Multi-step Prediction of Traffic Load with All-attention Mechanism

Predictive resource allocation can harness residual resources in wireless networks to serve non-real-time service, where the prediction of residual resources plays a key role, which can be converted into the problem of predicting traffic load of real-time service. In this paper, a new neural network structure proposed for natural language processing, which is completely based on attention mechanism, is introduced to time sequence prediction, in particular traffic load prediction. By training and testing with a real traffic load dataset measured in each second, the multi-step prediction method with all-attention mechanism is compared with other methods using recurrent neural network (RNN) or linear and non-linear predictors, in terms of complexity (measured with training and testing time), prediction accuracy (measured with mean relative percentage error) and prediction error statistics (measured with the mean value and standard derivation of the prediction error). Simulation results show that the primary advantage of the designed neural network with all-attention mechanism over RNN without attention mechanism lies in low training and test complexity. Due to the accumulative errors in multi-step prediction, its gain in terms of prediction error is not obvious.