The need for accurate local rainfall prediction is readily apparent when considering the many benefits such information would provide for river control, reservoir operations, forestry interests, flash flood watches, etc. While the data required to make such predictions has been available for quite some time, the complex, ever-changing relationships among the data and its effect on the probability, much less the quantity, of rain has often proved difficult using conventional computer analysis. The use of a neural network, however, which learns rather than analyzes these complex relationships, has shown a great deal of promise in accomplishing the goal of predicting both the probability and quantity of rain in a local area to an accuracy of 85%.
Using BrainMaker neural network software, Tony Hall (a hydrometeorologist from the National Weather Service in Fort Worth, Texas) has developed such a model. Nineteen meteorological variables (e.g. moisture, lift, instability, potential energy, etc.) were used to develope two networks for quantitative predictions--one for the warm season and and one for the cool season. Two additional networks for probability predictions were also generated. Another completely different program, written in C, was developed to allow both the quantitative and the probability networks to run simultaneously with the results appearing on the same computer monitor.
Results to date have been outstanding. In the quantitative model, five categories were used to group the rain fall data (0.01 to 0.49 inches, 0.5 to 0.99 inches, 1.0 to 1.99 inches, etc.) Different tolerances were allowed for each range. For example, the tolerance for the first category was +/-0.2 inches while the tolerance for the higher categories ranged from 0.25 to 0.5 inches. Predictions for the quantitative models have been accurate in a range of 74% to 100% for the five categories with an overall accuracy of 83%.
The probability model used the criteria that a prediction of 30% probability or higher had to correspond to a rainfall of 0.10 inches or more. Otherwise the network output would be considered in error. The accuracy achieved to date for this model is 94% which, when combined with the quantitative results, gives an overall accuracy of 85%.
Sensitivity analysis was performed on the input variables to determine which had the most effect on the output. This will allow the developers to refine the models and improve the accuracy. Since there are six additional sites in Texas that will be included in future studies, means of further automating both the data gathering and BrainMaker operations are being investigated to improve the cost and allow the technology to be used more economically.
Although only two years of training and testing data were available, the results achieved to date are believed to be reliable and consistent enough to be used for forecasting guidance. Since this was the original goal of the project, the use of BrainMaker neural networks to predict local rainfall is now expanding to locations in other parts of the country.