ABSTRACT

Drought is one of the important and costliest disaster all over the world. With the accelerated progress of climate change, its frequency of occurrence and negative impacts are rapidly increasing. It is crucial to initiate and sustain an early warning system to monitor and predict the possible impacts of future droughts. Recently, with the rise of data driven models, various case studies are conducted by using Machine Learning algorithms instead of using pure statistical approaches. The main goal of this paper is to conduct a drought forecasting study for a weather station located in Marmara Region. For that purpose, firstly, widely used univariate drought index, Standardized Precipitation Index is calculated for Bursa station. Thereafter, both the historical information retrieved from time series data and its wavelet transformation are considered to investigate Nonlinear Auto-Regressive and Nonlinear Auto-Regressive with External Input (NARX) type Neural Network (NN) models. According to a pool of Goodness-of-Fit (GOF) tests, the forecasting performance of the models with various number of hidden neurons are compared. The recent findings of the study showed that considering the data with its wavelet transformation under (NARX-NN) has benefits to increase the capacity of forecasting the drought index.

摘要

干旱是全世界最重要、代价最高的灾害之一。随着气候变化进程的加快，其发生频率和负面影响迅速增加。启动和维持预警系统以监测和预测未来干旱的可能影响至关重要。最近，随着数据驱动模型的兴起，各种案例研究都是通过使用机器学习算法而不是使用纯统计方法进行的。本文的主要目标是对位于马尔马拉地区的气象站进行干旱预报研究。为此，首先为布尔萨站计算了广泛使用的单变量干旱指数标准化降水指数。此后，考虑从时间序列数据中检索的历史信息及其小波变换来研究非线性自回归和非线性自回归与外部输入 (NARX) 类型的神经网络 (NN) 模型。根据拟合优度 (GOF) 测试池，比较了具有不同数量隐藏神经元的模型的预测性能。该研究的最新发现表明，考虑在（NARX-NN）下进行小波变换的数据有利于提高预测干旱指数的能力。