A Novel Interval Prediction Method based on Long Short-term Memory Networks with Adaptive Dropout

Aiming at the problem of interval prediction for key variables in the process industry, a Long Short-term Memory (LSTM) network based on adaptive Dropout is proposed. Firstly, in order to reduce the complexity of time series data and the mutual influence between time series data of different scales, Empirical Mode Decomposition (EMD) is used to decompose time series data into several Intrinsic Mode Functions (IMF) Weight and trend weight. Secondly, an adaptive Dropout method is proposed in the loop unit of the LSTM model. Some neurons are probabilistically stopped (equivalent to discarding neurons or corresponding weights), and the probability of stopping neurons is adaptively calculated according to the distribution characteristics of the data. Thereby, the LSTM model is improved that the model overfitting is reduced and the dependence between neurons is inhibited, so as to enhance the efficiency of information transmission and the model prediction accuracy. Thirdly, Bootstrap method is introduced to construct the prediction interval, and the comprehensive function of interval evaluation is used to estimate the prediction interval result. Finally, a simulation experiment is made on the Purified Terephthalic Acid (PTA) solvent system. The comparison results show that the proposed interval prediction model can effectively analyze the trend of key variables and has higher prediction accuracy and interval estimation ability.