Efficient machine learning techniques for hydrological data assimilation and forecast post-processing

点击翻译按钮获取中文摘要

My research program aims at developing new data assimilation and post-processing techniques that exploit machine learning (ML) to improve hydrological ensemble forecasts. ML techniques have been applied by researchers to solve hydrology-related problems for more than 25 years; the most common application, by far, is streamflow forecasting, where ML techniques (often neural networks) replace process-based hydrological models. Still, as noted in Abrahart et al. (2012), with very few exceptions, operational forecasting agencies do not use ML-based models for streamflow forecasting. Operational forecasters and water resource managers are held accountable for their decisions, and they must be able to support these decisions on sound reasoning based on hydrological processes. The black box nature of ML techniques may be one of the reasons for their lack of popularity among practitioners. Nevertheless, operational hydrological forecasting could benefit from the computing capabilities of ML techniques if they were used to improve portions of the forecasting chain that are strictly mathematical, where relationships with physical processes are not involved. Specifically, ML could improve (1) data assimilation as well as (2) bias correction, post-processing of meteorological forecasts and post-processing of streamflow forecasts. Neural networks (NN, a family of machine-learning techniques) can learn almost any non-linear relationship between inputs and outputs. It is assumed that they could learn the relationship between the simulated streamflow (from any hydrological model) and the corresponding state variables. Once learned, this relationship could be transposed to observed rather than simulated streamflow to obtain corrected state variables. This idea is at the core of the methodology developed for data assimilation in this proposal. Based on a similar principle, NN could also be used to correct bias and dispersion (post-processing) in forecasted streamflows. When fed with meteorological observations, a hydrological model produces streamflow simulations as outputs. When fed with meteorological forecasts, the hydrological model produces streamflow forecasts as outputs. If neural networks could be trained to learn the relationships between simulated and observed streamflow, then this relationship could be applied to forecasted streamflow for correcting model bias (post-processing). The proposed research will allow for more efficient data assimilation and post-processing to advance streamflow ensemble forecasting. The latter is increasingly important for society, as climate change is expected to modify the frequency and magnitude of extreme events.