Precipitation Forecasting Using Hybrid Data-Driven Modeling

Precipitation is one of the main hydrometeorological variables since it affects water resources, environment, and natural hazards like flood and drought. Due to the complexity and uncertainty of the precipitation, valid and reliable precipitation forecasting remains a challenge. This study aims to find the key features that are important in developing a valid model for rainfall forecasting. Multiple feature selection algorithms including ICAP, as an information theoretical-based algorithm, and fisher-score, as a similarity-based algorithm, are used to find principal features. In addition, the trend and cycle parts of the rainfall that are decomposed by the Hodrick-Prescott (HP) filter are simulated by the time series models and the machine learning algorithm, respectively. The hybrid model combining machine learning models (KNN, Random Forest) and time series models (AR, MA, ARIMA) is used to forecast rainfall. To find a proper set of features for precipitation forecasting model, different categories including, hydrological variables from NOAA and ECMWF, cloud properties from ISCCP, and large-scale atmospheric circulation from NOAA, are used to represent precipitation formation in different seasons. Indeed, different mechanisms of precipitation formation that varies in different seasons can be determined by a specific set of features. For instance, findings reveal that longwave radiation can be considered as a significant feature in fall season. Results show that although the key features vary in different periods due to different processes of precipitation formation in each season, large-scale circulation like the North Atlantic Oscillation (NAO) and the atmospheric pattern of ENSO (El Niño–Southern Oscillation) with cloud features are important in all seasons for the precipitation forecasting model. In addition, results indicate that the developed hybrid model for representing the trend (linear) and cycle (nonlinear) part of the rainfall achieves a high and satisfactory level of accuracy (R² = 0.8). The high accuracy of the model highlights the role of the key features in precipitation forecasting and importance of linear and nonlinear parts of the rainfall that need to be considered and modeled properly. The product of precipitation forecasting can be used as the input and driver of other models like hydrologic and ecosystem models. In addition, the developed model can be efficiently used in flood warning system to reduce damage and losses.