Bianca Howard

To support rapid urban decarbonization, building energy modeling must scale without sacrificing building-level accuracy. Current prototype-driven Urban Building Energy Models (UBEM) suffer from a severe data bottleneck, forcing reliance on default assumptions that drastically mischaracterize real-world performance. While a parallel body of work has applied generative AI to building energy modeling, these efforts have focused on automating low-level simulation syntax for individual, well-specified buildings, leaving the urban data bottleneck unaddressed. This study introduces an autonomous, multi-agent pipeline that uses Large Multimodal Models (LMMs) to ingest unstructured public data, ranging from Certificates of Occupancy (CO) to LiDAR point clouds, and directly pilot high-level simulation tools (e.g. Honeybee, OpenStudio), producing detailed, multi-zone models with explicit HVAC representation, demonstrated on 50 buildings in New York City (NYC).

Automated information extraction from construction drawings is a critical yet challenging task. While traditional approaches, from image processing to deep learning, have made great progress, they often suffer from inconsistent performance, extensive task-specific training, and narrow scope. Large Multimodal Models (LMMs) offer a promising alternative due to their vast knowledge bases and instruction-following capabilities, however, they struggle with spatial reasoning and dense high-fidelity images. This preliminary study evaluates the performance of two off-the-shelf Multimodal Generative AI (GenAI) workflows against two specialized architectures: an image-based Retrieval Augmented Generation (RAG) system and an image-based Multi-Agent workflow. By evaluating these workflows against a targeted set of benchmarking questions, this work provides a preliminary assessment of current GenAI capabilities in drawing interpretation and highlights specific areas for improvement in developing more robust automated workflows.

This paper proposes a socio-technical optimization approach for retrofitting the housing stock that mitigates fuel poverty while minimising both capital costs and carbon emissions, reconciling social, environmental, and economic sustainability. A two-stage optimization method identified appropriate retrofit measures. An exhaustive search is conducted to identify solutions that minimise greenhouse gas emissions and capital and operational costs, excluding solutions that result in fuel poverty. Building energy models were created for 597 building archetypes, as described by the English Housing Survey, and three different fuel poverty metrics were evaluated. Subsequently, the non-dominated solutions from the first stage were used to find stock level solutions, minimising greenhouse gas emissions and capital cost. Retrofit for 1.3 million UK dwellings was simulated. Results found that incorporating fuel poverty constraints limited the achievable decarbonisation, reducing its potential by 21–87 %. The order in which different EEMs were installed, from high to low emissions solutions, was consistent across both constrained and unconstrained solutions, with virgin roof insulation and cavity wall insulation always used first to reduce greenhouse gas emissions. The most costly measures, however, such as heat pumps, triple glazing, and solid wall insulation, were severely limited in their use even in the highest cost scenarios due to the potential for these measures to increase fuel poverty among households. This work demonstrates that using a typical techno-economic analysis without considering the impacts of fuel poverty would lead to an overestimation of the greenhouse gas emissions reduction potential and an underestimate of the need to address the costs of deep decarbonisation measures.

Building decarbonization, the aim to make buildings more energy efficient and switch to low-carbon fuels, is necessary to achieve climate change goals. Traditional approaches to create building decarbonization pathways often only examine cost and GHG emissions to determine viable retrofit solutions. If considered, impacts of broader social metrics of interest to tenants and elected officials are typically considered in subsequent analyses, hiding the explicit tradeoffs that occur across these, in some cases, competing goals.This work uses multi-objective optimization to explicitly explore how optimal retrofit scenarios change when considering objectives beyond cost and greenhouse gas emissions. The results of three optimizations with differing objectives are compared: a traditional techno-economic analysis, a techno-economic analysis extended to include estimates of energy burden, and a techno-economic analysis extended to include metrics for job creation. Analyzing eight buildings in West Harlem, NYC, with passive and active upgrades, the traditional techno-economic optimization prioritized low efficiency and low cost HVAC electrification, whereas the optimization considering energy burden considered measures to improve envelope and highly efficient HVAC electrification options in the most energy burdened buildings. The results show that incorporating these objectives during the optimization process leads to vastly different retrofit solutions meaning these metrics should be considered explicitly in decision-making tools.

