Optimizing power and performance in time series forecasting: A feature selection approach for green computing

In the information era, data availability is growing at an amazing speed, driving the evolution of machine learning across various domains. However, incorporating more data into machine learning algorithms does not always lead to better performance. Identifying the most relevant features that significantly impact model performance is crucial, especially for resource-efficient computing. Feature selection plays a key role in this process. Traditional methods often struggle with scalability or lack specialized techniques for specific data types, such as time series. This paper introduces a novel feature selection method for time series forecasting that addresses these challenges while promoting a green computing approach. Our method improves computational efficiency by identifying relevant features through a harmonic enhancement process that captures and highlights dominant frequencies. This reduces the need for excessive data processing. By optimizing feature selection, our approach minimizes computational overhead and lowers energy consumption without compromising predictive performance. Extensive experiments across diverse datasets and machine learning models demonstrate that, on average, our method achieves an ≈ 2 % error improvement being one of the faster methods studied, making it a sustainable and efficient solution for large-scale forecasting tasks. • Traditional methods struggle with scalability and computational cost, especially in time series forecasting. • Our proposal identifies and amplifies dominant frequencies, improving feature selection. • Processing time and energy consumption is optimized while maintaining predictive performance. • Results show strong results across various datasets and models.