Sales Forecasting Of Cigarette Distribution Center Based On BP Neural Network

With the development of market economy, precise sales forecasting which can help company make the decision becomes more and more important. But it is very difficult for the traditional forecasting methods to make a settlement for this problem. Artificial Neural Networks (ANN) is one of the most valuable methods for these problems at present.Cigarette distribution center is a nonlinear complex system. Through training parameters repeatedly, ANN can get the regulations among data and reflect optional complexity function. As simple and easy to be realized, Back-propagation Neural Network (BPNN) has currently been widely used in sales forecasting. However, there has been no clearly defined theory for calculating the ideal parameter settings. In this paper we wrote a forecasting programme for BPNN with C language which has good computation and performance. Then we discussed the details of the major impact factors of BPNN including the selection of input data, the processing of data, the numbers of hidden layer, learning velocity and momentum factor which influence the forecasting precision and convergence speed for cigarette sales forecasting.When year' s amount of cigarette was forecasted, we got better result in BPNN model by having the best number of hidden layer based on study of both convergence velocity and error. In BPNN model, the forecasting mean accuracy of training samples and test samples is 99.98% and 94.21%, respectively; Logarithm Regression Model has the highest mean accuracy 98.52% among all the traditional models. When month's sales amount was forecasted, we focused on analyzing impact factors of month' s sales amount including the inputs, the number of hidden layer, processing of dada, learning velocity and momentum factor. As the prediction model was decided according to all the above-mentioned factors, forecasting precision increased 2 percent on the whole and 18 percent in some especial months. The forecasting result had a great advantage to Exponent Regression Mode (90.33%) which is the best forecasting effect of traditional models and BPNN is suitable for the month's forecasting especially. When week's amount of cigarette was forecasted, we selected the proper samples and used festival-holiday factor adjust all data of the spring festival , the moon festival and the spring festival holidays, which helped to get the good forecasting mean accuracy as training samples and test samples is 98.09% and 95.32%, respectively. BPNN is better suitable for the week's sales forecasting of cigarette distribution center.On the whole, the optimization of network architecture parameters can obviously improve the network's capability and get better forecasting effect. BPNN is suitable for the sales forecasting of cigarette distribution center.