Time Series Forecast Of Car Cold-rolled Sheet Price On ARIMA-LSTM Model

Cold-rolled sheet is an important raw material in the stamping process of automobile manufacturing,and its price plays an extremely important role in the pricing of products in the automobile industry,measuring the market,and making production plans.The price of cold-rolled sheet is volatile and cannot be accurately priced due to complex and uncertain factors such as market changes and relevant policies.This article makes predictions from the perspective of time series.The time series is analyzed from the internal relationship of the data,without analyzing the many factors affecting the price,only the historical data can be fitted to achieve the purpose of prediction,and it is widely used.Firstly,the research status of time series is introduced,and then the identification of abnormal sequence data and the filling of vacancies are carried out.After this process,ARIMA model,neural network model and combination model are used to predict and evaluate the final prediction effect.(1)Data preprocessing.Judgment of outliers is performed before data interpolation to avoid the influence of outliers.Then,the outlier judgment is performed again on the new sequence.Outlier judgment includes kernel density map,nonparametric test,quartile box plot method,etc.Data filling methods include KNN,spline interpolation,forward and backward filling,etc.(2)A linear ARIMA model was established,including model stationarity judgment,white noise test,fitting effect judgment,model prediction,etc.First,make the data stable to meet the modeling premise,then fit the relevant model to the series and judge,and finally execute the prediction.(3)The nonlinear LSTM neural network model was built,including data set and the network structure.Considering the precocious local convergence of genetic algorithm,genetic algorithm and grid search method are used to determine the hyperparameters of neural network,and it is compared with the genetic algorithm to determine the hyperparameters of the network.The idea of weighted average is used to transform the network parameter learning amount,and it is used as a part of the fitness function to balance the network parameter learning amount and performance.Finally,individuals with high fitness and few network learning parameters are selected.(4)Establish two different combination models.The ARIMA is used to predict the linear part of the neural network to predict the nonlinearity of the residual,and then the two are added together as the combined model one;ARIMA and LSTM are combined with a certain weight as the model two,and the results of different prediction models are compared and evaluated.This dissertation studies from three aspects: traditional linear time series ARIMA model,nonlinear LSTM neural network and combined model.Stepwise prediction is used to improve prediction accuracy when building linear ARIMA modeling.The LSTM network hyperparameters are determined by genetic algorithm and grid search method,and the prediction results of network hyperparameters determined by genetic algorithm are compared,and better network hyperparameters are selected.The ARIMA prediction results and the neural network prediction results are used to establish a combination model according to the least square method,and compared with the traditional combination method.