Intelligent Wind Farm TechnologiesDr. Liu Yongqian
Professor, School of Renewable EnergyState Key Laboratory of Alternate Electrical Power System
with Renewable Energy SourcesNorth China Electric Power University, Beijing, China
Email: [email protected]
North China Electric Power University (NCEPU)
Key university that featured by energy and electricity in ChinaMOE(Ministry of Education) and top 16 electric companies in China are council membersEstablished in 195830,000 students, 2 campuses
• Beijing (Headquarter) • Baoding city of Hebei Province ( about 150 km from Beijing)
School of Renewable Energy, NCEPU 1st School of Renewable Energy in China(2007),1st undergraduate program on
Wind Energy and Power Engineering (2006). 1463 students, 100 teachers, 26 professors, 29 associate professors.
7 Research Centers
Solar Energy Research and Engineering CenterWind Power Research Center
Hydroelectric Energy & Engineering Research CenterNew Energy Materials and Photoelectric Technology Center
Biomass Energy Research CenterNew Energy Resources and Urban Environment Research Center
Hydropower Resettlement Research Center
Renewable Energy Science and EngineeringNew Energy Materials and Devices
Hydrology and Water ResourcesHydraulic and Hydro-Power Engineering
4 Undergraduate Majors
Outline
Part 1 Wind Power in China
Part 2 Intelligent wind farm technologies
Part 3 Our research
Part 4 Conclusions
Part 1: Wind power in China
Coal62.24%
hydro19.20%
wind9.21%
solar 7.33%
nuclear2.02%
Coal71.79%hydro
18.32%
wind4.54%
Solar1.49%
nuclear3.82%
Installed capacity mix in China, Dec 2017
Electric energy mix in China in 2017
Electric energy in China in 2017: 64,951 TWh 2, 950 TWh from wind, accounting for 4.5% ,
increased by 24.4%, the third largest electric power sources in China.
New installed capacity in 2017
Cumulative capacity Dec 2017 Increased by
Offshore wind power 1, 160 MW 2, 790 MW 97%
Wind turbine Export 641 MW 3, 205 MW 21%
Wind power in China
Wind power in China
China
AmericaGermanyIndia
SpainEngland
FranceBrazilCanada
Italy
Rest of the world
Top 10 cumulative capacity Dec 2017
China
AmericaGermany
IndiaSpain
England
FranceBrazilCanada Italy Rest of
the world
Top 10 newly installed capacity Jan-Dec 2017
Energy revolution to clean and low-carbon power system
Newly installed capacity in 2017 Cumulative capacity Dec 2017 Increase by
World 52.6GW 539.6GW 10.8%China 19.5GW 188.2GW 11.6%
Proportion (China/world) 37.1% 34.9%
Climate changeEnvironment pollutionFossil fuel resources
Drivers for wind
Challenge and OpportunityNo subsidies after 2020
Part 2: Intelligent wind farm technologies
Site-specific wind farm design
Customized wind turbine design
Precise wind turbine control
Integrated wind farm control
Optimized operational and maintenance strategy
Grid friendly support service control
Data-driven management of wind farm
2 Intelligent wind farm technologies2.1 Why intelligent wind farm technologies are prevailing?
Intelligent sensing device (smart sensors)
Advanced communication (IPv6 based AMI )
AI algorithms(Deep Learning)
Robotics
1. Decrease the Levelized Cost of Energy (LCOE) of wind power
Increase the wind farm production
Decrease the maintenance costs
Improve the grid support ability
2. IT and other enabling technologies Intelligent control
Big Data
Advanced visualization(VR, AR)
Computing power (GPU,TPU, Cloud Computing )
LCOE of wind energy
2 Intelligent wind farm technologies2.2 What is intelligence?
We consider the constrained and well-defined AI which could perceives its
environment and takes actions to maximize its chances of success.
We not consider the General intelligence or strong AI . It is a long-term
goal of AI research.
Definition by AI researches[1,2] :
interact with the environment
achieve a particular goal
adapt to different goals and environments
Intelligent Wind Farm
Functions?
goals?
low-level ability High-level ability ?
[1] Russell Stuart J and P. Norvig(2002). Artificial Intelligence: A Modern Approach. [2] Nilsson, Nils (2009). The Quest for Artificial Intelligence: A History of Ideas and Achievements
Environments?
