Traffic in high-speed networks shows distinct patterns at different timescales. This characteristic should be taken into account to address the error propagation in the multiple-step-ahead traffic prediction. Based on this idea, we proposed an algorithm in which traffic is modeled at different timescales using Gaussian process regression (GPR). The prediction at a timescale is made using the data of that timescale as well as the prediction results at larger timescales. Experiments performed on two public traffic data sets show that our algorithm has lower error propagation than other algorithms, including ARIMA, FARIMA, LSTM, and Convolutional LSTM.