Energy Prediction Based on Periodic Mechanical Impact in PVDF-Based Piezoelectric Energy Harvesting: A Comparative Analysis of Machine Learning Models

In this study, an experimental setup was established to measure the energy that can be harvested depending on the frequency of periodic mechanical pulses using Polyvinylidene Fluoride (PVDF) piezoelectric material. The measured data were employed in Artificial Neural Network (ANN), K-Nearest Neighbours (KNN), Support Vector Machine (SVM), and Random Forest (RF) algorithms, and the models were compared in terms of prediction performance and processing time. The performance of the models was evaluated using regression success metrics such as coefficient of determination (R²), root mean square error (RMSE), mean absolute error (MAE), Nash-Sutcliffe Efficiency (NSE), computation time and additionally Taylor diagram and Bland-Altman analysis were applied to further evaluate the accuracy and reliability of the models. According to the results, all models achieved high accuracy, with the highest R² value obtained from the RF model (R² = 0.9756), followed closely by SVM (R² = 0.9752) and ANN (R² = 0.9730). The RF model also showed the lowest error rates (MAE = 3.5874, RMSE = 8.7020) and a strong NSE value (0.9751), making it the most balanced and reliable model overall. The ANN model demonstrated competitive performance, though a noticeable gap was observed between training and testing errors (MSEtrain = 139.76 vs. MSEtest = 86.59), indicating potential overfitting tendencies. The KNN model, despite its very fast computation time (0.0066 s), exhibited lower predictive accuracy compared to the other models (R² = 0.9464, RMSE = 16.6871). These findings suggest that Random Forest provides the best compromise between accuracy, stability, and efficiency for piezoelectric energy harvesting prediction tasks. The developed predictive models will serve as a valuable reference for researchers conducting feasibility studies, particularly in applications such as powering CMOS circuits and sensors through periodic pulse-based energy harvesting.