Temperature Prediction Of Molten Salt Furnace Based On Recurrent Neural Network

Time series forecasting model is used to predict the future trend,by finding the timevariation of the historical data.The research on time series prediction model has a long history.There are many prediction algorithms,and they are frequently used in the field of transport,finance,market analysis and so on.But in industry,there are still some difficulties in predicting industrial process data.Because the industrial system is nonlinear,time-varying and large time delay and uncertain variables,variables are mutual coupling,the traditional prediction algorithm does not predict well.But the research and development of deep learning,provides a new research direction for time series prediction.Taking the heating process of molten salt furnace as the application background,this paper selects the most suitable algorithm through the study of recurrent neural network,and in consideration of the shortcomings of the algorithm and the characteristics of data,the algorithm is improved to predict the temperature of molten salt furnace,so as to optimize the process of industrial production.The main research content and research results of this paper are as follows:To understand the process principle and control strategy of the heating process of the molten salt furnace,the law and trend of temperature change and the main factors that affect the temperature change are known.A data acquisition system is established to get data related to the process,analyzing the characteristics of the data,determining the prediction algorithm according to the characteristics of the data,and provide a basis for the improvement of the algorithm and the setting of network parameters.Researched on the 3 kinds of recurrent neural network representative recurrent neural network in the development process,including the structure,calculation process,advantages and disadvantages of each algorithm,so as to select GRU with simple structure and good prediction effect.The selection of incentive function and optimizer in network configuration is also studied,and it lays a foundation for parameter selection and network configuration in the experimental part.According to the data of the large delay,nonlinear,phase change characteristics,combined with the self-attention model and dynamic learning to improve the GRU algorithm,makes the algorithm more easy to obtain the features of long distance correlation,updates and learns training set in real time so that the prediction results are more consistent with the current rules of temperature change.A modified prediction model,Self-attention Gated Recurrent Unit Based on Dynamic Learning is established to predict the temperature of the molten salt furnace.Compared with other prediction models,analyzing the prediction results,it shows that the algorithm not only improves accuracy and prediction precision,but also meets the trend of dynamic change of molten salt furnace temperature.