Deep Learning-Driven Hydro-Meteorological Forecasting: A Comprehensive Analysis and Model Development

Precise hydrological forecasting is critical for effective flood management in vulnerable regions, particularly in the flood-prone districts of Assam, India. This study presents a comparative deep learning-based approach using different networks to forecast river flow rates and water levels at high-risk locations of Assam. The primary objective was to evaluate the efficacy of each model in capturing short-term fluctuations in seasonal trends integral to the Brahmaputra River and then to hybridize the two models that outperformed all the other models. Historical daily flow rate and water level data from January 1, 2013, to July 11, 2022, were used as the core dataset. The data were pre-processed to reflect regional hydrological seasonality, with segmentation into monsoon and non-monsoon periods. A binary seasonality flag and sinusoidal encodings were introduced to capture cyclical behaviours. Evaluation metrics included Shelf-life analysis using Root Mean Squared Error (RMSE), MSE, and R ² values. From this study, it was observed that LSTM consistently demonstrated better performance in forecasting short-term seasonal patterns of water levels, effectively forecasting monsoonal peaks and dry season troughs across both sites. CNN showed strength in short-term pattern recognition of trends in flow rates with rainfall anomaly. Thus, these two models were chosen to make a hybrid CNN-LSTM algorithm for the same study sites. The same dataset of flow rates and water levels along with the rainfall data, was run through this hybrid model for all the sites to get better results. Graphical analysis of predicted versus actual flow rates and water levels confirmed these findings. In conclusion, the study highlights the potential of hybrid DL techniques in hydro-meteorological modeling. It supports the future application of hybrid models for enhanced flood early-warning systems and water resource planning in similar climatic zones.