Artificial intelligence modelling integrated with Singular Spectral analysis and Seasonal-Trend decomposition using Loess approaches for streamflow predictions

APAYDIN H., Sattari M. T., Falsafian K., Prasad R.

JOURNAL OF HYDROLOGY, cilt.600, 2021 (SCI-Expanded) identifier identifier

  • Yayın Türü: Makale / Tam Makale
  • Cilt numarası: 600
  • Basım Tarihi: 2021
  • Doi Numarası: 10.1016/j.jhydrol.2021.126506
  • Dergi Adı: JOURNAL OF HYDROLOGY
  • Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Academic Search Premier, PASCAL, Aerospace Database, Aqualine, Aquatic Science & Fisheries Abstracts (ASFA), Artic & Antarctic Regions, BIOSIS, CAB Abstracts, Communication Abstracts, Environment Index, INSPEC, Metadex, Pollution Abstracts, Veterinary Science Database, Civil Engineering Abstracts
  • Anahtar Kelimeler: Data-driven models, Deep learning, Attributes selections, Pre-processing approaches, ANN CNN LSTM, Gini index, INPUT VARIABLE SELECTION, SUPPORT VECTOR MACHINE, LEARNING-METHODS, SOLAR-RADIATION, NEURAL-NETWORKS, SOHU STREAM, WATER-LEVEL, RIVER, REGION, ANN
  • Ankara Üniversitesi Adresli: Evet

Özet

The nature of streamflow in the basins is stochastic and complex making it difficult to make an accurate prediction about the future river flows. Recently, artificial neural-based deep learning models with a nonlinear structure have become predominant in water engineering forecasting problems such as river flow predictions. In this study, we investigate the potential of Singular Spectral Analysis (SSA), Seasonal-Trend decomposition using Loess (STL) and attribute selection pre-processing approaches with the neural network methods in predicting monthly river streamflows in the Nallihan stream, Turkey. Antecedent measured streamflow, precipitation, relative humidity and temperature data between the years 1996 and 2016 from the observing stations in the basin boundaries were used as model inputs under different scenarios using the correlations between the past measured variables, to predict one-step-ahead flow data. To compare the newer hybrid model performances; evaluation metrics including coefficient of determination (R-2), Nash-Sutcliffe efficiency (NE), Willmott's Index of Agreement (WI), root mean square error, mean absolute error together with judicious plots; scatter plot, time series and the Taylor diagrams were utilized. The developed hybrid SSA-based models registered higher accuracy than other standalone Neural Networks models without pre-processing approaches. The R-2 for SSA-based models were higher ranging from 0.8300 to 0.9105, and the largest R-2 = 0.9105 was registered by the proposed SSA-ANN model. SSA-ANN models have also the highest NE and WI index values: 0.9045 and 0.9764, respectively. The outcomes revealed that based on NE, the SSA decomposition increased the monthly streamflow prediction accuracy by 24.11%, 18.40% and 5.11% of respective ANN, CNN and LSTM methods. The SSA preprocessing approach is able to unveil the embedded streamflow characteristics and could be further applied in basins with similar characteristics to attain more accurate predictions of river flows.