model of the wind driven generator blade into a two-dimensional model; step 4, constructing a wind driven generator blade icing quality monitoring function based on atmospheric parameters; step 5, reconstructing and synthesizing a three-dimensional icing shape of the wind driven generator blade; and 6, updating the icing quality function of the wind driven generator blade according to the real-time parameters.
The above technology does not give a specific guidance scheme on how to improve the accuracy of power forecasting under the situation that the fan is covered with ice.
Disclosure of Invention
The invention aims to provide a fan icing power generation power data enhancement method based on a TimeGAN deep learning method, which can improve the power prediction accuracy under the fan icing condition so as to solve the problem that the machine learning cannot learn the correct wind power generation characteristics and cannot accurately predict the power generation capacity under the icing condition due to unbalanced sample number under the wind power generation prediction wind power generator icing condition.
In order to solve the technical problems, the invention adopts the following technical scheme:
a fan icing power generation data enhancement method based on a TimeGAN deep learning method is characterized by comprising the following steps:
step 1, collecting original data from a wind farm, and arranging the data into a format of time stamp-wind speed-temperature-power;
step 3, using a function for the data marked in the step 2 according to the wind speed-power corresponding relation curve
Fitting classification is carried out, and data are further expanded by matching wind power curves similar to a plurality of power stations;
step 4, importing the data extracted in the step 3 into a TimeGAN network for training;
step 5, generating a new data set by using a generating network Generator, verifying the correlation of the generated data set on a verification set, retraining if the correlation is low, otherwise starting step 6;
step 6, the data generated in the step 5 are processed according to the following steps of 10-9: 1 and the data set arranged in the step 2 are spliced, the generated data time sequence can be directly connected with the icing data set ending time in the step 2, and the data set generated by the method needs to pay attention to the fact that the time sequence cannot be repeated with the time sequence of the real data set;
and 7, importing the new data set generated in the step 6 into a power prediction model for retraining, extracting icing data in the verification set, and carrying out a prediction test on the result, wherein if the accuracy of the fan icing early warning and the power prediction model is not lower than that of the power prediction model before retraining on the verification set, the new power prediction model can be used, otherwise, the step 4 is required to return to the step 4 to regenerate data by using different training iteration numbers.
In the above technical solution, the preferred technical solution may be that the expansion method in step 3 specifically includes the following steps:
step 3.1, deriving wind speed-power data of an original power station, and calculating characteristic variables a, b of a fitting curve through absolute error distances by using a function F (x);
and 3.2, substituting data of a new power station lacking icing observation data into a formula F (x), calculating absolute distance errors, and selecting data expansion original data sets with errors smaller than Cap multiplied by 0.1, wherein Cap is the installed capacity of the target power station.
In the above technical solution, the preferred technical solution may be that in the step 4, the data extracted in the step 3 is imported into the TimeGAN network for training, and the specific method is as follows:
step 4.1, timeGAN utilizes a feature encoding function
For static characteristics and time sequencesThe signs are encoded to obtain their low-order features, denoted as e,
wherein S represents original static characteristics, pi t ΧRepresenting the features of the original spatial vector,HSindicating the temporal characteristics after learning t HXRepresenting the space vector characteristics after learning;
step 4.2, utilizing the feature decoding function at the network output terminal TimeGAN
Restoring the potential features obtained in the step 4.1 into original static features and temporal features, wherein the result is expressed as r;
step 4.3, the Generator network Generator generates time series data in potential space by using a generating function on static characteristics and temporal characteristics of data selected from the original data;
step 4.4, the discriminator network uses the discrimination function to discriminate the true and false of the time series data generated in the step 4.3;
and 4.5, training a generator network, a discriminator network and a training loss balancing network by using an Adam optimizer, and simultaneously, training the generator network and the discriminator network by using a Supervisor monitoring network, so as to obtain a trained TimeGAN model through a plurality of iterative training.
In the foregoing technical solution, the preferred technical solution may be that, in the step 6, the data generated in the step 5 is preferably according to 9:1 and the data set sorted in the step 2 is spliced.
