Lightweight time convolution network for quick prediction of time series data

Abstract

The invention discloses a lightweight time convolution network for quick prediction of time series data. An overall thought framework and a specific structure of the lightweight time convolution network are improved, expansion convolution is replaced by one-dimensional step convolution, and a full convolution structure is removed; as the calculated amount of the lightweight time convolution network provided by the invention is greatly reduced, and all calculation results in the overall thought framework of the lightweight time convolution network are used in forward and backward propagation ofthe network, the calculation waste is reduced; and meanwhile, compared with the prior art, the calculated amount required by the network is also greatly reduced; and meanwhile, the depth of the network can be effectively reduced by using a relatively large convolution kernel and a relatively large convolution step length.

Description

technical field [0001] The invention relates to the technical field of lightweight temporal convolutional networks, in particular to a lightweight temporal convolutional network for rapid prediction of time series data. Background technique [0002] In recent years, dealing with timing modeling problems is often applied to a temporal convolutional network. The main structure of the temporal convolutional network includes causal convolution, dilated convolution, and residual connection. Experiments show that in terms of time series modeling, its ability exceeds traditional RNN, LSTM, GRU and other structures; temporal convolutional network can retain long historical information of data, and can automatically extract low-dimensional feature processing from original high-dimensional features; traditional The temporal convolutional network takes the sequence x0,...,xT as input and aims to predict the corresponding output y0,...,yT at each moment. The temporal convolutional netw...

AI Technical Summary

Problems solved by technology

[0003] However, when the temporal convolutional network performs causal convolution through a full one-dimensional convolutional structure, when the length of the input sequence is too large, due to the structure of the full convolution, the calculation of the temporal convolutional network is too large, and many of the calculation results are not applied to The forward and backward propagation of the network, so the disadvantages are also very obvious, so it is urgent to involve a lightweight temporal convolutional network for time series data, and reconstruct the network structure according to the core idea of ​​​​the temporal convolutional network, aiming to reduce the amount of calculation and improve the quality of the network. running speed

Images