Decarbonizing the built environment will likely involve an increase in renewable energy supply, efficient demand management and energy storage. An enabling technology required for this transition is for buildings to have more advanced HVAC controls capable of optimizing for different objectives such as cost saving, occupant comfort and peak load management. Reinforcement learning (RL) has shown promise in this area; however, ensuring good performance from RL agents requires either developing a realistic training environment, collecting a substantial amount of real-world data for offline training or a hybrid approach (Ding et al, 2024). Our work will focus on the development of realistic training environments and seek to answer the question: How close does the training environment have to match the real building to provide good performance? Specifically, we assess if the steady-state heat loss coefficient (HLC) can be used to calibrate building models to create training environments. Additionally, we explore how a mismatch between the training environment HLC and the real-world HLC might affect control performance. The steady-state HLC was calculated for each room in a real test building through a co-heating test. EnergyPlus was used to create a model of the building where the HLC of each room was calibrated to the measured values from the real building. An RL agent was trained in the simulated environment to control electric space heaters in three of the rooms to track a setpoint of 20°C during the day and 15°C at night. The agent was trained using Deep Q Learning, with each of the three room temperatures, the desired setpoint, and the outdoor air temperature as the input states, and the eight combinations of on/off control of the heaters in each room as the actions. The trained agent was tested in the real building for six days, with room temperatures recorded at one-minute intervals. The results showed that the air temperature was maintained well within the expected bounds, and there was no sim-to-real performance gap. This demonstrates that in this scenario, calibrating the training environment using the steady state HLC was suitable to ensure good performance of the RL agent. The training was repeated with separate RL agents using three building models in which the HLCs deviated slightly from the measured values. These agents were tested in the real building and also delivered good setpoint tracking showing that some mismatch between the training and real-world environment can still provide good performance.

This work seeks to contribute knowledge towards the question: How accurately do building simulators for training reinforcement learning (RL) agents need to be specified to result in good control performance? In this work four RL agents were trained using deep Q learning to maintain temperature set points in 4 different EnergyPlus-based environments whose steady-state heat loss coefficient varied by around 30%. These policies were then used to control a target environment in both simulation and in a real-world test bed. The results show that, for the conditions applied, the policies trained in different environments were able to maintain temperatures within ± 1°C of the setpoint in the target environment in the simulation and real-world test bed. This indicates that for temperature control, reasonable assumptions about building construction materials, insulation levels, and infiltration rates can generate an EnergyPlus model that is a sufficient environment for training RL agents to create high-performing control policies.

The complexity of performance-based building design stems from the evaluation of numerous candidate design options, driven by the plethora of variables, objectives, and constraints inherent in multi-disciplinary projects. This necessitates optimization approaches to support the identification of well performing designs while reducing the computational time of performance evaluation. In response, this paper proposes and evaluates a sequential approach for multi-objective design optimization of building geometry, fabric, HVAC system and controls for building performance. This approach involves sequential optimizations with optimal solutions from previous stages passed to the next. The performance of the sequential approach is benchmarked against a full factorial search, assessing its effectiveness in finding global optima, solution quality, reliability to scale and variations of problem formulations, and computational efficiency compared to the NSGA-II algorithm. 24 configurations of the sequential approach are tested on a multi-scale case study, simulating 874 to 4,147,200 design options for an office building, aiming to minimize energy demand while maintaining thermal comfort. A two-stage sequential process-(building geometry + fabric) and (HVAC system + controls) identified the same Pareto-optimal solutions as the full factorial search across all four scales and variations of problem formulations, demonstrating 100% effectiveness and reliability. This approach required 100,700 function evaluations, representing a 91.2% reduction in computational effort compared to the full factorial search. In contrast, NSGA-II achieved only 73.5% of the global optima with the same number of function evaluations. This research indicates that a sequential optimization approach is a highly efficient and robust alternative to the standard NSGA-II algorithm.