The invention provides a fan icing power generation data enhancement method based on a TimeGAN deep learning method, which is challenging for fan icing early warning and fan output prediction under icing conditions due to small sample size of wind power plant icing observation data. The method is used for improving the accuracy of the power generation prediction under the ice coating early warning and ice coating conditions of the fan. Firstly, collecting power generation data under the icing condition of a certain number of wind driven generators from a wind power plant, classifying and screening original data according to a wind speed-power generation function relation, then integrating the data, constructing a training model by using a neural network based on TimeGAN, inputting an original sample into the neural network for countermeasure training, and fusing a simulation data set generated by the neural network with the original data set to form a new data set. Dividing the new data set into a training set, a testing set and a verification set; inputting the data set into a wind driven generator power prediction model based on a random forest for testing so as to evaluate the performance of the new data set on the fan icing and the power generation power prediction model. Experiments prove that the method can balance sample data under the non-icing condition and the icing condition, is favorable for constructing a unified wind power prediction model under the non-icing condition and the icing condition, improves the accuracy of fan icing and the accuracy of wind power prediction under the icing condition, solves the problem that mechanisms such as new energy stations, power dispatching departments and the like trouble to early warn the high-precision fan icing and predict the accuracy of power generation under the icing condition to be low, can provide good guarantee for wind power generation, and is convenient for conversion and popularization. Compared with the prior art, the method improves the power forecasting accuracy from 75% to more than 84% under the condition that the fan is covered with ice. The experimental result is evaluated and calculated by adopting a daily comprehensive accuracy formula, wherein the calculation formula is as follows:
Drawings
Fig. 1 is an overall flowchart of data set generation by using the fan icing power generation data enhancement method based on the TimeGAN deep learning method.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the technical solutions of the present invention will be clearly and completely described below with reference to examples. It will be apparent that the described embodiments are only some, but not all, of the embodiments of the invention. All other embodiments, which can be made by one of ordinary skill in the art without undue burden, are within the scope of the present invention based on the following examples.
Example 1: as shown in fig. 1, reference numeral 1 in fig. 1 is a data acquisition flow, reference numeral 2 is a data generation countermeasure training flow, reference numeral 3 is a generation data integration flow, and reference numeral 4 is a generation result verification flow.
The fan icing power generation data enhancement method based on the TimeGAN deep learning method is implemented according to the following steps (specifically comprising the following steps):
step 1, collecting original data from a wind farm, and arranging the data into a format of time stamp-wind speed-temperature-power.
And 2, manually marking the fan icing time period for the data in the step 1, and arranging the data set under the condition of generating icing.
Step 3, for the data marked in step 2, the number is countedUsing functions according to the wind speed-power corresponding relation curve
Fitting classification is carried out, and data are further expanded by matching wind power curves similar to a plurality of power stations; the specific expansion method is as follows:
step 3.1, deriving wind speed-power data of the original power station, and calculating characteristic variables a, b of the fitting curve through absolute error distances by using a function F (x).
And 3.2, substituting data of a new power station lacking icing observation data into a formula F (x), calculating absolute distance errors, and selecting data expansion original data sets with errors smaller than Cap multiplied by 0.1, wherein Cap is the installed capacity of the target power station.
Step 4, importing the data extracted in the step 3 into a TimeGAN network for training, wherein the specific method is as follows:
step 4.1, timeGAN utilizes a feature encoding function
The static features and the timing features are encoded to obtain their low-order features, denoted as e,
wherein S represents original static characteristics, pi t ΧRepresenting the features of the original spatial vector,HSindicating the temporal characteristics after learning t HXRepresenting the space vector characteristics after learning;
step 4.2, utilizing the feature decoding function at the network output terminal TimeGAN
And (3) recovering the potential features obtained in the step 4.1 into original static features and temporal features, and expressing the result as r.
Step 4.3, the Generator network Generator generates time series data in potential space using a generating function on the static and temporal features of the data selected from the raw data.
And 4.4, the discriminator network uses the discrimination function to discriminate the true and false of the time series data generated in the step 4.3.
And 4.5, training a generator network, a discriminator network and a training loss balancing network by using an Adam optimizer, and simultaneously, training the generator network and the discriminator network by using a Supervisor monitoring network, so as to obtain a trained TimeGAN model through a plurality of iterative training.
Step 5, generating a new data set by using the generating network Generator, and verifying the correlation of the generated data set on the verification set, if the correlation is low, retraining, otherwise, starting step 6.
Step 6, the data generated in step 5 are processed according to 9:1 and the data set arranged in the step 2 are spliced, the generated data time sequence can be directly connected with the icing data set ending time in the step 2, and the data set generated by the method needs to pay attention to the fact that the time sequence cannot be repeated with the time sequence of the real data set.
And 7, importing the new data set generated in the step 6 into a power prediction model for retraining, extracting icing data in the verification set, and carrying out a prediction test on the result, wherein if the accuracy of the fan icing early warning and the power prediction model is not lower than that of the power prediction model before retraining on the verification set, the new power prediction model can be used, otherwise, the step 4 is required to return to the step 4 to regenerate data by using different training iteration numbers.
