Big Data Analysis in Wind Power Forecasting Pingwen Zhang School of Mathematical Sciences, Peking University Email: pzhang@pku.edu.cn Thanks: Pengyu Qian, Qinwu Xu, Zaiwen Wen and Junzi Zhang The Keywind Technology CO., LTD, Shengyang
Outline Background and Recent Advances Wind Speed Forecasting System Wind Power Forecasting System Conclusions
Outline Background and Recent Advances Development of Wind Power Generation Introduction of Forecasting System Current Advances and Challenges Wind Speed Forecasting System Wind Power Forecasting System Conclusions
Background Power station Controllable Users Power load is predictable Wind farm Uncontrollable Power grid Balance power generation and consumption
Development of Wind Power Generation in China 140000 120000 100000 80000 60000 40000 20000 0 Installed capacity in recent years (unit:mw) Year installed capacity Total installed capacity Global Installed Capacity China 73% Others 27% Global: 51477MW, 44% China: 23351MW, 45%
Why Do Wind Forecasting? Unforecasted wind fluctuations increase energy costs. Unforecasted ramp events may compromise system reliability. Good forecasts may lead to high economic value. Forecasts are essential for effective grid management with high wind penetrations (>5%) A 2million, 100metre-tall wind turbine caught fire in hurricane-force gusts at Ardrossan, North Ayrshire, Scotland (2011). http://www.dailymail.co.uk/news/article-2071633
Cost of Intermittent Wind Unforecasted spin-regulation waste 2.5-7.5% of total energy Results from Arizona Public Service (Acker et al., 2007) Typical range for all studies: $1.5 $4.5/MWh
Forecasting Systems Weather observations set the initial conditions; Numerical weather prediction (NWP) models forecast evolution of weather systems Post-Process Model remove the systematic biases in wind speed; Wind power forecasting models convert wind to power; Actual plant production data provide feedback to improve the statistical models Weather Observations Numerical Weather Prediction Models Post-Process Models and Wind Speed Output Wind Power Forecasting Models and Power Output Feedback
Review of Wind Power Forecasting System: China Name Developed By Time Test Sites Test Time Error(NRMSE) Horizons SPWF-3000 Sprixin Control System Technology Co. Ltd.: Beijing 72h < 20% Wind Power Forecasting System DXWIND Wind Power Forecasting System of China Electric Research Institute Nari Technology Co.Ltd. Dx-Wind Technology Co. Ltd. China Electric Research Institute Xi an Jiaotong University, Jonoon Group 72h <20% 24h Hongger Wind Farm, Neimenggu 2013 < 20% 24h ~ 15% 24h LingYang, JiangSu 2011 ~16.47% Note: Normalized Root Mean Square Error(NRMSE) is the official error standard in China.
Review of Wind Power Forecasting System: US and Europe Name Area Devoloped by Test Time Error Standard Test Places Error: 24h Error: 48h EWind USA TrueWind 2003 NMAE California 11.7% WEPROG German WEPROG 2007 NRMSE South Australia ECMWF England ECMWF Best in 11 operating forecasting systems) Europe Europe 12% Error :72h 2004 NRMSE Europe 17% 2000 NMAE Wusterhus 7% 7.5% 10% en,german 2001 NMAE Alaiz, Spain 19% Note: 1. Normalized Mean Absolute Error (NMAE) is co-used with NRMSE in some countries; 2. NRMSE is 2%-5% larger than NMAE
Challenges in Forecasting Wind Small pressure gradients over large distances: hard to forecast accurately Turbulent & chaotic processes are important: even harder to forecast Local topography can have strong influence: not in standard weather models Wind-power curves are highly nonlinear: small errors in wind = big errors in power Abnormal plants activities: malfunctions, downtime and sub-optimality
Outline Background and Recent Advances Wind Speed Forecasting System Meso-scale numerical weather forecasting system Post-process system Post-process: methods and results Wind Power Forecasting System Conclusions
Research Objectives meso/micro-scale wind speed/power forecasting Downscaling Large area wind farms Single wind farm Single wind turbine
Wind Speed Forecasting System Wind Speed Forecasting System based on Weather Research and Forecasting Model (WRF) Data Pre-Process WRF Real-Time Forecasting Wind Speed Post-Process Wind Speed Prediction Coarse NWP data Measured data from wind turbine Historical measured data Historical predicted data
Initialization Initial field from NCEP Measured Data from wind turbine and wind tower Data assimilation (Kalman filtering,4d-var, etc) Initial field
Model & Numerical Methods Euler equation under terrain following coordinate: where, u,v, w is vector of two horizontal speed and one vertical speed;,, is the speed under terrain following coordinate; is temperature potential,θ ; is pressure; is gravity potential;,, and Θ are external forces ) ( 0 ) ( 0 ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( 0 0 1 p R p p gw V V F V F p g Vw W F p p Vv V F p p Vu U d t t t W t V y y t U x x t Compressible Euler equation
Numerical Methods Finite difference method, Runge-Kutta method Stabilization & Numerical Filtering Boundary Condition
Limitation of WRF Mesh size 1:18km MAE = 0.8979 RMSE=1.0898 Mesh size 2:6km MAE = 0.8769 RMSE=1.0485 Mesh size 3:2km MAE = 0.8881 RMSE=1.0617