2 Intelligent wind farm technologies2.3 What is intelligent wind farm?
Definition:
Through intelligent Control-maintenance-
management platform, intelligent wind
farm can continuously improves itself to
maximize wind power generation,
minimize O&M costs and meet
requirements of the power system.Functions
goals
Environments Atmosphere (onshore and offshore)
Power system
Maximize the wind power generation
Minimize the O&M costs
Efficiently meet the power system requirements
Wind turbine generation systems
Sensing and communication
Accurate prediction and diagnoses
Integrated control-maintenance-management system
…
2 Intelligent wind farm technologies2.4 Philosophy of Intelligent wind farm
Life cycle Optimization
Integration of control, maintenance and management
Interaction and fusion of multi-source informationComprehensive sensing and monitoring of atmosphere information, wind farm information and power
system information, etc.
Real-time informational interaction between different devices, systems, and platforms.
Deep data fusion for multi-source unstructured data by big data analysis and machine learning techniques.
Wind farm design Operation and maintenance Retrofitting Decommissioning
Unified management of wind farm energy flow and information flowCoordinated organization of equipment, data and personnel in a wind farm
2 Intelligent wind farm technologies2.5 Characteristics of Intelligent wind farm
EvolutionAdaption
IntegrationDiagnosis
Self-learn form the history data
Optimization by humans
Responding to atmospheric environments
Responding to the power systems
Diagnosis and Health Management
Localization mechanism
remaining component life estimation+
+
System objectData monitoring
Maintenance
Operation-maintenance-
management integration platform
Management
Operation
Automatic decision
2.6 Framework of intelligent wind farm
Wind Farm Production ProcessInformation Flow
Energy Flow
Big Data Framework
Intelligence Applications
SCADA system DHM system Smart sensor WPF server
Energy management platform Intelligent robot Sparepart system Substation server
Storage framework Access framework Scheduling framework Database
Server, operating system Data backup Safety management Data management
Operation control application Maintenance strategy application Production Management Application
2 Intelligent wind farm technologies2.7 Intelligence applications
Intelligence applications for control
Intelligence applications for maintenance
Wind farm energy efficiency assessment based on big data analysis Self-organized wind farm management system Wind farm data quality control
Control strategy optimization based on operational data Combination of reinforcement learning and wind farm control Wind farm maximum wind power potential estimation based
on advance sensing and data analysis.
Multi-objective optimal control of wind plant Fault-tolerant control combined with fault diagnosis High precision wind power forecasting based on large scale
sensor network
Wind turbine Diagnosis and Health Management using Deep learning (supervised learning and transfer learning) Wind turbine component life estimation Preventive maintenance and opportunity maintenance strategy optimization combined with fault diagnosis system and
component life estimation
Intelligence applications for management
2 Intelligent wind farm technologies
• 1、 SMART strategy - National Renewable Energy Laboratory SMART strategy: With the support of The U.S. Department of Energy’s (DOE’s) Wind
Energy Technologies Office Atmosphere to Electrons (A2e) applied research program, the
National Renewable Energy Laboratory (NREL) proposed a strategy named "System
Management of Atmospheric Resource through Technology”.
Strategic objectives:SMART wind power plants will be designed and operated to achieve
enhanced power production, more efficient material use, lower operation and maintenance
and servicing costs, lower risks for investors, extended plant life, and an array of grid control
and reliability features. (A reduction of 50% or more from current cost levels )
Focus:The ability to truly understand, control, and predict the performance of the
future wind plant relies on understanding and tying together a range of physical
phenomena from regional weather systems to the wind flow that passes over
individual wind turbine rotor blades.
Major technical research areas:
( 1 ) Performance, risk, uncertainty, and financing. ( 2 ) High-fidelity modeling,
verification, and validation.(3)Wind power plant controls.(4)Integrated system
design and analysis.(5)Wind power plant reliability.
2.8 Current status and practice of intelligent wind farm
2 Intelligent wind farm technologies• 2、Long-term research challenges in wind energy a research agenda by the EAWE
The European Academy of Wind Energy (EAWE), united the important wind power research universities and institutions from 14 European countries,
discussed the long-term research challenges in the field of wind power, explained the current technology frontiers and technical limitations from 11
different research fields, and raised the issues that should be resolved first in the future development of wind power.[1]
11 research areas:1. Materials and structures 2. Wind and turbulence 3. Aerodynamics4. Control and system identification 5. Electricity conversion6. Reliability and uncertainty modelling 7. Design methods8. Hydrodynamics, soil characteristics and floating turbines9. Offshore environmental aspects10. Wind energy in the electric power system11. Societal and economic aspects of wind energy
[1] Kuik G. V, Peinke J. etc. (2016). Long-term Research Challenges in Wind Energy - A Research Agenda by the European Academy of Wind Energy. Wind Energ. Sci., 1, 1–39, 2016
Some point of views:
(1)The definition of CoE of future wind farms requires a multi-disciplinary approach.
