Enhancing Regional Quantitative Precipitation Forecasts Using Machine Learning in Western US Watersheds

Accurate and reliable precipitation forecasting is of paramount importance in informing hydrological decision-making and reservoir operations among many other applications. While traditional post-processing techniques have been long used, their limitations in accuracy have prompted the exploration of Machine Learning techniques.

Hu et al. [1] explored the application of the UNet convolutional architecture for probabilistic quantitative precipitation forecasts at a high spatial resolution. The probabilistic forecasts generated with this approach outperformed those from conventional methods including nonhomogeneous regression and mixed-type meta-Gaussian distribution [2, 3] in terms of deterministic and probabilistic skill scores including the root-mean-square error, the continuous ranked probability score, and the Brier skill scores. However, a prevailing challenge emerged: the generated precipitation forecasts displayed excessive smoothness, thus impeding the precise representation of detailed spatial features.

To address this limitation, this study delves into a comprehensive investigation employing a series of experiments. We propose to use a modified UNet that specifically uses basin scale features for improved prediction of extreme precipitation around regional watersheds (e.g., at the HUC6/8 levels). We also explore the sensitivity of the UNet predictions to different sizes of input domains since spatial features that are outside of the immediate watershed region can be helpful in improving the quality of forecasts. Specifically, this research leverages the 34-yr reforecast dataset based on the West Weather Research and Forecasting (West-WRF) mesoscale model, developed by the Center for Western Weather and Water Extremes. Here, we demonstrate improvements using this deep learning method in the quality of daily accumulated precipitation forecasts relative to leading dynamical models (GEFS, ECMWF) over key watersheds of interest in the western US, namely the Russian River, Yuba-Feather, and Santa Ana watersheds.

Reference:

[1] Hu, W., Ghazvinian, M., Chapman, W. E., Sengupta, A., Ralph, F. M., & Delle Monache, L. (2023). Deep Learning Forecast Uncertainty for Precipitation over the Western United States. Monthly Weather Review, 151(6), 1367-1385.

[2] Ghazvinian, M., Zhang, Y., & Seo, D. J. (2020). A nonhomogeneous regression-based statistical postprocessing scheme for generating probabilistic quantitative precipitation forecast. Journal of Hydrometeorology, 21(10), 2275-2291.

[3] Ghazvinian, M., Zhang, Y., Seo, D. J., He, M., & Fernando, N. (2021). A novel hybrid artificial neural network-Parametric scheme for postprocessing medium-range precipitation forecasts. Advances in Water Resources, 151, 103907.