Accurate cooling load forecasting in high-efficiency chiller plants with ice storage systems is essential for intelligent control, energy conservation, and maintaining indoor comfort. However, conventional forecasting methods often struggle to model the complex nonlinear dependencies among influencing variables, limiting their predictive performance. To address this, this paper introduces Time-LLM, a novel time series forecasting framework that leverages a frozen large language model (LLM) to improve the accuracy and generalization of cooling load forecasting. Time-LLM extracts features from historical data, reformulates them as natural language prompts, and uses the LLM for temporal sequence modeling; a linear projection layer then maps the LLM output to final predictions. To enable lightweight deployment and improve temporal feature prompting, we propose ETime-LLM, an enhanced variant of Time-LLM. ETime-LLM significantly reduces deployment costs and mitigates the original model’s response lag during trend transitions by focusing on possible turning points. Extensive experiments demonstrate that ETime-LLM consistently outperforms or matches state-of-the-art baselines across short-term, long-term, and few-shot forecasting tasks. Specifically, in the commonly used 24 h forecasting horizon, compared with the original model, ETime-LLM achieves an approximately 17.3% reduction in MAE and a 19.3% reduction in RMSE. It achieves high-quality predictions without relying on costly external data, offering a robust and scalable solution for green and energy-efficient HVAC system management.
Zhu et al. (2025) studied this question.