Research On Multivariable Time Series Short Term Forecasting Model Based On Neural Network

With the continuous development of intelligence and automation of data acquisition equipment,data collection is more convenient,the content of data collection is more diversified,the data dimension is increasing,and the frequency of data collection is also increasing.At present,most meteorological data are collected in minutes.The meteorological department invests a large amount of money in the renewal and transformation of meteorological monitoring equipment every year.Its purpose is not only to record data,but also to dig out some meteorological laws through collecting data,so as to better serve the economic and social needs.In recent years,with the development of deep neural network technology,researchers have done a lot of work in using time series to forecast,and made remarkable progress.However,the accuracy and timeliness of time series forecasting is still a hot spot in the research of forecasting models.Therefore,in this paper,two multivariable time series prediction models based on neural network and a dynamic prediction mechanism are proposed:(1)Bi LSTM-AM Short-time Series Prediction Model Based on Multivariable Input The model consists of two layers of multi-dimensional variable bi-directional long and short term memory network,an Attention allocator composed of an Attention Mechanism(AM),and a fully connected network composed of two single hidden layers.The advantage of the model is that the bi-directional learning ability of BI-LSTM is used to improve the prediction accuracy of the model,and then the attention allocator is used to selectively extract the key information at different time steps.(2)Hybrid Neural Network Sequence Prediction Model Based on CNN-BLSTM In order to better learn the correlation between different variables and target variables and solve the correlation problem between different variables,the model extracts the parallel features among the multi-dimensional variables in the input data by adding a Convolutional Neural Network(CNN)submodule on the basis of the BI-LSTM Network.And Bi-LSTM is used to extract and encode the data features along the time axis.(3)In order to solve the problem of large amount of data and constant dynamic change of data,and realize the timeliness and high precision of real-time online forecast,a Dynamic Recurrent Learning Prediction Mechanism(DRPM)with real-time learning and real-time Prediction is proposed,which is used in the above two models.The above two models and the DRPM mechanism were evaluated on the road temperature dataset of Jiangsu expressway network from 2016 to 2019.The main evaluation indexes were MAE,MRE and RMSE.The two models are compared with ARIMA,NARX RNN,CNNLSTM and DA-RNN in each evaluation index,the experimental results show that the two models proposed in this paper are superior to the models compared in each index,indicating that the prediction accuracy and stability of the model proposed in this paper are improved significantly.