Improving Monthly Rainfall Forecast in a Watershed by Combining Neural Networks and Autoregressive Models

The main aim of the rain forecast is to determine rain occurrence conditions in a specific location. This is considered of vital importance to assess the availability of water resources in a basin. In this study, several methods are analyzed to forecast monthly rainfall totals in hydrological basins. The study region was the Almendares-Vento basin, Cuba. Based on Multi–Layer Perceptron (MLP), Convolutional Neural Network (CNN) and Long Short–Term Memory (LSTM) neural networks, and Autoregressive Integrated Moving Average (ARIMA) models, we developed a hybrid model (ANN + ARIMA) for rainfall prediction. The input data were the one year lagged rainfall records in gauge stations within the basin, sunspots, the sea surface temperature and time series of nine climate indices up to 2014. The predictions were also compared with the rainfall records of a gauge station network from 2015 to 2019 provided by the Cuban National Institute of Hydraulic Resources. Based on several statistical metrics such as mean absolute error, Pearson correlation, BIAS, Nash–Sutcliffe efficiency and Kling–Gupta efficiency, the CNN model showed higher ability to forecast monthly rainfall. Nevertheless, the hybrid model was notably better than individual models. Overall, our findings have proved the reliability of using the hybrid model to predict rainfall time series for water management and can be extensively applied to this sort of application. In addition, this work proposes a new approach to enhance the planning and management of water availability in watershed for agriculture, industry and population through improving rainfall forecasting.