With the increased use of AI across several domains, from manufacturing, to banking, to intelligent transport systems, Europe urgently needs to develop new data management solutions that will harness the transformative potential of AI, whilst meeting the European Green Deal objectives.1<\/sup><\/p>\n Addressing the energy demands of AI services extends beyond model training, encompassing inference, data storage, retrieval, and data centre cooling.<\/p>\n A key mitigation strategy involves designing AI algorithms with lower energy consumption. Energy efficiency, quantified by the ratio of energy consumed to output or work produced (e.g., energy per prediction, per transaction), requires a multifaceted evaluation considering various factors and metrics.<\/p>\n Measuring, for example, the energy required to train an AI model involves instrumenting the hardware, or using power models that provide energy consumption estimates based on hardware specifications, collecting data on energy consumption during the training process. Then the data is normalised based on the size and complexity of the model, and the results are reported along with other performance metrics.<\/p>\n Operational energy consumption can be monitored using energy monitoring tools, power meters, or specialised software, facilitating estimation of the service\u2019s carbon footprint based on energy source.<\/p>\n Additionally, energy-efficient AI systems must be designed with scalability in mind, allowing for flexible adjustments to demand without a significant rise in energy consumption. Balancing efficiency and performance is the key to progress. It entails ensuring that a service is both energy-efficient and maintains or surpasses performance standards while consuming less energy. This guarantees acceptable levels of accuracy and speed.<\/p>\n Techniques for improving the energy efficiency of algorithms include hardware acceleration, model optimisation, and data compression.2-10<\/sup> AI developers now have a responsibility to mitigate the environmental impact of their technology.<\/p>\n In this context, the GREEN.DAT.AI<\/a><\/u> project11<\/sup> aims to channel the potential of AI towards Europe\u2019s sustainability goals, by developing novel Energy-Efficient Large-Scale Data Analytics Services, ready-to-use in industrial AI-based systems while reducing the environmental impact of data management processes.<\/p>\n A number of challenges are posed by large-scale data analysis, including data transfer (bringing the data and computational resources together), controlling access to the data, managing the data, standardising data formats, and integrating data of multiple different types to accurately model industrial systems.<\/p>\n The project vision is to design smart AI-powered applications that allow computing to move from data centres to edge devices. Shifting computation from the cloud to devices reduces the flow and potential leakage of sensitive data, leads to faster inference with a shorter reaction time and drives innovation in applications where these parameters are critical.<\/p>\n Core to the project innovation are federated learning mechanisms, which utilise distributed data for inferring information based on decentralised collaborative modelling algorithms. The ambition is to enable data sharing while minimising data transfer. To achieve this, a novel Toolbox of reusable energy-efficient AI services<\/a> is being developed including:<\/p>\n The project will also develop a benchmarking and evaluation framework for measuring and comparing the energy efficiency of different AI services, addressing the need for more accurate energy consumption models and the development of energy-aware AI algorithms.<\/p>\n The energy-efficient AI services will be assessed across four industries \u2013 Smart Energy, Smart Agriculture, Smart Mobility, and Smart Banking \u2013through six distinct pilots, all leveraging the potential of data spaces.<\/p>\n PILOT 1: Renewable energy<\/strong><\/p>\n In the renewable energy sources domain, data are experiencing unprecedented growth, generated by diverse energy systems. These data hold immense potential for applications in reliability and performance monitoring, predictive maintenance, and asset management of power plants.<\/p>\n Additionally, they play a crucial role in advancing power forecasting methods. The pilot envisions implementing large-scale collaborative analytics across data from various owners. This will enable short-term time series forecasting and will contribute to the optimisation of power plant performance.<\/p>\n PILOT 2: Electric vehicle charging<\/strong><\/p>\n The concurrent charging of multiple electric vehicles (EVs) in urban areas can strain the electrical grid. Implementing intelligent charging strategies becomes imperative to regulate charging rates and schedules, drawing insights from local data as well as external sources like driving patterns and data gathered from crowdsensing platforms. Energy-efficient AI systems have the potential to enhance battery and charging management, optimising vehicle charging during periods of high renewable energy supply.<\/p>\n The adoption of edge AI on its own may lead to an increase in the cost of EV chargers, owing to high computational requirements. To mitigate this, the pilot will employ frugal AI solutions across diverse data sources, fostering distributed intelligence. This approach aims to refine forecasting techniques and charging management strategies, ultimately fostering greater reliance on renewable energy sources for EV charging.<\/p>\n PILOT 3: Smart farming<\/strong><\/p>\n In the agricultural sector, AI applications rely on data sourced from remote sensing technologies such as satellites, drones, and in-situ sensors. These applications encompass a wide array of tasks, including the detection of water stress, monitoring crop diseases, and mapping weeds, all the way to predicting crop yields.<\/p>\n This is achieved through: Real-time generation of fertilisation maps tailored to soil and crop requirements for optimising fertilisation; early detection of plant pests and diseases through (near) real-time object assessment, achieved by fusing satellite imagery and drone inspections; and prognostication of soil health status and determining optimal harvesting times. The pilot will integrate federated learning techniques with digital twin solutions, reducing the volume of necessary training data and shortening model training times.