John W Zack AWS Truepower, LLC 185 Jordan Rd Troy, NY 12180

©2013 AWS Truepower, LLC

ALBANY • BARCELONA • BANGALORE

463 NEW KARNER ROAD | ALBANY, NY 12205awstruepower.com | [email protected]

A CUSTOMIZED RAPID UPDATE MULTI-MODEL FORECAST SYSTEM FOR RENEWABLE ENERGY

AND LOAD FORECASTING APPLICATIONS IN SOUTHERN CALIFORNIA

SoCal Atmospheric Modeling MeetingJune 3, 2013

Monrovia, CA

JOHN W ZACKAWS TRUEPOWER, LLC

185 JORDAN RDTROY, NY 12180

©2013 AWS Truepower, LLC©2011 AWS Truepower, LLC

Overview

• Current Modeling System• Output Products and Applications• Near-term Plans for Modeling System Upgrades


The Modeling System


Prediction Tools at AWST

• Numerical Weather Prediction (NWP)- Weather Research and Forecasting (WRF) Model- Advanced Regional Prediction System (ARPS)- Mesoscale Atmospheric Simulation System (MASS)- WRF- Data Assimilation Research Testbed (WRF-DART)

• Non-NWP- Wide range of statistical tools applied to:

• Model Output Statistics (MOS)• Geospatial statistical models• Weather-dependent application models

Example: Wind power plant output model

- Feature detection and tracking


Current SoCal Modeling System Overview

• Numerical Weather Prediction (NWP)- Continental-scale EnKF medium res ensemble- SoCal downscaled NCEP/EC models- SoCal rapid update models

• Advanced Model Output Statistics (MOS)- Dynamic screening multiple linear regression- Other methods in development and testing

• Non-NWP models- Cloud vector model based on satellite images being

implemented for solar forecasting- Geospatial statistical models being implemented


SoCal Modeling SystemJune 2013

MASS6-72

WRF6-72

MASS6-72

MASS12-72

WRF6-72

ARPS2-12

MASS2-12

NAM GFSGEM RR

PIM-C0.25-3

EnKF

GEOSP0.25-3

MOS MOS MOS MOS MOS MOS MOS

Optimized Ensemble Algorithm

Composite Forecast Products

MOS

HRRR

MOS MOS MOS MOS

MOS


Ensemble Kalman Filter (EnKF) Ensemble

• Objectives- Provide potential alternative to NCEP/EC larger scale

models for higher-res model initial and boundary conditions

- Provide flow-dependent spatial error covariances for high-res data assimilation

- Provide indication of forecast sensitivity patterns• Current Use• Experimental configuration – not used in forecast

production


EnKF Configuration

• 24 WRF members- 51 km outer grid- 17 km inner grid

• 84-hr forecast every 12 hours

• 12-hour data assimilation cycles via DART

• Limited data assimilation on inner grid at present- Radiosonde- Satellite-derived winds- ASOS- Mesonet & buoy

• GFS outer grid BCs (perturbed)


Statistical NWP Forecast Sensitivity

• Based on a statistical analysis of an ensemble of “perturbed” NWP forecasts

• Needs an ensemble of statistically significant size

• Maps the relationship of a change in the forecast at the target site to changes in initial condition variables at the the time of forecast initialization

• Case-specific

Forecast Site


Downscaled NWP• Objective:

- Add higher frequency features to larger scale NCEP/EC forecasts due to surface properties and non-linear interaction of atmospheric features

• Approach- Nested grid with 5 km inner

resolution- 72 hour forecast - 6 hour update for NCEP models- 12-hour update for EC GEM

model- 3 MASS model runs- 2 WRF model runs


Rapid Update NWP• Objective:

- Improve 0-12 hour forecasts by frequently assimilating local and regional data with high resolution NWP model in rapid update mode

• Approach:- 2-hour update cycle- 5-km resolution- MASS with 4-hr pre-forecast

observation nudging cycle (4DDA)

- ARPS with 3DVAR data assimilation


Rapid Update Local Data Assimilation

Inferred moisture from satellite visible and infrared imagery

Winds and temperature from SCE met sensor network in the Passes

ASOS, Mesonet & buoy

Winds and temperature from AQMD profilers/RASS network

Temperature, water vapor and cloud water from SCE radiometer @ LAX

VAD winds and reflectivity from NWS 88D radars

1 2 3

45 6


Model Output Statistics (MOS)

• Objective- Reduce the magnitude of systematic errors in

NWP forecasts for specific variables of interest• Approach

- Screening multiple linear regression- Dynamic 30-day rolling training sample- Advanced statistical approaches under

development


Non-NWP Models: Atmospheric Feature Vector Model

• Objective:- Short-term (0-3 hrs) forecasts of weather features

on time scales for which it is difficult to obtain value from the NWP approach

• Approach- Pyramidal Image Matcher (PIM)

• Possible applications- Cloud vector (Currently operating for Hawaii and

being implemented for SoCal)- Radar reflectivity vector (Under development)- Other feature vectors (Under development):


Pyramidal Image Matcher Attributes

• Development history- Originally developed for stereographic video processing.- Adapted by Zinner et. al. (2008) for satellite image processing.

• Multi-scale approach enables the PIM to capture the motion and development/dissipation of clouds at a wide range of scales of motion.

