Buildings account for approximately 30–40% of global final energy consumption and nearly one-third of greenhouse gas emissions, making intelligent energy management a critical pathway toward decarbonization. Model predictive control (MPC) has emerged as a powerful framework for building energy management due to its ability to optimize energy cost, thermal comfort, and operational constraints under uncertainty. However, the effectiveness of MPC is highly dependent on the quality and temporal consistency of load and temperature forecasts, while access to sufficient real building data is often limited by privacy and availability constraints. This paper proposes an integrated building energy management framework that combines multi-scale synthetic data generation, a modern state-space sequence forecaster, and MPC. A multi-scale hierarchical generative adversarial network (MS-HGAN) is employed to generate privacy-preserving synthetic trajectories that augment limited real training data for forecasting, rather than acting as a control surrogate. Forecasting is performed using a Mamba-based state-space model with a Random Forest (RF-Tree) output regressor, and the resulting predictions are embedded into an MPC architecture with probabilistic constraints and a safety filter. The proposed approach is evaluated on a large commercial multi-zone shopping mall (retail center) operating in a temperate climate region (Tehran-like, Köppen Csa). To support augmentation under privacy constraints, MS-HGAN is trained on auxiliary building energy time-series datasets and is used to enrich the training distribution of the forecaster via a controlled real-only, synthetic-only, and mixed real–synthetic study. Comprehensive evaluations are conducted using both short-term (30-day) and long-term (one-year) closed-loop simulations. Results show that mixed real–synthetic training improves forecasting robustness and translates into superior MPC performance in terms of energy savings, reduced comfort violations, and peak-load shaving, compared with deterministic MPC and strong learning-based baselines, including LSTM/GRU, ARX, S4, and diffusion-based predictors. The findings provide direct evidence that privacy-motivated synthetic augmentation can play a complementary and measurable role in improving predictive control performance for commercial buildings through the forecasting-to-MPC pathway.

Reliable prediction of tunneling-induced ground settlement in coastal environments is challenged by coupled hydro-mechanical processes, site heterogeneity, and limited labeled data in new projects. This study presents a physics-informed and transferable deep learning framework that embeds dual physical constraints (Peck-type settlement envelope and Darcy-based pore-pressure consistency) within a dynamic transfer learning architecture and an actionable explainable-AI (XAI) layer. The model fuses heterogeneous, time-synchronized streams from tunnel boring machines, geotechnical sensors, and groundwater monitoring systems, adapting across segments and sites through similarity-aware transfer. Applied to a coastal shield-tunneling project (~ 80 000 samples), the framework achieved \({R}^{2}\)= 0.95, RMSE = 1.5 cm, MAE = 1.2 cm, outperforming empirical regressions and unconstrained deep baselines. Ablation studies confirm that physics regularization enhances calibration and stability, while transfer learning improves cross-site generalization. For field deployment, the end-to-end pipeline operates in a 30 s cycle with median latency ≈ 1.6 s and peak inference time ≈ 0.05 s/sample (≈ 20 predictions/s). The XAI layer translates feature attributions into actionable engineering guidance—such as adjusting shield pressure, advance rate, or dewatering—to enable interpretable, auditable decision-making. These findings highlight physics-regularized, explainable transfer learning as a practical route toward trustworthy AI for geotechnical risk prediction in coastal tunneling.