在煤矿智能化升级的大环境下，从海量的工作面监测数据中挖掘高质量的信息来构建科学的模型从而提高预测时长和精度是防范上隅角瓦斯浓度超限的关键。然而，上隅角瓦斯浓度影响因素众多，海量数据利用匮乏，瓦斯浓度预测精度高但时长较短，仅为0～30 min，而中长时30～60 min预测精度低、泛化能力差。为了解决这个问题，笔者以山西某矿回采工作面为研究对象，对该工作面的煤层瓦斯含量进行动态提取，组建煤层瓦斯含量、瓦斯浓度、采煤机、风速的特征集合，并对该特征集合进行预处理，通过相关性分析对不同特征进行筛选，进一步构造相关特征的短时趋势、长时趋势、周期趋势以及拼接特征，首先构建基于频率增强分解Transformer(FEDformer)的瓦斯浓度预测层，构建基于轻量梯度增强机(LGBM)的残差修正层，然后引入自适应阈值(AT)技术构建阈值感知层，最终组成3层的瓦斯超限预测模型架构，对未来60 min内上隅角瓦斯浓度进行预测，并通过召回率，误报率，平均绝对误差(MAE)以及平均绝对百分误差(MAPE)对预测性能进行考察。研究结果表明：所构建的基于FEDformer-LGBM-AT架构的上隅角瓦斯浓度预测模型的短时召回率为0.956，误报率为0.035，MAE为0.033，MAPE为0.183；长时预测的召回率为0.940，误报率为0.035，MAE为0.047，MAPE为0.262；与传统的灰色模型(GM)、支持向量机(SVM)、反向传播(BP)、门控循环单元(GRU)、粒子群优化的长短期记忆(PSO-LSTM)、Transformer等模型的长时预测能力相比，FEDformer-LGBM-AT架构模型具有更好的长时预测精度和泛化能力，自适应阈值感知使得模型对高值瓦斯浓度敏感。该架构模型弥补短期预测局限性和泛化性，支撑现场瓦斯超限防治措施，可为回采工作面瓦斯浓度智能预测提供一定的借鉴和参考。

In the context of the intelligent upgrading of coal mines, mining high-quality information from massive monitoring data of working faces to construct scientific models that enhance prediction duration and accuracy is crucial for preventing excessive gas concentration in the upper corner. However, there are many factors that affect the gas concentration in the upper corner, and there is a lack of utilization of massive data. The prediction accuracy of gas concentration is high but the duration is short, only 0～30 minutes, while the prediction accuracy and generalization ability are poor for medium to long duration 30～60 minutes. In order to solve this problem, this article takes a coal mining face in Shanxi Province as the research object. Firstly, the coal seam gas content of the face is dynamically extracted, and a feature set of coal seam gas content, gas concentration, coal mining machine, and wind speed is constructed. Then, the feature set is preprocessed, and different features are screened based on correlation analysis. Further construct short-term trends, stable trends, periodic trends, and concatenated features of relevant features. Firstly, a gas concentration prediction layer based on frequency enhanced decomposition transformer (FEDformer) is constructed, and a residual correction layer based on lightweight gradient boosting machine (LGBM) is constructed. Then, adaptive thresholding (AT) technology is introduced to construct a threshold perception layer. Finally, a three-layer gas concentration prediction model architecture is formed to predict the gas concentration in the upper corner within the next 60 minutes, and the prediction performance was investigated by recalling rate, false positive rate, MAE and MAPE. The research results indicate that the short-term recall rate of the upper corner gas concentration prediction model based on the FEDformer-LGBM-AT architecture is 0.956, the false alarm rate is 0.035, the MAE is 0.033, and the MAPE is 0.183; The recall rate of long-term prediction is 0.940, the false positive rate is 0.035, the MAE is 0.047, and the MAPE is 0.262; Compared with traditional models such as Grey Model (GM), Support Vector Machine (SVM), Backpropagation (BP), Gated Recurrent Unit (GRU), Particle Swarm Optimized Long Short Term Memory (PSO-LSTM), Transformer, etc., the FEDformer-LGBM-AT architecture model has better long-term prediction accuracy and generalization ability. The adaptive threshold perception makes the model sensitive to high-value gas concentrations. This architecture model compensates for the limitations and generalization of short-term prediction, supports on-site gas exceedance prevention measures, and can provide certain reference and guidance for intelligent prediction of gas concentration in mining face.