The spatiotemporal attention logging parameter prediction model based on historical feature separation of prior knowledge

In oil and gas exploration and development, logging is fundamental for reservoir identification and decision-making, relying on data about the physical properties of geological formations. Accurate analysis of logging reservoir parameters is critical, but existing data-driven prediction models often lack the capability to effectively extract reservoir parameter features, leading to weak generalization in practical scenarios.This paper addresses these limitations by proposing a novel model, SMHDS, specifically designed to predict high-information-density data, such as logging parameters. The model partitions the data into historical and current data, using the latter as prior knowledge and focusing on input parameters most relevant to the prediction target. It improves upon traditional attention mechanism by embedding spatial and temporal information into the attention mechanism, reallocating parameter weights for enhanced prediction accuracy.For the above method, this paper conducts three sets of experiments on actual data in an exploration area, with porosity as the target parameter. The performance of the model is tested from three perspectives: the rationality of the model structure, the accuracy of model prediction, and the generalization ability of the model. It is compared with three similar algorithms (CNN-LSTM, LSTM, MLP).The results indicate that the spatiotemporal model based on historical feature separation of prior knowledge, exhibits rational weight allocation for regression problems with logging parameters as the target. It effectively utilizes the relevant features of input parameters, resulting in smaller prediction errors compared to actual results, and demonstrates good data generalization capability.