Time series prediction system and method based on PT-LSTM

Abstract

The invention provides a time series prediction system and method based on PT-LSTM and the method specifically comprises the steps: obtaining the representation of time sequence data based on positioncoding, so as to not only enable time information to be added to input data explicitly, but also weaken the interference of a sliding window in an original data partitioning process; the invention provides a method capable of utilizing time sequence characteristics, namely a position coding and time gate expansion LSTM (Position Encoding and Time Gateway LSTM), and the method is abbreviated as PT-LSTM. The method is innovative in two aspects, one aspect is input data representation, the time series data representation provided by the invention is used as input, the other aspect is model structure optimization, a new network structure, namely T-LSTM, is designed, a specially designed time gate is added in the T-LSTM, the time gate is mainly controlled by time attributes, and the model structure is optimized. Partial control effects are achieved in the input data acquisition stage, the internal state updating stage and the result output stage of the model; a prediction framework based on PTLSTM is designed.

Description

technical field [0001] The present invention relates to the fields of time series forecasting and deep learning, and relates to a PT-LSTM-based time series forecasting system and method, which are used to solve the problem that the recurrent neural network cannot utilize time series characteristics. Background technique [0002] Time series prediction has important applications in many fields. In recent years, with the popularity of deep learning, many people have begun to introduce recurrent neural network (RNN) into time series prediction problems, and have made great achievements in their respective fields. good result. Compared with the differential integrated moving average autoregressive model (ARIMA), RNN has a strong nonlinear fitting ability, can capture the correlation between sequences well, and is suitable for various types of time series. Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) are two common RNN structures. According to the different proc...

AI Technical Summary

Problems solved by technology

In addition, the length of the actual time series data is often too long to be directly input into the RNN, and the time series needs to be divided into multiple subsequences through a sliding window

However, different times have different impacts on the sequence values, which makes the sequence values ​​in different windows vary greatly, which may cause the actual content learned by the model to be inconsistent with the expected content

Images