BERT and transfer learning-based CO2 storage prediction method

Abstract

The invention belongs to the technical field of carbon sequestration, and particularly relates to a BERT and transfer learning-based CO2 sequestration prediction method, which comprises the following steps of: constructing a multi-source data knowledge base, and preprocessing source domain and target domain data of industrial solid waste sequestration CO2 into BERT feature sequences; pre-training the BERT model of the source domain, and optimizing the weight by adopting a self-adaptive mask to obtain the universal feature extraction capability; and migrating the pre-training BERT to a target domain, freezing bottom layer parameters, finely tuning a top layer, and adding regularization constraint differences to obtain a CO2 sequestration prediction model. A model library is constructed for different scenes, parameters are input to predict the storage capacity, and optimal parameters are output through a backward reasoning and optimization algorithm. Newly-added data are collected for verification, fine adjustment is conducted again when errors exceed a threshold value, and dynamic optimization is achieved. The problem of insufficient model training caused by scarcity of industrial solid waste storage field data is effectively solved, and the prediction precision is remarkably improved compared with that of a traditional machine learning model.

Images