Performance of short, medium, and long-term streamflow forecasting using machine learning

Wong, Wai Kit (2023) Performance of short, medium, and long-term streamflow forecasting using machine learning. Final Year Project, UTAR.

Preview

PDF Download (1173Kb) | Preview

Abstract

Accurate streamflow forecasting is imperative for efficient water resources management. Yet, precise prediction of streamflow is difficult owing to the underlying complex relationships between variables, as well as the accompanying nonstationary and nonlinearity of the problem. Therefore, application of the hybrid ANN model has been given much consideration for its capabilities in delivering high predictive accuracy in hydrological forecasting. This study aims to assess the performance of short, medium and long term streamflow prediction by hybrid ANN model, after conducting thorough investigation on wavelet decomposition and developing the most accurate model. Historical streamflow data obtained from the Department of Irrigation and Drainage was separated into training and testing data set by a 80%:20% ratio, respectively. Data pre-processing was conducted by stationary wavelet transformation. In total, 12 different models with various combinations of 2 scenarios, 2 cases and 3 wavelets, including sym5, db5 and coif5 were used to predict streamflow. The improvement of hyperparameter tuning and hybrid model were verified, where tuning improved the models by a range of 0.93% to 68.17%, whereas hybrid models received 1082.78% to 1612.64% improvement as compared to the standalone model. By majority, S1 models have better performance than S2 models, while C1 models have better performance than C2 models. No wavelet was observed to exhibit apparent advantage. The best performing model was identified by the visualizations through the Taylor diagram and the Violin diagram and it was the S1*C1sym5 model that stood out. Prediction with this model had revealed that short term prediction is the most accurate, followed by the medium and long term predctions. By comparison then, the loss of accuracy in terms of (R2 , RMSE, MAE) for the medium and long term respectively, are (95.18%, 1923.37%, 2070.52%) and (95.56%, 4811.91%，3920.38%). Applications of different ANN learning algorithms such as grid search and random search; implementation of quantitative analysis to assess the similarity between wavelet and input time series; comparison of DWT and SWT; as well as combination of decomposed wavelets from different wavelet families; are but some of the recommended for future similar works.

Actions (login required)