A hybrid approach to improving rainfall forecasts

High-resolution rainfall forecasting— the ability to predict rainfall intensities at enhanced resolution in space and time—is one of the most difficult, but also most useful, problems in applied meteorology and hydrology. Such forecasting could help scientists predict floods and manage water resource operations. Despite its well-known benefits,1 however, scientists have made little progress over the years.2,3 Although better understanding of atmospheric physics and faster computing speeds have resulted in improved numerical models of the weather,4 the weather system’s chaotic nature and the lack of knowledge of rainfall physics at higher resolutions2 limit the predictability of rainfall. Statistical extrapolations of radar measurements at enhanced resolutions are useful for short-term (about oneto two-hour lead time) distributed forecasts but have limited use at longer lead times.3 The complexity of the problem and the potential benefits have led researchers to explore statistical and mathematical models, parameterized physics combined with Kalman filters, Gaussian mixtures, and even multifractal models. Investigative studies have demonstrated that neither available physics nor traditional statistical models, nor even data-mining tools such as neural networks, can improve forecasts by themselves.1 Previous results using neural networks have been mixed.5–7 In this article, I present a hybrid strategy for high-resolution rainfall forecasting combining weather physics, traditional statistics, and a neural network-based approach.1 The strategy is able to draw on all available information, account for and use aspects of the domain physics that are better understood, and exploit the strengths of the available data-dictated tools.