Data-driven multi-time-step ahead daily rainfall forecasting using singular spectrum analysis-based data pre-processing

Abstract

Abstract Accurate forecasting of rainfall, especially daily time-step rainfall, remains a challenging task for hydrologists' invariance with the existence of several deterministic, stochastic and data-driven models. Several researchers have fine-tuned the hydrological models by using pre-processed input data but improvement rate in prediction of daily time-step rainfall data is not up to the expected level. There are still chances to improve the accuracy of rainfall predictions with an efficient data pre-processing algorithm. Singular spectrum analysis (SSA) is one such technique found to be a very successful data pre-processing algorithm. In the past, the artificial neural network (ANN) model emerged as one of the most successful data-driven techniques in hydrology because of its ability to capture non-linearity and a wide variety of algorithms. This study aims at assessing the advantage of using SSA as a pre-processing algorithm in ANN models. It also compares the performance of a simple ANN model with SSA-ANN model in forecasting single time-step as well as multi-time-step (3-day and 7-day) ahead daily rainfall time series pertaining to Koyna watershed, India. The model performance measures show that data pre-processing using SSA has enhanced the performance of ANN models both in single as well as multi-time-step ahead daily rainfall prediction.