 | 
|  |
For companies whose bottom line is affected by weather, short-term climate forecasts can be an important tool in operational planning and risk management activities. With the launch of AIR's ClimateCast™, you can now obtain probabilistic short-term climate forecasts based on the latest climate modeling and prediction technologies. We are all familiar with weather forecasts, but what exactly are short-term climate forecasts? Traditional weather forecasts predict the details of specific weather events, typically over a time frame of one day to one week. Climate forecasts, on the other hand, describe the probability of average weather over the course of months and seasons. For example, a short-range climate forecast (a forecast for the season ahead) might indicate there is a probability of a warmer than normal winter ahead, or that there will be more frequent heat waves next summer. Such information can be extremely valuable. If warmer than normal conditions are expected for the Northeast, for example, a major retailer may decide to keep a lower inventory of winter clothing in its Northeast warehouses. An electric utility can plan its energy trading strategy for the upcoming winter according to expected climate conditions in different regions. If there is a forecast of warmer than normal conditions in its service area, the utility can also plan a hedging strategy for its weather-related demand risk. In order to use a climate forecast effectively, one needs access to the following information: the probability distribution derived from the forecast output, and a quantitative assessment of the skill, or accuracy, of the underlying forecasting system. Based on that information, the user would be able to determine the appropriate weight to be assigned to the forecast. Key features of ClimateCast include forecast maps, station-specific full probability distributions, model uncertainty information, forecast skill assessment, a ClimateCast Guide, and the Forecast Description. These features enable the user to correctly interpret the forecasts for various business decisions. ClimateCast forecasts are updated every month and are readily accessible through AIRWeather’s ClimateCast website. Behind ClimateCast ClimateCast includes a number of intelligently engineered components that have been integrated to create an advanced, state-of-the-art, climate forecasting system, shown schematically in Figure 1. Figure 1: The AIRWeather™ CimateCast™ Forecasting System
The Real-time Data Acquisition System ingests several gigabytes of environmental data 24 hours a day, 365 days a year. Within the environmental data are fields representing the observed state of the atmosphere, ocean, and land on a global scale. These data are used to generate initialization fields for the atmosphere, ocean, and land components of the model, the starting conditions for global climate model simulation runs. At the core of the system is a dynamical climate model, the Community Climate Model version 3.6 (CCM3.6), developed at the National Center for Atmospheric Research (NCAR). The CCM3.6 is the culmination of years of research by many of the world’s top climate scientists and represents the state of the art in climate modeling. At AIR, the CCM3.6 has been specifically tuned for short-term climate forecasting. Enhancements include improved land surface initialization and increased model resolution, corresponding to a global grid of approximately 100-mile resolution. ClimateCast utilizes an ensemble forecasting methodology, similar in concept to that used by the US National Weather Service (NWS) and other forecasting centers in Europe. Running the CCM3.6 model in ensemble mode means that the model is run numerous times, starting from slightly different initialization fields derived from oceanic and atmospheric observations. The multiple runs generate an ensemble of possible states of the future climate. Each model run is known as an ensemble member of the forecast. Every month, the AIR's Climate model is run on a cluster of advanced Linux-based multi-processed computers to generate 60 to 80 ensemble runs. Over 600 gigabytes of climate model output is then processed to create probabilistic regional forecasts that extend six months out. The AIR Meteorological team has developed an advanced technique to generate station-specific forecasts derived from the gridded ensemble runs. This technique incorporates grid-to-station location downscaling together with a model-to-station adjustment. These regional and station-specific forecasts are available on the AIR's ClimateCast website. Using high quality historical weather data is critical to generating a station-specific forecast derived from the ensemble of gridded model outputs. AIR’s downscaling process leverages our high quality reconstructed data, which are cleaned data that have been de-shifted and de-trended. The reconstructed time-series is continuous and temporally homogeneous, and behaves as if the station had observed its entire history of weather from its current location and configuration. Reconstructed data are carefully tested for accuracy in generating estimates of temperature [For details see the AIRWeather Risk Report, February 2002 issue]. The test results provide a measure of confidence that the reconstructed data are robust and most appropriate for generating the station-specific forecasts from the gridded ensemble runs. To produce station-specific forecasts, AIR computes regression equations that characterize the relationship of the station data to the climate model climatology for the station location (based on 20 years of CCM3.6 climatology and 20 years of reconstructed station data). Figure 2 illustrates the benefits of using reconstructed data for the downscaling process. The time-series use plots of January averaged minimum temperature over a period of 20 years. The maroon line represents observations for a station in Florida. The green line represents climate model response corresponding to that station location. In the left panel, the station data undergoes a shift resulting from a station move. Because of that shift, the resulting regression equation is poorly formed. In the right panel, the station data has been reconstructed to remove the shift. This correction enables more accurate regression equations to be generated. Figure 2: Benefits of Using High Quality Reconstructed Historical Data in Generating Station-specific Forecasts
In addition to understanding the methodology and key assumptions involved in generating a forecast, backtesting and verification of the underlying forecasting system provide information that enables the user to assign the correct weight to the forecast information for a business decision. Systematic quantification of the forecasting system’s “skill” is not, however, readily available from forecast providers. The process is highly resource intensive and time consuming. Backtesting involves a comprehensive evaluation of the accuracy of the system based on a set of previously made forecasts, or hindcasts. In climate prediction, skill studies are often conducted with hindcast data because it would take several years to generate and evaluate a sufficient number of forecasts that are based on real-time data. Typically, forecast skill is evaluated by comparing how well the hindcasts performed compared to the climatological distribution, in explaining what actually occurred. AIR has dedicated extensive resources to quantify the skill of ClimateCast. A testing procedure was designed and implemented to measure the skillfulness of hindcasts for every other year during the last 20 years. Hindcasts are produced using the same objective approach as forecasts, except that the initialization data comes from the past. The results from the hindcast runs are then scored against historical observations. Scoring is based on the Linear Error in Probability Space (LEPS) skill-scoring method. LEPS scores have been shown to produce robust, equivalent skill assessments that do not suffer from non-linearity complications, such as those experienced using the Rank Probability Skill Score (RPSS). The LEPS score is bounded between -1 and 1, with scores greater than zero indicating that the forecast is more skillful than climatology. Scores less than zero indicate that the forecast is less skillful than climatology. Figure 3 depicts the ClimateCast LEPS skill scoring methodology. Initialization data from historical years are used to generate hindcasts of the months and seasons forward from the initialization date. These hindcasts are then tested via the LEPS scoring method to determine whether the forecasts are more skillful than climatological distributions. Figure 3: LEPS Hindcast Skill Scoring Process
ClimateCast forecasters are available to address your specific questions via email at ClimateCast@air-worldwide.com.
|
