Integrating hybrid deep learning and CSI for multi-interval hydrological data in enhanced flood prediction

Abstract

Flood prediction accuracy is often constrained by heterogeneous and asynchronous hydrological data collected at different time intervals. This study proposes a hybrid deep learning–based flood prediction framework that integrates long short-term memory (LSTM), convolutional neural network (CNN), and cubic spline interpolation (CSI) to address these challenges. Rainfall, river discharge, and water level data representing upstream, midstream, and downstream conditions of the Bengawan Solo watershed were utilized. CSI was applied as a preprocessing step to harmonize multi-interval data, reduce noise, and recover missing observations, thereby improving data consistency. The experimental results show that the proposed hybrid LSTM–CNN model enhanced with CSI outperforms baseline LSTM and non-interpolated hybrid models, achieving a mean absolute percentage error (MAPE) of 5.84%, root mean square error (RMSE) of 0.125 m, mean absolute error (MAE) of 0.082 m, and R² of 0.948. The integration of spatio-temporal feature learning with data harmonization enables more accurate flood level prediction and supports timely flood early warning systems. The proposed approach demonstrates strong potential for improving flood risk management and disaster preparedness in flood-prone regions.