为了提高矿井粉尘浓度预测精度，针对煤矿粉尘浓度数据的时序特征，提出了一种基于多元时间序列分析的煤矿粉尘浓度预测方法。采用变分模态分解(VMD)将粉尘浓度时序信号分解为趋势、周期和随机波动3个维度；分别利用灰色模型(GM(1, 1))、霍尔特-温特斯(Holt-Winters)三次指数平滑法及自回归移动平均(ARMA(p, q))模型对各维度进行预测，并将预测结果进行融合生成最终预测值。利用现有矿井监测数据对提出的粉尘浓度预测方法进行了验证。实验结果表明，基于多元时间序列的煤矿粉尘浓度预测方法的平均绝对误差(MAE)为0.009 4，均方误差(MSE)为0.000 1，均方根误差(RMSE)为0.010 4，最大相对误差为0.48%。将基于多元时间序列的煤矿粉尘浓度预测方法与经典单一或复合方法进行比较，其在MSE、RMSE及最大相对误差等关键指标方面均优于经典方法，验证了该方法的有效性。

To enhance the accuracy of dust concentration prediction in coal mines, a prediction method of coal mine dust concentration based on multivariate time series analysis was proposed according to the time series characteristics of coal mine dust concentration data. The method began by employing variational mode decomposition (VMD) to decompose the dust concentration time series signal into three components: trend, periodic, and random fluctuations. Subsequently, each component was predicted using different models: the grey prediction model (GM(1, 1)), the Holt-Winters triple exponential smoothing method, and the autoregressive moving average (ARMA(p, q)) model. The prediction results from these models were then combined to produce the final forecast value. The proposed method of dust concentration prediction was verified by using the existing mine monitoring data. Experimental results indicate that the method based on multivariate time series achieves an average absolute error (MAE) of 0.009 4, a mean square error (MSE) of 0.000 1, a root mean square error (RMSE) of 0.010 4, and a maximum relative error of 0.48%. Compared with the classical single and fused method, the dust concentration prediction method based on multivariate time series is superior to the classical method in MSE, RMSE and maximum relative error, which verifies the effectiveness of the method.