Monthly Precipitation Predictions for Flood-Prone Cities in Vietnam: A Deep Learning Hybrid Solution

Protecting lives and property in Vietnam requires reliable rainfall forecasts, given the country’s vulnerability to extreme weather events such as heavy rainfall, flash floods, and storms. On the other hand, forecasts enable authorities to implement preventive measures and warn the public. Hence, research on rainfall forecasting in Vietnam is essential. While deep learning has shown promise for rainfall forecasting, the specific application of hybrid recurrent architectures combining Gated Recurrent Units (GRU) with Long Short-Term Memory (LSTM) and Bidirectional LSTM has not been systematically evaluated for Vietnam's tropical monsoon climate. This study introduces a methodological innovation by comparing GRU-LSTM and GRU-BiLSTM hybrid models against Random Forest (RF) for monthly rainfall prediction in Long Xuyên and Thái Nguyên (2000–2023), using a parsimonious input structure of only three precipitation lags, with 80% used for training and 20% for testing, a significant simplification over covariate-intensive approaches. To avoid excessive complexity and maintain simple modeling, three precipitation lags were considered, with 80% used for training and 20% for testing. The results were analyzed using visual graphs and evaluation criteria of coefficient of determination (R2), root mean square error (RMSE), and Nash-Sutcliffe coefficient (NSE). The results showed that the GRU-BiLSTM model was more accurate in predicting monthly precipitation than other models in Long Xuyen city, with evaluation criteria of R²=0.821, RMSE=13.895 cm, and NSE=0.796, and in Thái Nguyên city with R²=0.946, RMSE=8.96 cm, and NSE=0.906. It can also be found that a three-month lag in forecasting with hybrid models provides results with acceptable accuracy. According to these results, the hybrid model can be used in regions similar to Vietnam to improve decision-making for disasters.