Abstract: To address the challenge of predicting gas outbursts and spontaneous coal combustion in goaf, and to shift from real-time alerting to proactive prediction, this study proposes an edge-computing-oriented multi-variable temporal fusion prediction model for gas monitoring in goaf. This model tackles the key limitations of existing methods: their narrow predictive dimensions and overly complex architectures that hinder edge deployment. The model adopts a time-frequency collaborative architecture. Specifically, it decomposes the time series data into trend and seasonal components. In the time domain, an enhanced lightweight Transformer encoder captures the macroscopic trends of gas concentrations, while in the frequency domain, a compact network extracts periodic patterns, enabling multi-gas information fusion. Experiments using field monitoring data from Kongzhuang Coal Mine demonstrate that, under the multi-variable prediction mode, the model achieves significantly higher CO prediction accuracy than the single-variable baseline, validating the effectiveness of multi-gas fusion modeling. On a coal spontaneous combustion dataset, the proposed model improves prediction accuracy by 18.1% and 22.4% compared to iTransformer and Mamba, respectively, highlighting the substantial benefit of multivariate collaborative information for predictive performance. The model attains an average inference speed of 30 milliseconds, with GPU memory consumption reduced by approximately 45% relative to the baseline (iTransformer). This efficiency advantage arises from the near-linear complexity enabled by time-frequency collaborative computation and the effective mitigation of memory bottlenecks via dual-path computation. In summary, the proposed model achieves a favorable balance among prediction accuracy, computational efficiency, and edge deployment feasibility, offering a practical technical pathway for intelligent early warning of underground coal mine disasters.