The lighting system accounts for 8% of the total electricity consumption in commercial buildings in the United States and 12% of the total electricity consumption in public buildings globally. This consumption share can be effectively reduced using the demand-response control. The traditional lighting system control method commonly depends on the real-time occupancy data collected using the passive infrared (PIR) sensor. However, the detection inaccuracy of the PIR sensor usually results in false-offs. To diminish the false-error frequency, the existing lighting system control simply deploys a delayed reaction period (e.g., 5 to 20 min), which is not sufficiently accurate for the demand-response operation. Therefore, in this research, a novel data-driven model predictive control (MPC) method that is based on the temporal sequential-based artificial neural network (TS-ANN) is proposed to overcome this challenge using an updated historical occupancy status. Using an office as case study, the proposed model is also compared with the traditional lighting system control method. In the proposed model, the occupancy data was trained to predict the occupancy pattern to improve the control. It was found that the occupancy prediction mainly correlates with the historical occupancy ratio and the time sequential feature. The simulation results indicated that the proposed method achieved higher accuracy (97.4%) and fewer false-offs (from 79.5 with traditional time delay method to 0.6 times per day) are achieved by the MPC model. The proposed TS-ANN-MPC method integrates the analysis of the occupant behavior routine into on-site control and has the potential to further enhance the control performance practice for maximum energy conservation.

照明系统占美国商业建筑总用电量的8％，全球占公共建筑总用电量的12％。使用需求响应控制可以有效地减少这种消耗份额。传统的照明系统控制方法通常取决于使用无源红外（PIR）传感器收集的实时占用数据。但是，PIR传感器的检测误差通常会导致误报。为了减少错误率，现有的照明系统控制仅部署了延迟的反应周期（例如5至20分钟），这对于需求响应操作而言不够准确。因此，在这项研究中 提出了一种基于基于时间序列的人工神经网络（TS-ANN）的新型数据驱动模型预测控制（MPC）方法，以使用更新的历史占用状态来克服这一挑战。以办公室为案例研究，将提出的模型与传统照明系统控制方法进行了比较。在提出的模型中，训练了占用数据以预测占用模式以改善控制。发现占用预测主要与历史占用率和时间序列特征相关。仿真结果表明，通过MPC模型，该方法具有较高的准确率（97.4％）和较少的误报（从传统时延方法的误报率从79.5降低到每天0.6次）。