Performance assessment of rainfall forecasting models for urban Guwahati City using machine learning techniques and singular spectrum analysis

Abstract

Rainfall forecasting is pivotal in improving the lead time for issuing flood warnings and flood management. Machine learning (ML) models are popular as they can effectively manage extensive data and non-stationarity of the data series with improved performance and cost-effective solutions. However, more studies are required to understand the dynamic characteristics of rainfall. This study proposes a hybrid model and demonstrates its efficiency in improving the daily rainfall forecast. Singular spectrum analysis (SSA) was used as a data pre-processing technique (successfully removing and identifying the nature of noise) and coupled with ML models (artificial neural network (ANN) and support vector machine (SVM)) improving daily scale forecast. Since the current response of the hydrological system depends on previous responses, rainfall at the next time step was derived with the previous 2, 3, 5 and 7 days of rainfall. Study shows that the first eigen vector derived through SSA is the trend component which has a maximum contribution of 18.75%, suggesting it can explain 18.75% of the given rainfall series. The 16.42% (eigen vector 2–9) contributes to periodicity, with period of 1 year, 6 months, and 4 months within the data. Conclusively, the hybrid SSA–ML model outperformed the single model for daily rainfall forecasts.