Abstract:
Bangladesh is a flood-prone country due to its low-lying geography, heavy monsoon rainfall, and frequent cyclones, particularly in the coastal and tidal regions. Severe flooding in recent decades has resulted in many causalities and food scarcity. As a result, reliable and timely flood forecasting and warning are recognized as crucial factors to decrease flood-related damage and human suffering. Unfortunately, the current flood forecasting system provides fairly accurate forecast for shorter lead times (upto 3 days) only (FFWC 2021). Therefore, improving medium to long-range flood forecasting with 5 to 10 day lead times has become essential for better flood preparedness in the country. The aim of this research was to explore the potential of artificial intelligence for medium-range flow forecasting with a 5-day lead time at the Bahadurabad station in the Brahmaputra basin, using data such as precipitation, precipitable water, soil moisture storage, and satellite-derived river water levels. Artificial Neural Network (ANN) and Support Vector Machine (SVM) were utilized for this purpose. This approach presents a promising alternative to traditional streamflow forecasting methods using a hydrologic model, with the results compared against the forecast from the Flood Forecast and Warning Centre (FFWC) of the Bangladesh Water Development Board. For both the ANN and SVM models, 70% of the available data was used for model training, 15% of data was used for model testing, and the rest of the data was used for independent validation. Various input data combinations (called ‘model scenarios’ here) are considered for the ANN and SVM models to simulate discharge. In addition, each data combination was tested using the currently available data, which has a 5-day latency, real-time data, and forecast data, as not all variables are yet available in real-time or forecasts, but are expected to be in the near future.
The model that considers all input variables (scenario 01) and the model that uses a combination of water level, precipitation, and soil moisture as inputs (scenario 04) were found as the best models for both ANN and SVM. In this study, three models (Model A, Model B, and Model C) were developed to address data latency issues, considering different data availability scenarios. Model A uses ERA5 data, which currently has a 5-day latency period. Model B assumes the availability of real-time data (used ERA5 data ignoring latency). Model C simulates how discharge forecasts could be improved if reliable climate forecast data were available. The performances of the ANN and SVM methods demonstrated a comparable level of accuracy to the observed flow and, in some cases, surpasses the predictive accuracy of the FFWC model, which forecasts water levels that converted to flow via a rating curve. Here, the SVM methods generally outperformed the ANN methods. Specifically, for the SVM methods of Model A at which scenario A-01 achieved a Root Mean Square Error (RMSE) of 6802.82 m³/s, a coefficient of determination (R²) of 0.87, and a Nash–Sutcliffe Efficiency (NSE) of 0.87, while Model scenario A-04 resulted in a RMSE of 7236.50 m³/s, a R² of 0.85, and a NSE of 0.85. In case of evaluating the RMSE value by comparison in a percentage scale, it was observed that best SVM model scenario C-01 demonstrated an improvement of approximately 58.22% over FFWC Prediction of discharge via a rating curve. This research demonstrates that artificial intelligence, particularly SVM, offers a promising alternative to traditional flood forecasting methods, with improved accuracy in predicting flow at Bahadurabad station in the Brahmaputra basin. Its ability to utilize satellite-derived data enhances flood forecasting and contributes to more reliable predictions, which can significantly improve flood preparedness and risk management in Bangladesh.