3/2/2011

A Turbine Hub Height Wind

Speed Consensus Forecasting System

NCAR’s Wind Energy Prediction System

Developed for Xcel Energy

Bill Myers

Seth Linden

National Center for Atmospheric Research

Xcel Energy Service Areas

Wind Farms (50+)

~3000 Turbines (growing)

~ 3.75 GW (Wind)

~10% Wind

3.4 million customers (electric)

Annual revenue $11B Copyright 2010 University Corporation for Atmospheric Research

NCAR Wind Energy Prediction System Xcel Energy Configuration

Copyright 2010 University Corporation for Atmospheric Research 3

WRF RTFDDA System

NCEP Data NAM GFS RUC MOS

Observations

Wind Farm Data Nacelle wind speed Generator power

Node power Met tower Availability

Ensemble RTFDDA System

Environment Canada

GEM Global GEM Regional

Operator GUI

Meteorologist GUI

WRF Model Output

Wind to Energy Conversion Subsystem

Dynamic, Integrated Forecast System

(DICast®)

CSV Data

3/2/2011

Xcel DICast® System

Overall Xcel system goal:

Generate accurate forecasts of power

Approach taken by NCAR :

First generate accurate forecasts of wind speed at

the hub height of each turbine

Then derive power from individual turbines from

wind speed

As a consensus point forecast system, DICast was a

logical choice to generate the HH wind speed

forecasts

DICast tries to predict Nacelle wind speed sensor

value

3/2/2011

Xcel DICast ® System Diagram

Wind Speed

Forecasts

GFS DMOS

NAM DMOS

RUC DMOS

Data

Ingest

Integrator

RT-FDDA WRF DMOS

Post

Processing

Ensemble Mean

. . GEM DMOS

DMOS means Dynamic Model Output Statistics

3/2/2011

Xcel DICast® System

Hub Height Wind Speed not explicitly

predicted by NWP models

Like other DICast variables (POP, etc)

Predictors relevant to HH Wind Speed

must be derived from NWP data

All attempt to directly predict HH Wspd

Use 10m Wspd and P-level Wspds

3/2/2011

Hub Height Wind Speed Predictors

1 : Interpolate between 10m and first P-

level above winds

2 : Extrapolate* up from 10m winds

3 : Extrapolate* down from next P-level

above winds

etc

* log-based extrapolation with roughness

dependent on land-use

Hub Hgt

First P-level above

First P-level below

10m wind level

Second P-level above

3/2/2011

Dynamic MOS

• Linear regression-based statistical method

• Similar to NWS MOS, but regressions built

dynamically

• New equations generated each week

• System learns based on recent data

• Can be applied to any NWP forecast model fairly

easily

• Tuned forecasts can be generated quickly for new

sites

• Uses “default equations” if statistical model fails

3/2/2011

NAM-DMOS Performance

3/2/2011

Xcel DICast ® System Diagram

Wind Speed

Forecasts

GFS DMOS

NAM DMOS

RUC DMOS

Data

Ingest

Integrator

RT-FDDA WRF DMOS

Post

Processing

Ensemble Mean

. . GEM DMOS

DMOS means Dynamic Model Output Statistics

3/2/2011

Forecast Integrator Objectives

• Discovers the “best” combination of forecast modules for a given forecast time and location.

• Computationally simple and robust.

• Can easily adapt to the addition of new modules or removal of obsolete modules.

• System learns as weights are updated daily • Weights nudged in direction of gradient in weight space

To combine forecasts from a set of models:

3/2/2011

Forecast Integrator Forecast error as function of W1 & W2

W2

W1 W1(i)

W2(i)

Integration Step

0

1

1 0

3/2/2011

Integrator Performance

Copyright 2010 University Corporation for Atmospheric Research

3/2/2011

Conclusions DICast reduces forecast errors with multiple

optimization steps

DMOS provides optimized forecasts from

individual NWP models

Integrated forecasts better than forecasts

from each individual NWP model

Summer predictions more difficult than

fall/winter

High resolution modeling by itself is not the

optimal solution to wind energy forecasting

DICast is a robust technology providing

documented forecast improvements