A Hybrid Wavelet-Machine Learning Approach for Short- and Long-Term Streamflow Forecasting in Western U.S. by Using Local and Global Climate Patterns

This research is focused on a hybrid computational intelligence modeling approach based on wavelet techniques and multivariate machine learning regression to produce short- and long-term streamflow forecasts. The streamflow data are obtained from selected USGS gauge stations located in rivers across the Western U.S. The machine learning model is based on a multivariate Bayesian approach for regression. The inputs of the model utilize past information of streamflow, precipitation, temperature, snow water equivalent and Pacific sea surface temperature. These inputs are decomposed into meaningful components formulated in terms of wavelet multiresolution analysis and used to improve the forecasting potential of the machine learning model. The proposed hybrid modeling approach can incorporate important information from trends of the local and global climate time series into models that learn these patterns to produce improved streamflow predictions at different time scales. A bootstrap analysis is used to explore the robustness of the proposed modeling approach.