Abstract: Abstract: Load sequences in coal mines simultaneously contain long-term trends, periodic patterns, and short-term disturbances, making unified representation difficult for existing forecasting models and resulting in unstable prediction performance during operating-condition transitions. To address this issue, a forecasting framework based on adaptive multi-scale exponential moving averages and multi-path signal feature fusion (MF-AMES) was developed. Discrete wavelet decomposition was first used to extract trend and disturbance components at different scales, while operating-condition labels derived from clustering, daily and weekly time encoding, and multiple exponential moving average sequences were incorporated to construct multi-path structural representations that jointly characterize trend, periodicity, and disturbance information. A two-layer long short-term memory network was then applied to hierarchically capture long-term dependence and local dynamics. An adaptive multi-scale EMA fusion module (AMES) was finally introduced at the output stage, where fusion weights of different smoothing scales were dynamically adjusted according to local fluctuation intensity, thereby improving trend stability and disturbance responsiveness. Experiments on real coal-mine load data from Yunnan Province showed that, compared with the long-horizon forecasting model Patch-TST, MF-AMES reduced the mean absolute error from 47.52 to 32.95, decreased the root mean square error from 65.54 to 45.84, and increased the coefficient of determination (R2) from 0.8860 to 0.9442. These results indicate that MF-AMES maintains both forecasting accuracy and stability under strongly non-stationary and multi-component conditions, providing reliable support for abnormal electricity-use identification and safety-oriented coal-mining operation management.

Key words: Abstract: coal-mine load forecasting, non-stationary time series, abnormal electricity-use identification, feature fusion, adaptive multi-scale EMA.

摘要： 摘 要: 针对煤矿负荷序列中长期趋势、周期性与短期扰动多成分并存导致现有模型难以统一表征、在工况切换阶段预测稳定性不足等问题，提出一种基于自适应多尺度EMA与多路径信号特征融合的煤矿负荷预测方法(MF-AMES)。首先，利用离散小波分解提取不同尺度的趋势与扰动分量，并结合聚类得到的运行模式标签、日内与周内时间编码以及多档指数移动平均平滑序列构建多路径结构化特征，以统一表征趋势、周期与短时扰动特征；其次，构建双层长短期记忆网络分层建模长期依赖于局部动态；最后在输出端引入自适应多尺度EMA融合模块(AMES)，根据局部波动强度动态调整各尺度平滑预测的融合权重，以提升趋势段的稳定性与扰动段的响应性。在云南省煤矿真实负荷数据上的实验结果表明，与经典时序预测方法 Patch-TST 相比，该方法的平均绝对误差从47.52下降至32.95，均方根误差从65.54下降至45.84，决定系数(R2)由0.8860提升至0.9442实验结果表明，该方法能够在多成分强非平稳工况下兼顾预测精度与稳定性，为煤矿用电异常识别与安全生产提供可靠支撑。

关键词: 关键词: 煤矿负荷预测, 非平稳时间序列, 异常用电识别, 特征融合, 自适应多尺度EMA