Network Traffic Prediction Based on Short- Time Fourier Transform and Koopa

Predicting network traffic accurately is essential for the network's stable operation. The challenge of predicting network traffic lies in sudden changes and short-term fluctuations. A hybrid prediction model based on short-time Fourier transform and Koopa is proposed in this paper. Initially, data is split into time-invariant and time-variant components using the short-time Fourier transform. The Koopa model based on Koopman theory, predicts the time-invariant and time-variant datasets separately. Finally, the prediction outcomes are combined to form the model's forecast results. The data set comprises network traffic from two districts. Considering the risks of predictions being lower than actual values during elastic scaling of computational resources, the risk-adjusted weighted mean absolute percentage error (RAWMAPE) is used as an evaluation metric. Experimental results show that STFT -Koopa reduces the mean squared error (MSE) to 0.024, mean absolute percentage error (MAP E) to 0.117, and RAWMAPE to 0.112, indicating higher prediction accuracy and lower risk of the proposed model.