<\/p>\n There are three primary objectives, firstly to enhance the learning of various plant pests and disease detection models by drawing insights from different users with diverse crops and pest\/disease challenges. Secondly, to generate a comprehensive soil health status map by integrating data from various sources and finally, to optimise fertilisation strategies based on different crops and soil conditions.<\/p>\n PILOT 4: Water management<\/strong><\/p>\n AI models can be used to optimise irrigation by assessing soil moisture levels, after integrating satellite imagery with precipitation data. Integrating accurate weather information is also crucial for effective irrigation, as it impacts evapotranspiration.<\/p>\n In farming, there is a pressing need for precise water measurement and treatment, including disinfection and oxygenation. Novel AI techniques are needed for determining suitable water types for different crops and optimising treatment towards increased water value. The pilot will develop AI-powered services for:<\/p>\n PILOT 5: Urban mobility infrastructure planning<\/strong><\/p>\n Emerging mobility models, including electric bike-sharing services, necessitate AI-driven applications for optimising the operation of docking and charging infrastructures. This involves implementing intelligent reservation systems for controlled charging.<\/p>\n The pilot is centred around fine-tuning the energy demand within an electric bike station network through advanced prediction techniques. Understanding usage patterns will yield crucial insights for prioritising charging at specific stations.<\/p>\n The envisioned application will provide information from data analysis encompassing routes, distances, frequency, and average durations of bike usage for achieving network stability and aid in strategically locating new charging stations. AI services will draw from various data sources including fleet management, recharging habits, journey analysis, and vehicle lifetimes to enable accurate prediction of energy demand and enhance infrastructure capacity and shared vehicle availability while mitigating overconsumption peaks during recharging.<\/p>\n PILOT 6: Smart banking<\/strong><\/p>\n The banking sector\u2019s prediction systems face challenges due to a reliance on human intervention and the diverse, sensitive nature of customer data. These factors make it difficult for existing systems to effectively discern default patterns. Moreover, any biases present in the training data can be perpetuated within the algorithms.<\/p>\n To address these issues, the pilot focuses on enhancing fraud detection systems. It aims to achieve this by employing trustworthy and interpretable feature-learning methods for both classification and regression tasks, thereby mitigating bias (Explainable AI).<\/p>\n Data spaces can serve as a digital enabler for the deployment of data-driven collaborative services. They provide a trustworthy framework for sharing operational data across decentralised networks, whilst introducing semantic interoperability. The latter is particularly important for establishing a common understanding of data meaning and context across diverse sources, domains and AI services.<\/p>\n Through data spaces data sovereignty, transparency and fairness are ensured, creating the ground for stakeholders\u2019 adoption. Data connectors facilitate seamless data exchange services while incorporating robust data security enforcement mechanisms.<\/p>\n GREEN.DAT.AI will combine data spaces with the Toolbox of AI Services, yielding significant benefits.<\/p>\n Smarter data exchange and integration<\/strong><\/p>\n Data spaces provide a unified framework enabling diverse data from various sources to be discovered, aggregated, integrated, and analysed. When coupled with AI, integrated data become a powerful resource for generating insights and making informed decisions.<\/p>\n Improved predictive capabilities<\/strong><\/p>\n AI algorithms excel at recognising patterns and making predictions. By leveraging data spaces, AI models have access to a broader and more comprehensive dataset. This leads to more accurate and reliable predictions, which can be invaluable in fields like forecasting, recommendation systems, and anomaly detection.<\/p>\n Increased efficiency and automation<\/strong><\/p>\n Data spaces combined with AI enable the automation of tasks that would be time-consuming or impractical to perform with traditional methods. This includes tasks like data cleansing, feature selection, decision-making processes, and distributed data processing.<\/p>\n Such architectures leverage the computing capabilities distributed over the edge-fog-cloud stack to cope with extreme-scale data, by storing and processing the data as close as possible to the data sources. This may also apply to data pre-processing for AI applications, involving more energy-efficient normalisation, feature extraction, handling missing values, and dealing with outliers.<\/p>\n GREEN.DAT.AI envisions demonstrating an AI-ready data space, thus a data management framework designed to support the use of energy-efficient AI techniques. This infrastructure will enable the efficient processing, analysis, and sharing of data across different organisations in a way that is compatible with AI workflows.<\/p>\n The key features of an AI-ready data space include interoperability, data quality, data security and privacy and data discovery. Overall, an AI-ready data space is an important component of an organisation\u2019s AI strategy, as it provides a foundation for the effective use of energy-efficient AI techniques in data-driven workflows. Together, these technologies are expected to unlock new potential for existing and future AI applications that address emerging business and societal challenges.<\/p>\n The project consortium comprises a diverse and multidisciplinary group of 17 partners<\/a> hailing from ten different countries. These partners bring together a wealth of complementary technical and scientific expertise, spanning various sectors including banking, mobility, energy, and agriculture.<\/p>\n Project Coordinator:<\/strong> Ioanna Fergadiotou, Ioanna.fergadiotou@inlecomsystems.com<\/a><\/p>\n Project Manager: <\/strong>Despina Brasinika, despina.brasinika@inlecomsystems.com<\/a><\/p>\n Communications: <\/strong>Panagiotis Katrakazas, p.katrakazas@aegisresearch.eu<\/a><\/p>\nEnergy demand of AI services<\/h3>\n
Developing energy-efficient AI<\/h3>\n
\n
Real-life industrial pilots<\/h3>\n
\n
Green AI-ready data spaces<\/h3>\n
About the team<\/h3>\n
\n
References<\/h4>\n
![\"\"](\"https:\/\/www.innovationnewsnetwork.com\/wp-content\/uploads\/2023\/10\/EU-flag-Horizon-Europe-1.jpg\")
<\/p>\n