Federated transformer-based ROP prediction under geological heterogeneity and privacy constraints

• Proposed a Transformer-based ROP prediction model (FLSA) under a federated learning framework without sharing raw data. • Incorporated multi-head self-attention and feature-weighted aggregation to address non-IID data across heterogeneous reservoirs. • Achieved up to a 72% reduction in mean squared error for long-sequence forecasting compared to traditional local models. . This study presents a Transformer-based framework for intelligent rate of penetration (ROP) forecasting in oil and gas drilling, with a focus on addressing the challenges of cross-regional generalization and data privacy constraints. The proposed approach integrates a multi-head self-attention mechanism within a transformer architecture to effectively capture long-range dependencies and complex temporal patterns in heterogeneous drilling time-series data. Built upon a horizontal federated learning paradigm, the framework enables collaborative model training across multiple oilfields while ensuring that raw data remains locally protected. Experiments were conducted using real-world datasets from the Volve field, XJ oilfield, and BH oilfield, which represent diverse geological and operational conditions. A sample-weighted federated aggregation algorithm was employed to mitigate performance degradation caused by non-independent and identically distributed data. Results show that the proposed model significantly outperforms traditional local and centralized baselines, reducing mean squared error by up to 72% in long-sequence forecasting tasks. The improvements are particularly evident in data-scarce single-well scenarios, where the model demonstrates enhanced stability and generalization. By leveraging the representational power of Transformer architectures and the secure collaboration enabled by federated learning, this framework offers a scalable, privacy-preserving, and generalizable solution for drilling performance prediction across geologically heterogeneous environments.