Slide 1

Using Forecast Analogs as Historical Medium Range Guidance for Winter Storm Events Chad M Gravelle and Charles E Graves Saint Louis University Department of Earth and Atmospheric Sciences NWSFO St. Louis, MO Winter Weather Workshop 19 November 2008 Slide 2 A blend of knowledge and experience is critical for effective interpretation of model output. Yet, even seasoned forecasters sometimes fail to recognize significant weather events due to (over)reliance on model QPF values and less attention paid to the causative factors. Composite analysis and forecast analogs (perfect prog approach) can assist with this problem because weather events have quasi- repeatable atmospheric fields and results. The perfect prog approach, which is utilized in this research, has one major weaknesses: the uncertainty that is inherent with deterministic NWP model forecasts (which this methodology depends on for pattern recognition skill). By utilizing forecast analogs in the medium range (HWO phase), a forecaster can gain historical experience and become familiar with the meteorological patterns associated with an event. Motivation Slide 3 The basis for analog guidance is to search a climatological dataset for maps that resemble the current forecast, and then assume that the atmosphere will evolve similar to the historical analogs (adapted from Wilks 1995). Search the 28-yr North American Regional Reanalysis (NARR) dataset against the model forecast (GFS212-40km) for potential analogs. 6 months over the winter season (OCT - MAR) 6 h temporal resolution 20,160 potential analogs Remove duplicate times by choosing the best analog over a 24-h period. 1984011512, 1984011518, 1984011600, 1984011606 Refine and rank the resulting analogs to create products that are useful for medium range guidance. Methodology The Big Picture Slide 4 Determining what constitutes an analog is done statistically using the following techniques: Pattern Correlation Mean Absolute Error Root-Mean-Square Error Anomalies Methodology What is an analog? 300 mb HGHT COR 500 mb HGHT COR 850 mb TMPC MCOR 850 mb TMPC MMAE PMSL MCOR 850 mb HGHT MCOR During the first pass through the NARR dataset, the statistics are computed on the following domains (REGN and MESO). If certain thresholds are not exceeded, the date/time is not considered a potential analog. Slide 5 But what about the thresholds? How were the values of the thresholds determined? A control run was developed that focused on heavy snowfall (6) in the LSXCWA. Over the past 28 winters (80/81 07/08) in the LSXCWA, 30 heavy snow events occurred with snow swaths oriented from southwest to northeast. Methodology The Control Run These events (members) were composited in a system-relative sense based on when the center of the 850 mb low crossed the 91 st meridian (t=0h). Finally, statistics were generated comparing the composite fields against the 30 member database. COOP snowfall for the 48-h period ending at 1200 UTC 19870110 Slide 6 Methodology The Control Run Slide 7 Slide 8 Slide 9 Slide 10 To reduce the 20,160 potential analogs, threshold values were determined based on the control run for the following fields: 300 HGHT COR 0.85 REGN 500 HGHT COR 0.83 REGN 850 TMPC MCOR 0.88 MESO 850 TMPC MMAE 3.8 MESO PMSL MCOR 0.83 MESO 850 HGHT MCOR 0.60 MESO* Once the 20,160 analogs are reduced, duplicate times are removed from the potential analogs. Duplicate times occur due to the variability in system speed (e.g., a slow historical system may exhibit similar patterns to the forecast over a longer period of time). Therefore, the best analog is