Dam water levels prediction using advanced hybrid deep learning model based on Bayesian Optimization approach

Abstract

Effective dam water level prediction is of critical importance for the optimization of hydroelectric power generation, flood risk reduction and sustainable water resources management. In this study, a hybrid deep learning model is proposed for short-term water level prediction. In addition to deep learning models such as LSTM, BiLSTM, GRU and CNN, hybrid versions of these models (CNN-LSTM, CNN-BiLSTM, CNN-GRU) are also evaluated. The dataset used is based on daily hydrological data recorded between 2014 and 2023 of Deriner Dam, one of the strategically important dams of Turkey. The modeling process is supported by the Bayesian Optimization approach, which is one of the Neural Architecture Search (NAS) approaches, in order to minimize human intervention in hyperparameter selection. The NAS-optimized versions of each model are developed and compared separately. The highest accuracy was achieved with the proposed CNN-GRU Unified (CGU) hybrid model with a score of R² = 0.9941. The proposed CGU model combines spatial feature extraction and temporal dependencies modeling in the same structure, and better performance results are obtained with this model compared to state-of-the-art models and their hybrid versions. The high model accuracy and low error rate in the study show that the CGU architecture is a successful and reliable solution that can be integrated into real-time dam management systems. These findings have brought a new and scalable modeling approach to the literature, showing the usability of NAS-supported hybrid models in strategic water management applications.