Accurate building energy prediction methods have become a key factor in achieving energy-saving goals. Traditional methods for building energy multi-step prediction often use recursive or direct strategies to address time series prediction problems, which may neglect the data sequence correlation and result in the cumulative error. To solve the above problem, this paper proposed a Temporal Feature Decomposition Fusion Network (TFDFNet) model for building energy consumption multi-step prediction, with an encoder-decoder architecture. In the encoder, feature fusion layers is employed to consider the influence of different feature sequences on the predicted sequence. Through the decomposition of historical load data sequences, different hierarchical sequence structures are constructed to enhance the interpretability and predictability of the data. In the decoder, a simple and efficient network is constructed using MLP to decode the encoded information and obtain the prediction results. Experimental results show that, the proposed model achieves higher prediction accuracy and more stable convergence than the other five comparable methods, which also indicates the potential of achieving excellent building energy consumption multi-step prediction results with a simple model design.

中文翻译：

---

用于建筑能源多步预测的时间特征分解融合网络


准确的建筑能耗预测方法已成为实现节能目标的关键因素。传统的建筑能源多步预测方法往往采用递归或直接策略来解决时间序列预测问题，这可能会忽略数据序列的相关性并导致累积误差。为了解决上述问题，本文提出了一种时间特征分解融合网络（TFDFNet）模型，用于构建能耗多步预测，具有编码器-解码器架构。在编码器中，采用特征融合层来考虑不同特征序列对预测序列的影响。通过对历史负荷数据序列的分解，构建不同的层次序列结构，增强数据的可解释性和可预测性。在解码器中，利用MLP构建简单高效的网络来对编码信息进行解码并获得预测结果。实验结果表明，该模型比其他五种可比方法具有更高的预测精度和更稳定的收敛性，这也表明通过简单的模型设计即可获得优异的建筑能耗多步预测结果的潜力。