Multivariate Time Series Imaging for Short-Term Precipitation Forecasting Using Convolutional Neural Networks

In this work, we explore the use of Convolutional Neural Networks to forecast discretized rainfall intensity by transforming multivariate time series data into image representations using Recurrence Plots (RP). A Convolutional Neural Network (CNN) architecture which takes these image representations in order to produce classification forecasts is proposed. Experimental results in classifying DOST-PAGASA' ‘s color-coded rainfall advisory yield 96.27% 10-fold cross validation accuracy with a configuration of 24-hour lag and 6-hour forecast horizon.