AI-Driven Solutions for Efficient Generation and Optimisation of Substation Designs in Brownfield Context.

The paper begins by outlining the critical challenges faced in brownfield substation design, including spatial restrictions, equipment obsolescence, interdependencies between new and existing assets, compliance challenges based on original versus current standards, and the necessity for continuous operation during the execution of works. Traditional CAD-based design approaches often fall short in managing complexity and staging requirements, leading to inefficiencies, delays, and increased cost and operational risk.

This study proposes a comprehensive AI-driven framework integrating generative design algorithms, reinforcement learning, constraint-based modelling, and computer vision to support planning and layout optimisation in brownfield substations. Historical design data, operational constraints, environmental parameters, and equipment specifications are leveraged to generate multiple viable design alternatives. Evolutionary algorithms and machine learning techniques evaluate and rank these options based on performance metrics such as cost, compliance, safety, reliability, construction and commissioning staging, and maintainability.

A major contribution is the development of an intelligent design assistant fully integrated in

CAD tools, capable of suggesting circuits and layouts that account for existing infrastructure while ensuring regulatory compliance and utility-specific practices. The AI system identifies conflicts between new and legacy components, predicts operational bottlenecks, and proposes staging strategies to reduce outage duration and extent. Feedback loops allow the model to learn from expert adjustments, improving future recommendations.

1 Reinforcement learning further guides design refinements using domain-specific heuristics and feedback. A reward function penalises violations (clearance breaches, non-compliance, outage duration) and incentivises optimal space utilisation and reduced staging complexity. Computer vision modules interpret as-built drawings, including older raster-type data, or drone imagery, enabling rapid digital twin creation and real-time site condition mapping.

Australian case studies validate the methodology. Results demonstrate significant time savings in conceptual and detailed design phases, improved spatial efficiency, and reduced design iterations compared to traditional methods. In one example, the AI-driven approach reduced layout development time by 30% and improved space utilisation by 25% while maintaining security of supply during implementation.

To contextualise these improvements, the AI-assisted design workflow for a representative 132 kV brownfield substation was completed in approximately 14 working days from initial site data ingestion to final layout recommendation, compared to an estimated 20 working days using conventional 2D/3D CAD-based design methods for the same scope.

The time saving is primarily attributable to the automated generation and evaluation of candidate layouts, which eliminates several manual iteration cycles. Similarly, the 25% improvement in spatial utilisation was measured by comparing the total equipment footprint and maintenance access area achieved by the AI-optimised layout against the best manually produced alternative for the same site, normalised to the available site area.

The paper concludes by discussing implementation considerations, including data availability and security, system integration challenges, and the need for human-in-the-loop oversight.

Future research directions include integrating AI tools with Building Information Modelling

(BIM) platforms and enhancing interpretability of AI-generated designs for engineering teams.