1. Alan F. Hamlet, Andy Wood, Nathalie Voisin, Dennis P. Lettenmaier Center for Science in the Earth System Climate Impacts Group and Department of Civil and Environmental Engineering University of Washington February, 2005 PDO/ENSO-Based Streamflow Forecasts in the PNW: Climatological Foundation to Quasi- Operational Products

2. Climatological Foundation

3. A history of the PDO warm cool warm A history of ENSO 1900 1910 1920 1930 1940 1950 1960 1970 1980 1990 2000 Pacific Decadal OscillationEl Niño Southern Oscillation

5. Shifts in the location of tropical convection result in an altered jet stream and a shift in the position of the dominant winter storm track. Cool ENSO Storm Track Warm ENSO Storm Track Elevation (m)

6. SWE Feb 1, 2005

7. Effects of the PDO and ENSO on Columbia River Summer Streamflows Cool Warm high low PDO Red = warm ENSO Green = ENSO neutral Blue = cool ENSO

8. Naturalized Summer Streamflow at The Dalles

9. In 6 out of 8 test years, accurate categorical ENSO forecasts (warm, neutral, cool) have been available in June preceding the water year. By October simple persistence gives an accurate forecast. 1998 ✔ 1999 ✔ 2000 ✔ 2001 X 2002 ✔ 2003 ✔ 2004 X 2005 ✔

10. Numerical Nino3.4 Forecasts for Water Year 2005 Forecasted range for Jan. Nino3.4 anomaly is ~ 0.2 to 1.2 C

11. Figures from Neumann et al., 2003, ENSO forced variability of the Pacific Decadal Oscillation, J. Climate, 16 (23), p 3853 The annual PDO index is predictable with about a six month lead time (i.e. ENSO forecast lead time)

12. Streamflow Forecasts

13. April 1 SWE (mm) By January 1 measurements of basin snowpack are sufficient to produce a useful summer streamflow forecast using simple regression-based techniques Summer Streamflow Volume Regression Equation

14. Snow Model Schematic of VIC Hydrologic Model and Energy Balance Snow Model PNW CA CR B GB

15. Validation of VIC monthly time step streamflow simulations of naturalized flow in the Columbia River at The Dalles, OR raw simulation w/ statistical bias correction

16. Background: Forecast System Schematic NCDC met. station obs. up to 2-4 months from current local scale (1/8 degree) weather inputs soil moisture snowpack Hydrologic model spin up SNOTEL Update streamflow, soil moisture, snow water equivalent, runoff 25 th Day, Month 0 1-2 years back LDAS/other real-time met. forcings for spin-up gap Hydrologic forecast simulation Month 6 - 12 INITIAL STATE SNOTEL / MODIS* Update ensemble forecasts ESP traces (40) CPC-based outlook (13) NCEP GSM ensemble (20) NSIPP-1 ensemble (9) * experimental, not yet in real-time product

17. 1916-2002 Naturalized Streamflow at The Dalles, OR Historic Water Years vs January Nino 3.4 Anomalies warmcool

18. 1992 Actual Initial Conditions as of Nov 1

19. 1992 Actual Initial Conditions as of Jan 1

20. 1992 Actual Initial Conditions as of Mar 1

21. Range =16.7% of ensemble summer mean On October 1 Initial soil moisture accounts for about 16% of the range of flows in the subsequent summer. For normal and high flow years the timing is not significantly altered. Role of Initial Soil Moisture in Fall

22. In dry years, streamflow timing is also affected. Range =16% of ensemble summer mean

23. Forecast Skill and Error Characteristics

24. Definitions: Forecast Value: Economic (or intangible) value of the forecast in the context of some management decision process or other systematic response by a particular forecast user. Metrics of this type are designed to show the value to a particular forecast user. Forecast Skill: Forecast performance as measured by a quantitative skill metric relative to some standard (e.g. climatology). Skill metrics are designed to objectively compare different forecasts.

25. October 1 PDO/ENSO Forecast Red = observed Blue = ensemble mean January 1 ESP forecast

26. A Simple Skill Metric: Take the central tendency of the forecast (ensemble mean) and compare to the long term mean of observations (climatology) using the MSE of the forecast relative to the observed value being forecast. Skill_1 = 1 - MSE (ensemble mean)/ MSE(climatology) or when evaluating over the historic period: Skill_1 = 1 - MSE (ensemble mean)/ [variance of observations] Where MSE is the mean squared error relative to the observed value being forecast. Note that a skill of 1.0 is a perfect forecast, a skill of 0.0 is equivalent to climatology, and a negative value means that climatology is a better forecast.

