Research And Application In Time Series Forecasting Model Of Mobile Service

With the rapid development of mobile communication technology and the popularization of intelligent equipment,the mobile service volume has shown explosive growth.The analysis shows that China's mobile service distribution is centralized in time and uneven in space.In order to meet the demand of large-scale users iin hot spots who can use mobile service fluently in a concentrated period of time,it is necessary to accurately model and predict the mobile business volume.In the dimension of space and time,mobile business prediction can be modeled as time series.Due to the mobile network environment has the characteristics of diversity,complexity and time-varying,the mobile business data has large fluctuations and noise.In order to construct a time series model with high precision,excellent real-time prediction and robust performance,a comprehensive use of fuzzy time series and recurrent neural network is considered.Among them,the fuzzy time series model has better prediction performance for the time series with strong randomness and noise,and is suitable for abstract expression of complex mobile business data.With strong learning ability and high real-time performance,recurrent neural network is suitable for fast analysis and prediction of massive business data.In this paper,firstly,we proposed an improved LSTM model with robustness and fast training based on standard LSTM model and the gated recurrent unit structure.This model reduces the number of parameters involved in the gate operation,only two control gates are used,and the operation of the control gate is simplified as a sum of long-term memory and current input.The model performance was verified by manually generated periodic,non-periodic sequences and MNIST handwritten recognition data sets.Compared with LSTM model.RFLSTM has obvious advantages in prediction accuracy and training speed.Secondly,we proposed dynanmic interval division algorithm,whiclh takes data distribution of sample data into account,and adopts FCM algorithm to obtain non-uniform interval by self-adaptation.The determination of interval boundary is related to the degree of data concentration,whiclh improves the accuracy and interpretability of the interval.By applying dynamic interval division algorithm to RFLSTM model,the performance of the model was verified.Compared with RFLSTM model of multi-scale ratio,the proposed model gets higher accuracy.Finally,the above two algorithms proposed in this paper are applied to the prediction of mobile traffic.The measurement reports reported by 10 cell's users in the current network are selected as data sources and were selected to divide the virtual grid.And the time series are constructed according to the granularity of a certain time.According to the experimental results,the proposed model gets better performance in mobile business forecasting.