RUFCO: A Deep Learning Framework to Postprocess Subseasonal Precipitation Accumulation Forecasts

Abstract

Abstract Postprocessing is a critical step in attaining calibrated and reliable probabilistic forecast output from numerical weather prediction models. A novel deep learning framework is proposed to postprocess 20 years of 7- and 14-day precipitation accumulation reforecasts from the Global Ensemble Forecast System at subseasonal time scales (week 1, week 2, and combined weeks 3–4 forecasts) over the contiguous United States. The network builds upon previous studies and is a combination of three parallel-trained components suitable for subseasonal prediction. The first is a ResUnet architecture which learns nonlinear relationships between binned observed precipitation and input images of weather and geographical variables. The second conditions the network to the month-of-year via a feature-wise linear modulation (FiLM) layer. The third helps the network learn when to revert the forecast to that of climatology. The RUFCO (named for its components ResUnet, FiLM, and Climatological-Offramp) forecasts are compared against raw and climatological forecasts as well as those from a state-of-the-art distributional regression postprocessing model, “censored, shifted gamma distribution (CSGD),” and a simple bias-corrected model. At week 1, every method exhibited a competitive advantage over climatological forecasts. At week 2, RUFCO generated forecasts with statistically significant improvement over climatology at 82%–94% of the domain, beating CSGD’s coverage of 76%–90% of the grid points. At week 3, RUFCO’s skillful coverage was 65%–85%, while CSGDs dropped to only 12%–37%. At the longer lead times, RUFCO achieved the highest domain-averaged skill scores across seasons. However, the network tends to “smooth” forecast skill, making it less competitive with CSGD in limited areas with strongly spatially varying biases. Significance Statement Precipitation accumulation forecasts 1, 2, and 3–4 weeks in advance are increasingly in-demand for a variety of decision-making applications around hydrologic forecasting, flood and drought awareness, and wildfire preparedness. However, raw forecasts from numerical weather prediction systems have errors that hinder skill. Postprocessing methods remove those errors and provide more reliable and skillful forecasts. We show that a new neural network technique is an effective and competitive postprocessing tool compared to more traditional techniques.