Comparative analysis of data-driven models and signal processing techniques in the monthly maximum daily precipitation prediction of El Kerma station Northeast of Algeria

• Published in: Soft computing (Berlin, Germany) Vol. 28; no. 17-18; pp. 10751 - 10765
• Main Authors: Katipoğlu, Okan Mert, Keblouti, Mehdi
• Format: Journal Article
• Language: English
• Published: Berlin/Heidelberg Springer Berlin Heidelberg 01.09.2024
• Subjects: Artificial Intelligence; Computational Intelligence; Control; Engineering; Mathematical Logic and Foundations; Mechatronics; Neural Networks; Robotics; Seybouse watershed; Missing data; Empirical mode decomposition; Wavelet transform; Monthly precipitation; Machine learning
• ISSN: 1432-7643; 1433-7479
• DOI: 10.1007/s00500-024-09860-3

The availability of a climate database is an essential requirement for modeling and mapping hydrological and environmental processes. Regardless of the nature and structure of these models, most of them require a complete and reliable dataset on a spatiotemporal basis. Unfortunately, the measurement of hydrological variables (precipitation, flow, etc.) can be affected by systematic errors, gaps and random data. The watershed of Seybouse located in the North-East of Algeria, has a network that has gaps in the monthly data with an average error percentage of 20.46% over the period of 1970–2008 on all the stations in operation. The current study combines support vector regression, artificial neural networks, boosted tree, bagged tree, gaussian processes regression and linear regression techniques with empirical mode decomposition and discrete wavelet transform techniques to fill the missing precipitation gaps. This study aims to determine which kernel function, regression type, tree and network structure and data decomposition technique will produce the best outputs in predicting missing rainfall. For this purpose, it aims to comprehensively evaluate various soft computing and signal processing models and use strengthened hybrid approaches by combining them. While modeling, 70% of the data was used for training and the rest for testing. Mean absolute error (MAE), root mean square error (RMSE), determination coefficient (R 2 ), Taylor diagram and Violin plots were used to find the best model. As a result of the analysis, it was concluded that the SVM model produced superior outputs compared to other machine learning models. The single LSVM model had an R 2 value of 0.78, RMSE of 8.18, and MAE of 5.90. The W-LSVM model had an R 2 value of 0.78, RMSE of 8.10, and MAE of 5.90. The EMD-CGSVM model had an R 2 value of 0.69, RMSE of 10.43, and MAE of 7.10. These results indicate that the W-LSVM models provided the most accurate rainfall forecast outputs. The results showed that the wavelet-based linear support vector regression model in precipitation prediction is the most successful artificial intelligence model. Finally, using the W-LSVM model to predict rainfall in the study area can significantly improve water resources management and flood strategies and help optimize water use and structures.