Bayesian Optimized CNN-RNN Hybrid Model for Predicting Streamflow in Potomac River Basin

• Published in: 2024 16th International Conference on Advanced Computational Intelligence (ICACI) pp. 121 - 126
• Main Authors: Zhang, Nian, Rouamba, Stephanie, Robinson, Gavin, Deksissa, Tolessa
• Format: Conference Proceeding
• Language: English
• Published: IEEE 16.05.2024
• Subjects: Computational modeling; convolutional neural networks (CNNs); Correlation; Data models; deep learning; Discharges (electric); long short-term memory Networks (LSTM); machine learning; Predictive models; Recurrent neural networks; recurrent neural networks (RNNs); Time series analysis; Time series prediction; water quantity prediction
• DOI: 10.1109/ICACI60820.2024.10537003

Accurately estimating streamflow quantity within a watershed is essential for anticipating potential severe drought conditions in the future. This paper introduces a hybrid model that combines Convolutional Neural Networks and Recurrent Neural Networks (CNN-RNN) to predict streamflow discharge using historical data from the Potomac River Basin. The experimental outcomes showcase the model's proficiency in successfully predicting streamflow discharge values for the upcoming month. Additionally, a new sequence, derived from a one-step-ahead approach, is constructed to forecast streamflow discharge values beyond the testing data horizon. Notably, the CNN-RNN model demonstrates superior performance with a lower Symmetric Mean Absolute Percentage Error (sMAPE) at 0.048592 compared to the RNN method at 0.067662. Moreover, the prediction accuracy of the CNN-RNN model surpasses that of the RNN model. The correlation between the test sequence and the CNN-RNN prediction is notably high at 0.97753, while the RNN correlation lags slightly at 0.96694. Furthermore, the results affirm the CNN-RNN model's potential capability to forecast unobserved values beyond the time series horizon when observed data is unavailable.