Accurate building energy forecasting is essential for efficient energy management and sustainable building operation. While recent studies suggest that large language models (LLMs) exhibit strong potential for time-series forecasting, their application to building energy prediction remains limited by high computational cost, inefficient parameter utilization, and inadequate modeling of multivariate dependencies. To overcome these challenges, this paper proposes MaPL-LLM, a lightweight LLM-based forecasting framework that integrates multivariate prompt fusion and temporal patching. MaPL-LLM adopts a frozen LLaMA-1B backbone and introduces two complementary modules: (1) a multivariate prompt-based embedding module that encodes temporal context and statistical characteristics into structured textual prompts, and (2) a multivariate patching-based numerical embedding module that captures local temporal patterns and cross-variable interactions. Only lightweight input transformation and output projection layers are trained, significantly improving computational efficiency while maintaining stable performance. Extensive experiments on multiple building types from the BDG2 dataset demonstrate that MaPL-LLM consistently outperforms state-of-the-art methods, including TimeXer, PatchTST, and TFDFNet. For a 24-step forecasting horizon, MaPL-LLM achieves an MAE of 0.166, an MSE of 0.072, and an R² of 0.946, reducing forecasting error by over 10% compared with the strongest baseline. Moreover, training and inference time are reduced by more than 60% relative to existing LLM-based methods, highlighting its suitability for scalable and real-time building energy management applications.

Short-term load forecasting (STLF) is crucial to building energy optimization. Among existing approaches, forecasting methods face challenges in improving predictive accuracy, achieving cross-building transferability, and meeting engineering compliance requirements, among others. To address these challenges, we propose Building-MoE, a sparse Transformer-based Time-series Foundation Model (TSFM) for building STLF that integrates a Mixture-of-Experts (MoE) into an encoder–decoder framework to enhance cross-building transferability under heterogeneous and non-stationary building loads. At inference, only about 98M active parameters are engaged (compared with Time-MoE ≈198M and Chronos-Bolt ≈ 203M), delivering higher parameter and compute efficiency. However, sparse MoE training is prone to expert imbalance or collapse, which reduces effective capacity and generalization. Therefore, we design a Closed-Loop Routing Scheduler (CLRS) that continuously monitors routing entropy and the maximum expert share, combines stagewise temperature and noise scheduling with feedback corrections and short pulse interventions, and revives long-inactive experts through bias updates. We also introduce a token-weighted Load Balancing Loss (LBL) that normalizes by routed tokens to suppress long-term imbalance, and we use the Huber loss to improve robustness to anomalies and noise. Under the domain adaptation setting, the proposed Building-MoE achieves state-of-the-art (SOTA) performance, attaining average CVRMSE 23.21%, NMAE 15.29%, and NMBE 2.95%, with the vast majority of scenarios satisfying the ASHRAE Guideline 14 hourly thresholds (CVRMSE ≤ 30%, |NMBE| ≤ 10%). Under the direct transfer setting, Building-MoE also achieves SOTA performance, with averages of CVRMSE 27.22%, NMAE 18.26%, and NMBE 3.80%, and most scenarios likewise satisfying the same thresholds, demonstrating stable cross-building generalization.