Federated spatiotemporal transformer for cross-facility carbon emission prediction in dynamic scheduling of low-carbon workshops

Amid the global shift toward low-carbon manufacturing, workshop-level dynamic scheduling has become a key approach for reducing carbon emissions due to its direct influence on production processes. However, most existing studies focus on single-workshop optimization and rarely explore emission reduction through cross-facility collaboration. Meanwhile, traditional carbon emission prediction methods struggle to build global carbon footprint models because of data silos, which limits the development of integrated emission reduction strategies. To overcome these challenges, this study proposes a cross-facility carbon emission prediction and dynamic scheduling framework based on a federated spatiotemporal transformer. Federated learning is employed to construct a distributed prediction architecture that enables global modeling while preserving data privacy, thereby mitigating prediction bias caused by isolated datasets. A spatiotemporal transformer module integrating multihead attention and graph neural networks is further developed to capture complex spatial and temporal dependencies of carbon flow across facilities. Based on prediction results, a closed-loop optimization mechanism is established by combining real-time forecasting with an improved whale optimization algorithm to generate low-carbon scheduling strategies. Experimental results across five representative industrial scenarios show that the proposed framework improves carbon emission forecasting accuracy and supports more effective low-carbon scheduling decisions. The contribution of the study lies in constructing a privacy-preserving closed-loop framework for cross-facility low-carbon manufacturing, in which federated learning, spatiotemporal carbon emission modeling, and prediction-driven scheduling are coupled within a unified optimization pipeline.