Abstract:

In the production process of polyester fiber, accurate control of esterification temperature is crucial. However, due to the differences in production equipment and process parameters, traditional prediction methods are difficult to meet personalized needs, and there are privacy leakage and communication pressure problems in the data sharing process. The federated personalized adaptive time series forecast algorithm on esterification temperature is proposed, which protects privacy and reduces data communication pressure. Introducing a joint forecasting mechanism, the algorithm assigns private and shared models to each client, integrates model forecasting with learnable weights to enhance forecasting accuracy and generalization capability. An adaptive stage adjusts model parameters intelligently based on client data distribution and learns unique joint forecasting weights for each client to obtain personalized forecasting outputs. In addition, the adoption of Bayesian optimization algorithms over traditional methods like random search and grid search addresses issues related to high dimensionality, complexity, and resource constraints. This approach enables rapid identification of the optimal hyperparameter combinations for various algorithms. Experiments comparing the proposed algorithm with seven federated algorithms using real raw data collected from the distributed control systems of three domestic polyester fiber manufacturers show that the forecasting performance of the algorithm in this study excels across five different forecasting models. The proposed algorithm can provide support for newly established factories, reduce their production debugging costs, enhance production efficiency and product quality, and promote the intelligent development of the industry.

Key words: time series forecasting, security and privacy, federated learning, personalized, polymerization, esterification, neural networks

摘要：

在聚酯纤维生产过程中，酯化温度的精确控制至关重要。然而，由于生产设备和工艺参数的差异，传统预测方法难以满足个性化需求，且数据共享过程中存在隐私泄露和通信压力问题。提出了一种基于联邦学习的个性化自适应酯化温度时间序列预测算法。采用联合预测机制为每个客户端分配私有和共享模型，设置自适应阶段根据客户端数据分布智能调整模型参数，并学习客户端独有的联合预测权重，实现个性化预测输出。采用贝叶斯优化算法解决高维复杂和资源限制问题，快速高效地得到了最佳超参数组合。在三家聚酯纤维生产厂家的真实数据集上的广泛实验结果表明，算法在五种预测模型下均取得了最佳预测性能，有效提高了酯化温度预测的准确性。

关键词: 时间序列预测, 安全隐私, 联邦学习, 个性化, 聚合, 酯化, 神经网络