(2)Wind power stations have to fulfil the “L3 conditions”: low cost, long-lasting, and low service requirement.
(3)Specific technical solutions should be proposed because of the multi-scale characteristics of wind farms.
(4) Tremendous data can help the design and operation of a wind power station, data filtering and quality control is very important.
(5)The entire wind energy conversion process should be calculated by well-validated computer programs covering the lifetime.
2 Intelligent wind farm technology• 3、GE's digital wind farm
(1) Wind farm design service based on GE's Predix platform
(2) Digital wind farm optimization service during the operation phase of the wind farm
(3) Operation and maintenance decision service based on operational mode and abnormal state analysis
• 4、Gold wind's smart Wind Farm Solution (1) Centralized monitoring for wind farm cluster and regional service sharing
(2) Preventive maintenance with intelligent fault diagnosis and big data analysis
(3) End-to-end wind farm cluster performance management
(4) Centralized power prediction and energy management in the wind farm cluster level
• 5、Vestas (1) Aerodynamics Upgrades, Extended Cut Out, Power Performance Optimization and Power
Uprate.
(2) Asset optimization management services including monitoring and preventive diagnostics for wind turbines, weather stations and substations
(3) Power prediction, weather prediction and icing prediction services.
3 Our research
3.1 Condition Diagnosis and Health Management of Wind Turbine
Definition• Condition diagnosis and health
management is a technology that provide intelligent health monitoring of equipment to avoid serious failures in the future. This technology is the core technology of predictive maintenance, which can determine the maintenance schedule by sensor data and analysis algorithm.
• Mainly includes:• Early fault diagnosis• Remaining useful life prediction
Observed data RUL
Early Fault Diagnosis
Remaining Useful Life Prediction
Failure Threshold
Con
ditio
n In
dica
tors
Source: NASA Ames Research Center
Condition diagnosis and health management technique bring the following advantages to wind turbine maintenance tasks:
• By identifying the early failures of key components of wind turbine and predicting the remaining useful life, maintenance tasks and spare parts can be prepared in advance. The efficiency of maintenance work is improved, and the downtime of wind turbine is reduced.
• It can assist in determining the root cause of the failure, enabling maintenance staff to perform appropriate operations without consuming too much resources to determine the cause of the failure.
• Avoid unnecessary maintenance work and expenses.
Advantages
Procedure
1. Data Acquisition• Real-time operational data of the wind
turbine under normal condition.• Real-time operational data of the wind
turbine under failure condition.• Maintenance data(condition data)
2. Data Preprocess• Mark operational data based on
condition data• Eliminate the outliers in the data
Acquire DataMaintenance
Data
Sensor Data
Develop Diagnosis Model or Prediction Model
Condition Feature
Extraction
TrainModel
Preprocess Data
Deploy & Integrate
3. Diagnosis or prediction model• Fault diagnosis model• Remaining useful life prediction model
4. Deployment • Cloud deployment• Embedded deployment• Hybrid deployment
Condition Diagnosis and Health Management
Fault models and dynamics model of wind turbine drive system
Condition diagnosis of wind turbine drive train based on vibration analysis
Fault diagnosis of wind turbine based on SCADA data
Fault diagnosis and health management research and system development ofwind turbine
Main research fields:
Research achievement: 4 dissertations (2 PhD dissertations) 11 papers, including 6 SCI and EI 1 invention patent Fault diagnosis and health management system of wind
turbine Second Award of the National Science and Technology
Progress Award in China
Example analysis: Fault diagnosis of wind turbine bearing
Gaussian mixture model for wind turbine
drivetrain performance degradation prediction
Fault Diagnosis of Wind Turbine Bearing
Normal
Ball Fault
Inner Fault
Outer Race Fault@3
Outer Race Fault@6
Outer Race Fault@12
Motivation: The vibration data of
different conditions have very different time domain and frequency domain waveforms.
Combining image recognition methods, using convolutional neural network to achieve fault diagnosis.
