Machine learning algorithms for streamflow forecasting of Lower Godavari Basin

Abstract

The present study applies three Machine Learning Algorithms, namely, Bi-directional Long Short-Term Memory (Bi-LSTM), Wavelet Neural Network (WNN), and eXtreme Gradient Boosting (XGBoost), to assess their suitability for streamflow projections of the Lower Godavari Basin. Historical data of 39 years of daily rainfall, evapotranspiration, and discharge were used, of which 80% applied for the model training and 20% for the validation. A Random Search method was used for hyperparameter tuning. XGBoost performed better than WNN, and Bi-LSTM with an R2, RMSE, NSE, and PBIAS of 0.88, 1.48, 0.86, and 29.3% during training, and 0.86, 1.63, 0.85, and 28.5%, during validation, indicating the model consistency. Therefore, it was further used for projecting streamflow from climate change perspective. Global Climate Model, Ec-Earth3 was employed in the present study. Four Shared Socioeconomic Pathways (SSPs) were considered and downscaled using Empirical Quantile Mapping. Eight decadal streamflow projections were computed – D1 to D8 (2021–2030 to 2091–2099) – exhibiting significant changes within the warm-up period. They were compared with three historical time periods of H1 (1982–1994), H2 (1995–2007), and H3 (2008–2020). The highest daily streamflow projections were observed in D1, D3, D4, D5, and D8 in SSP245 as per XGBoost analysis. HIGHLIGHTS eXtreme Gradient Boosting (XGBoost) model displays an exceptional performance with R2 = 0.88, RMSE = 1.48, NSE = 0.87, and PBIAS = 29.3% in training. These values in validation are 0.86, 1.48, 0.85, and 28.5%. XGBoost has the edge over the four variants of WNN and a Bi-LSTM model.; The highest daily streamflow projections were observed in D1, D3, D4, D5, and D8 in SSP245 as per XGBoost analysis.;

Keywords

• bi-lstm
• hyperparameter tuning
• ssp
• streamflow
• wnn
• xgboost