Machine learning approaches for thermal updraft prediction in wind solar tower systems

Abstract Wind solar towers constitute a fairly new scheme for harvesting renewable energy from solar and wind energy sources. In such a tower, solar radiation is collected and hot air is enforced to go fast through the tower, a process called thermal updraft, which fuels a wind turbine to generate power. Using vortex generators at the top of the tower creates a pressure difference, which increases the thermal updraft. In this work, we describe the setup of a wind solar tower system established at Kyushu University in Japan. Then, we demonstrate how data was collected from this system in order to train regression models for thermal updraft prediction. The feature selection process was guided by sensitivity analysis. After that, several machine learning models were investigated and the most suitable model was selected based on quality and time metrics. The linear regression model was particularly examined in detail, and was shown to have a satisfactory high accuracy of thermal updraft prediction graphically and numerically with a coefficient of determination of R2 = 0.981. We also evaluated a reduced prediction model based on the six most essential features, which could be a reduced model description for the WST. This reduced model showed little performance degradation (R2 = 0.974), with significant reduction in the needed effort and resources, as well as data collection requirements.