Research on Low-Correlation Sequence Data Prediction Based on Flip Network
In some practical applications,there are often no other dimensional variables that highly correlate with the predicted time variable,or these dimensional variables are difficult to collect.At the same time,time series data with lower correlations are prevalent,which have more important significance for the improvement of forecasting results.Therefore,this study proposes a low-correlation multi-dimensional time series data prediction model based on attention flipping network.First,for low-correlation time series data,the correlation changes with time and a batch sliding window is introduced to remove the interference caused by time changes to better capture the dimensional correlation.Second,in view of the problem that the traditional Gated Recurrent Unit(GRU)network discards a large number of low-correlation samples,a flipped GRU network to screen low-correlation multi-dimensional data for the first time is established,the number of multi-dimensional data transmitted in the network is controlled,dimensional variables being discarded due to low-correlation is avoided,and survival time of the multi-dimensional data with low-correlation in the model is improved.Simultaneously,a dimension-based attention mechanism is used to adaptively adjust the importance of different dimension sequences in the correlation extraction process.Finally,a square Long Short-Term Memory(LSTM)network is established to fit the weighted data and determine the influence of the correlation on the predicted parameters in more detail.The experimental results show that the determination coefficient of the proposed model can reach 0.95 and its predictive performance is superior to traditional neural network models such as GRU and LSTM.
time series datadeep learningcorrelationattention mechanismLong Short-Term Memory(LSTM)networkGated Recurrent Unit(GRU)
国家科技期刊平台
NSTL
万方数据
维普
