Improvement of Voting Feature Interval using Empiric Distribution for Prediction of Rainfall Categories on Sub-Seasonal to Seasonal Scale

Abstract

Predictions of rainfall categories on the sub-seasonal to seasonal (S2S) scale based on the Global Climate Model (GCM) output from the European Center for Medium-Range Weather Forecasts (ECMWF) have a prediction time range of up to 46 days ahead. The S2S predictions for the 46 days ahead need 46 models with one model per day. One of the methods for S2S rainfall prediction is the convolutional neural network (CNN) method. However, the CNN requires extensive computational resources. By contrast, the Voting feature interval (VFI) classification algorithm requires few resources, because the VFI training is performed once using all training data. However, the VFI algorithm has low prediction accuracy for classification problems with continuous features such as GCM data. This is due to determining the number of feature intervals based on the number of classes in the VFI training. Furthermore, the vote for each feature interval in the VFI training depends on the number of data in each interval only. This study proposes a modification of the standard VFI algorithm namely the Kernel VFI (KVFI). In KVFI, the vote for each feature interval is determined by the distribution of all data from that feature using the Kernel method. In addition, the proposed method allows for a more flexible determination of the number of feature intervals based on experiments. The newly proposed KVFI algorithm is then utilized to predict the rainfall categories in the Indramayu Regency, Indonesia. The prediction results show that the KVFI algorithm produces higher accuracy than the standard VFI.

Keywords

• Global Climate Model
• Kernel
• Rainfall
• Sub-Seasonal To Seasonal
• Voting Feature Interval