Rainfall Forecasting using a Bayesian framework and Long Short-Term Memory Multi-model Estimation based on an hourly meteorological monitoring network. Case of study: Andean Ecuadorian Tropical City

Rainfall forecasting is a challenging task due to the time-dependencies of the variables and the stochastic behavior of the process. The difficulty increases when the zone of interest is characterized by a large spatio-temporal variability of its meteorological variables, causing large variations of rainfall even within a small zone such as the Tropical Andes. To address this problem, we propose a methodology for building a group of models based on Long Short-Term Memory (LSTM) neural networks using Bayesian optimization. We optimize the model hyperparameters using accumulated experience to reduce the hyperparameter search space over successive iterations. The result is a large reduction in modeling time that allows the building of specialized LSTM models for each zone and forecasting time. We evaluated the method by forecasting rain events in the urban zone of Cuenca City in Ecuador, a city with large spatio-temporal variability. The results show that our proposed model offers better performance over the trivial forecaster for up to 9 hours of future forecasts with an accuracy of up to 84.4%. The model was compared to its equivalent LSTM model without optimization.