27. What’s wrong with the simple metric in the previous example? Metrics that use only the central tendency of each forecast pdf will fail to distinguish between red, green, and aqua forecasts, but will identify the purple forecast as inferior. Example metric: MSE of ensemble mean compared to MSE of long term mean of observations (variance of obs.) Obs Value PDF 1 1 1 2

28. A More Sophisticated Skill Metric: In this case the MSE of the forecast relative to the observation is calculated for each of the ensemble members, and the average of these squared errors over all ensemble members is calculated. Note that in this case both the forecast and the climatology are treated as an ensembles as opposed to a single deterministic trace. Skill_2 = 1 - [  (forecast - obs) 2 /N /  (climatology - obs) 2 /M ] where N is the number of forecast ensemble members and M is the number of climatological observations. The difference between this skill metric and the simple one is that this metric rewards accuracy, but also punishes spread. So a forecast with a tighter distribution and the same central tendency as the climatology will achieve a higher skill than climatology using this metric.

29. Obs Value PDF More sophisticated metrics that reward accuracy but punish spread will rank the forecast skill from highest to lowest as aqua, green, red, purple. Example metric: average RMSE of ALL ensemble members compared to average RMSE of ALL climatological observations. 3 2 1 4

30. 1) A Comparison of Forecast Skill between: a) January 1 ESP forecasts (no climate forecast) b) ENSO-Based Forecasts for October 1, November 1, and December 1 (hydrologic initial conditions plus climate forecast) 2)Assessment of Changes in January ESP Forecast Skill when Forecast is Composited to Reflect : a) ENSO b) ENSO Transition c) ENSO/PDO (epoch) d) ENSO/PDO (annual) 2 Experiments:

31. OctNov Dec Jan ESP Red = Obs Blue = Ensemble Mean

32. OctNov Dec Jan ESP Red = Obs Blue = Ensemble Mean

33. OctNov DecJan ESP Red = Obs Blue = Ensemble Mean

34. Comparison of Skill For Warm ENSO Years

37. Comparison of Skill For Cool ENSO Years

38. Comparison of Skill For ENSO Neutral Years

39. December Winter Climate Forecasts Dominate Hydrologic State Variables Dominate JuneMarch Range =16.7% of ensemble summer mean April 1 SWE (mm)

40. Improvements from Adding Climate Information to January ESP forecasts

41. Changes in Forecast Skill Associated with Climate Information (positive means improved skill) 32 of 48 improved or unchanged 26 of 48 improved or unchanged 35 of 48 improved or unchanged 33 of 48 improved or unchanged

42. Conclusions Fall and early winter ENSO-based long-lead forecasts for the Columbia basin (based on resampling methods) typically have a lower skill than January 1 ESP forecasts based on persistence of hydrologic state (soil moisture and snow), but frequently have higher skill than climatology. Prior to December 1 the ENSO based forecasts are considerably less robust than the Jan 1 ESP forecasts. For ENSO neutral years the skill metrics are questionable due to the distribution of flows within the ensemble. Adding ENSO, ENSO transitions, and interannual PDO climate information to January ESP forecasts improves or leaves the forecast skill unchanged about 70% of the time. An interannual formulation of PDO appears to be preferable to an epochal formulation.

43. Moving from Pilot Applications to Quasi- Operational Forecasting Products

44. 1997-2002: One long-range forecast per year, typically delivered in June or July, for a few streamflow locations in the PNW. Retrospective forecast evaluation and development of water management applications Emphasis is on demonstrating feasibility, useful lead time, and skill of climate-based forecasts. http://www.ce.washington.edu/~hamleaf/DallesForecast.html 2003-present: West-wide forecasts, updated monthly. Greater emphasis on real-time data assimilation designed to improve estimates of initial state variables (snow and soil moisture) and on data processing steps needed to mimic operational forecasts. Greater coordination with operational forecasting agencies (particularly the NRCS) http://www.hydro.washington.edu/Lettenmaier/Projects/fcst/

