建筑照明插座能耗异常检测能够有效提高建筑能源效率，对实施建筑能源优化措施、实现建筑节能管控的研究具有重要意义。由于建筑照明插座系统能耗很大程度上受到建筑内部人员随机行为的影响，针对照明插座时间序列数据存在噪声较多和特征难以提取的问题，该文提出了一种结合工况划分与深度学习的无监督异常检测方法，旨在提升能耗异常识别的精度与鲁棒性。首先基于决策树方法对能耗数据按工作日与非工作日、上班与非上班时段等属性划分工况，然后针对不同工况分别构建基于长短期记忆神经网络-自编码器（LSTM-AE）的异常检测模型。该模型通过对正常数据的重构学习，计算重构误差，并设定差异化阈值，实现无标签数据下的能耗异常检测。以夏热冬暖地区某办公建筑578 d的照明插座逐时能耗数据为研究对象，开展数据建模与超参数优化实验。结果显示：迭代次数、神经元数和激活函数均对模型性能有显著影响；工作日工况下的能耗数据稳定性优于非工作日，异常检测精度相对更高；所提方法在4类工况下的平均精确率、召回率、F₁分别为91.23%、90.87%、90.80%，能够有效实现建筑照明插座系统能耗异常检测。

Anomaly detection of energy consumption in building lighting and socket systems can effectively improve energy efficiency. It holds significant importance for the implementation of building energy optimization measures and the realization of energy-saving management and control. Since the energy consumption of building lighting and plug load systems is heavily influenced by the random behavior of building occupants, and given the challenges posed by noisy time-series data and difficulty in feature extraction, this study proposed an unsupervised anomaly detection method that integrates operating condition classification with deep learning, aiming to enhance the accuracy and robustness of energy consumption anomaly identification. First, the decision tree algorithm was employed to classify the energy data based on attributes such as working days vs. non-working days and working hours vs. non-working hours. Then, for each identified condition, a long short-term memory autoencoder (LSTM-AE) model was constructed to detect anomalies. This model learns to reconstruct normal data and calculates the reconstruction error. By setting differentiated thresholds, it enables energy consumption anomaly detection under unlabeled data conditions. Using 578 days of hourly lighting and socket energy consumption data from an office building located in a hot-summer and warm-winter region, the study conducted model training and hyperparameter optimization experiments. Results indicate that the number of iterations, the number of neurons, and the activation function have significant effects on the model’s performance. Energy data during working days demonstrate greater stability than those on non-working days, resulting in higher detection accuracy. The proposed method achieves average precision, recall, and F₁ of 91.23%, 90.87%, and 90.80%, respectively, across four typical operating conditions, demonstrating its effectiveness in detecting energy anomalies in building lighting and socket systems.