A drilling overflow prediction algorithm based on a TCN-Transformer-LSTM fusion network

To increase the lead time of overflow warnings and achieve more accurate risk assessment, this article proposes a drilling overflow prediction algorithm based on a temporal convolutional network (TCN)-Transformer-long short-term memory (LSTM) fusion network (TTL network). The TTL network replaces the positional encoding layer with TCN to capture robust temporal dependencies and local features; utilizes the Transformer encoder based on the self-attention mechanism to model complex global dependencies; and employs LSTM instead of the Transformer’s decoder to learn long-term patterns and short-term fluctuations in time series data. To more precisely assess the risk of overflow, we divide the overflow risk into five levels. Experimental results demonstrate that, using the time of issuing a Level A warning as the reference point, the TTL network advances overflow warning lead times by 60 s (vs. TCN-Transformer), 219 s (vs. LSTM), 227 s (vs. GRU), 94 s (vs. Transformer-LSTM), and 129 s (vs. Attention-LSTM), achieving a 10-min advance in early warning. The TTL network performs well in terms of evaluation metrics, achieving an accuracy of 0.840 and a recall of 1.000. This study proposes an innovative method for predicting drilling overflow, which is significant for ensuring the safety of drilling operations.