Early Overflow Prediction Method Based on Genetic Algorithm-Transformer-Gated Recurrent Unit Algorithm

Summary To increase the early warning time for overflow, improve prediction accuracy, and reduce missed alarms and false alarms, we proposes a method based on the genetic algorithm (GA)-transformer-gated recurrent unit (GRU) framework (GA-Transformer-GRU). This framework constructs a new decoder module by replacing the original transformer decoder structure with a GRU network and a fully connected layer. We introduce the focal loss function to replace the traditional binary cross-entropy loss function, effectively addressing the imbalance of positive and negative samples in time-series prediction. At the same time, a GA is used to optimize the hyperparameters of the Transformer-GRU model. The experimental results of actual drilling overflow prediction show that the GA-Transformer-GRU method can issue overflow warnings 8 minutes in advance. Compared with other models such as GRU, bidirectional GRU (Bi-GRU), bidirectional long short-term memory (Bi-LSTM), and bat algorithm (BA) optimized random forest (RF) (BA-RF), recall rates increase by 19.65%, 20.18%, 19.93%, and 7.78%, respectively. F1 scores improve by 12.17%, 12.20%, 12.39%, and 6.16%, respectively. Overall, the GA-Transformer-GRU improves prediction accuracy and overflow warning time, helping to enhance the safety of the drilling system and advance the automation of the drilling process.