Predictive modelling from monthly rainfall patterns using imputation approaches combined with multivariate analysis

This study identifies torrential rainfall patterns in Yogyakarta, Indonesia using multivariate and univariate approaches to propose a statistical model for solving associated issues. First, addressing its long-gap missing rainfall data (approximately 52.8%) is crucial. Therefore, classical imputation methods were enhanced by combining them with the bootstrap algorithm. The hybrid of Random Forest (RF) and the bootstrap algorithm, with the lowest Root Mean Square Error (RMSE) of 7.96 and Mean Absolute Error (MAE) of 0.29, is the best statistical method for imputing Yogyakarta rainfall data missing values. Cluster analysis then classified stations into different rainfall regimes; hierarchical clustering analysis (HCA) recognised four distinct, homogenous regions. The multivariate approach, principal component analysis (PCA), homogenised the rainfall series and optimally reduced long-term rainfall data to validate the HCA analysis results. From the 75% cumulative variation, 14 factors for the Dry and the Rainy seasons and 12 factors for the Intermonsoon season were extracted using varimax rotation. Subsequently, the rainfall pattern forecasting was done using a univariate approach, Singular Spectrum Analysis (SSA). Recurrent Forecasting and Vector Forecasting-Singular Spectrum Analysis (RF-SSA and VF-SSA) were proposed by establishing length ( ) and eigentriple ( ) parameters. This forecasting model effectively discriminated noise in a time series trend, producing significant forecasting results. Overall, the best performances are from and for the Rainy and Dry seasons and the Intermonsoon season, respectively, based on the lowest MAE and Mean Forecast Error (MFE). For the dry season, has the lowest MAE (2.7116); the MFE tends to over forecast monthly rainfall data by 0.0126%. For the Intermonsoon season, has the lowest MAE (3.3940); the MFE tends to under forecast by -0.0621%. All results utilised the RF-SSA algorithm. Concisely, the hybrid of the bootstrap algorithm and SSA improves the forecasting results. Different data classifications may advance climate warning systems.