Operational Partitioning of Precipitation Phase Using Machine Learning
Abstract
Precipitation phase is often determined using simple approaches (such as a temperature threshold) because of the lack of automated in-situ observations, resulting in an overly simplified, yet crucial step of operational hydrological modelling. At near freezing temperatures, the distinction between rain, snow or mixed precipitation is challenging. Seasonally, erroneous partitioning can greatly affect simulated runoff during snowmelt and the state of the snowpack, underlying a need to partition precipitation properly over the winter. At the same time, there is an increase in availability of radar disdrometers, low energy consumption instruments that identify the precipitation phase according to incoming particle characteristics, such as fall speed and size of the hydrometeors. Thus, a network of disdrometers can provide valuable automatic measurements to improve the diagnostic of phase partitioning. The goal of this study is to present a machine learning approach for the precipitation phase partitioning using input variables commonly available in an operational context such as air temperature and relative humidity, as well as disdrometer data. This study uses data from 31 sites on a 1000 km transect, between the 47°N and 53°N. The sites are located mainly in the Côte Nord region, north of the Saint-Lawrence River. The annual average temperature at the sites is of -1.71°C and average peak snow water equivalent is of 261 mm. Each of these sites provides measurements of air temperature, relative humidity, and precipitation rate and phase. The precipitation phase is measured with the WS100 radar disdrometer and identifies solid, liquid, and mixed precipitation phases, at a 15-min time step, which is used as targets for the machine learning. The model is built by using a Random Forest Regression algorithm and is compared to common partitioning models of varying complexity, namely single and double temperature threshold, minimum and maximum temperature function, and psychrometric energy balance. Finally, the performance of the machine learning model is evaluated based on the input variables provided, informing on the necessary variables for the modelling of different precipitation phases.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2022
- Bibcode:
- 2022AGUFM.H22J..06B