Short-Term Rainfall Forecasting with E-LSTM Recurrent Neural Networks Using Small Datasets

Abstract

This paper proposes an ensemble of forecasting methods based on neural networks/recurrent neural networks (E-LSTM). The aim of the algorithm is to help organizing the planting cycle using short-term rainfall forecasts when the data are taken from a single observation point. The computational models are carried out for univariate rainfall time series by of multi-step prediction horizons in combination of nonlinear autoregressive models (NAR) modified by several approaches such energy associated to series, subsampling methods and their combinations, which are heuristically modified by Bayesian inference and statistical roughness in the learning process. The study analyses and compares the relative advantages and limitations of each algorithm against the aforementioned to forecast rainfall from 1 to 6 months ahead. Simulation results illustrate the effectiveness of the E-LSTM approach through different series classified by their statistical roughness in both, the learning process and the validation test using the SMAPE and RMSE metrics. Comparisons also are made by adding fractional Gaussian noise to highlight the performance and constraints of the ensemble approach.