Time Domain Frequency Domain
Source: Case Western Reserve University Bearing Experiment Data
Fault Diagnosis of Wind Turbine Bearing
• One-dimensional convolutional neural network diagnosis model
• Extract fault characteristics directly from time domain signals
• Two-dimensional convolutional neural network diagnosis model
• Time spectrum of the vibration signal obtained by short-time Fourier transform
• Extract fault characteristics from time spectrum
Proposed two diagnosis models with different structures
5 Conv1D, 16
5 Maxpooling1DRelu
5 Conv1D, 32
5 Maxpooling1D
Relu
5 Conv1D, 64
5 Maxpooling1DRelu
5 Conv1D, 128
5 Maxpooling1D
Dense, 256
Dropout, 0.5Relu
Dense, 64
Relu
Relu
Dense, 6SoftMax
Time domain vibration sample
2×2 Conv2D, 16
2×2 Maxpooling2DRelu
2×2 Conv2D, 32
2×2 Maxpooling2D
Relu
2×2 Conv2D, 64
2×2 Maxpooling2DRelu
2×2 Conv2D, 128
2×2 Maxpooling2D
Dense, 256
Dropout, 0.5Relu
Dense, 64
Relu
Relu
Dense, 6SoftMax
Frequency domain spectrum
Yuanchi Ma, Yongqian Liu, Zhiling Yang, et al. Multi-channel deep convolutional neural network for wind turbine fault diagnosis method: China, 201710662249.2.[P]. 2017.8
Comparison of results:
Fault Diagnosis of Wind Turbine Bearing
2D CNN1D CNN The darker the color
on the diagonal, the better the diagnostics.
The effect of 2D convolutional neural network on outer race fault identification is better than that of 1D convolutional neural network.
Fault Diagnosis of Wind Turbine Bearing
Compared with the classic model(feature extraction + SVM )
Classic model is less accurate in identifying specific fault locations.
Procedure:• The prediction procedure is divided
into offline part and online part.• The offline part extracts the condition
characteristics(mean, kurtosis) from the health wind turbine, and uses the maximum likelihood method to estimate the parameters of the Gaussian mixture model.
• The online part deploys a well-trained model, using the negative log-likelihood function value as the decline index to predict the decline trend of the drivetrain.
Drivetrain Performance Degradation Prediction
Vibration signal of wind turbine (healthy)
Extracting features
Training GMCM
Real-time vibration data
Extracting features
GMCM
Negative log-likelihood probability
Decline index
Offline part Online part
Xu Qiang. Research on state diagnosis method of wind turbine drive chain. PhD thesis, North China Electric Power University, Beijing, 2015.
Characteristic:
• Extracting fault characteristics of key components of wind turbine based on fault mechanism.
• Suitable for real-time prediction of health condition of high ratio drivetrain.
• Compared with the traditional early warming methods for monitoring the variation of kurtosis distribution, the prediction of degradation trend is over 60% faster.
• Improve the safety and reliability of wind turbines in complex environments.
Performance Degradation(Our method Vs Classic method)
Drivetrain Performance Degradation Prediction
Xu Qiang. Research on state diagnosis method of wind turbine drive chain. PhD thesis, North China Electric Power University, Beijing, 2015.
3 Our research
3.2 Wind Power Forecasting
Research fieldsa. NWP wind speed correctionb. Statistic WPF methodsDeterministic WPF• Data driven• Similar days• WTs groupingProbabilistic WPF
c. Physical WPF methods• CFD pre-calculated flow fields
Research achievements 10 academic dissertations, including 3 PhD
dissertations 55 papers, including 49 SCI or EI 7 invention patents 1 intelligent wind power forecasting system
3.2 Wind Power Forecasting
3.2 a. NWP wind speed correction
3 models (multiple linear regression, radial basis function neural network and Elman neural network) are established for correcting the NWP wind speed error.
Frequency distribution for NWP wind speed error
MonthMean value of wind
speed (m/s)Error of NWP wind
speed Correlation coefficient
NWP Measured RMSE MAEJan. 6.41 5.61 3.33 2.75 0.52Feb. 8.32 6.77 3.64 2.78 0.68Mar. 7.58 6.17 2.95 2.29 0.76Apr. 9.08 7.59 3.27 2.48 0.73May. 7.68 6.01 3.55 2.74 0.62Jun. 8.22 6.97 2.89 2.27 0.68Jul. 6.69 5.53 3.35 2.58 0.44Aug. 7.08 5.43 3.20 2.50 0.59Sep. 6.39 5.10 2.93 2.25 0.58Oct. 6.37 6.04 2.40 1.70 0.71Nov. 5.48 5.69 1.24 0.94 0.92Dec. 11.21 9.1 3.79 2.97 0.77
Comparison of NWP and measured wind speed over the sample year
NWP error characteristics
Results
Annual RMSE variation of corrected and original NWP wind speed (based on last 10 days of each month)
Little NWP error may incur huge error in wind power forecasting because of the cubic relationships between wind speed and wind power.
