Energy-aware forecasting of thermoelectric generator output in agricultural microgrids using recurrent neural networks

Abstract Thermoelectric generators offer a promising solution for sustainable, low-maintenance energy harvesting in smart agriculture, especially for powering off-grid sensor nodes and microgrid systems. However, their inherently variable output under changing environmental conditions challenges the consistency of an energy supply. This study proposes a data-driven forecasting framework utilizing long short-term memory and gated recurrent unit neural networks to predict real-time TEG output. Models were trained on actual voltage data collected under operational conditions, with current estimated via Ohm’s law. A sliding window approach with 60-step input sequences was adopted, and model performance was rigorously evaluated using multiple regression metrics, including RMSE, R ² , MSE, MAE, MAPE, and the Index of Agreement. The results demonstrate consistently high predictive accuracy across multiple runs, with average R ² values exceeding 0.996 and RMSE values on the order of 10 ⁻⁴ to 10 ⁻³ for both voltage and current outputs, indicating strong agreement between predicted and measured values. These findings underscore the models’ capability to capture complex temporal patterns in thermoelectric generation. The study marks the first application of deep recurrent neural networks for forecasting TEG behavior in agriculture and lays the groundwork for integrating predictive intelligence into energy-aware farm systems. Future research should focus on hybrid AI integration, field validation with live TEG sensors, and real-time deployment in precision agriculture applications.