High Precision Prediction Model of Wind Speed for Underground Mine Tunnel by Using ASTF Net

ABSTRACT An Adaptive Spectral‐Temporal Fusion Net (ASTF Net) is proposed in this paper to improve wind speed prediction accuracy in underground coal mines. The model first decomposes the wind speed time series data into three sub‐components: trend term, seasonal term, and residual term through STL decomposition. Subsequently, it employs Temporal Convolutional Network (TCN), Bidirectional Long Short‐Term Memory (Bi‐LSTM), and Convolutional Neural Network (CNN) models to predict each component separately, which greatly improves prediction accuracy. Furthermore, the residual component is analyzed via Compact Negative Selection Algorithm (CNSA) to detect and label anomalies, which increases the robustness of the proposed model. Simultaneously, spectral features extracted through Short‐Time Fourier Transform (STFT) are incorporated as auxiliary inputs to the dynamic ensemble module. In contrast to static weighting models, the proposed ASTF Net introduces a feature‐wise attention mechanism within a Multi‐Layer Perceptron (MLP), which adaptively adjusts the aggregation weight of each sub‐model. Experimental results demonstrate that the proposed model outperforms all traditional models, achieving the lowest Root Mean Square Error (RMSE), Mean Squared Error (MSE), Mean Absolute Error (MAE), Mean Bias Error (MBE), and the highest R ‐squared ( R 2 ). The proposed ASTF Net offers practical support for intelligent ventilation control.