The complexity of performance-based building design stems from the evaluation of numerous candidate design options, driven by the plethora of variables, objectives, and constraints inherent in multi-disciplinary projects. This necessitates optimization approaches to support the identification of well performing designs while reducing the computational time of performance evaluation. In response, this paper proposes and evaluates a sequential approach for multi-objective design optimization of building geometry, fabric, HVAC system and controls for building performance. This approach involves sequential optimizations with optimal solutions from previous stages passed to the next. The performance of the sequential approach is benchmarked against a full factorial search, assessing its effectiveness in finding global optima, solution quality, reliability to scale and variations of problem formulations, and computational efficiency compared to the NSGA-II algorithm. 24 configurations of the sequential approach are tested on a multi-scale case study, simulating 874 to 4,147,200 design options for an office building, aiming to minimize energy demand while maintaining thermal comfort. A two-stage sequential process-(building geometry + fabric) and (HVAC system + controls) identified the same Pareto-optimal solutions as the full factorial search across all four scales and variations of problem formulations, demonstrating 100% effectiveness and reliability. This approach required 100,700 function evaluations, representing a 91.2% reduction in computational effort compared to the full factorial search. In contrast, NSGA-II achieved only 73.5% of the global optima with the same number of function evaluations. This research indicates that a sequential optimization approach is a highly efficient and robust alternative to the standard NSGA-II algorithm.

Abstract Reinforcement learning (RL) for controlling building heating, ventilation, and air conditioning (HVAC) systems has shown promise in delivering energy-efficient, energy-flexible and resilient buildings. However, training RL agents in real-time in buildings is generally considered time and cost prohibitive. RL agents can be trained offline in simulated environments, reducing the time to performance once online, but complex models for simulation can be difficult to generate. What level of complexity is required in an offline training model to produce acceptable online results is unknown. This work seeks to determine how simplified models can be used to train RL agents for building HVAC control. The work uses the EnergyPlus simulation environment paired with HVAC systems modeled in Modelica to act as the real-world analogue environment. Simplified reduced-order models of the building were developed using the lumped capacitance method. The results show that an RL agent trained using Deep Q Learning on a simplified building model can be transferable to a more complex building simulation environment with comparable performance. In virtually all cases, the RL agents were able to deliver improved performance over the baseline proportional controller. Online learning gave mixed results in terms of enhancing the policy transfer process. The implementation in this paper was more suitable for improving the performance of poor policy transfer rather than fine-tuning already well-performing policies.

There is an increase in renewable energy sources connected to the electricity grid due to recent drives to achieve grid decarbonization milestones. However, such expansions cause grid balancing issues due to the renewable sources intermittency. Thus, grid operators introduced demand response <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">(DR)</b> schemes to mitigate this problem by controlling consumer load demands in exchange for incentives. Industries have enormous electricity demand making them ideal candidates for such programs. Nonetheless, non-intrusive demand load flexibility assessment for an industry's potential in DR programs remains a challenge. In this paper, a data-driven framework for quantifying the DR potential of an industrial consumer is proposed. The framework uses smart electricity meter data to identify operational patterns to derive a flexibility boundary that quantifies the flexibility in the industrial consumer's system. The framework also evaluates the <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">DR</b> participation scenario to quantify the net benefit of trading the identified flexibility. A Case-study has been carried out for two industrial consumers (i.e., an electronics factory and a poultry feed factory). Initial energy behavioral analysis indicates three different energy use patterns for the electronics factory and six energy use patterns for the poultry feed factory. Evaluating the operational flexibility boundary for the clusters, the framework found two feasible clusters with DR potentials for the electronics factory and three feasible clusters for the poultry feed factory. The cost-benefit analysis indicates a potential energy cost reduction in the region of <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$5\%-8\%$</tex-math></inline-formula> for passive participation and as much as <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$12\%-24\%$</tex-math></inline-formula> for active participation. The framework could be adopted to evaluate wide scale industrial consumers' flexibility potential.