Microsoft’s Azure AI Agent Services powers innovative campus energy management system

A collaborative project between Imperial College London and Microsoft develops a novel multi-agent conversational system for university energy sustainability, highlighting advancements in autonomous AI ecosystems and hybrid data retrieval techniques.

A proof-of-concept conversational multi-agent system for campus energy management has been developed and evaluated using Microsoft’s emerging Azure AI Agent Services. This initiative addresses the increasing need for smart energy solutions in universities striving to meet net-zero emission targets, blending commercial incentives with research into autonomous large language model (LLM) frameworks. The project, led by MSc Applied Computational Science students from Imperial College London in collaboration with Microsoft, focuses on creating an adaptive, user-centric chatbot to engage stakeholders across campus — students, faculty, and administrators — with energy management information and tools.

The system exemplifies a novel integration of specialised AI agents orchestrated through a triage coordinator agent. Four domain-specific agents handle different functions: one retrieves general campus energy information; another caters to administrators with more detailed, confidential data; a chart plotter generates visual summaries and prognostics; and a feedback agent collects and preprocesses student inputs. These agents are supported by Azure AI Language Services for conversational understanding and custom question answering, alongside Azure AI Search enabling a hybrid retrieval-augmented generation (RAG) approach to efficiently provide contextual answers by combining semantic and keyword search techniques. Synthetic energy data, based on real monitoring schemas and supplemented by mocked environmental and timestamped data, is stored in Azure Cosmos DB and Blob Storage, facilitating dynamic updates and document indexing.

Evaluations reveal that GPT-4o outperforms other base models in specialised agent tasks, while larger LLMs like Llama-3.3 70B do not necessarily provide superior results, highlighting the complexities and nuances in model selection for domain-specific applications. Despite testing various prompting strategies such as few-shot examples and reasoning trajectories, the system’s response quality and workflow directionality depend more heavily on agent type—action-based feedback agents benefit from reasoning prompts that foster sequential actions, whereas information retrieval agents produce more comprehensive responses under such prompts. The triage agent was found critical in synthesising answers concisely, avoiding excessive verbosity by effectively routing queries among agents.

The project faced challenges typical of composite AI workflows, including the difficulty of tuning multiple integrated Azure services—such as Custom Question Answering, Conversational Language Understanding, and RAG—for optimal performance. Additionally, limitations in the evaluation framework, which used a small set of 50 queries scored on a coarse scale, restrict the capacity to measure nuanced agent contributions and user experience impacts. The synthetic dataset lacks real-world variability, such as long-term seasonal or academic calendar patterns, and safety mechanisms around sensitive data access require more thoughtfully crafted domain-specific scenarios beyond default cloud service instructions.

Looking ahead, the research team suggests expanding synthetic datasets by incorporating longer-term energy simulations using tools like EnergyPlus, broad user-testing across stakeholder groups to refine workflows, and developing targeted safety attack scenarios to robustly evaluate system security—such as simulated prompt injections designed to breach access controls. The modular multi-agent design, integration of language understanding services, and hybrid retrieval approach are considered foundational strengths that can underpin future enhancements toward generalist, adaptive campus energy management solutions.

This development aligns with broader emerging trends in AI-driven energy management. Parallel research, including projects like GridMind, integrates LLM-powered agents with traditional engineering solvers to deliver conversational scientific computing with numerical precision for power systems. Other initiatives deploy multi-agent reinforcement learning frameworks to detect and manage grid violations in real-time, illustrating the expanding role of autonomous agents in energy infrastructure. Microsoft’s ecosystem, featuring Azure AI Foundry and the AI Agent Service, provides a comprehensive platform enabling the secure design, deployment, and scaling of such AI agents across complex data environments, including integrations with Bing, Microsoft Fabric, and third-party APIs.

Real-world applications of Microsoft’s AI technologies in energy management are already improving customer engagement and operational efficiency for utilities, as evidenced by Aydem Energy’s successful use of Azure OpenAI Service in handling customer service requests. Meanwhile, Azure Data Manager for Energy supports energy companies in modernising data infrastructures to enhance analytics and decision-making capabilities.

This campus-focused conversational agent project thus represents a leading-edge demonstration of AI’s potential to move beyond traditional monitoring systems towards interactive, multi-faceted energy management tools tailored for institutional settings. Continual advancements in AI agent frameworks, cloud integration, and domain-specific modelling will be crucial to realise the scalable, adaptive energy ecosystems universities aspire to deploy.

📌 Reference Map:

Source: Noah Wire Services