Time series data pre-training model fine-tuning method and device, time series data prediction model

Abstract

The application discloses a time series data pre-training model fine-tuning method and device and a time series data prediction model, and comprises the following steps: obtaining a pre-training model and input time series data, the pre-training model comprising an encoder and a decoder, and the encoder being used for extracting input time series features of the input time series data; performing linear calculation on the input time series features by using a linear layer to generate corresponding dynamic prompt features; determining enhanced time series features in combination with mask features, dynamic prompt features and input time series features; inputting the enhanced time series features into the decoder for decoding, and performing prediction of time series data of a future moment based on the future moment to be predicted. Through implementation of the application, for each input time series feature, the dynamic prompt features are generated by considering the implicit context knowledge thereof, the dynamic prompt features are used as instance-level prompt information for model parameter fine-tuning of a downstream task, the overfitting problem in the fine-tuning process of the time series pre-training model is effectively avoided, and the prediction accuracy of the downstream time series task is improved.

Images