A scalable deep learning framework for job completion time prediction in cloud-edge environments

Abstract

Accurate makespan prediction is essential for optimizing scheduling and operational decisions in collaborative manufacturing environments, particularly under conditions of uncertainty and dynamic system behavior. However, conventional prediction models often fall short in capturing complex non-linear spatial–temporal dependencies, adapting to evolving shop-floor states, and handling noisy inputs or service disruptions commonly encountered in practice. To overcome these limitations, this study presents a robust hybrid deep learning framework that combines an attention-based configurable convolutional neural network with an attention-enhanced bidirectional long short-term memory network. The proposed architecture is designed to selectively emphasize salient machine–task interactions and model temporal dependencies in both forward and backward directions, thereby enhancing resilience to incomplete data and transient service failures. To evaluate its effectiveness, extensive experiments were conducted under diverse scheduling scenarios, including metaheuristic, reinforcement learning, and agent-based strategies. Results demonstrate that the proposed model consistently outperforms a range of traditional deep learning baselines and advanced neural architectures, achieving up to a 24.2% reduction in mean squared error and a 24.7% improvement in mean absolute percentage error. These findings underscore the model’s robustness and practical value for real-time, reliable makespan prediction in complex manufacturing settings.