found over a 24-h period by using the following formula: SUM(COR) - SUM(MAE/3) Methodology Reducing the Potential Analogs Slide 11 After the potential analogs are reduced, the program is rerun to find statistics on the following variables: 300 HGHT COR REGN 500 HGHT COR REGN 700 FRNT COR REGN 850 HGHT COR MESO 850 TMPC COR MESO 850 TMPC MAE MESO 850 FRNT COR MESO 850 THTEADV COR MESO 2m TMPC MAE MESO PMSL COR MESO PWTR COR MESO After new statistics are determined, a results score is computed using the following formula: 850HGHTCOR*3 + PMSLCOR*2 + SUM(COR) - SUM(MAE/3) Methodology Ranking the Analogs Slide 12 Finally, in order to catch possible system direction, statistics are computed for 12 h from the time of the best analog using the matching forecast. -12 h Analog: GFS 084h FCST Analog: GFS 096h FCST +12 h Analog: GFS 108h FCST Using statistics from the 12 h times, results scores are also computed. The final analog rank is determined by using the results scores in the following formula: AVERAGE(m12ANALOG, ANALOG, p12ANALOG) Methodology Ranking the Analogs Slide 13 Winter storm affected the CONUS east of the Rocky Mountains from 31 January through 2 February 2008. Band of >2 snow fell from Kansas/Oklahoma east into the mid- Mississippi Valley and northeast into the Great Lakes region. An axis of >6 snow fell from STL (00Z-12Z 1 February 2008) northeastward into lower Michigan. Results The Sample Case Event was relatively well forecast 3-5 days out. Deterministic models had both good agreement in mass fields and fairly good run-to- run consistency. COOP snowfall for the 72-h period ending at 1200 UTC 20080202 Slide 14 To test the forecast analog approach, the 096h GFS212 forecast was utilized. Results GFS212 20080128/0000F096 Slide 15 The GFS212 20080128/0000F096 was compared against 20,160 potential analogs. After the thresholds were applied: 387 potential analogs 300 HGHT COR 0.85 REGN 500 HGHT COR 0.83 REGN 850 TMPC MCOR 0.88 MESO 850 TMPC MMAE 3.8 MESO PMSL MCOR 0.83 MESO 850 HGHT MCOR 0.60 MESO After duplicate times were removed: 187 analogs 25 out of 30 SW/NE oriented STL heavy snow cases were included within the 187 analogs. Results GFS212 20080128/0000F096 Analogs Slide 16 DATEH300corH500corH850corT850corT850maeF850corTA850corF700corT2corT2maePMSLcorPWTRcorf096TOTALf084TOTALf108TOTALAVG 20080201/00000.9960.9940.8200.9691.10.0980.4120.2750.9432.90.9190.9638.6159.0369.2698.973 20071215/18000.9540.9510.8650.9382.00.3590.5940.0110.9433.20.9460.8658.3697.4115.9237.234 20031214/00000.967 0.7410.8652.90.1910.478-0.0380.9371.90.7540.9347.4327.3976.3817.070 20001214/00000.9650.9690.7540.8793.10.4120.5550.3120.9184.30.9620.9457.6746.5735.4396.562 19871215/06000.9200.8910.8790.9003.40.1120.4090.4590.9202.30.8670.9037.9855.4885.5876.353 19870119/06000.9830.9720.7450.9082.20.1220.4300.2420.9203.40.8450.9337.5683.3997.9636.310 20021224/18000.9170.9050.7090.9671.30.0000.5440.0830.9491.60.8560.9358.1724.0336.4386.215 19930226/00000.9130.8910.7920.8303.40.3170.4630.1230.8473.70.8880.8527.0217.4633.3635.949 19990102/12000.9690.9490.8930.8624.2-0.0930.4310.0210.9434.30.9400.7736.5817.6933.4575.910 19970109/06000.9480.9490.8720.8882.5-0.0270.499-0.0230.8393.70.8120.9327.1784.7755.7225.892 19950107/00000.9770.9550.2240.8873.10.0130.2360.0560.8812.30.5770.9074.9386.1616.2485.783 19911203/00000.9490.9630.4830.9343.50.0750.276-0.0750.9402.70.7970.9245.9625.4235.4095.598 19890221/00000.9790.9640.4580.8344.30.1120.0820.1690.8944.60.7690.8394.8187.1694.7115.566 20020119/12000.9320.9370.1040.7893.30.2390.183-0.0540.9112.30.6730.7424.4706.7155.3375.507 