Adaptive meta-aggregation stochastic configuration network for adjacent well distance prediction

Abstract Precise adjacent distance prediction between relief wells and accident wells is critical for drilling safety in the oil and gas industry. The current ranging methods that rely on mathematical models face two key problems. As drilling depth grows, magnetic field data becomes non-stationary, making traditional models unsuitable for complex drilling conditions and resulting in unstable ranging accuracy. Moreover, the significant variability of magnetic field information among wells challenges the generalization ability of existing neural network models. To this end, we propose a novel Adaptive Meta-Aggregation Stochastic Configuration Network (AMA-SCN) for accurate adjacent well distance prediction. First, for each distinct well type in the training set, we train separate meta-SCN models based on their specific data characteristics. Then, through a greedy aggregation strategy, we fully utilize the diverse well-related information in the dataset to construct an initial model with enhanced generalization capabilities. Subsequently, we develop an elastic error threshold mechanism to guide the online updates of the SCN structure. Leveraging error gradients across consecutive data batches and constrained by adjacent well distance factors, this mechanism enables adaptive model updates during drilling, enhancing prediction accuracy for new data. Experimental results demonstrate that AMA-SCN outperforms previous approaches in terms of both accuracy and generalization ability.