Liu Y, Wang Y, Li L, et al. Numerical weather prediction wind correction methods and its impact on
computational fluid dynamics based wind power forecasting[J]. Journal of Renewable & Sustainable
Energy, 2016, 8(3):770-778.
3.2 a. NWP wind speed correction
Proposed a multi-to-multi NWP correction method based on stacked denoising auto-encoder (SDAE).
NWP patterns
0 25 50 75 100 125 150 175 200 225 250 275 288
Time/10min
0
5
10
15
20
win
d sp
eed
(m
/s)
NWP wind speed
measured wind speed
mean measured wind speed
mean NWP wind speed
Time series of NWPs and wind speed measured at multi-sites
Multi-to-multi network of SDAE for NWP correction
Comparison and improvement of the proposed SDAE-m-m method and benchmarks
Correction network
Results
Advantages Captured more features and labels of wind speed Ability to learn spatial correlation Correction accuracy improved by 15% and 18%
compared to NN and SVM (RMSE)
Yan J, Zhang H, Liu Y, et al. Forecasting the High Penetration of Wind Power on Multiple Scales Using
Multi-to-Multi Mapping[J]. IEEE Transactions on Power Systems, 2017, PP(99):1-1.
One-site NWP wind speedcorrection methods can not considerthe micro-scale effects, such asterrain, wake, roughness andobstacles, the correction accuracy islimited.
3.2 b. Statistic WPF methods
Framework and architecture of the ensemble SDAEs
WPF framework WPF results
Advantages Applicable for the high penetrated and correlated wind
power in a region Captured the variable correlation patterns of wind and
power output with respect to a range of local factors Higher wind power forecasting and greater improvement
when the correlation of wind power outputs is strong
Comparisons of different regional WPF methods
Proposed a multi-to-multi wind power forecasting method based on ensemble SDAEs.
Yan J, Zhang H, Liu Y, et al. Forecasting the High Penetration of Wind Power on Multiple Scales Using Multi-to-Multi Mapping[J]. IEEE Transactions on Power Systems, 2017, PP(99):1-1.
Wind power forecasting methodswhich only consider the in-siteinformation, can not utilize thetemporal - spatial dependencybetween wind farm cluster, theforecasting accuracy is limited.
NRMSE WF1 WF2 WF3 WF4 WF5 WF6 WF7 Avg
SDAE-m-m 0.15 0.15 0.16 0.15 0.17 0.14 0.15 0.15
NN-1-1 0.19 0.17 0.18 0.17 0.18 0.16 0.16 0.17
SVM-1-1 0.18 0.17 0.18 0.17 0.17 0.16 0.16 0.17
RF-1-1 0.19 0.17 0.19 0.18 0.18 0.17 0.16 0.18
—Data driven
Proposed a dynamic wind power forecasting method based on wind process recognition.
Theoretical power curve and empirical power generation scatters
Comparisons of different WPF methods
Width of power curve Slope of power curve
Wind speed Wind direction Wind characteristics
NWP error characteristics
Clustering method
Running condition of wind turbine generator
K-means/Spectral cluster
Process 1# Process 2# Process N#…
Results
Structure of clustering model
Advantages Considered external changing conditions Forecasting models can be recognized and mapped in real time Improved forecasting accuracy
3.2 b. Statistic WPF methods
Most traditional wind power forecasting methods are static without considering external changing conditions.
—Similar days
11.8%13.5%13.7% 14.9%14.6% 15.1%
0%
4%
8%
12%
16%
RVM ANN
RM
SE
K-means Spectral cluster Direct forecastingYan J, Liu Y, Zhang H, et al. Dynamic Wind Power Probabilistic Forecasting Based on Wind Scenario Recognition[J]. Modern Electric Power, 2016.
Proposed a method to directionally select training samples with similar wind speed features in a given day based on the wind speed cloud model.
