Abstract

Rainfall forecasting is an important input for decision-making in multiple areas, such as water resource planning and management associated with agriculture, hydropower generation, hydration in soils, and reducing vulnerability and risk in the integration of the corresponding systems. However, due to the spatial-temporal variability of rainfall amounts, it is very difficult to achieve high precision in the forecasts. This research addresses the challenge of rainfall forecasting by proposing the application of a methodology based on a combination of techniques within the framework of wavelet decomposition principles, the machine learning approach, and a lagged regression model. We implemented wavelet decomposition in a preprocessing phase followed by the use of a long short-term memory network (LSTM) and proposed a prediction enhancement phase in which the outputs were optimized by algorithms for monthly rainfall forecast corrections. The methodology was implemented at four weather stations in Venezuela, and it was compared with transfer function models, multiple regression and other powerful forecasting methods. The research results suggest that our approach improved the performance accuracy by correcting rainfall forecasting biases, achieving adjusted coefficients of determination greater than 0.76 and normalized mean absolute error (NMAE) values less than 0.31.