The Cross-Domain Enhanced Spectrum Prediction Framework Using Knowledge Distillation

Spectrum prediction plays an important role in spectrum management. In practical scenarios, the deployment of spectrum prediction models faces the issue of insufficient data for training, which will lead to decreased prediction accuracy. Regarding this challenge, this letter proposes an innovative method called the Cross-Domain Enhanced Spectrum Prediction Framework using Knowledge Distillation (CESP-KD), which combines LSTM, BiLSTM and CNN for prediction model training. Our proposed method primarily utilizes models trained on a station with sufficient data to guide the training of models at another station lacking sufficient data, to enhance the accuracy of spectrum prediction for the latter station. We performed experiments utilizing a real-world spectrum dataset, and assessed the method introduced in our letter compared to other spectrum prediction methods. The experimental results demonstrate that our method achieves better performance in cross-domain spectrum prediction tasks.