The growing global demand for computational resources, particularly in Artificial Intelligence (AI) applications, raises increasing concerns about energy consumption and its environmental impact. This study introduces a newly developed energy consumption index that evaluates the energy efficiency of Deep Learning (DL) models, providing a standardized and adaptable approach for various models. Convolutional neural networks, including both classical and modern architectures, serve as the primary case study to demonstrate the applicability of the index. Furthermore, the inclusion of the Swin Transformer, a state-of-the-art and modern non-convolutional model, highlights the adaptability of the framework to diverse architectural paradigms. This study analyzes the energy consumption during both training and inference of representative DL architectures, including AlexNet, ResNet18, VGG16, EfficientNet-B3, ConvNeXt-T, and Swin Transformer, trained on the Imagenette dataset using TITAN XP and GTX 1080 GPUs. Energy measurements are obtained using sensor-based tools, including OpenZmeter (v2) with integrated electrical sensors. Additionally, software-based tools such as CarbonTracker (v1.2.5) and CodeCarbon (v2.4.1) retrieve energy consumption data from computational component sensors. The results reveal significant differences in energy efficiency across architectures and GPUs, providing insights into the trade-offs between model performance and energy use. By offering a flexible framework for comparing energy efficiency across DL models, this study advances sustainability in AI systems, supporting accurate and standardized energy evaluations applicable to various computational settings.

Electroencephalography (EEG) signal classification is a challenging task due to the low signal-to-noise ratio and the usual presence of artifacts from different sources. Different classification techniques, which are usually based on a predefined set of features extracted from the EEG band power distribution profile, have been previously proposed. However, the classification of EEG still remains a challenge, depending on the experimental conditions and the responses to be captured. In this context, the use of deep neural networks offers new opportunities to improve the classification performance without the use of a predefined set of features. Nevertheless, Deep Learning architectures include a vast number of hyperparameters on which the performance of the model relies. In this paper, we propose a method for optimizing Deep Learning models, not only the hyperparameters, but also their structure, which is able to propose solutions that consist of different architectures due to different layer combinations. The experimental results corroborate that deep architectures optimized by our method outperform the baseline approaches and result in computationally efficient models. Moreover, we demonstrate that optimized architectures improve the energy efficiency with respect to the baseline models.

In this study, the energy efficiency of time series forecasting algorithms is addressed in a broad context, highlighting the importance of optimizing energy consumption in computational applications. The purpose of this study is to compare the energy efficiency and accuracy of algorithms implemented in different frameworks, specifically Darts, TensorFlow, and Prophet, using the ARIMA technique. The experiments were conducted on a local infrastructure. The Python library CodeCarbon and the physical energy consumption measurement device openZmeter were used to measure the energy consumption. The results show significant differences in energy consumption and algorithm accuracy depending on the framework and execution environment. We conclude that it is possible to achieve an optimal balance between energy efficiency and accuracy in time series forecasting, which has important implications for developing more sustainable and efficient applications. This study provides valuable guidance for researchers and professionals interested in the energy efficiency of forecasting algorithms.