Machine Learning-based prediction of the performance of a wind-excited piezoelectric energy harvester deployed in urban environment

The growing urgency to mitigate climate change has driven significant interest in renewable energy solutions, including energy harvesting from wind-induced vibrations from piezoelectric materials. While most prior research has optimized energy harvester designs through controlled wind tunnel experiments and computational fluid dynamics (CFD) simulations, their performance under real-world, long-term conditions remains largely unexplored. This study addresses this gap by deploying a piezoelectric energy harvester in an urban environment and analysing its performance using one month of ambient wind data. Machine learning (ML) models are developed to predict the output voltage of the harvester based on wind speed, azimuth, and elevation angles, as well as diurnal/nocturnal variations. The results revealed that wind speed magnitude influences voltage output, with clear sensitivity to directional and elevation components, which is of relevance in urban environments, where wind interacts with the surrounding structures. Among the tested ML models, Random Forest (RF) demonstrated the highest predictive accuracy, outperforming Gradient Boosting Regression Trees (GBRT) and Decision Tree Regression (DTR). This work underscores the potential of ML-driven approaches to improve the operational efficiency of piezoelectric wind-excited energy harvesters deployed in complex urban environments.