Application of Gated Multimodal Learning in Building Energy Efficiency Performance Prediction and Retrofit Analysis

This paper proposes a gated multimodal learning framework that integrates EPC tabular data, assessment texts, and GIS spatial features to achieve interpretable prediction of building energy performance and analysis of retrofit scenarios. The framework aims to address the issues of poor timeliness, high cost, and limited coverage in traditional Energy Performance Certificate (EPC) assessments, providing a scalable, low-cost, and accurate decision support tool for building energy efficiency retrofits.

Achieving resilient and sustainable cities is a key global challenge, with decarbonization of residential buildings being particularly critical—residential buildings account for 20% of greenhouse gas emissions in the UK and 25% of energy-related emissions in the EU. Traditional EPC assessments rely on on-site inspections and have three major limitations:

• Poor timeliness: Unable to perform timely assessments at the urban scale
• High cost: Significant human and financial investment required for large-scale deployment of inspectors
• Limited coverage: Old or remote buildings struggle to get timely assessments Therefore, developing scalable, low-cost, and accurate prediction methods is of practical significance.

The gated multimodal framework proposed in this study integrates three types of heterogeneous data sources:

1. EPC tabular variables: Structured data such as building type, construction year, and heating system
2. Assessor free text: Qualitative descriptions of building conditions and potential issues
3. GIS spatial features: Geometric information such as floor area, height, orientation, and shape The core innovation is the sample-level gating mechanism, which dynamically adjusts the weight of each modality (e.g., new residential buildings rely on tabular data, old buildings rely on text, and buildings with complex geometry rely on GIS); an auxiliary grade classification head is also introduced to stabilize training.

Results from the Westminster, London case study:

• SAP score prediction: MAE=4.03 points, R²=0.757
• EI score prediction: MAE=4.76 points, R²=0.748
• Overall average MAE=4.39 points Ablation experiments show that full-modal fusion (tabular + text + GIS) achieves the best performance, followed by bimodal fusion, and unimodal fusion is the weakest, verifying the complementary effect of multi-source information.

The framework provides rich interpretability:

• Gating weights: The model relies heavily on assessor texts, reflecting the value of professional observations
• SHAP features: Main fuel type, building form, and construction year group are key influencing factors
• Text occlusion: Descriptions related to roofs and walls are the most informative
• Spatial features: Building height, floor area, and shape (irregular shapes have higher heat loss) are dominant factors

The framework is applied to three types of retrofit scenarios:

1. Wall insulation upgrade: Reduces heat conduction loss, improves SAP/EI scores, and lowers energy costs and CO₂ emissions
2. Roof insulation upgrade: High cost-effectiveness, quantifies the effects of different schemes
3. Window glass upgrade: High-performance low-emissivity glass reduces radiant heat loss and improves comfort Each scenario outputs a comparison of key indicators before and after retrofit, supporting investment decisions.

Research significance: Provides tools for policymakers/planners to support retrofit priority ranking, optimal resource allocation, and net-zero housing transition. Technically, the gated multimodal design can be extended to other heterogeneous data fusion tasks. Future directions: Explore more complex gating architectures, introduce time-series data for dynamic prediction, and expand to more cities and building types.