C.i He
Texas A&M University, Texas, United States
Keywords: Spectrum Prediction, Dynamic Prediction, Spectrum Sharing, Wireless Spectrum Allocation
Wireless networks are crucial to modern communication but face increasing challenges due to spectrum congestion, resulting in jitter, packet loss, latency, connection failures, and signal dropout, all of which degrade the user experience. Addressing spectrum congestion is difficult with current methods. Our research proposes a new deep neural network (DNN) approach for predicting spectrum utilization across various wireless environments. Unlike traditional methods that assess entire spectrum blocks, our model evaluates subspectrum spaces, allowing for efficient spectrum sharing. This enables the integration of diverse signals, like radar or satellite communications, into underutilized subchannels. Our approach uses historical waveforms processed by an encoder enhanced through contrastive learning, encompassing multiple wireless environments. A diffusion model with adjustable denoising balances prediction diversity and accuracy, outperforming static parameter models. Our method demonstrates adaptability and efficiency across different wireless scenarios, advancing spectrum prediction by providing a detailed, scalable solution for optimizing network management. Supporting prediction across 30 wireless environments and 32 future time frames, the model achieves an Average Displacement Error (ADE) and Final Displacement Error (FDE) of 0.047. Its performance is validated by an ROC curve with an AUC of 0.84, indicating robust and accurate spectrum prediction.