340 likes | 433 Views
Near real-time predictions of salinity intrusion in a river-dominated estuary: tales and implications of a challenging cruise. Baptista, Y. Zhang, G. Law, J. Needoba, N. Hyde, S. Frolov, P. Turner, M. Wilkin, C. Seaton, B. Howe, D. Hansen. Modified from a presentation to the
E N D
Near real-time predictions of salinity intrusion in a river-dominated estuary: tales and implications of a challenging cruise Baptista, Y. Zhang, G. Law, J. Needoba, N. Hyde, S. Frolov, P. Turner, M. Wilkin, C. Seaton, B. Howe, D. Hansen Modified from a presentation to the Unstructured Grid Workshop, Halifax, Sep 2008
Outline The “mature” observatory The “inconvenient” cruise The short term “fix” Jul 2008 since 1996 Jul 2008 “retrospective “analysis Open benchmark Skill metrics Aug-Sep 2008 Sep 2008 Sep 2008
Conclusions • The river-dominated CR river-to-ocean system provides major scientific and management challenges • The end-to-end observatory SATURN offers a modern and comprehensive monitoring and modeling infrastructure • Under-predicted SIL in a recent cruise has challenged the SATURN modeling skill, leading to a new benchmark • SELFE has met most of the benchmark challenges through added resolution. But, will other codes do better? • Allied with Opendap-CF standards, an Open CR benchmark could offer a stringent snapshot of modeling skill across leading-edge models, with automated updates • The goal is to unite and cross-inform (not divide) the multiple unstructured-grid model communities We invite broad participation!
SATURN: an end-to-end observatory Observation network Cyber-infrastructure Researchers Daily forecasts Educators Simulation databases Students Scenarios Managers Network optimizations … Stakeholders Modeling system
Observation network • CORIE stations • SATURN “endurance” stations • SATURN “pioneer” stations • Land-based remote sensing • Context networks: • SATURN mobile platforms • CMOP cruises 1 Slocum glider 2 REMUS-100
Circulation modeling system Function • Support cruise planning, execution and analysis • Characterize processes • Characterize long-term variability • Characterize and anticipate change • Re-design observation network Mechanisms • Daily forecasts (multiple) • Multi-year simulation databases (multiple; since 1999) • Scenario simulations • Climate • Human activities • Plate displacement Redundancy (models/simulations) as philosophy Codes (past): QUODDY, ADCIRC, POM Codes (current): ELCIRC, SELFE
What makes SELFE the current default model • Robustness • Ability to represent complex circulation processes and features, as required by CMOP research • Computational efficiency • MPI SELFE v2.0g • Intel Xeon 2.3GHz cluster (canopus) with GBit connection • ~27K horizontal nodes; 24 S levels; ~30m minimum equiv. diameter • with 30s step: ~9x faster than real time • with 50s step: ~15x faster than real time • ** See Joseph Zhang’s presentation, Friday afternoon **
Blind retrospective cruise analysis – estuary Salinity LMER - observations SELFE simulation Cruise data courtesy D. Jay psu psu … shows ability to represent complex and episodic features June 1999
Blind retrospective cruise analysis – plume ● Cruise data X SELFE simulation RMSE=2.64 psu correlation = 0.80 Data courtesy D. Jay (RISE project) Pt Sur path (surface )
Coarse scale cruise planning/analysis Minimum surface salinity in the plume over cruise period Maximum bottom salinity in the estuary over cruise period Cruise data courtesy L. Herfort and M. Smit Total RNA content from the Aug 2007 CMOP cruise
Forecast skill: prediction of plume location Cruise data courtesy B. Crump
Goal: validate simulation of SIL (Salinity Intrusion Length) SIL has a clear response to river discharge, and is being consider as a possible “sentinel” for CR variability and change
SIL: difficult to measure … (a) Data collected by David Jay on LMER and NOAA cruises Chawla, Jay, Baptista, Wilkin and Seaton, CSR 2008
… and difficult to simulate (forecast; fDB16; July 13) 07:23 08:41 09:00 09:32 10:09 Cruise data courtesy J. Needoba
Exploring options (in forecast mode, during the cruise) • Data assimilation (DA) • Method of Frolov et al. 2008 • Model-independent • Reliant on fast model surrogates (SVD decomposition, machine-learning trained … • Grid refinement • nchannel • schannel
Grid refinement Refined grid (nchannel) grays cbnc3 mottb fDB16
Bottom salinity (forecasts; July 17) July 17 0:30am Tide (at grays ) 1.6m -1.5m July 16 July 17 fDB16 July 17 0:30am July 17 0:30am DA goes here nchannel DA trained on DB16
CMOP July 2008 cruise: Real-time forecast July 17 2008 09:59 fDB16 DA nchannel da
Salinity at challenging stations (forecasts, July 16-17) mottb fDB16 nchannel DA cbnc3 nchannel da nchannel DA fDB16
Grid refinement # nodes: 27416 # elements: 53314 # levels 24 min element area: 942 m^2 max element area: 89834 m^2 Refined grid (“hires”) grays eliot fDB16
“hires” hindcasts (eliot; Oct 2004) DB16 DB16 hires t=30sec hires t=30sec
“hires” hindcasts (eliot; Oct 2004) hires t=30sec DB16 hires t=50sec hires t=75sec
Forecast (grays; Sep 15-16 2008) fDB16 RMSE= 5.2 psu Salinity ` fhires; t=20sec RMSE= 1.6 psu
Definition of Skill Assessment metrics See : http://www.ccalmr.ogi.edu/~cseaton/tmp/dec06/pub/index_page.html
Forecast skill assessment (fhires; Sep 15-16, 2008) Correlation skill IOA Biofouled sensor Degraded sensor Biofouled sensor Stations RMSE N Telemetry interrupts
Hindcast skill assessment (sandi; salinity; correlation) tide Correlation skill
CR context and issues 1997 Q (m3/s) 2002 2001 • Climate forcing • Pacific Decadal Oscillation & ENSO (precipitation, ocean climate) • Global climate change • (sea level rise, snow pack) Winter 01 courtesy J. Barth N W E Barnes et al. 1972 S N E W E S Summer 01
System response to forcing: estuary Salinity (psu) Tide range (m) Q (m3/s) Salinity intrusion am169
CR open benchmark • Similar to NOAA’s Delaware Bay “model evaluation environment”, in that it enables cross-model comparisons • Distinct in estuary type (river-dominated estuary) and philosophy • Enable continuous enhancement of multiple models and exploration of diverse modeling strategies • Maximize value-added expertise of model developers/expert users, while minimizing their time investment • Dynamic timeframes (blending controlled hindcasts with continuous blind forecasts) • Focus on unstructured grid models • Implementation phases • CMOP-driven SELFE pilot (on-going) • CMOP-assisted pilots for other lead models with by-invitation participation of the respective developers / expert users (a ~12 month effort) • Open to community (early 2010) and consider exporting (2011) • Enablers • CMOP’s SATURN modeling system & Rapid Deployment Forecasting System • OpenDAP-CF standards for unstructured grid models (synergistic effort led by Rich Signell, with participation of at least the FVCOM, ADCIRC, SELFE, ELCIRC communities)
Code registration Registration of modeling strategy Reference static benchmark ? Refine modeling strategy All static benchmarks Operational forecasts ? Simulation databases ? ? Scenario simulations ? ? Forecast benchmark