45. Background: Forecast System Schematic NCDC met. station obs. up to 2-4 months from current local scale (1/8 degree) weather inputs soil moisture snowpack Hydrologic model spin up SNOTEL Update streamflow, soil moisture, snow water equivalent, runoff 25 th Day, Month 0 1-2 years back LDAS/other real-time met. forcings for spin-up gap Hydrologic forecast simulation Month 6 - 12 INITIAL STATE SNOTEL / MODIS* Update ensemble forecasts ESP traces (40) CPC-based outlook (13) NCEP GSM ensemble (20) NSIPP-1 ensemble (9) * experimental, not yet in real-time product

46. Seasonal Climate Forecast Data Sources ESP ENSO/PDO ENSO CPC Official Outlooks Seasonal Forecast Model (SFM) CAS OCN SMLR CCA CA NSIPP-1 dynamical model VIC Hydrolog y Model NOAA NASA UW

47. Background: Estimating Initial Conditions SNOTEL assimilation Assimilation Method weight station OBS’ influence over VIC cell based on distance and elevation difference number of stations influencing a given cell depends on specified influence distances spatial weighting function elevation weighting function SNOTEL/ASP VIC cell distances “fit”: OBS weighting increased throughout season OBS anomalies applied to VIC long term means, combined with VIC-simulated SWE adjustment specific to each VIC snow band

48. Background : Streamflow Forecast Locations in development: Colorado R., Upper Rio Grande Columbia R. basin California Snake R. basin

49. 1916-2002 Historic Water Years vs January Nino 3.4 Anomalies

50. Oct 1 forecast for The Dalles, OR All Years from 1950-2003 for which J. Nino3.4 >= 0.2 AND <=1.2

51. Oct 1 forecast for The Dalles, OR All Years from 1977-2003 for which J. Nino3.4 >= 0.2 AND <=1.2 (Warm PDO)

52. http://www.hydro.washington.edu/Lettenmaier/Projects/fcst/

53. SWE Feb 1, 2005

54. Soil Moisture Feb 1, 2005

56. Applications to Hydropower Operations

57. HoursDaysMonthsYearsDecades Long-term reservoir operations Strategic planning and budgeting Maintenance and Operations Power marketing Short-term reservoir operations Flood forecasting Climate and land use change Immediate operations Next hour power marketing Flood forecasting Relevant Time Scales

58. Naturalized Summer Streamflow at The Dalles

59. Water Management Model VIC Hydrology Model Retrospective VIC forecasts are linked to the ColSim reservoir model to quantify the value of the forecasts. Hamlet, A.F., Huppert, D., Lettenmaier, D.P., 2002, Economic Value of Long-Lead Streamflow Forecasts for Columbia River Hydropower, ASCE Journal of Water Res. Planning and Mgmt, 128 (2), pp 91-101

60. Probability of Exceedance for Spring Surplus Energy Resources in the PNW (Strong relationships to natural gas usage in CA)

61. Nino3.4 Anomaly vs Avg NDJF Peak Electrical Demand

62. Constructing the Status Quo Energy Content Curve (wet year) No Forecast Information Used Forecast Information Used

63. Expected Wet Year Cool PDO/La Nina Increased Available Water in Fall over Status Quo Adjustments to Energy Content Rule Curves

64. Expected Dry Year Warm PDO/El Nino Decreased Available Water in Fall over Status Quo

65. Economic Implications of Shifting Generation to Fall Non-firm marketing with forecasts increases here Traditional non-firm marketing is focused here Alternatives permit release of more water for energy production when value is higher than traditional markets, increasing revenue

66. Economic Value of Long-Range Streamflow Forecasts Alternate Energy Targets Without Forecast With Forecast

67. Energy Production and Revenue from ColSim model simulations Increasing Fall Energy Production

68. Reliability of System Objectives

69. A Forecast Timeline Jun Aug Jun On ~June 1 we have: Current Reservoir Contents ( PNW and CA ) Current Summer Streamflow Forecast (PNW and CA) ENSO forecast PDO forecast CA Demand Surplus PNW Energy for Coming Spring PNW Winter Demand Jan Forecasts:

70. Seattle Public Utilities (in-house statistical forecasts) BPA has recently (last two years) incorporated an ESP forecasting system developed by the PNWRFC and has assimilated the forecasting techniques described here. Seattle City Light (Skagit River) Forecasts provided by private consultant Tacoma Power (Cowlitz and Nisqually Rivers) Forecasts provided by private consultant PacificCorp (Lewis River) Forecasts provided by private consultant U.S. Army Corps of Engineers (statistical forecasts for Dworshak and Libby) Early Adopters of PDO/ENSO Streamflow Forecasts