To address the challenges of complex load fluctuations and lagging resource adaptation within the digitized production processes of the textile industry, this paper proposes an AI-driven intelligent decision-making scheme for load prediction and elastic scaling in multi-cloud heterogeneous environments. This approach ensures that the high-frequency data generated by automated weaving and dyeing systems are processed with optimal efficiency, allowing computational resources to dynamically align with the volatile demands of modern textile manufacturing. An improved extended Long Short-Term Memory (xLSTM) load prediction model is proposed, which enhances the adaptability to heterogeneous data and the ability to capture burst loads through multi-source feature fusion, adaptive normalization, and lightweight optimization. A decision-making mechanism based on improved Proximal Policy Optimization (PPO) is designed, integrating a multi-objective optimization function, fuzzy comprehensive evaluation for adaptability assessment, and a dynamic balance strategy. A hybrid simulation platform combining OpenStack and Amazon Web Services (AWS) is built, and comparative experiments are conducted under three types of load scenarios. The results show that the proposed prediction model achieves a Mean Absolute Percentage Error (MAPE) below 2.5% in normal, sudden, and fluctuating load scenarios. Compared with traditional methods, the decision-making mechanism achieves a cost optimization rate of 18.3%, improves the Service Level Agreement (SLA) compliance rate to 99.2%, and reduces the decision delay to 280ms. This research establishes a closed loop prediction-decision-scheduling framework, providing technical support for the efficient management of multi-cloud heterogeneous resources in digitized textile manufacturing.

AI-driven, multi-cloud heterogeneous, load prediction, elastic scaling, textile manufacturing