Forecasting the average monthly rainfall in the northwest of Iran using teleconnections and machine learning

Forecasting monthly and seasonal rainfall is very important from a hydrological point of view, including the water resources management, as well as from climatic and atmospheric hazards perspective. Monthly and seasonal forecasts are widely used in sectors such as energy and agriculture. On the other hand, it has been proven that large-scale atmospheric teleconnections and indices have an effect on the atmospheric conditions of different parts of the earth. Machine learning methods are widely used to predict different atmospheric variables. The reason for using these methods is the high modeling ability of these algorithms when faced with a large amount of data. At the same time, these algorithms have many other capabilities such as the ability to model uncertainty.    In this article, three machine learning methods, namely artificial neural network, gradient boosting and random forest are used for prediction. The performance of these algorithms will be compared with each other in monthly and seasonal forecasting scales. The studied area is a part of northwestern Iran, including the area of Lake Urmia basin. Three evaluation criteria, namely Nash-Sutcliffe, correlation coefficient and root mean square error were used to evaluate the results. Moreover, by using the Nino 3.4, WP, and NAO indices, the regional average monthly rainfall in the region is predicted with the help of three mentioned machine learning methods. The reason for choosing these three indicators is their effectiveness on the rainfall of the studied area. Two monthly and seasonal modeling methods are compared. Since teleconnections may be effective on precipitation with a delay, different delays of zero to three months were considered for all three studied teleconnections and the best results were obtained. Three methods of machine learning, i.e., random forest, neural network and gradient boosting were used. In the monthly mode, the performance of the gradient boosting method (with a three-month delay for the WP index and the other two indices without delay) was better than the random forest method, and the neural network ranked last in terms of performance. In terms of estimating the extreme precipitation, the gradient boosting method has an acceptable performance. In seasonal modeling, for the three seasons of spring (March-April-May), autumn (September-October-November) and winter (December-January-February), the performance of random forest and for the summer season (June-July-August), the performance of the gradient boosting method were superior to other methods.