Post-Process System Errors in NWP systems: model error complexity of the terrain boundary layer phenomenon Solution: post-process using statistic methods Input: Predicted wind speed, temperature, terrain parameters, et al. Post-Process System Output: Corrected wind speed Feedback and parameters identification
Parameters for Experiments Model Parameters: 1 Place: Nantong,Jiangsu; 2 Test time: Sept 01--Dec. 30 3 mesh size:5km 4 Boundary: Yonsei University Wind turbine parameters: Number:61 Product model:ge1500-77 Rated wind speed:15m/s Error definitions for experiments: 1 MAE N 1 RMSE N N i 1 i V a V N i 1 i m i V a V i 2 m MAE CAP N i i V V N i a 1 m rated wind speed Original MAE: MAE of uncorrected wind speed
Post-Process Models: V a f V, T, P,...) ( p p p V a : Analyzed wind speed ; is random error. V p,t p, P p are predicted wind speed, temperature and pressure. 1 2 3 Linear methods Kalman Filter Least Square Support Vector Machine
Model Output Statistic ---Weighted Linear Regression Model: V a b av p Question: Determine parameters from historical measured data Parameters vary with time; Too much of data will hide the time-variation of the parameters; Too little of data will cause random error. weighted least square regression min n i1 w k V k a av b How to choose the weight coefficients? k p 2
Model Output Statistic ---Weighted Linear Regression Determination of weight for each data: --- The nearest data has more impact on the model than the old data. Measured historical wind speed time series: v,, v, v n 1 v,,v N 0 1 are treated as test dataset. are treated as regression dataset. v, v n,, v 1, v0, v1, N Empirical weight function: w( t) ( t0 t), 0 Such that min N j1 V a ([ a, b], t j ) v a, b are regression coefficients with respect to weight function w(t). j
Model Output Statistic ---Weighted Linear Regression Results with a=0.8, and a training set of 25 days: MAE RMSE MAE Cap Original MAE 24hours 1.379 1.746 9.19% 2.387 48hours 1.524 1.925 10.16% 2.769 72hours 1.539 1.964 10.26% 3.166
Model Output Statistic ---Kalman Filter Principal of Kalman filter: X(k): actual wind speed. Z(k): predicted speed from meso-scale NWP system. Φ(k+1,k) and H(k+1): calculated by linear regression.
Model Output Statistic ---Kalman Filter Results by Kalman filter: MAE RMSE MAE Cap Original MAE 24hours 1.448 1.819 9.65% 2.387 48hours 1.651 2.071 11.01% 2.769 72hours 1.789 2.282 11.93% 3.166
Model Output Statistic ---Least Square Support Vector Machine Least squares support vector machines (LS-SVM) are least squares versions of support vector machines (SVM), which are a set of related supervised learning methods that analyze data and recognize patterns, and which are used for classification and regression analysis. --- Suykens; Vandewalle (1999)
Model Output Statistic ---Least Square Support Vector Machine Results from LS-SVM: MAE RMSE MAE Cap Original MAE 24hours 1.418 1.809 9.45% 2.387 48hours 1.661 2.115 11.07% 2.769 MAE:mean average error: 72hours 1.721 2.231 11.47% 3.166
Outline Background and Recent Advances Wind Speed Forecasting System Wind Power Forecasting System Structure of Wind Power Forecasting System Wind Power Forecasting System: Methods and Results New Improvements under Research Conclusions
Wind Power Forecasting System Preprocess the data for model training Adjust the model Data Cleaning Model Update Model Training Model Evaluati on Linear regression, ANN, time series, etc. RMSE, NMAE Wind Power Forecast
Data Cleaning Missing data handling: linear interpolation Outlier handling: hard threshold Prepare data according to the choice of model
Wind Power Forecasting Model --Linear Regression p t 1 t f ( wt 1 t, pt, T, P, S) Features include: wt 1t : p t T : P : S : : Forecasted wind speed (and high order terms) Historical wind power Temperature Pressure Periodic terms
Results of Wind Power Forecasting Standard error(2014-09-01 2014-09-30): NRMSE (Real Wind Speed*) NRMSE (Predicted Wind Speed**) 24 hours 3.78% 10.87% 48 hours 3.66% 12.59% 72 hours 3.77% 13.21% *: Wind power prediction with real wind speed. **: Wind power prediction with predicted wind speed.
Results Example (2014-08-29)
Results Example (2014-09-12)
New Improvements under Research Linear Regression t-penalty Regression Motivation: Heavy-tailed distribution can reduce the effect of large outliers Details: 1. Linear regression coefficients as initial value 2. instead of as error measure 3. Nonlinear unconstrained optimization algorithms Problems: Parameter selection + algorithmic efficiency
Linear Regression Results(9.14)
t-penalty Regression Results(9.14)
Linear Regression Results(9.15)
t-penalty Regression Results(9.15)
Future Improvements Data Cleaning 1. Linear Interpolation Data Augmentation 2. Hard Threshold Weighted Regression 1. Common Piecewise Regression Model Model 2. Model Selection: Sample Error Minimizing (AIC, BIC, VC dimension) Others New Variables, Wind Farm-Wind Turbine 2 Scale Model, New Data Restoration Structures
Conclusions Wind speed forecasting system is assembled base on WRF Predicted wind speeds are obviously improved after post-process Wind power forecasting models are studied and forecasting system are assembled The error of predicted wind power is smaller than claimed error by many companies in China
Q& A Thanks for your attention!