Predicted day Similar day 1 Similar day 2 Similar day 3 Similar day 4 Similar day 5
Win
d sp
eed
(m/s
)
4
6
8
10
12
14
Box-plot of wind speed among predicted day and similar days
Forecasting error comparison of the proposed method and traditional method in random days of each month
Time series (15min)
0 50 100 150
Pow
er (M
W)
0
0.4
0.8
1.2
1.6
2.0Predictive power based on similar daysTraditional predictive powerReal power
Forecasting results of two methods based on measured data
MonthRMSE(%) MAE(%)
Similar Random Similar Random
Jan. 18.1% 20.8% 17.7% 21.1%
Feb. 12.5% 15.5% 18.4% 21.9%
Mar. 14.4% 17.2% 14.9% 17.3%
Apr. 13.3% 16.6% 16.0% 18.3%
May. 9.9% 11.6% 15.6% 18.2%
Jun. 6.2% 10.4% 8.3% 11.8%
Jul. 9.8% 13.0% 10.9% 13.9%
Aug. 7.9% 9.6% 9.9% 11.9%
Sep. 7.6% 9.4% 10.9% 13. %
Oct. 9.4% 11.3% 10.4% 12.8%
Nov. 13.7% 16.9% 13.9% 17.0%
Dec. 14.9% 17.8% 19.6% 21.7%
Year 11.48% 14.18% 13.88% 16.58%
Two situations of similar day selection based on cloud model
Advantages Considered the complex diversity and ambiguity of the actual weather system Improved the learning ability in capturing the randomness Improved the fuzziness of wind speed in designated time period Improved forecasting accuracy
Results
Training sample selected
3.2 b. Statistic WPF methods
YAN Jie, XU Chengzhi, LIU Yongqian, et al. Short-term wind power prediction based on daily similar wind speed cloud model[J]. Automation of Electric Power System, 2018, 42(06): 53-59.
—Similar days
Proposed a grouping method for wind turbines by considering the flow correlation.
Flow chart of self-organizing feature mapping wind turbine grouping
Wind speed
Elevation
Prevailing wind direction coordinate system
Wind speed correlation
Grouping model of wind turbine generations
Grouping results of wind turbine generations
Reference points for wind power forecasting
Prevailing wind direction of wind farm
Boundary of wind farm
Location coordinates of wind turbine generations
Structure of wind turbine grouping
Results
Technical route RMSETesttime
Train time
Wind tower as reference point
16.73% 1.4s 288.2s
Representative WTs as reference points
15.67% 9.4s 2304.4s
Each WT as a reference point
15.09% 46.3s 9504.6s
Results of WPF for different technical routes
Conversion diagram of prevailing wind coordinate
Advantages Reflected the flow characteristics of a particular wind farm Facilitated the combination of flow characteristic with the wind
turbines grouping method and wind power forecasting technique Improved the forecasting accuracy and reduced the operation time at
the same time
3.2 b. Statistic WPF methods —WTs grouping
Liu Y, Gao X, Yan J, et al. Clustering methods of wind turbines and its application in short-term wind power forecasts[J]. Journal of Renewable &
Sustainable Energy, 2014, 6(5):474-482.
Only one representative WT→ reduce forecasting accuracyEach WT is forecasted individually→increases the computational cost
Probabilistic forecasting of VVRVM on a day in December Intervals of the wind power for different confidence levels
Proposed a probabilistic wind power forecasting method based on varying variance relevant vector machine.
Advantages Ability to predict deterministic future wind power as well as its
fluctuation range under given confidence level Could well simulate the power generation process of wind under
variable meteorological conditions Just required small amount of relevance samples
Modeling steps
Modeling steps of probabilistic wind power forecasting model
Probabilistic wind power forecasting
Operation data of wind farm
Numerical Weather Prediction Confidence level
Created training samples for each month
Grouped according to NWP error grade
Screened training samples
Normalization processing
Initialized model parameters
Calculated the posterior distribution of weight coefficients
Calculated the mean and variance of the posterior distribution
Stopped the iteration, outputted parameters
Periodically updated
Corrected variances
YES
NO
Yan J, Liu Y, Han S, et al. Wind power grouping forecasts and its uncertainty analysis using optimized relevance vector machine[J]. Renewable & Sustainable Energy Reviews, 2013, 27(6):613-621.
3.2 b. Statistic WPF methods —Probabilistic WPF
ResultsThe probability of forecasting error occurrence in deterministic WPF is almost 100%, which has strong uncertainty.
3.2 c. Physical WPF methods
Terrain model Roughness model
Inflow wind condition discretization
Establishment of CFD simulation method
CFD simulation for flow fields in discrete inflow wind conditions
Wake model
Power curve Database establishment of wind speed and direction
NWP input data
Database establishment of wind power in different inflow conditions
Power prediction of wind farm
Power prediction of single wind turbine
Statistics of the operating wind turbines
Proposed a physical approach of the wind power prediction based on the CFD pre-calculated flow fields.
Structure diagram of wind power prediction
Prediction result when power output decreases from installed capacity to zero
Frequency distribution histogram of forecasting wind power error
Prediction result when power output changes drastically
Advantages Independent of the historical data Applied to newly-built wind farms Short computation time Captures the local air flows more precisely Great performance: 15.2% (RMSE), 10.8% (MAE)
Modeling steps
Results
LI Li, LIU Yong-qian, YANG Yong-ping, et al. A physical approach of the short-term wind power prediction based on CFD pre-calculated flow fields[J]. Journal of Hydrodynamics, 2013, 25(1):56-61.
