| With so many TV brands today,if you want to gain a foothold in a fierce market environment,you can’t do without spending forecasts.At present,most enterprises rely on the decision-makers’ knowledge of products and previous experience in product sales forecasting.They have some blindness and randomness,and cannot reasonably arrange production and purchase of materials.This thesis intends to use recurrent neural network and attention mechanism to reasonably predict the TV category sales data of an enterprise.Recurrent neural network as a main method in time series data prediction,has been used in many data prediction fields.This thesis selects the sales data of a certain company’s TV category for preprocessing,removes outlier data and performs normalization processing,constructs three models of RNN,LSTM,and BLSTM in the recurrent neural network,and adjusts and designs the super parameters in the neural network,use the same batch of TV sales data for forecast comparison.The mean square error and the mean absolute percentage error are used as reference indicators to measure the prediction ability of each model,and it is concluded that the prediction result of BLSTM is better,the accuracy is higher,and the stability is stronger.The thesis also conducts a certain research on the attention mechanism.For the selfattention mechanism and the multi-head attention mechanism,an algorithm flowchart for this experiment was designed,and multiple experiments were carried out.The BLSTM model under the multi-head attention mechanism has a good accuracy in the prediction of TV sales.Provide an important reference basis for the future development of the enterprise.The experimental results prove that the attention mechanism in the field of natural language processing and image processing can also contribute to the field of time series data prediction,and the successful citation of the attention mechanism in time series data prediction provides new ideas for the study of attention mechanism. |
PDF Full Text Request