Date of Award

2024

Document Type

Open Access Master's Thesis

Degree Name

Master of Science in Mechanical Engineering (MS)

Administrative Home Department

Department of Mechanical Engineering-Engineering Mechanics

Advisor 1

Gordon Parker

Committee Member 1

Jungyun Bae

Committee Member 2

William Endres

Abstract

Due to the unpredictable nature of large bodies of water, wave energy can be a difficult renewable resource to rely on. One way to make Wave Energy Converters (WECs) more efficient is to apply a control strategy. In many control solutions, it is assumed that the wave excitation force is known into the future. In many instances, especially with complex waveforms, this is simply not the case. Simulation studies have shown the promise of wave force prediction using neural networks. This study demonstrates this experimentally and aims to characterize the important factors when designing such a network. Several wave elevation measurement factors are considered, including: quantity, their location relative to the buoy, and their configuration. The relationship between forecast horizon and the number of measurement backvalues is also evaluated along with both the wave form complexity and the performance impact of including instantaneous buoy acceleration. A 14.2 cm buoy, constrained to vertical motion, was subjected to 30, 60 second tests using regular and irregular waves in a wave tank. For each test its vertical motion was recorded along with an array of twelve upstream and downstream wave elevation measurements. Neural networks were trained using subsets of the data to examine the effect of the factors mentioned above on prediction performance. The results showed that upstream measurements were the most important, where the distance between the measurement and the buoy is critical. A diamond-shaped configuration of elevation measurements performed nearly the same as using all twelve measurements illustrating the importance measurement topology. It was also found that the number of past measurements used had a significant impact on performance. Specifically, performance was best when the ratio of the prediction horizon time to the number of backvalues was one. 1 gauge far upstream, 2 gauges immediately upstream, and 1 gauge to the rear performed just as well as a full set of 12 gauges. Including acceleration as an input appeared to lower the error of most of these cases as well. It was discovered that a ratio of the forecast horizon to the number of backvalues allows the network to perform its best as this ratio is near or less than 1. Further testing is required to obtain a more complete view of the impact of waveform complexity on the results of the network.

Share

COinS