According to the invention, after data matching expansion is performed through wind speed-power characteristics, a TimeGAN neural network is used for performing countermeasure enhancement training, and a new data set is generated by performing splice fusion on a generated data set and a true data set. The original wind speed-power curve equation of the target power station is fitted through a function, and then wind speed-power characteristics of other power stations are substituted into the equation to calculate the similarity. Station data with similarity higher than a threshold value is extracted to augment the original station data.
As a further improvement of the technical solution of the present invention, in the step 3, a multi-head feature extraction network is adopted to replace a feature coding network in an original TimeGAN network, and an improved LSTM network is used to replace a GRU generation network in the original TimeGAN network. Experiments prove that the adaptation of the TimeGAN network to the wind power generation scene can be improved by the modification.
The embodiment of the invention provides a fan icing power generation power data enhancement method based on a TimeGAN deep learning method, which improves the power prediction accuracy under the condition of fan icing, and solves the problem that the machine learning cannot learn the correct wind power generation characteristics and cannot accurately predict the power generation capacity under the icing condition due to unbalanced sample number under the condition of wind power generator icing in wind power generation prediction. Compared with the prior art, the invention improves the power forecasting accuracy by more than 9% under the condition of ice coating of the fan.
Claims (3)
1. A fan icing power generation data enhancement method based on a TimeGAN deep learning method is characterized by comprising the following steps:
step 1, collecting original data from a wind farm, and arranging the data into a format of time stamp-wind speed-temperature-power;
step 2, manually marking the fan icing time period for the data in the step 1, and arranging and generating a data set under the icing condition;
step 3, using a function for the data marked in the step 2 according to the wind speed-power corresponding relation curve
Fitting classification is carried out, and data are further expanded by matching wind power curves similar to a plurality of power stations;
the expansion method in the step 3 is specifically as follows:
step 3.1, deriving wind speed-power data of an original power station, and calculating characteristic variables a, b of a fitting curve through absolute error distances by using a function F (x);
step 3.2, substituting data of a new power station lacking icing observation data into a formula F (x), calculating absolute distance errors, and selecting data expansion original data sets with errors smaller than Cap multiplied by 0.1, wherein Cap is the installed capacity of a target power station;
step 4, importing the data extracted in the step 3 into a TimeGAN network for training;
step 5, generating a new data set by using a generating network Generator, and verifying the correlation of the generated data set on a verification set;
step 6, the data generated in the step 5 are processed according to the following steps of 10-9: 1 and the data set arranged in the step 2 are spliced to generate a data time sequence which is directly connected with the ending time of the ice-covered data set in the step 2, and the data set generated by the method needs to pay attention to the fact that the time sequence cannot be repeated with the time sequence of the real data set;
and 7, importing the new data set generated in the step 6 into a power prediction model for retraining, extracting icing data in the verification set, and carrying out a prediction test on the result, if the accuracy of the verification fan icing early warning and the power prediction model on the verification set is not lower than that of the power prediction model before retraining, using the new power prediction model, otherwise, returning to the step 4, and regenerating data by using different training iteration numbers.
2. The fan icing power generation data enhancement method based on the TimeGAN deep learning method according to claim 1 is characterized in that in step 4, the data extracted in step 3 is imported into a TimeGAN network for training, and the specific method is as follows:
step 4.1, timeGAN utilizes a feature encoding function
The static features and the timing features are encoded to obtain their low-order features, denoted as e,
wherein S represents original static characteristics, pi t ΧRepresenting the features of the original spatial vector,HSindicating the temporal characteristics after learning t HXRepresenting the space vector characteristics after learning;
step 4.2, utilizing the feature decoding function at the network output terminal TimeGAN
Restoring the potential features obtained in the step 4.1 to the original static stateFeatures and temporal features, the result being denoted r;
step 4.3, the Generator network Generator generates time series data in potential space by using a generating function on static characteristics and temporal characteristics of data selected from the original data;
step 4.4, the discriminator network uses the discrimination function to discriminate the true and false of the time series data generated in the step 4.3;
and 4.5, training a generator network, a discriminator network and a training loss balancing network by using an Adam optimizer, and simultaneously, training the generator network and the discriminator network by using a Supervisor monitoring network, so as to obtain a trained TimeGAN model through a plurality of iterative training.
