【摘要】 准确预测煤炭需求对于保障国家能源安全、稳定市场价格及制定宏观经济政策具有至关重要的作用。然而，影响煤炭需求的因素众多，其相关数据往往来源于不同部门，具有日度、旬度、月度等多样的采集频率，给传统预测模型带来了巨大挑战。为解决该问题，提出一种融合多频率特征的深度学习模型——多频时间序列Transformer(MFT-Former)，用于煤炭需求预测。该方法首先通过一套系统化的数据处理流程，将多源异构的原始数据清洗、对齐并重采样为3个时间同步的高、中、低频特征矩阵。随后，将此3个矩阵作为并行输入，送入一个特殊设计的多输入Transformer网络。该网络包含3个独立的编码器分支，分别捕捉各频率下的时间依赖模式，并通过一个融合层将提取到的深层特征进行整合，实现对未来煤炭需求的预测。利用包含多个经济与行业指标的真实数据集，以过去12个月的数据预测未来6个月的需求为任务，对模型预测表现进行评估。实验结果表明，MFT-Former模型能够有效融合不同时间尺度的信息，其在测试集上的平均绝对百分比误差达到6.24%，证明了该方法在处理复杂、多频时间序列预测问题上的有效性和准确性。更多还原

【Abstract】 Accurately predicting coal demand is crucial for ensuring national energy security, stabilizing market prices, and formulating macroeconomic policies. However, numerous factors influence coal demand, and the relevant data often originates from different departments with varied collection frequencies such as daily, ten-day, and monthly, posing significant challenges to traditional forecasting models. To address this issue, this paper proposes a deep learning model that integrates multi-frequency features—the Multi-Frequency Time-series Transformer(MFT-Former) —for coal demand prediction. The method first employs a systematic data processing pipeline to clean, align, and resample multi-source heterogeneous raw data into three time-synchronized high-, medium-, and low-frequency feature matrices. These three matrices are then used as parallel inputs and fed into a specially designed multi-input Transformer network. This network contains three independent encoder branches that capture temporal dependency patterns at their respective frequencies, and integrates the extracted deep features through a fusion layer to achieve future coal demand prediction. Using a real dataset containing multiple economic and industry indicators, this paper evaluates the model’s predictive performance with the task of forecasting the next six months’ demand based on data from the past twelve months. Experimental results show that the MFT-Former model can effectively integrate information from different time scales, achieving a Mean Absolute Percentage Error of 6. 24% on the test set,demonstrating the method’s effectiveness and accuracy in handling complex, multi-frequency time series forecasting problems.更多还原