KAN-Enhanced LSTM for Accurate and Scalable Flood Forecasting: A Case Study of the Mahanadi Basin

Floods are globally catastrophic, with profound impacts on India’s environment, agriculture, and infrastructure. Between 1953 and 2010, floods annually affected 7.2 million hectares and 3.2 million people. Odisha, particularly the Mahanadi basin, ranks seventh in flood vulnerability, with over 90\% of its annual rainfall occurring during the monsoon. Synoptic systems from the Bay of Bengal exacerbate rainfall, causing frequent and severe floods. Despite existing forecasting systems, downstream regions remain inadequately protected, highlighting the need for more accurate predictive mechanisms.

Recent advancements in Machine Learning (ML) and Deep Learning (DL) offer new opportunities for flood forecasting. Long Short-Term Memory (LSTM) models excel at capturing intricate temporal patterns in rainfall and streamflow data, outperforming conventional methods. Innovations like hybrid LSTMs and Spatio-Temporal Attention (STA) mechanisms enhance their performance, while novel architectures such as Temporal Convolutional Networks (TCNs) and Bi-LSTMs improve long-term predictions.

This study introduces a Kolmogorov-Arnold Network (KAN)-enhanced LSTM model for five-day-ahead flood prediction in the Mahanadi basin. KAN leverages learnable activation functions and spline representations, improving accuracy, interpretability, and computational efficiency. Evaluated against traditional LSTMs using metrics such as Nash-Sutcliffe Efficiency (NSE), time-to-peak prediction, and convergence speed, the KAN-enhanced model consistently outperformed standard LSTMs. It demonstrated a 12\% improvement in NSE, superior peak timing, and 20\% faster convergence, offering crucial advancements for early warning systems.

By integrating KAN's ability to model non-linear relationships with LSTM’s strength in sequential data analysis, this framework addresses complex hydrological dynamics, providing reliable flood forecasts. These findings underscore the potential of KAN-enhanced architectures to revolutionize flood prediction, offering scalable and interpretable solutions for flood-prone regions.