Enhancing Affordable Wind Power Forecasting through Temporal Convolutional Networks: A Comparative Study with LSTM Architectures

Resumen

The global transition toward low-carbon energy systems demands increasingly accurate tools to address the inherent variability of renewable resources. In particular, wind power forecasting plays a crucial role in improving energy planning, enhancing grid stability, and enabling cost-efficient operation of wind-based systems. At the same time, the democratization of clean energy technologies requires forecasting methods that are not only reliable, but also computationally efficient and suitable for deployment on affordable hardware platforms. In this context, this work evaluates the effectiveness of advanced deep learning architectures for short-term wind power forecasting, with a particular focus on Temporal Convolutional Networks (TCN). Building upon previous research employing Long Short-Term Memory (LSTM) networks on low-cost devices, this study provides a rigorous comparative analysis between both approaches. The evaluation examines forecasting accuracy and real-time performance through metrics such as MAE, RMSE, R², predictions per second, and inference time per sample. The results demonstrate that TCNs can significantly reduce inference time—achieving up to an order of magnitude faster prediction speed than LSTMs—while maintaining competitive accuracy across all scenarios and outperforming recurrent models in specific feature configurations. These findings highlight the suitability of convolution-based architectures for real-time forecasting and their potential integration into low-cost, edge-computing solutions for small-scale or distributed wind energy systems.