Application of artificial intelligence algorithms for hourly river level forecast: A case study of Muda River, Malaysia

• Published in: Alexandria engineering journal Vol. 60; no. 4; pp. 4015 - 4028
• Main Authors: Nur Adli Zakaria, Muhamad, Abdul Malek, Marlinda, Zolkepli, Maslina, Najah Ahmed, Ali
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01-08-2021; Elsevier
• Subjects: ANFIS, MLP-NN; Flood forecasting; River level; Short-term forecasting; ANFIS, MLP-NN; Short-term forecasting; Flood forecasting; River level
• ISSN: 1110-0168
• DOI: 10.1016/j.aej.2021.02.046

A reliable river water level model to forecast the changes in different lead times is vital for flood warning systems, especially in countries like Malaysia, where flood is considered the most devastating natural disaster. In the current study, the ability of two artificial intelligence (AI) based data-driven approaches: Multi-layer Perceptron Neural Networks (MLP-NN) and An Adaptive Neuro-Fuzzy Inference System (ANFIS), as reliable models in forecasting the river level based on an hourly basis are investigated. 10-year of hourly measured data of the Muda river's water level in the northern part of Malaysia is used for training and testing the proposed models. Different statistical indices are introduced to validate the reliability of the models. Optimizing the hyper-parameters for both models is explored. Then, sensitivity analysis and uncertainty analysis are carried out. Finally, the capability of the models to forecast the river level for different lead times (1, 3, 6, 9, 12, and 24-hours ahead) is investigated. The results reveal that a high accuracy was achieved for the MLP-NN model with 4 hidden neurons with RMSE (0.01740), while for ANFIS, a model with three G-bell shaped membership functions outperformed other ANFIS models with RMSE (0.0174). MLP-NN and ANFIS achieved a high level of performance when two input combinations were used with RMSE equal to 0.01299 and 0.0130, respectively. However, MLP outperformed ANFIS in terms of running time and the uncertainty analysis test, in which the d-factor is found to be 0.000357.