Daily Arrivals Forecasting Of Call Center Based On Combined Model Of SARIMA And CNN-LSTM

With the continuous progress of economic society and science and technology,the competition among enterprises has already changed from product competition to service competition.As a window for enterprises to provide services to customers directly,call center plays an increasingly important role.How to accurately predict arrivals of call center has become the focus of administrators and academic circles.Based on the existing research results at home and abroad,this paper introduces a time series prediction model to consider the time inertia of short-term arrivals,and a deep neural network model to consider the time correlation of long-term arrivals.Based on the designed variable weight of time in forecasting cycle,a variable weight combination model combining the time series forecasting model and the deep neural network model is constructed,which can effectively improve the daily arrivals forecasting performance of call center in long time cycle.First of all,on the basis of clear call center and other related concepts,the domestic and foreign research on arrivals of call center prediction is reviewed,and the main prediction models of this paper are determined as time series prediction model and deep neural network model,which are holtWinters,SARIMA,stacked LSTM and CNN-LSTM models based on multiplication respectively.Secondly,in order to further improve the performance of the cycle for a long time,in view of a single model of different prediction effect and the advantage of combination model,is proposed and constructed based on the cycle time variable weight combination forecast model,the weight through solving the linear constrained quadratic programming model,to ensure the overall minimum error prediction period.Then,taking the data of a province of electric industry call center as an example,the paper makes descriptive statistics of daily arrivals from two aspects of time sequence and influencing factors,and determines the key influencing factors of daily arrivals,including time information,date type,weather conditions,temperature and precipitation.Finally,the time series prediction model,deep neural network model,equal weight combination model and variable weight combination model are used to predict the daily arrivals in the call center of electric industry,and the prediction period is 1 day,14 days and 21 days respectively.The prediction results show that:(1)SARIMA model shows better prediction accuracy than the other three single models in a short prediction period(1 day).On average,MAPE and RMSE are reduced by 17.98% and 13.50%,and standard deviation is reduced by 11.99% and 6.44%,respectively.In the long prediction cycle(14 and 21 days),CNN-LSTM model showed better prediction stability,with average MAPE and RMSE decreased by 24.80% and 25.31%,respectively,and standard deviation decreased by 57.64% and 40.88%,respectively.(2)The prediction performance of the variable weight combination model is significantly better than SARIMA,CNN-LSTM and equalweight combination model in both short and long time periods.When the prediction period is 1 day,MAPE and RMSE are 13.71% and 1049.83,respectively,which are 8.31% and 7.70% lower than those of the other three models,and their standard deviations are 8.91% and 6.66% lower than those of the other three models.When the prediction period was 14 days,MAPE and RMSE were 15.68%and 1577.73,decreasing by 15.99% and 12.95% on average,and standard deviation decreased by28.65% and 20.84% on average.When the prediction period was 21 days,MAPE and RMSE were17.10% and 1779.20,decreasing by 11.97% and 10.88% on average,and standard deviation decreased by 22.57% and 14.00% on average.In conclusion,the variable weight combination model proposed in this paper can significantly improve the accuracy and stability of daily arrivals prediction in the whole prediction cycle.