Static pressure prediction method for CO2 flooding oil reservoirs based on time series partitioning Transformer model

Reservoir’s static pressure is an essential basic data in the development and research of oil and gas fields. Its acquisition conditions are strict， and the sample number is extremely small. Static pressures are estimated with empirical methods based on dynamic pressure data during the production process； however， data errors are significant. To address the above issues， a static pressure prediction method for CO2 flooding oil reservoirs based on the time series partitioning Transformer model was proposed， utilizing deep learning theory. Model parameters were selected based on correlation analysis， and iterative interpolation was used to fill in samples to construct a static pressure prediction sample set. According to the principle of channel independence， the multivariate time series was divided into univariate time series， and a time series partitioning mechanism was introduced to divide the time series into subsequential blocks to capture local features. Based on the Transformer model architecture， a multi-head self-attention mechanism was utilized to extract features， and a self-supervised learning mechanism was employed to enhance the ability to capture complex dynamic characteristics， achieving the prediction of reservoirs’ static pressure. The research results indicate that the proposed model can accurately predict the static pressures at the middle of the oil reservoir of each well in the active well group， significantly improving prediction accuracy.