AI Planner Turns UK Public Building Retrofits into Decision-Ready Whole-Life Costing

Deciding which public buildings to retrofit first, and in what order, has long been one of the most stubborn problems in UK public-sector asset management. Councils, hospital trusts and schools must balance whole-life cost, carbon reduction, service continuity and auditability under limited data and constrained capital budgets. A new study published in Complex & Intelligent Systems by Ziru Tao, Paul Baguley, Rashid Maqbool and Obuks Ejohwomu of the University of Manchester proposes an answer that blends environmental, social and governance priorities with uncertainty-aware artificial intelligence into a single decision-ready workflow for retrofit appraisal.

The researchers built their workflow around a scenario library constructed from national non-domestic Energy Performance Certificate and Display Energy Certificate data, delivery evidence from the Public Sector Decarbonisation Scheme, and stakeholder inputs. This library captures public-building archetypes, the retrofit actions available to each, and the appraisal constraints that decision-makers face. Rather than treating these data sources as separate silos, the framework integrates them so that every candidate retrofit sequence can be evaluated against realistic budget and performance conditions.

A central technical challenge was translating stakeholder priorities into something an algorithm could use. The team gathered priorities through interviews, a questionnaire and a co-design workshop, then converted them into transparent weights using a fuzzy analytic hierarchy process. F-AHP extends the classical pairwise-comparison method by allowing decision-makers to express judgements with fuzzy values, acknowledging that human preferences about cost versus carbon versus service disruption are rarely crisp numbers. These weights then parameterise a reinforcement learning planner, meaning the AI’s optimisation objective directly reflects what stakeholders said mattered most.

To evaluate retrofit policies repeatedly without running expensive building simulations each time, the researchers developed a surrogate model combining a Bayesian convolutional neural network with a long short-term memory network. The BCNN–LSTM surrogate provides uncertainty-bounded trajectories of energy use, carbon emissions and cost over time. The Bayesian component is crucial: it attaches probability distributions to its predictions rather than single point estimates, so planners can see how confident the model is about each projected outcome. The LSTM captures temporal dependencies, tracking how energy and cost performance evolves across a building’s life as retrofit measures take effect.

The decision-search stage of the workflow uses a deep Q-network, a reinforcement learning algorithm that learns to select retrofit actions by estimating the long-term value of each possible choice. For 1000 matched archetype–budget cases, the trained DQN achieved a 30.03-fold speedup over traditional scalar feasible-sequence enumeration. That speedup matters because retrofit sequencing is a combinatorial problem: with many buildings, many possible measures and limited budgets, the number of feasible sequences grows explosively. Enumeration quickly becomes intractable, while a trained network can search the space efficiently once it has learned the structure of the problem.

Performance results show that the workflow supports auditable retrofit sequencing across archetype and budget conditions and improves composite ESG performance relative to conventional baselines. The team compared their approach against whole-life costing only, £/tCO2e ranking, and static multi-criteria decision analysis. Each baseline represents a common current practice: pure cost optimisation, carbon-per-pound ranking, or fixed weighting schemes that cannot adapt to context. The integrated workflow outperformed all three, suggesting that combining ESG weights with uncertainty-aware AI produces better-balanced retrofit portfolios than any single criterion alone.

Auditability is a recurring theme in the paper. Public-sector spending in the UK is subject to scrutiny, and decisions that cannot be explained are decisions that cannot be defended. The workflow is designed to be explainable: stakeholder weights are transparent, the surrogate’s uncertainty bounds are explicit, and the reinforcement learning planner’s choices can be traced back to the priorities and constraints that parameterised it. This contrasts with black-box AI approaches that might achieve good results but leave auditors unable to verify why a particular building was sequenced before another.

The authors are careful about the limits of their validation. The current evidence remains based on public datasets and scenario-level analysis, so the findings demonstrate what the workflow can support in principle rather than proving performance on live portfolios. Future work, they note, should extend the framework using project-specific metered data, digital twins and live portfolio deployment. Digital twins, virtual replicas of buildings updated with real-time sensor data, could replace scenario-level estimates with measured performance, tightening the uncertainty bounds the surrogate provides.

The implications reach beyond the UK public sector. Many countries face the same structural problem: large building stocks, decarbonisation targets, tight budgets and stakeholders with conflicting priorities. The Manchester team’s contribution is an integrated and explainable whole-life costing, ESG and AI workflow for early-stage retrofit decision support, a template that could be adapted wherever decision-makers need defensible, data-driven sequencing of capital investments. By showing that fuzzy stakeholder weights, Bayesian surrogates and reinforcement learning can be stitched into one auditable pipeline, the study moves retrofit planning from intuition and spreadsheets toward something closer to engineering discipline.

For the researchers, the next step is deployment. Moving from public datasets to live portfolios will test whether the 30-fold search speedup and improved ESG composites survive contact with messy real-world data, from incomplete meter readings to shifting budget allocations mid-programme. If they do, the workflow could become a practical tool for the public bodies spending billions on decarbonisation, giving them a way to spend each pound where it delivers the most verified, explainable value across a building’s whole life.

Subject of Research: AI-driven whole-life costing for decision-ready UK public building retrofits

Article Title: AI-driven whole-life costing for decision-ready UK public building retrofits

Article References: Tao, Z., Baguley, P., Maqbool, R., & Ejohwomu, O. (2026). AI-driven whole-life costing for decision-ready UK public building retrofits. Complex & Intelligent Systems. https://doi.org/10.1007/s40747-026-02516-9

Image Credits: AI Generated

DOI: 10.1007/s40747-026-02516-9

Keywords: artificial intelligence, reinforcement learning, whole-life costing, building retrofit, public sector decarbonisation, ESG, Bayesian deep learning, UK public buildings, AI-driven, whole-life, costing, decision-ready