Multi-Zone Office HVAC Dataset with Real External Signals and Multi-Horizon Price Forecasts: A Four-Year Benchmark

This dataset was developed for research on forecast-assisted HVAC control, building energy management, demand response, thermal comfort, and indoor air quality in multi-zone office buildings. The modeled building consists of eleven heterogeneous zones, including eight office rooms, one hall/corridor, one meeting room, and one conference room, all served by a centralized thermal-conditioning and ventilation system. The dataset covers four consecutive years, from January 1, 2020, to December 31, 2023, at a 30-minute temporal resolution. It combines real external signals, including CAISO real-time electricity prices and NSRDB weather and solar-resource variables, with simulated office occupancy, meeting and conference events, and control-oriented metadata describing the thermal and ventilation characteristics of each zone. Six-step electricity-price forecasts, covering the next 30 minutes to three hours, are provided for every timestamp. These forecasts were generated using a Peak-Aware Multi-Horizon Dual-Attention LSTM model (PA-MHDA-LSTM), which combines an LSTM encoder–decoder architecture with encoder self-attention, temporal context aggregation, future-time embeddings, horizon embeddings, and decoder-to-history cross-attention. The forecasting pipeline follows a strictly chronological and leakage-free evaluation procedure. Data normalization and model fitting use only the training period, model selection is performed on the evaluation period, and the test-period forecasts are generated without access to future electricity-price targets. Only historical observations and known future calendar features are used to predict each forecasting horizon. The data are divided chronologically into training data for 2020–2021, evaluation data for 2022, and test data for 2023. Each subset contains the office operating data, six-horizon electricity-price forecasts, and zone metadata. The dataset is intended for reproducible evaluation of reinforcement-learning, model-predictive, rule-based, forecasting-assisted, and other intelligent HVAC control methods.本数据集专为多区域办公楼的预测辅助暖通空调（Heating, Ventilation and Air Conditioning, HVAC）控制、建筑能源管理、需求响应、热舒适及室内空气质量研究而开发。所建模的办公楼包含11个异质区域，涵盖8间办公室、1处大厅/走廊、1间会议室以及1间报告厅，全部由集中式温控通风系统提供服务。 本数据集覆盖2020年1月1日至2023年12月31日连续四年的时间序列数据，时间分辨率为30分钟。数据集整合了真实外部信号与模拟数据：真实外部信号包含加州独立系统运营商（CAISO）实时电价、国家太阳能辐射数据库（NSRDB）的气象与太阳能资源变量；模拟数据涵盖模拟办公人员驻留情况、会议与活动事件，以及面向控制的元数据，该元数据可描述各区域的热工与通风特性。 每个时间戳均提供6阶电价预测，覆盖时长为未来30分钟至3小时。此类预测采用峰值感知多视野双注意力长短期记忆网络（Peak-Aware Multi-Horizon Dual-Attention LSTM, PA-MHDA-LSTM）生成，该模型结合了长短期记忆（Long Short-Term Memory, LSTM）编码器-解码器架构，集成编码器自注意力、时间上下文聚合、未来时间嵌入、视野嵌入以及解码器-历史跨注意力机制。 该预测流程遵循严格的时序化且无数据泄露的评估准则：数据归一化与模型拟合仅使用训练期数据，模型选择基于评估期数据完成，测试期预测生成过程不会引入未来电价目标信息；仅依靠历史观测数据与已知的未来日历特征，即可完成各预测视野的电价预测。 数据集按时序划分为三部分：2020-2021年训练集、2022年评估集以及2023年测试集。每个子集均包含办公楼运行数据、6阶电价预测数据以及区域元数据。 本数据集可用于可复现地评估强化学习、模型预测控制、基于规则、预测辅助及其他智能暖通空调控制方法。