Research And Application Of Neural Network In Time Series Forecasting

Time series data refers to a set of data collected at a constant sampling frequency in chronological order.Normally,it refers to the observation values at equidistant time nodes.Time series data widely exists in finance,weather,transportation,construction,etc.The research on time series data is of great significance to social development.Time series data prediction tasks are mainly divided into short-term real-time prediction and long-term prediction: short-term real-time prediction is to predict the next moment data based on a number of historical data values;long-term prediction is to predict the data value in the next time period based on a number of historical data values.This paper studies the application of neural networks in time series data forecasting,including short-term real-time time series forecasting and long-term time series forecasting.The main research work includes:(1)For the deep echo state network,the size of each reserve pool and the sparsity of the weight matrix need to be determined empirically,and the size,sparsity,input scaling factor,spectral radius and other parameters of each reserve pool are all determined.There is a problem that the reserve pool is initialized randomly,and the internal network nodes are randomly connected to cause its performance to be unstable.This paper combines the deep architecture of the deep echo state network and the topological structure of the deterministic cyclic hopping state network,and proposes a deep deterministic cyclic hopping state network(deep cycle reservoir with regular jumps,DCRJ),and then proposed a modeling method to build a DCRJ model using the butterfly optimization algorithm(BOA);in order to speed up the training speed of the DCRJ model,this article improves the BOA algorithm and proposes a flight based on Levi And reverse learning butterfly optimization algorithm(OLBOA),and then test the convergence performance and optimization performance of the OLBOA algorithm on multiple benchmark functions.Experiment shows that the accuracy and convergence speed of the OLBOA algorithm are greatly improved compared with the BOA algorithm;finally use the OLBOA algorithm to build the DCRJ model for short-term real-time data prediction,and combine the DCRJ model with others The model is simulated on multiple data sets to verify the performance of the DCRJ model.Experiment shows that compared with the DESN model,the stability of the DCRJ model is greatly improved without loss of prediction accuracy.(2)Because most of the time series data does not have stable and linear characteristics,the use of a single traditional statistical forecasting and neural network forecasting method cannot achieve the desired effect,and in the process of long-term forecasting of the time series,this instability increases with The increasing influence of time is getting bigger and bigger.For this reason,this paper proposes the CEEMDAN-DCRJ model for long-term prediction of time series data.First,use the CEEMDAN algorithm to decompose the original time series,and then analyze the components obtained by the CEEMDAN decomposition through the permutation entropy and Pearson correlation coefficient,and then use the obtained components to construct the DCRJ model for long-term prediction of time series data.The model is compared with the temporal convolutional neural network on the real data set to verify the performance of the model.Experiment shows that after CEEMDAN decomposition,and then build different prediction models for different components,and finally perform modal superposition to obtain the prediction results,the prediction method can improve the prediction accuracy,and CEEMDAN-DCRJ has a good effect in long-term time series prediction,and the prediction accuracy is higher than CEEMDAN-TCN model.(3)The DCRJ model is used in the trend prediction module of the bridge intelligent management simulation system.First,the relevant basic knowledge of the bridge and the measurement method of bridge deflection and natural frequency are introduced,and then the deflection and vibration sensor data of a certain bridge are selected for simulation experiment to verify The availability of the DCRJ model in the prediction of bridge deflection and natural frequency data,and then proposed the architecture design plan of the bridge intelligent management simulation system,and finally described and demonstrated some of the functions of this system.