Traffic space-time sequence multi-step prediction method and system and storage medium

Abstract

The invention discloses a traffic space-time sequence multi-step prediction method and system, and a storage medium, and the method comprises the steps: 1, calculating an improved historical mean value based on historical data X1:T; 2, carrying out the downsampling of the historical data X1:T and the improved historical mean value through a 3D-CNN module; 3, performing multi-step prediction from asequence to a sequence model to obtain an encoder and a decoder of a model, wherein the encoder and the decoder respectively consist of ConvLSTM, fusing an attention mechanism into the decoder, and further performing up-sampling by using transposed 3D convolution; 4, performing multi-model fusion to obtain a final prediction result. The prediction method provided by the invention is higher in precision.

Description

technical field [0001] The invention relates to a spatiotemporal sequence prediction method based on 3D convolutional neural network and multi-task learning, in particular to a multi-step traffic spatiotemporal sequence prediction method. Background technique [0002] The spatiotemporal sequence forecasting problems can include traffic forecasting, weather forecasting, people flow forecasting, etc. In essence, these problems are similar in that they all predict relevant spatiotemporal states within a certain time range in the future based on historical sequences. [0003] Time series methods, especially ARIMA-like models, were first applied to spatiotemporal series forecasting. Hamed et al. (M.M.Hamed, H.R.Al-Masaeid, and Z.M.B.Said, "Short-term prediction of trafficvolume in urban arterials," J.Transp.Eng., vol.121, no.3, pp.249–254, 1995 .) proposed to use the ARIMA model to predict the traffic volume on urban arterial roads. Starting with this, the researchers combined ...

AI Technical Summary

Problems solved by technology

Both IMS and DMS have different problems: on the one hand, the IMS model has no "global concept", and it only needs to be responsible for the accuracy of the next timestamp, and the prediction accuracy of the previous step will greatly affect the prediction accuracy of the next step. , which ultimately affects the overall prediction accuracy

On the other hand, the optimization goal of the DMS model involves multiple future time stamps, and the loss function needs to balance the accuracy of multiple time stamps, which is "out of reach" and difficult to train

Images