JP1.10
A Comparison Between Neural Network and Linear Regression Approaches to a Short-Range Quantitative Precipitation Forecasting Problem
Yerong Feng, China Meteorological Administration, Guangzhou, China; and D. Kitzmiller
The National Weather Service has developed a short-range (0-3 h) quantitative precipitation forecast (QPF) system. This automated system utilizes input from remote sensors (radar reflectivity, lightning, infrared satellite) and operational numerical weather prediction model output to produce rainfall probabilities and rainfall amount forecasts for small areas. The probabilities are for rainfall exceeding 0.1, 0.5, 1, and 2 inches (~2.5, 12.5, 25, and 50 mm) during the 3-h period immediately following observation time. The rainfall amount forecasts are for five categories: < 0.1, 0.1-0.49, 0.5-0.99, 1-1.99 and >=2 inches. The probabilities are produced from equations derived through forward-selection linear screening regression. Presently, the rainfall amount forecasts are derived by comparing the probabilities to pre- determined thresholds.
This threshold technique initially sets the forecast to the lowest amount category (< 0.1 inch). If the probability of rainfall >=0.1 inch is 27% or higher, then the amount forecast is increased to the 0.1-0.49 inch category. The forecast is increased further if the probabilities of 0.5, 1, and 2 inches exceed 20%, 15%, and 9%, respectively. The thresholds were derived by finding the values that yielded a bias of ~1.5 within the data sample used to develop the probability equations. They yield categorical amount forecasts that capture a reasonably large portion of the heavier 1- and 2-inch rainfall events without grossly overforecasting the areal extent of the heavy rainfall.
An alternative approach, utilizing a neural network developed by the back-propagation algorithm, is being tested. The predictors presented to the network included all those appearing in the probability equations, plus the probabilities themselves. We conducted an experiment to determine if a neural-network (NN) approach could yield an improvement over this threshholding technique. The accuracy of both approaches was compared within the developmental (training) data sample, containing data from the period 1996-1998, and another sample from 1999. Results of the comparison will be included in the preprint and shown at the conference.
Joint Poster Session 1, Ensemble Forecasting and Other Topics in Probability and Statistics (Joint with the 16th Conference on Probability and Statistics in the Atmospheric Sciences and the Symposium onObservations, Data Assimilation,and Probabilistic Prediction)
Wednesday, 16 January 2002, 1:30 PM-3:00 PM
Previous paper Next paper