Brahmaputra River (Pandu Station) Flow Prediction Using MLR, ANN, and RF Models Combined with Wavelet Transform

Abstract

In the current work, a DWT (Discrete Wavelet Transform) was linked to ANN, MLR, and RF to develop hybrid models WANN, WMLR, and WRF, respectively, for Brahmaputra River flow forecasting. We used ten-year daily flow data from Pandu Station, which was decomposed (up to five levels) into multiresolution time series using DWT and Daubechies wavelets db1, db2, db3, db8, and db10. The predicted discharge values for multiple lead times (2, 3, 4, 7, and 14 days) have been then obtained by feeding multiresolution time series data as the input to MLR, ANN, and RF. It was discovered that the WMLR-db10 model outperformed the WANN and WRF models for all lead times. Throughout the testing phase, the values of Nash-Sutcliffe efficiency (NS) along with RMSE (shown in bracket) for the WMLR-db10 model for lead times 2, 3, 4, 7 and 14 days have been observed to be, respectively, 0.998 (415.18 m³/s), 0.998 (514.21 m³/s), 0.996 (713.62 m³/s), 0.991 (1030.83 m³/s), and 0.977 (1638.64 m³/s). Additionally, it has been observed that WANN performed better for low-order wavelets (db1, db2, db3), WMLR performed better for high-order wavelets (db8, db10), and WRF performed worst of all the wavelets.