The CFD method has not been successfully applied to the wind power forecasting system so far because it takes too much time to calculate.
Intelligent wind power forecasting system
Functions
Ultra-short-term WPFShort-term WPFUncertainty analysis of
forecasting resultsAutomatic remote update
of the modelMulti-point NWPRamps forecasting
Advantages
Mathematical theories Adaptabilities Uncertainty analysis
7 ultra-short-term WPF statistic models
4 short-term WPF statistic models
Physical model based on CFD flow field pre-calculation
ARIMABP、RBF
PSVMORVM
GA-PSVMHHT-ANN
Established or new WFPlain or hilly ground WFOnshore or offshore WF
4 uncertainty analysis methods
Power interval or probabilistic distribution under given confidence level
3 Our research
3.3 Optimal operation of wind farm
3.3 Optimal operation of wind farms
3.3 a. Objective 1: Increase power production by considering the wake effects
3.3.b Objective 2: Extend the life of wind farm by considering the fatigue damage
Fast wake distribution calculation
Wind farm optimal dispatch based on GA and PSO
Wind farm optimal dispatch based on multi-agents
Unit commitment optimization in wind farm with the target of reducing fatigue damage
of wind turbines Research achievements : 7dissertations (2 doctoral
dissertations) 9 papers(SCI and EI) 2 invention patent
Research projects: National High-techDevelopment Plan (863) Project"Research and Application of WindFarm-Photovoltaic Power StationCluster Control System"
3.3 a. Fast wake distribution calculation method
1. The wake superposition model of multiple wind turbines(a)Calculation method for wake superposition area
(b)wind speed calculation of downstream unit
2. Fast calculation of wind speed distribution in wind farm under different wind direction
(a)determine the arrange order of wind turbines (b)coordinate transformation (c)Fast wind speed distribution calculation
Intersection are between wake and wind rotor
Gu Bo. Wake Fast Calculation and Power Optimal Dispatch for Wind Farms. PhD thesis, North China Electric Power University, Beijing, 2017.
3.2 a. Wind farm optimal dispatch based on GA and PSO
For a wind farm with multiple wind turbines, the overall power output of wind farm can be increased by choosing an appropriate set of axial inducing factors or thrust coefficients.
Fig.1 layout of wind farm Horns Rev Fig.2 Wake wind speed distribution of wind turbines at the row 4.
Fig.3 Wake wind speed distribution of wind turbines at the row 10.
Fig.4 Optimization control calculation process of PSO algorithm
wind speed 8.5m/s, wind direction 270°Speed calculated by wake model
Measured wind speed
wind turbine
Before Optimization
After Optimization
wind turbine
wind speed 8.5m/s, wind direction 222°
Iteration times
Win
d sp
eed(
m/s)
Win
d sp
eed(
m/s)
Pow
er o
utpu
t
Advantages Accurate and fast wake distribution
calculation to meet the control requirements Increase the wind farm output with different
wind conditions
Axial induction factor or thrust coefficient
Wind farm speed distribution
Wind farm power output
Improvement rate9.96%
Gu Bo. Wake Fast Calculation and Power Optimal Dispatch for Wind Farms. PhD thesis, North China Electric Power University, Beijing, 2017.
3.3 a. Wind farm optimal dispatch based on multi-agents
Multi-agent grid
Ag1,1 …… Ag1,2 Ag1,col
Ag2,1 Ag2,2 Ag2, col
⁞ ⁞
⁞ ⁞
Agrow,1 Agrow,2 …… Agrow,col
1 2 3 4 5 6 7 8 9 105
5.1
5.2
5.3
5.4
5.5
5.6
5.7x 107
迭代次数
风电
场整
体输
出功
率
(W)
MA 优化过程输
未优化前输出功率
风向270度,风速 8.5 m /s
1 2 3 4 5 6 7 8 9 106.45
6.5
6.55
6.6
6.65
6.7x 107
迭代次数
风电
场整
体输
出功
率
(W)
MA 优化过程输
未优化前输出功率
风向222度,风速 8.5 m /s
Objective: Maximize the wind farm productionAssumption: Each wind turbine is an intelligent agent interaction between wind turbines through
the Wake effectsAgent design Domain competition operator Mutation operator Self-learning operator
Advantages reduce the dimension of the solution space make the optimal dispatch solution by
designed agent and Fast wake distribution calculation method.
