Deep learning–based long short-term memory recurrent neural networks for monthly rainfall forecasting in Ghana, West Africa

Abstract

Ghana has witnessed major flood and drought events as a result of unpredictable climate conditions, affecting numerous sectors of the economy such as agriculture, mining, hydroelectric power, and so on. Forecasting rainfall on monthly or seasonal timescales will be critical in averting or mitigating the consequences of these natural disasters, particularly given the present danger posed by climate change. The suitability of long short-term memory (LSTM) recurrent neural networks for monthly rainfall forecasting in Ghana has been investigated in this study. Both univariate and multivariate LSTM models are examined. The numerical results showed that the proposed models performed well in forecasting monthly rainfall in the selected locations across the country. There was no significant difference in performance between the two LSTM models. The values of coefficient of correlation (R), Nash-Sutcliffe efficiency coefficient (NS), root mean squared error (RMSE), and mean absolute error (MAE) ranged from 0.604 to 0.894, 0.352 to 0.679, 50.21 to 75.89, and 36.53 to 44.56, respectively, for the univariate LSTM model, and 0.693 to 0.890, 0.357 to 0.655, 42.10 to 77.04, and 33.07 to 58.60, respectively, for the multivariate LSTM model. The LSTM models’ performance was compared to two standard machine learning models: support vector machine and random forest. Overall, the proposed LSTM recurrent networks outperformed the standard SVR and RF forecasting models. However, all the models (LSTM, SVR, and RF) considered in this study can be used to forecast monthly rainfall across the country with reasonable accuracy.