UBC Theses and Dissertations
Post-processing precipitation forecasts in British Columbia using deep learning methods Sha, Yingkai
Medium range precipitation forecasts are a crucial input of hydrology models that provide streamflow information for water resource management and flood risk assessments. Generating accurate and timely precipitation forecasts has been a long-standing challenge in British Columbia (BC), Canada, because of its complex terrain and a paucity-of-data problem. In this dissertation, a novel precipitation forecast post-processing routine for BC is developed to convert raw ensembles into bias-corrected, probabilistically calibrated, and downscaled spatiotemporal sequences out to 7 days. The post-processing routine features a hybrid of conventional statistical methods and state-of-the-art Convolutional Neural Networks (CNNs). In the bias-correction and calibration stage, raw ensembles are converted to an Analog Ensemble (AnEn) first and then reconstructed to physically realistic spatiotemporal sequences using the Minimum Divergence Schaake Shuffle (MDSS). These sequences are further bias-corrected by a CNN that considers climatology and terrain information. In the downscaling stage, a CNN pre-trained with high-quality, high-resolution precipitation analysis in the continental US is applied and transferred to BC without acquiring extra training data. It downscales post-processed precipitation sequences into 4-km grid spacing, which resolves small-scale terrain features. Additionally, for operating the post-processing methods on a near-real-time basis, a CNN-based precipitation observation quality control procedure is developed. It removes suspicious observations and returns clean observations that can be used to measure and verify post-processed precipitation forecasts. This post-processing routine is developed for the Global Ensemble Forecast System (GEFS) 3-hourly precipitation forecasts, and it is tested by the GEFS reforecasts from 2017 to 2019. Station-observation-based verification indicates that the post-processed precipitation ensembles are skillful in the BC South Coast, Southern Interior, and Northeast---watersheds with diverse climatological conditions. Compared to conventional statistical post-processing, the methods in this dissertation achieved roughly a 10% increase of Continuous Ranked Probability Skill Score (CRPSS) in all lead times. The Brier Skill Scores (BSS) of heavy precipitation events are increased up to 60% for both 3-hourly lead times and 7-day accumulated totals. In summary, this dissertation pioneers the combination of conventional statistical post-processing and neural networks, and is one of only a few studies pertaining to precipitation ensemble post-processing in BC.
Item Citations and Data
Attribution-NonCommercial-NoDerivatives 4.0 International