Improvement rate10.74%
Wind turbines
Wind farm controller
wind speed 8.5m/swind direction 270°
wind speed 8.5m/swind direction 222°Improvement rate
2.93%Po
wer
out
put
Pow
er o
utpu
t
Iteration times Iteration times
Before Optimization
After Optimization
Before Optimization
After Optimization
Gu Bo. Wake Fast Calculation and Power Optimal Dispatch for Wind Farms. PhD thesis, North China Electric Power University, Beijing, 2017.
3.3 b. Wind turbine fatigue load and fatigue damage
Based on GH-Blade and IEC61400-1 standards, the fatigue load of wind turbine and fatigue damage of important components of wind turbine are calculated.
3.83E-08, 0< 934.78E-08, 93< 3261.01E-07, 326< 7531.32E-07, 753< 15001.63E-07, 1500
1.18E-09,0< 15001.28E-09, 1500
2.04E-08,0< 15001.01E-08, 1500
2.41E-12
PP
a PP
P
Pb
P
Pc
Pd
≤ ≤= ≤ ≤
>≤
= >≤
= >=
Equivalent fatigue damage expression of a wind turbine
a- Fatigue damage of wind turbines during normal operation condition.
b - Fatigue damage of wind turbine when starting.
c- Fatigue damage of wind turbine when stopping.
d - Fatigue damage of wind turbine when idling.
P - power output of wind turbine.
Determining the fatigue load condition of wind turbine
Wind turbine model and wind model
Rain flow and Miner fatigue cumulative damage theory
Equivalent fatigue load and relative fatigue damage
Equivalent fatigue load and fatigue damage of blade roots Equivalent fatigue load and fatigue damage of hub center Equivalent fatigue load and fatigue damage of tower bottom Equivalent fatigue load and fatigue damage of tower top
Zhang Jinhua. Research on Unit Optimal Dispatch in Wind Farm. PhD thesis, North China Electric Power University, Beijing, 2014.
3.3 b. Unit commitment optimization for reducing the fatiguedamage of wind turbines
optimization objective1 1
1 1 1 1 1 1 1 1min( ( ) (1 ) (1 ) ( (1 ) ))
T N T N T N T Nj j j j j j j j j j
i i i i i i i i i ij i j i j i j i
F a u t b u u c u u d u t− −
= = = = = = = =
= ⋅ + − + − + − ⋅∑∑ ∑∑ ∑∑ ∑∑
Constraints:(1) unit upper and lower limit constraint(2) load dispatch constraint(3) maximum power change rate constraint
During the whole dispatching period, the starting andstopping of the wind turbines can be selected tominimize the fatigue damage of the whole wind farmunder the constraints.
0 100 200 300 400 5001.26
1.28
1.3
1.32
1.34
1.36
1.38
1.4
1.42
1.44x 10-4
遗传代数
解的变化
进化过程
Convergence of the genetic algorithm
Uints 1 2 3 4 Units 1 2 3 4
1 0 1 0 0 18 1 1 0 0 2 1 1 1 0 19 1 0 0 1 3 0 0 0 0 20 0 0 0 1 4 0 1 0 0 21 1 0 0 0 5 0 1 1 1 22 1 0 1 0 6 0 0 0 1 23 1 0 0 1 7 0 1 1 1 24 1 0 1 0 8 0 0 0 1 25 1 1 0 0 9 0 0 1 0 26 1 1 0 1 10 0 1 1 1 27 1 0 1 1 11 0 0 1 1 28 1 0 0 1 12 0 0 1 1 29 1 1 0 0 13 1 0 1 0 30 0 1 0 0 14 0 0 0 1 31 1 1 0 0 15 0 0 1 0 32 1 1 1 0 16 0 0 1 1 33 1 1 0 1 17 0 1 0 0
The optimal unit commitment
According to the wind power prediction and the dispatching instructions ofthe power system, the optimal unit commitment scheme can be solved byreducing redundant operation and avoiding frequent start-up andshutdown of units.
Optimized by GA
Fitn
ess v
alue
Iteration times
Zhang Jinhua. Research on Unit Optimal Dispatch in Wind Farm. PhD thesis,North China Electric Power University, Beijing, 2014.
Part 4: Conclusions• The ultimate goal of intelligent wind farm technologies is to lower LCOE.
• The integrated framework for Intelligent Control, Maintenance and technical Management System for Wind Farm (ICMMS for wind farm) is presented.
• Our reach on intelligent wind farm technologies are introduced, include Optimal operation control, wind power forecast, condition diagnosis and health management.
• There are still a long way to go for the Intelligent wind farm technologies.
Intelligent wind farm technologies: Bright future!
Thank you!