• Estimates coarse cloud motion vector field a larger scales using visible satellite images averaged to coarse resolution.

• Refines cloud motion vector field at successively finer scales until the full resolution image is reached.

• Estimates future images by propagating current image forward in time using the motion vector field.


1330HST

1400HST

Full 1 km Resolution Image

8 km Averaged Image

Step 2: Compute Motion Vectors at 8 km resolution.

Step 3: Use motion vectors to estimate1400 HST 1 km from 1330 image.

Step 4: Average estimated 1 km image to 4 km.

Step 5: Estimate correction to motion vectors using 1330 HST observed 4 km and estimated 1400 HST observed 4 km images.

Step 6: Repeat steps 2-4 at 2 km and 1 km scales.

Pyramidal Image Matcher MethodFirst Phase: Estimate Motion Vectors of Clouds

Step 1: Compute 8-km averaged images.


Pyramidal Image Matcher MethodSecond Phase: Estimate Future Locations of Clouds

• Use motion vectors computed in first phase to estimate future cloud locations• Wind forecasting: Use feature identification techniques to identify potential

ramp-causing cloud features (such as outflow from rain showers) and predict their arrival at wind farms.

• Solar Forecasting: Apply PIM to solar irradiance derived from visible satellite images to predict future solar irradiance.

Observed 60 minute forecast


Non-NWP Models: Geospatial Statistical Models

• Objective:- Very short-term (0-2 hrs) forecasts of weather

variables of interest on time scales for which it is difficult to obtain value from the NWP approach

• Approach- Identify and use time-lagged statistical

relationships- May have simple linear components and complex non-linear

components


Geospatial Statistical Model Example:Time-lagged Spatial Correlations

Clear Sky Factor Estimated from Satellite Brightness Data

Forecast Site Forecast Site

60-Minute Time-lagged Correlation 150-Minute Time-lagged Correlation


Output Products and Applications


Products to Support Load Forecasting


Individual Model Output: Key Application Variables Hourly Regional 2-m Temperature

ImagesAnd

Animations:0-72 hrs

MASS-NAM


Ensemble Composites: Key Application Variables Hourly Regional 2-m Temperature

Images and Animations:0-72 hrs

Ensemble Mean Ensemble Standard Deviation


Ensemble Member Point Data:Day-ahead 2-m Temperature

Ensemble member temperature forecasts for CQT for today(from yesterday afternoon’s runs)


Individual Model Output: Support Variables Boundary Layer Height

WRF-NAM WRF-GFS


Individual Model Output: Support Variables Marine Layer Height

WRF-NAM WRF-GFS

Definition of marine layer based on max RH in PBL and vertical RH gradient


Individual Model Output: Support Variables Marine Layer – Marine RH in the PBL

WRF-NAM WRF-GFS


Individual Model Output: Support Variables Cloud Variables – Global Solar Irradiance

WRF-NAM WRF-GFS


MOS-derived Support Variables: LA Basin MSLP Table


MOS-derived Support Variables: LA Basin MSLP Difference Table


Products to Support Renewable Energy Production Forecasting


Individual Models: 50-m WindsZoomed Images and Animations of the Passes

WRF-NAM WRF-NAM

Tehachapi Pass San Gorgonio Pass


Wind Power Forecasts: TabularAggregated wind power

forecasts (kW)@ SCE substations

Ensemble Composite Wind Forecasts @ SCE met tower sites


Near-term Plans for Upgrades:


Background Upgrades

• Model updates as they become available• Data assimilation system upgrades as they become

available• Assimilation of additional data as it becomes

available in real-time


Advanced Rapid Update Data Assimilation

• Objective: Improve impact of assimilated data on forecast performance

• Approach• Implement GSI with 2-hr update WRF run• Use flow-dependent spatial model error covariance

estimates- From EnKF run? NCEP ensemble? Time-lagged AWST

forecast ensemble?- Use to derive flow-dependent nudging coefficients (i.e.

weights for nudging terms)- Hybrid 3DVAR


Advanced MOS:Decision Tree Regression

• Objective: More effectively reduce the magnitude of systematic model errors in NWP forecasts of specific variables of interest especially for infrequent of extreme events

• Approach- Employ decision tree methods in place of screening

multiple linear regression• More potential to identify and correct non-linear error patterns• Demonstrated to be among the best statistical prediction

techniques for a variety of applications- Use larger training samples where possible

• Advanced non-linear approaches tend to exploit larger samples more effectively


Advanced MOS:Analog Ensemble

• Objective: More effectively reduce the magnitude of systematic model errors in NWP forecasts of specific variables of interest especially for infrequent of extreme events

• Approach- Employ analog ensemble concept

• Compare current NWP forecast to all NWP forecasts in an historical archive with respect to a set of “matching parameters”

• Identify the the N closest forecasts matches• Compile an N-member ensemble of the observed outcomes for the N best matches • Use the outcomes to generate a deterministic (e.g. ensemble mean) or

probabilistic (e.g. ensemble distribution) forecast

- Effectively customizes to MOS to each forecast scenario- Preliminary result suggest it may perform much better

than typical MOS approaches for infrequent or extreme events (with an appropriate sample)

Documents

John W Zack AWS Truepower, LLC 185 Jordan Rd Troy, NY 12180