Advanced prediction of river discharge: a systematic comparison of seven models and eight machine learning methods
Environmental Research Communications, cilt.8, sa.10, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 8 Sayı: 10
- Basım Tarihi: 2026
- Doi Numarası: 10.1088/2515-7620/aeb18b
- Dergi Adı: Environmental Research Communications
- Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus
- Anahtar Kelimeler: air temperature, empirical cumulative distribution function, precipitation, SHAP, support vector regression
- Açık Arşiv Koleksiyonu: AVESİS Açık Erişim Koleksiyonu
- Bursa Uludağ Üniversitesi Adresli: Evet
Özet
This study investigates the predictability of daily river discharge (Q, using eight machine learning (ML) methods: multilayer perceptron (MLP), Kolmogorov–Arnold network, convolutional neural network, extreme gradient boosting, random forest, k-nearest neighbors (KNN), support vector regression (SVR), and linear regression (LR). Precipitation (P), air temperature (T), cumulative precipitation (PC), and chronological order of months (α) were used as input parameters. The prediction models with different input combinations were developed and evaluated for three gaging stations in the Broad River Basin, South Carolina, USA. Model performance was assessed using various performance metrics and the empirical cumulative distribution function (ECDF). Furthermore, the SHAP technique was applied to examine the relationship between the input parameters and Q. The model, which incorporates multiple lagged P inputs along with T, PC, and α, consistently achieved the lowest errors and the highest Nash–Sutcliffe efficiency coefficient values (0.546–0.785) across all stations in all datasets. Whilst SVR is the most successful method, the other methods (except for LR) have achieved results very close to SVR’s. In the ECDF analysis, for all datasets and gaging stations, the SVR and MLP methods demonstrated high consistency between observed and predicted Q distributions at high values, whilst the MLP and KNN methods showed high consistency at low values. Nevertheless, when different flow regimes and bootstrap-based confidence intervals are considered, the SVR model has demonstrated more robust performance compared to other models. Based on SHAP values, PC and α are the most influential features. This study demonstrates that ML methods can be effectively used for daily Q prediction and provide a validated basis for use with future meteorological inputs under climate change scenarios.