Hybrid machine learning models for precipitation at Sakarya Station: An evaluation of Tunable Q-Factor Wavelet Transform decomposition technique

Çıtakoğlu H., İlkentapar M., Akşit S.

5th International Eurasian Conference on Science, Engineering and Technology, Ankara, Türkiye, 26 - 28 Haziran 2024, cilt.1, sa.1, ss.77

  • Yayın Türü: Bildiri / Özet Bildiri
  • Cilt numarası: 1
  • Basıldığı Şehir: Ankara
  • Basıldığı Ülke: Türkiye
  • Sayfa Sayıları: ss.77
  • Abdullah Gül Üniversitesi Adresli: Evet

Özet

In this study, precipitation data between 1999 and 2023 of the meteorological station located within the borders of Sakarya province, one of the important industrial cities of Türkiye, were utilized. Precipitation data between 1999 2017 was used as training data, and precipitation data between 2018‒2023 was used as test data. Precipitation at exit time t was estimated using precipitation values at lag times t-1, t-2, t-3, t-4 as input variables. The purpose of this study is to compare the performance of hybrid machine learning models in predicting precipitation. Adaptive neuro-fuzzy inference system (ANFIS), support vector machine regression (SVR), and Gaussian process regression (GPR), algorithms were utilized as machine learning methods. Tunable Q-Factor Wavelet Transform was utilized as a pre-processing technique to develop hybrid models. Mean absolute error, relative root mean squared error, coefficient of determination, Nash−Sutcliffe efficiency coefficient, and Kling-Gupta efficiency metrics were used to evaluate model performance. According to performance criteria, no clear superiority was observed between the TQWT-GPR and TQWT-ANFIS hybrid models. Additionally, the performance of the TQWT-SVR hybrid model remained weaker. In addition to classical performance criteria, the performances of the three hybrid models were compared visually using Taylor, Violin, and Error boxplot diagrams. According to these diagrams, TQWT-GPR and TQWT-ANFIS hybrid models gave similar results. There is no superiority between these two methods. One-tailed Wilcoxon signed rank test was performed to determine whether the TQWT-GPR hybrid model predicted more accurately than the TQWT-ANFIS hybrid model. As a result of this test, it was tested that there was no statistically significant difference between the two hybrid models. At the end of the study, the Kruskal–Wallis test determined that the measured data and the estimated data came from the same distribution. The efficiency of the recommended methods for comparisons was proven at the end of this study.

Keywords: Precipitation, GPR, ANFIS, SVR, Türkiye