Application of Long Short-Term Memory (LSTM) Network for seasonal prediction of monthly rainfall across Vietnam

Abstract

Seasonal rainfall forecasting is important for water resources management, agriculture, and disaster prevention. Our study aims to provide an automated deep learning method for the seasonal prediction of monthly rainfall at stations in seven climatic sub-regions in Vietnam with lead times of up to 6 months. An appropriate set of predictors was selected based on numerous climate indices and neighbor station data for the period 1980–2020. We developed an adapted deep learning pipeline for both short- and long-term analysis. The predicted rainfall was verified against the observed data using mean absolute error (MAE), root mean squared error (RMSE), and Pearson correlation coefficients. The results showed that our model generally captured well observed data reflected by low error (MAE and RMSE < 0.2) and high correlation (at 0.8–0.9) for all climatic sub-regions. For the leadtimes of 1–3 months, the rainfall predictionsmade using climate indices as predictors were outperformed by those using neighbor stations data; while for longer leadtimes (4–6 months), the climate indices themselve were able to improve the performance. The rainfall predictions of our methods on all three lead times climatological predictions by factoring additional values. However, there is room for improvement in predicting extreme and abrupt shifts in time series patterns.