An integrated sequence-to-sequence framework for well production prediction

Machine learning models, especially sequence prediction algorithms, have been widely used to predict oil production. However, most of the existing algorithms have low prediction accuracy and poor stability. Here, a new production prediction model is proposed. The model utilizes a convolutional neural network (CNN) for feature extraction of static data, a Sequence to Sequence (Seq2Seq) to capture the sequence relationship of dynamic data, and an attention mechanism (AT) to enhance the ability of Seq2Seq to learn the location and spatial relationship of sequences. Kolmogorov–Arnold networks (KANs) are then introduced to integrate static and dynamic data, replacing conventional “weights” with learnable one-dimensional functions to strengthen the nonlinear fitting relationship with the target variable. Furthermore, the whale optimization algorithm (WOA) is applied to optimize hyperparameters and the network structure, ensuring the model remains in an optimal state. Finally, the WOA-CNN-Seq2Seq-AT-KAN model is established. The research results show that the WOA-CNN-Seq2Seq-AT-KAN model has high prediction accuracy and good stability. It can alleviate the problem of error accumulation in the long-term future step size, and its prediction accuracy can be improved by 31.2% compared with CNN-long short-term memory. Meanwhile, WOA-CNN-Seq2Seq-AT-KAN has good convergence, effectively captures the relationship of long-distance data, and has strong nonlinear fitting ability.