A decomposed streamflow non-gradientbased artificial intelligence forecasting algorithm with factoring in aleatoric and epistemic variables / Wei Yaxing

Wei , Yaxing (2024) A decomposed streamflow non-gradientbased artificial intelligence forecasting algorithm with factoring in aleatoric and epistemic variables / Wei Yaxing. PhD thesis, Universiti Malaya.

	PDF (The Candidate's Agreement) Restricted to Repository staff only Download (147Kb)
	PDF (Thesis PhD) Restricted to Repository staff only until 31 December 2026. Download (2801Kb)

Abstract

The dissertation aims to develop an effectively decomposed time-series nongradient- based artificial intelligence model for forecasting a time-series regression machine learning task. Real-world optimizations, such as forecasting streamflow, are a complicated process that is highly non-linear and multi-modal, demanding the use of a suitable modeling tool, with an emphasis on artificial intelligence algorithms, to get befitting forecast results. Although artificial intelligence algorithms are of primary importance in today's media information and technology, with substantial research in streamflow, handling the limitations must be done appropriately in real-world optimization. The deterministic approach, which utilizes the gradient information in the search process, is prone to trapping at local minima when dealing with complex streamflow forecasting, primarily due to saddle points and local minima in the nonconvex objective function in an artificial neural network. Consequently, the study involved exploiting optimization techniques to enhance the training artificial intelligence algorithm for streamflow forecasting from a gradient-based to a stochastic population-based approach in several aspects, including solution quality, computational effort, and parameter sensitivity on streanflow in Johor, Malaysia. Empirical studies of metaheuristic algorithms performance demonstrated that the hybrid metaheuristic algorithms-artificial neural network outperformed the gradient-based artificial neural network (RMSE=113.92 m3/s) for streamflow forecasting, notably with the firefly approach, with an average RMSE=96.06 m3/s. However, not all accepted metaheuristic algorithms are compatible with enhancing the ANN for streamflow forecasting, demanding a thorough analysis due to performance differences across cases. While the firefly algorithm solution is superior, it has a higher time complexity compared to other algorithms used when there are more hidden layers and neurons. The firefly algorithm remains a feasible alternative for shallow architectural network models, while metaheuristic algorithms such as the Particle swarm algorithm and Bat algorithm are better options for deeper architectural network models. To summarise, metaheuristic algorithms can give a superior optimization approach than the traditional artificial neural network method, providing the computing time is within an acceptable range. Besides, the wavelet transform, which decomposed the original series into unsophisticated sub-series, requires assessing the additional algorithm's capacity, which is the parameter setting sensitivity to perform well in various contexts while tackling a specific problem. Given the multitude of components to manage, streamflow forecasting is preferable to employ an algorithm with low sensitivity to parameter variations. Firefly algorithm outperformed the other metaheuristic algorithms used to solve this proposed hybrid artificial intelligence model regarding parameter sensitivity. A comparison of deep learning convolutional neural network and artificial neural network algorithms was also performed, with findings revealing that convoluted input formation was less stochastic than feedforward formation, particularly for a more complicated series and vice versa, due to its capacity to attract features. The contradictory behavior could be due to the feature attraction ability of convolution neural network to over-fit the simpler sub-series. Finally, one key drawback of estimating streamflow outlined above is that it does not account for variability. As illustrated by the supplementary projected standard deviation and mean variables, generating the forecast in the Aleatoric and Epistemic forms might help improve one's impression of the predicted results.

Actions (For repository staff only : Login required)