Towards Sustainable AI: Monitoring and Analysis of Carbon Emissions in Machine Learning Algorithms

The advent of Artificial Intelligence (AI) has brought advancements in many domains, reshaping the way we live and work. The rapid expansion of AI technologies has raised concerns in terms of energy consumption, particularly regarding their environmental footprint. Thus, the urgency of developing AI systems that not only deliver high performance but also minimize their carbon footprint is growing. Recent studies only highlight AI's strides in enhancing model accuracy. However, these achievements have driven a surge in computational resource demands, limiting accessibility to the broader research community. As application domains and Machine Learning (ML) models have grown in complexity, the necessity of a vast volume of data and extended training durations has become consumption-critical. These factors result in increased energy consumption for data storage and greater demand for computational power during extended training sessions. A comprehensive analysis is needed to define clear guidelines for algorithm selection and parameter adjustments to reduce energy consumption. This thesis explores the concept of sustainable AI, focusing on ML algorithms and examining key factors contributing to this issue, such as large-scale data training and hyperparameter tuning. To explore sustainable AI, this research employs an empirical approach. It investigates energy consumption patterns of existing ML algorithms when applied to critical data, such as network data, which presents unique challenges due to its real-time nature, necessitating the handling of streaming data and periodic training. An ad-hoc library, CodeCarbon, is used to monitor the carbon emission quantities produced by the training phases. The experimentation uses two different real datasets in order to compare the results: they include Traffic Data from an Italian Mobile Network Operator and PVWatts Energy Estimate Data in Turin, both within the time-series domain. The data stability is a key differentiator, with photovoltaic (PV) panel production exhibiting more predictability than dynamic traffic data. Various implementations of the same algorithm from libraries like PyTorch and TensorFlow are compared to identify the best starting point for further enhancements. Architectural model design is explored, with models used to handle time-series data, such as Long Short Term Memory (LSTM). An analysis of both architectural design (which includes layer size and quantity) and training parameters, such as the number of epochs and input sequence length, is carried on in order to define their impact on sustainable AI. Then, various strategies are employed by manipulating these factors to reduce energy consumption without significantly compromising performance. Different dataset configurations are also analyzed, for instance, aggregating information derived from different sources. The experiments confirm the importance of formulating efficient strategies for reducing carbon emissions, including parameter manipulation and problem formulation. Moreover, the case study gives an example where the approach prioritizes efficiency over accuracy, aiming to provide general guidelines applicable to any research.