HYBRID REGRESSION APPROACH FOR PREDICTION OF RENEWABLE ENERGY FROM HUMAN FOOTSTEP POWER SYSTEMS

The increasing global emphasis on renewable and sustainable energy sources has driven researchers to explore innovative methods of power generation. Footstep power harvesting, which utilizes human walking motion to generate electricity, offers a promising avenue for powering low-energy devices in smart cities and IoT applications. According to recent studies, an average adult produces between 3–8 W of instantaneous mechanical power per step, and large-scale foot traffic in urban areas can potentially yield over 500 kWh annually from a single busy location. However, accurately predicting the electrical output from such systems remains a challenge due to variations in human weight, walking patterns, and environmental factors. Existing methods, including the Huber Regressor and Support Vector Regression (SVR), provide reasonable predictions but often struggle with capturing the complex nonlinear relationships between input parameters such as voltage (V), current (μA), weight (kg), location, and resulting power (mW). These traditional approaches are also sensitive to outliers and fail to fully exploit patterns in multidimensional datasets, leading to suboptimal forecasting accuracy. To address these limitations, this study proposes a hybrid deep learning–machine learning framework combining a Feed Forward Neural Network (FFNN) with an Extra Tree Regressor (ETR) to enhance prediction reliability. The methodology begins with rigorous data preprocessing, including normalization of numerical features, one-hot encoding of the categorical location attribute, and outlier detection to improve model robustness. The existing Huber Regressor and SVR models are implemented as benchmark baselines to assess comparative performance. The proposed FFNN-ETR model leverages the FFNN’s capability to learn intricate feature interactions and nonlinear dependencies, while the ETR enhances prediction stability through ensemble-based decision boundaries. This dual-stage approach not only achieves superior accuracy but also improves resistance to noise and environmental variability in footstep energy harvesting datasets.