3. The fan icing power generation data enhancement method based on the TimeGAN deep learning method according to claim 1, wherein in step 6, the data generated in step 5 is calculated according to 9:1 and the data set sorted in the step 2 is spliced.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310292758.6A CN115994325B (en) | 2023-03-24 | 2023-03-24 | Fan icing power generation data enhancement method based on TimeGAN deep learning method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310292758.6A CN115994325B (en) | 2023-03-24 | 2023-03-24 | Fan icing power generation data enhancement method based on TimeGAN deep learning method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN115994325A CN115994325A (en) | 2023-04-21 |
| CN115994325B true CN115994325B (en) | 2023-05-30 |
Family
ID=85993886
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310292758.6A Active CN115994325B (en) | 2023-03-24 | 2023-03-24 | Fan icing power generation data enhancement method based on TimeGAN deep learning method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115994325B (en) |
Families Citing this family (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113837002B (en) * | 2021-08-18 | 2024-04-23 | 西安理工大学 | Small sample data fault diagnosis method based on improved TimeGan model |
| KR102617312B1 (en) * | 2023-09-07 | 2023-12-27 | (주)바이브컴퍼니 | Method for forecasting temperature in underground utility tunnel based on CA-TimeGAN and residual CNN-LSTM |
| CN117638926B (en) * | 2024-01-25 | 2024-04-05 | 国能日新科技股份有限公司 | New energy power prediction method and device based on icing and power coupling modeling |
Family Cites Families (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20210197335A1 (en) * | 2019-12-26 | 2021-07-01 | Dalian University Of Technology | Data Augmentation Method Based On Generative Adversarial Networks In Tool Condition Monitoring |
| CN113449847B (en) * | 2021-06-17 | 2022-08-19 | 江苏方天电力技术有限公司 | Offshore wind power rolling prediction method considering second-level time series wind speed change |
| CN115345222A (en) * | 2022-07-27 | 2022-11-15 | 西安理工大学 | Fault classification method based on TimeGAN model |
| CN115423810B (en) * | 2022-11-04 | 2023-03-14 | 国网江西省电力有限公司电力科学研究院 | Blade icing form analysis method for wind generating set |
-
2023
- 2023-03-24 CN CN202310292758.6A patent/CN115994325B/en active Active
Also Published As
| Publication number | Publication date |
|---|---|
| CN115994325A (en) | 2023-04-21 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN115994325B (en) | Fan icing power generation data enhancement method based on TimeGAN deep learning method | |
| CN109635928A (en) | A kind of voltage sag reason recognition methods based on deep learning Model Fusion | |
| CN109359697A (en) | Graph image recognition methods and inspection system used in a kind of power equipment inspection | |
| CN108491877A (en) | A kind of classification convection weather probability forecast method and device | |
| CN116229380B (en) | Method for identifying bird species related to bird-related faults of transformer substation | |
| CN112925824A (en) | Photovoltaic power prediction method and system for extreme weather type | |
| WO2024051524A1 (en) | Joint prediction method and apparatus for hydraulic, wind and photovoltaic generation power | |
| CN114169445A (en) | Day-ahead photovoltaic power prediction method, device and system based on CAE and GAN hybrid network | |
| CN114021822A (en) | Clean energy power generation power prediction method and system | |
| CN111967675A (en) | Photovoltaic power generation amount prediction method and prediction device | |
| CN115829145A (en) | Photovoltaic power generation capacity prediction system and method | |
| CN112084860A (en) | Target object detection method and device and thermal power plant detection method and device | |
| CN114386654A (en) | Multi-scale numerical weather forecasting mode fusion weather forecasting method and device | |
| CN112327190B (en) | Method for identifying health state of energy storage battery | |
| CN116910633B (en) | Power grid fault prediction method based on multi-modal knowledge mixed reasoning | |
| CN112327189B (en) | Comprehensive judging method for health state of energy storage battery based on KNN algorithm | |
| CN104834816A (en) | Short-term wind speed prediction method | |
| CN117033923A (en) | Method and system for predicting crime quantity based on interpretable machine learning | |
| CN110543675A (en) | Power transmission line fault identification method | |
| CN113689053B (en) | Strong convection weather overhead line power failure prediction method based on random forest | |
| CN114519293A (en) | Cable body fault identification method based on hand sample machine learning model | |
| CN113779892A (en) | Wind speed and wind direction prediction method | |
| CN117688367B (en) | Wind power generation ultra-short term power prediction method and device based on instant learning | |
| CN117477551A (en) | Photovoltaic power generation power prediction method and system based on domain knowledge embedded model | |
| CN117131783B (en) | Multi-mode learning-based power transmission line risk prediction model, method and system |
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 |