Attention based multi-component spatiotemporal cross-domain neural network model for wireless cellular network traffic prediction

Abstract Wireless cellular traffic prediction is a critical issue for researchers and practitioners in the 5G/B5G field. However, it very challenging since wireless usually shows high nonlinearities complex patterns. Most existing methods lack abilities of modeling dynamic spatial–temporal correlations data, thus cannot yield satisfactory results. In order to improve accuracy network prediction, an attention-based multi-component spatiotemporal cross-domain neural model (att-MCSTCNet) proposed, which uses Conv-LSTM or Conv-GRU neighbor daily cycle weekly data modeling, then assigns different weights three kinds feature through attention layer, improves their extraction ability, suppresses information that interferes with time. Finally, combined timestamp embedding, multiple fusion, jointly other models assist prediction. Experimental results show compared models, performance proposed better. Among them, RMSE att-MCSTCNet (Conv-LSTM) on Sms, Call, Internet datasets improved by 13.70 ~ 54.96%, 10.50 28.15%, 35.85 100.23%, respectively, models. The (Conv-GRU) about 14.56 55.82%, 12.24 29.89%, 38.79 103.17% higher than respectively.