Improving Short-term Daily Streamflow Forecasting Using an Autoencoder Based CNN-LSTM Model

• Published in: Water resources management 2024-12, Vol.38 (15), p.5973-5989
• Main Authors: Kumshe, Umar Muhammad Mustapha, Abdulhamid, Zakariya Muhammad, Mala, Baba Ahmad, Muazu, Tasiu, Muhammad, Abdullahi Uwaisu, Sangary, Ousmane, Ba, Abdoul Fatakhou, Tijjani, Sani, Adam, Jibril Muhammad, Ali, Mosaad Ali Hussein, Bello, Aliyu Uthman, Bala, Muhammad Muhammad
• Format: Article
• Language: English
• Subjects: Artificial intelligence; Artificial neural networks; Atmospheric Sciences; Civil Engineering; Datasets; Drought; Earth and Environmental Science; Earth Sciences; Environment; Environmental management; Flood control; Flood forecasting; Flood management; Geotechnical Engineering & Applied Earth Sciences; Hydroelectric power; Hydroelectric power generation; Hydrogeology; Hydrology/Water Resources; Long short-term memory; Neural networks; Performance evaluation; Quality management; Root-mean-square errors; Statistical models; Stream discharge; Stream flow; Streamflow forecasting; Water management; Water quality; Water quality management; Water resources; Water resources management
• ISSN: 0920-4741; 1573-1650
• DOI: 10.1007/s11269-024-03937-2

Streamflow forecasting is vital for managing water resources, such as flood control, agriculture planning, hydropower generation, environmental management, drought management, and water quality management. Motivated by the success of artificial intelligence models for hydrological applications, this study proposes a model that integrates an autoencoder, the Convolutional Neural Networks (CNN), and the Long Short Term Memory (LSTM) networks. Thirty years daily dataset were served to the Autoencoder Convolutional Neural Network Long Short Term Memory (AE-CNN-LSTM) and the baseline models. To evaluate the model's accuracy, 80% of the dataset was used for training and the remaining 20% was used to test the performance of these models. Statistical metrics, for instance, the Root Mean Squared Error (RMSE), the Mean Absolute Error (MAE), the Mean Absolute Percentage Error (MAPE), the Nash–Sutcliffe Efficiency (NSE), and the Coefficient of determination (R 2 ) were employed to evaluate the model’s performance. In terms of train RMSE, test RMSE, train MAE, test MAE, train MAPE, test MAPE, train NSE, test NSE, train R 2 , and test R 2 , the proposed model significantly obtained the best results with values of 2.6299, 2.7971, 0.1676, 0.1881, 16.76, 18.81, 0.98, 0.97, 0.98, and 0.96, respectively.