ABSTRACT

Accurate drought forecasting is necessary for early warning of drought hazards, water resources and eco-environmental management. This study assessed the capability of the Prophet model to forecast meteorological drought based on the standardized precipitation index (SPI). A comparative assessment was conducted among Prophet, support vector regression (SVR) and multiple linear regression (MLR) models to test the applicability of the new model. The SPI was computed at multiple time scales ($\mathrm{S}\mathrm{P}{\mathrm{I}}_3$, $\mathrm{S}\mathrm{P}{\mathrm{I}}_6$, $\mathrm{S}\mathrm{P}{\mathrm{I}}_{12}$ and $\mathrm{S}\mathrm{P}{\mathrm{I}}_{24}$) for 38 meteorological stations located in semi-arid areas in western India. Minimum redundancy maximum relevance (mRMR) was applied to select input variables prior to the model development. The results reveal that the Prophet model yielded acceptable accuracy for drought forecasting, whereas SVR and MLR models showed greater error for short-term forecasting in terms of coefficient of determination (${\mathrm{R}}^2$) and Nash-Sutcliffe efficiency (NSE). Therefore, the Prophet model is recommended as a new robust model for drought forecasting.

摘要

准确的干旱预报是干旱灾害预警、水资源和生态环境管理的必要条件。本研究评估了 Prophet 模型基于标准化降水指数 (SPI) 预测气象干旱的能力。对 Prophet、支持向量回归 (SVR) 和多元线性回归 (MLR) 模型进行了比较评估，以测试新模型的适用性。SPI是在多个时间尺度上计算的（$\mathrm{小号}\mathrm{磷}{\mathrm{我}}_3$,$\mathrm{小号}\mathrm{磷}{\mathrm{我}}_6$,$\mathrm{小号}\mathrm{磷}{\mathrm{我}}_{12}$和$\mathrm{小号}\mathrm{磷}{\mathrm{我}}_{24}$) 用于位于印度西部半干旱地区的 38 个气象站。在模型开发之前应用最小冗余最大相关性（mRMR）来选择输入变量。结果表明，Prophet 模型对干旱预报的准确性可以接受，而 SVR 和 MLR 模型在决定系数方面对短期预报的误差较大（${\mathrm{R}}^2$) 和纳什-萨特克利夫效率 (NSE)。因此，Prophet 模型被推荐为干旱预报的新的稳健模型。