Study on VMD-LSTM-based mine dust concentration prediction

To address the problem of insufficient accuracy in traditional prediction models caused by the nonlinear, non-stationary, and strong noise characteristics of underground coal mine dust concentration data, a hybrid mine dust concentration prediction method integrating Variational Mode Decomposition (VMD) and a Long Short-Term Memory Network (LSTM) was proposed. The raw dust concentration time series data were fed into VMD. Under the set conditions for the number of modes K and the constraint factor α, VMD decomposed the raw data into K mode components with different frequency characteristics, with each component corresponding to amplitude information in different frequency bands. The component data were then fed into LSTM and trained using a selective forgetting/input gate algorithm to output the component prediction results. The component prediction results were superposed and reconstructed to produce the final prediction result. The dust concentration data from a working face in the Sandaogou coal mine were used to analyze the effects of the constraint factor α on the VMD decomposition performance and the number of modes K on the prediction performance. The analysis results showed that: when K=5, the samples were completely decomposed by VMD, and each mode component contained detailed frequency information, allowing for a clear and intuitive analysis of the overall signal's composition; when α=2 000, the profiles of each mode component were complete and fully separated, whereas an excessively small α led to more redundant information in the independent components, and as the value of α increased, the bandwidth of the mode components continuously decreased while the resolution improved. The experimental results showed that: with K=5 and α=2 000, the error between the VMD-LSTM's predicted results and the measured values was minimal, and its MAE, MSE, RMSE, and MRE were all superior to those of other models. The VMD-LSTM model exhibits strong generalization ability and robustness for predicting nonlinear, non-stationary, and high-noise dust concentrations under complex environmental conditions.