Time sequence prediction method based on Radam-DA-NLSTM

Abstract

The invention relates to the technical field of time sequence prediction methods, in particular to a RAdam-DA-NLSTM-based time sequence prediction method, which adopts a new internal LSTM unit structure as a memory cell of an LSTM through a nested LSTM neural network, enables a model to be capable of instructing memory forgetting and memory selection, improves the prediction precision of the model, and improves the prediction accuracy of the model. Then constructing an auto-encoder network based on a double-stage attention mechanism, selecting input features and hidden state features of a time sequence by adopting an encoder based on an input attention mechanism and a decoder based on a time attention mechanism, improving the problem of attention dispersion defects of a traditional LSTM, and finally solving a target function by adopting an RAdam optimizer, according to the method, an Adam optimizer and an SGD optimizer can be dynamically selected according to the variance dispersity, and rectifier items are constructed, so that the adaptive momentum is fully expressed, the stability of the model is enhanced, and the method has higher prediction precision and stability.

Description

technical field [0001] The present invention relates to the technical field of time series prediction methods, in particular to a time series prediction method based on RAdam-DA-NLSTM. Background technique [0002] The world of the Internet of Things generates massive amounts of data, how to make good use of these data is a key technical issue in today's social and economic development. Therefore, scholars have been studying methods based on artificial intelligence to process and analyze massive data in the Internet of Things. Among them, the time series forecasting method is one of the research hotspots. [0003] LSTM (Long Short Term Memory, long short-term memory) neural network can fully mine the information contained in massive data, and has the ability of deep learning. Therefore, one of the main methods of the current time series method is to use the LSTM neural network. However, most time series in life present high-dimensional and complex features. When the predi...

AI Technical Summary

Problems solved by technology

However, most time series in life present high-dimensional and complex features. When the prediction model based on the traditional LSTM neural network processes these data, it will have poor robustness due to weak network feature aggregation, model instability and generalization caused by overfitting. poor performance

Images