Application of multimodal machine learning in building ecological assessment

Traditional building ecological assessment methods often rely on single-source data or static evaluations, failing to capture the dynamic, multi-dimensional nature of construction lifecycles. While recent BIM + IoT and multimodal approaches offer improvements, they struggle with critical challenges including spatiotemporal misalignment, modality dominance, and inconsistent semantics, leading to suboptimal fusion and limited predictive accuracy.To address these gaps, we propose MMEANet (Multimodal Environmental Awareness Network). Distinct from existing studies that typically employ simple concatenation or early fusion, MMEANet introduces a framework featuring: (1) a dynamic graph modeling mechanism with explicit spatiotemporal alignment to unify heterogeneous data streams; (2) a late attention fusion strategy that adaptively weights modalities based on lifecycle stages; and (3) an end-to-end multi-task learning architecture for simultaneous prediction of carbon emissions, energy intensity, and environmental risks. Extensive validation on a large-scale commercial project demonstrates that MMEANet significantly outperforms state-of-the-art baselines. The model achieves a Mean Squared Error (MSE) of 0.042 and a Mean Absolute Error (MAE) of 0.158 for implicit carbon prediction, representing an 18.5% improvement in MSE and 21.3% in MAE over the best-performing baseline. Furthermore, it attains an F1-score of 0.88 in anomaly detection with a mean absolute percentage error (MAPE) of 5.2%. These results confirm MMEANet’s capability to provide robust, real-time, and interpretable ecological assessments across the building lifecycle.