Performing accurate building energy prediction (BEP) is one of the most important foundations for achieving energy resource allocation and developing energy efficiency measures. Buildings are diverse and operate under complex conditions leading to distribution differences of energy-related data among different buildings. Owing to such differences, the already reported data-driven BEP models exhibit poor cross-building prediction performance since they only use insufficient operation data of a single building. Although several deep transfer learning (DTL) strategies have been applied with improved cross-building prediction performance, there is still a lack of comparison of various DTL strategies, which would help to determine the optimal DTL strategy for different scenarios. Hence, three DTL strategies were compared: network-based Fine-tune, adversarial-based domain adversarial neural network (DANN), and mapping-based domain adaptive neural network (DaNN). The usefulness of the DTL strategy was validated using the open source dataset Building Data Genome Project 2 in the scenario of insufficient available data for real buildings. The influences of several factors were analysed, such as the available data volumes of the source and target domains within the training set. The applicability of different DTL strategies was discussed considering both accuracy and computational cost. Results show that the three DTL strategies outperform the traditional long short-term memory (LSTM) with an average BEP performance improvement ratio (PIR) of 0.75. For the extremely limited amount of available training data scenario, Fine-tune was recommended for the next-few-weeks prediction owing to its good balance between time cost and prediction performance. For the data shortage scenario of BEP tasks for nearly a year, DANN was recommended owing to its outperforming prediction accuracy. This provides insights for practical applications of DTL during the development of BEP models for cross-building BEP tasks without sufficient operation data.

执行准确的建筑能源预测 (BEP) 是实现能源资源配置和制定能源效率措施的最重要基础之一。建筑物种类繁多，在复杂的条件下运行，导致不同建筑物之间能源相关数据的分布差异。由于这些差异，已经报道的数据驱动的 BEP 模型表现出较差的跨建筑物预测性能，因为它们只使用了不充分的单个建筑物的运行数据。尽管已经应用了几种深度迁移学习（DTL）策略并提高了跨建筑预测性能，但仍然缺乏对各种 DTL 策略的比较，这将有助于确定不同场景的最佳 DTL 策略。因此，比较了三种 DTL 策略：基于网络的微调，基于对抗的域对抗神经网络（DANN）和基于映射的域自适应神经网络（DaNN）。在真实建筑物可用数据不足的情况下，使用开源数据集 Building Data Genome Project 2 验证了 DTL 策略的有用性。分析了几个因素的影响，例如训练集中源域和目标域的可用数据量。考虑到准确性和计算成本，讨论了不同 DTL 策略的适用性。结果表明，三种 DTL 策略的性能优于传统的长短期记忆（LSTM），平均 BEP 性能改进率（PIR）为 0.75。对于可用训练数据量极其有限的场景，由于时间成本和预测性能之间的良好平衡，建议对未来几周的预测进行微调。对于BEP任务近一年的数据短缺场景，推荐使用DANN，因为它的预测精度优于DANN。这为在没有足够操作数据的情况下跨构建 BEP 任务的 BEP 模型开发过程中 DTL 的实际应用提供了见解。