Research On E-commerce Demand Forecasting Based On LSTM Neural Network

With the rapid development of the e-commerce industry in recent years,the operation of e-commerce companies has continued to be refined.One factor that is important to operational considerations is the future sales of goods.For e-commerce companies,there are many types of goods,and the use of traditional forecasting methods may not necessarily achieve the precision of fine-grained operations.Providing accurate demand data forecasts for multiple SKUs has become an important requirement for e-commerce refinement operations.This article takes the historical sales data of e-commerce industry as the research object and explores the possibility and method of using LSTM neural network to predict the future sales.In order to achieve this goal,this paper constructs LSTM neural network,uses S company's historical sales data as training data,predicts future sales and evaluates the accuracy,and proposes a new training method to optimize the accuracy of neural network.The main research content is as follows:(1)For the prediction of whether LSTM neural network can be used for the future sales of e-commerce companies,this paper constructs LSTM neural network,designs the data input form,selects the activation function,discusses ways to avoid over-fitting,and finally determines the use of Random gradient descent as a training method for supervised learning.(2)The LSTM neural network was trained and tested using S company's 12 SKUs historical data.Seasonal and conventional commodities were included to test the effect of LSTM neural networks on different types SKUs.The parameters of the neural network were manually adjusted during the experiment,including the number of LSTM layers,the number of hidden neurons in the LSTM unit,and the number of training epochs.The influence of different parameter settings on the neural network was explored.At the same time,the experiment has included the use of the BP neural network and compared the performance with LSTM neural networks in this scenario.(3)Due to the unreliability of manual adjustment of parameters,this paper proposes a training optimization method of LSTM neural network,which makes it adaptive and can dynamically adjust the number of hidden neurons in LSTM units.The data mentioned was retrained to verify its improvement in neural network accuracy.This paper constructs LSTM neural network to forecast the future sales of S company's products.The neural network has achieved good results in the test set.The average prediction error is below 10%,which is significantly lower than that of the BP neural network.The adaptive training method for the LSTM neural network is also effective compared to manual adjustment of parameters,and most SKUs have achieved lower errors.Experiments show that LSTM neural network is more suitable for sales forecast of e-commerce company.