摘要： 准确的生成多变量时间序列数据为解决数据规模不足问题提供了有效途径，对电力负荷预测和风光发电评估等下游任务至关重要。然而，现有方法难以同时捕捉长短期依赖与变量间关联，且缺乏可解释性，无法满足能源系统分析需求。为此，本文提出了一种基于可解释扩散模型的多变量时间序列生成方法（Interpretable Diffusion Model for Multivariate Time Series，IDMTS）。首先，在扩散模型的去噪网络中引入了包含三重注意力（Triplet Attention，TA）的Transformer架构，以捕捉长短期依赖与变量特征关联；接着，结合多尺度趋势季节分解，利用双向LSTM（Bidirectional Long Short-Term Memory，BiLSTM）和傅里叶注意力（Fourier Attention，FA）分别建模趋势项和季节项，提升生成数据的准确性和可解释性；同时，通过多尺度自适应最大均值差异（Adaptive Maximum Mean Discrepancy，Ada-MMD）损失函数优化生成质量。实验结果表明，IDMTS在4个公开数据集上的生成准确性显著优于基线方法，其中Context-FID得分、相关性得分、判别得分和预测得分分别降低了51.5%-84.5%、4.1%-26.8%、24%-68.8%、0.3%-40%。同时，在可解释性实验以及条件插补和预测实验中，IDMTS展现出良好的可解释性和泛化能力。

Abstract: Accurate generation of multivariate time series data provides an effective way to solve the problem of insufficient data scale, and is crucial for downstream tasks such as power load forecasting and wind and solar power generation evaluation. However, the existing methods are difficult to capture the long and short term dependence and the correlation between variables, and lack of interpretability, which can not meet the needs of energy system analysis. Therefore, an Interpretable Diffusion Model for Multivariate Time Series (IDMTS) is proposed. Firstly, Transformer architecture containing Triplet Attention (TA) is introduced into the denoising network of diffusion model to capture long and short term dependencies and variable feature associations. Then, combined with the multiscale trend seasonal decomposition, the trend term and the season term are modeled respectively by using Bidirectional Long Short-Term Memory (BiLSTM) and Fourier Attention (FA). Improve the accuracy and interpretability of generated data; At the same time, the generation quality is optimized by multiscale Adaptive Maximum Mean Discrepancy (Ada-MMD) loss function. The experimental results show that the generation accuracy of IDMTS on the four public data sets is significantly better than that of the baseline method, in which the Context-FID score, correlation score, discriminant score and prediction score are reduced by 51.5% to 84.5%, 4.1% to 26.8%, 24% to 68.8%, and 0.3% to 40%, respectively. At the same time, IDMTS shows good interpretability and generalization ability in interpretability experiment, conditional interpolation experiment and prediction experiment.