Abstract
Piezoelectric energy harvesting (PEH) enables maintenance-free micro-power generation for autonomous sensing and ultra-low-power electronics by converting ambient mechanical excitation into electrical energy. Despite substantial progress in piezoelectric materials and device structures, forecasting PEH electrical outputs remains difficult because the response is nonlinear, excitation is stochastic, and performance can drift under repeated loading. This paper proposes a hybrid deep learning architecture that integrates Long Short-Term Memory (LSTM) and Artificial Neural Network (ANN) components to forecast voltage, current, and power from footstep-driven PEH time-series data. The dataset is constructed by sampling the harvester voltage under controlled walking-induced excitation and organizing the continuous signal into supervised samples using a sliding-window scheme; features are normalized and paired with future targets for multi-output regression. The model is trained and evaluated against standalone LSTM, standalone ANN, and classical forecasting baselines using RMSE, MAE, MSE, and R2. Experimental results show high voltage prediction accuracy (R2=0.9896, RMSE = 0.0035, MAE = 0.0022), while current and power are predicted with acceptable performance consistent with their higher noise sensitivity and nonlinear coupling. These findings indicate that combining temporal memory with nonlinear regression improves forecasting stability for PEH outputs within the defined experimental setting and provides a practical basis for energy-aware scheduling and monitoring in self-powered sensing applications. Future work will extend the dataset to broader excitation conditions and incorporate uncertainty-aware modeling for robust edge deployment.