19920114/00000.9780.9640.4210.9362.1-0.0860.436-0.0900.9503.60.7810.8965.9093.2427.1925.448 19860206/18000.9500.9340.8950.9033.30.1070.413-0.1290.9435.50.9550.7146.4975.2644.5695.443 20071228/12000.9330.8880.3710.8992.70.0170.231-0.2250.8962.40.8210.8115.5056.2924.4235.407 19820212/18000.9420.9230.0190.8263.00.0030.2570.0460.9212.60.7570.7534.3755.8705.7205.322 19831204/00000.9730.9660.6680.9256.70.2960.4670.1040.9484.00.9610.9135.9514.5515.3565.286 19971114/00000.9120.8760.7160.8632.90.1290.3790.1260.8894.30.7760.8346.3086.4003.0885.265 20070122/00000.9120.8950.0210.8472.6-0.1830.122-0.0670.9332.60.6080.8283.8336.1735.7775.261 20000214/00000.9640.9650.3800.8064.30.3550.2210.1480.9166.40.8670.8854.5675.8855.2945.249 19940309/18000.9530.9430.2870.9502.20.2160.649-0.1180.7424.00.7260.9135.4944.2605.8995.218 19850110/18000.9050.8740.4020.8275.3-0.1040.346-0.0310.9412.00.8840.9085.2076.8313.5045.180 19960126/18000.9640.9510.3070.7884.40.0080.367-0.0670.8034.30.4150.8703.5357.4674.5255.176 20021204/18000.9400.8930.3610.7843.90.1960.239-0.0540.8882.30.7470.7425.1387.8732.0985.037 19831206/06000.9680.9700.1800.9133.50.1040.3890.0630.9073.40.5650.9304.6142.5627.6214.932 19930216/12000.9660.9630.5160.8154.7-0.1690.1770.0610.9483.70.7020.9264.8396.0423.7784.886 20020302/12000.9420.9220.3900.7544.9-0.1310.308-0.0860.8453.10.7390.7324.2677.3832.9654.872 20000311/12000.9530.9410.4810.9521.5-0.0130.4450.2570.9492.90.8290.9287.0463.3524.1094.836 20070213/18000.9640.9290.6530.8813.60.0300.3540.0770.9653.10.8990.7986.5226.2381.7044.821 In STL heavy snow climo Heavy snow occursHeavy snow does not occur Results GFS212 20080128/0000F096 Analogs Slide 17 Results The Best Analog (NARR 20071215/2100) Slide 18 Results The Best Analog (GFS212 20080128/0000F096) Slide 19 Results The Best Analog (COOP EVENT SNOW 20071217) COOP snowfall for the 96-h period ending at 1200 UTC 20071217 Slide 20 Results The Best Analog (COOP EVENT SNOW 20080202) COOP snowfall for the 72-h period ending at 1200 UTC 20080202 Slide 21 Results Potential Products Slide 22 Slide 23 Slide 24 Slide 25 Slide 26 Slide 27 Slide 28 Slide 29 Slide 30 Slide 31 The goal of the analog forecast approach is NOT to make a forecast; but to provide medium-range guidance for events by using a historical dataset. In addition, a forecaster can quickly gain historical experience and become familiar with the meteorological patterns associated with certain events. The current approach is independent of QPF yet can still provide precipitation results (i.e., we already have the answers). Product possibilities are almost endless, this is only a sampling of what is possible. The analog forecast approach can be applied to any meteorological event as long as a control run can be created. Summary and Future Research Slide 32 The GFS212 is being run at hour 120. When an event approaches the domain it is tracked from 120 h to 96 h to 72 h. This will potentially show consistency in the analogs and associated products. Coming Soon: 2 mesoscale domains in the Midwest and 2 mesoscale domains on the East Coast. Using the GEFS to find the analogsanalogs of potential solutions may produce better results. Reassess how well the analog system performed at the end of the winter and make changes for next year. What are we doing now and what will we be doing that is different? Slide 33 CIPS Analog Webpage: www.eas.slu.edu/CIPS/ANALOG/analog.php CIPS Analog Webpage Slide 34 Questions or comments? gravelle@eas.slu.edu or gravesce@slu.edu You can find this presentation online at: http://www.eas.slu.edu/CIPS/presentations.html Questions