Comparison of Classic Time Series Forecasting Methods and Deep Learning Models for Reference Evapotranspiration Forecasting
Abstract
Reference evapotranspiration (ETo) is a meteorological variable determined by the energy available to vaporize water from a reference crop surface and is a measure of the atmospheric evaporative demand. An accurate estimate of ETo is important in various fields of study, including hydrology, climatology, and agriculture. In addition to real-time estimation, an accurate and reliable forecast of ETo is crucial for water resources management and irrigation scheduling. Although there are different methods available for time series forecasting, only a few studies are comparing their advantages and limitations in the case of ETo, and importantly their forecast reliance on various co-variates is understudied. In this study, a variety of well-established and cutting-edge time series forecasting methodologies have been employed for monthly ETo forecasting. Meteorological data from 114 standardized weather stations across California is divided into four categories based on the length of time series to investigate the effects of the data availability on the performance of forecasting approaches. Moreover, various forecasting horizons from one- to six-month ahead are analyzed, using univariate and multivariate forecasting strategies. Our findings show that SARIMAX (Seasonal Auto-Regressive Integrated Moving Average with eXogenous factors) is the most accurate model for one-step-ahead monthly ETo forecasting, while also being more interpretable than deep learning-based approaches. Nonetheless, SARIMAX cannot be employed for multi-step ahead forecasting, and deep learning models can outperform SARIMA (univariate Seasonal Auto-Regressive Integrated Moving Average) for this task. However, deep learning models generally require more training data and are less interpretable.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2022
- Bibcode:
- 2022AGUFM.H42K1419A