Modeling and predicting rainfall time series using seasonal-trend decomposition and machine learning

Abstract

This study presents a hybrid approach that integrates seasonal-trend decomposition and machine learning (termed STL-ML) for predicting the rainfall time series one step ahead based on the historical rainfall and other meteorological (e.g., temperature, humidity, etc.) data. The proposed hybrid STL-ML approach mainly consists of three steps: (1) The seasonal-trend decomposition is used to firstly decompose the rainfall time series into the trend, seasonal, and remainder components; (2) Three different machine learning (ML) models, namely Gated Recurrent Unit (GRU) network, multi-time-scale GRU network, and Light Gradient Boosting Machine (LightGBM) model, are developed for modeling and predicting the three components, respectively; (3) The predicted rainfall is eventually acquired by adding up the predicted values of the three components, and several metrics are used to evaluate the model performance. To verify the applicability and validity of the proposed approach, a case study is conducted on the daily meteorology data collected in Cairns, Australia, from 1st Jan 2000 to 31st Dec 2020. The case study results imply that: (1) Through the seasonal-trend decomposition of the rainfall time series, various patterns and information beneath the rainfall time series can be fully extracted and explicitly demonstrated in its three components, which is beneficial to an accurate rainfall prediction. (2) The GRU network, multi-time-scale GRU network, and LightGBM model can well fit and predict the trend, seasonal, and remainder components, respectively. (3) By adding up the predicted values of the three components, the predicted rainfall shows satisfactory agreement with the ground truth, and a reliable one-step-ahead prediction can be achieved even if an sudden extreme rainfall occurs. (4) The comparisons with three baseline methods further justify the rationality and effectiveness of the hybrid STL-ML approach. The novelty of the proposed STL-ML approach lies in its capabilities of (1) fully extracting and utilizing the information in every regard to predict rainfall; (2) providing a good one-step-ahead rainfall estimation for sudden heavy rainfall events. Therefore, it can be used as an essential complement to numerical rainfall prediction and thus can play a crucial role in flood prediction and hydrological disaster control.