Abstract: The precise prediction of ship fuel oil consumption (FOC) at sea plays a crucial role in protecting the marine environment and reducing operational costs in the shipping industry.However,the data privacy of maritime vessels and statistical heterogeneity of heterogeneous ships pose limitations on the predictive performance of conventional machine learning (ML) methods.To address it,a method called personalized federated learning (PFL) with CatBoost was proposed.Firstly,data from various sources,including ship information and sea condition data,were merged and cleaned to enhance data quality.Secondly,CatBoost,a gradient boosting method for categorical features,was applied to perform feature selection on local data,removing redundant information.Next,the Federated Learning with Personalization Layers (FedPer) framework was introduced,incorporating a personalized layer to build a predictive model for heterogeneous ship FOC while ensuring data privacy.Furthermore,the basic layer's weight matrix was aggregated using the Federated Averaging Algorithm (FedAvg) for parameter updates and feedback,and the weight matrix for the personalized layer was optimized locally by client-side Deep Feedforward Neural Networks (DFNN) to mitigate the impact of data heterogeneity and improve prediction accuracy.Finally,comparative experiments were conducted using real-world examples of heterogeneous ship fuel oil consumption.The results demonstrate that the proposed method achieves higher prediction accuracy compared to other models.The proposed method has practical significance for reducing the heterogeneous ship FOC.

Key words: heterogeneous ship fuel oil consumption prediction, personalized federated learning, categorical boosting, federated averaging algorithm, deep feedforward neural networks

摘要： 船舶航行油耗的精准预测,对保护海洋环境、减少航运业运营成本起关键作用,但航运业船舶的数据私密性、及异构船舶的数据异质性,导致常规机器学习方法的预测效果有限。为此,提出一种基于类别型特征的梯度提升（Categorical Boosting,CatBoost）联合个性化联邦学习(personalized federated learning,PFL)预测方法。首先,对本地不同数据源的船舶信息数据及海况数据进行数据融合和清洗过滤,以提高输入数据质量;其次,对本地融合数据用CatBoost进行特征选取,以去除冗余数据;随后,引入带个性化层的联邦学习（Federated Learning with Personalization Layers,FedPer）框架,建立异构船舶航行油耗预测模型,以保证异构船舶的数据私密性;进一步,对基本层权重矩阵采用联邦平均算法（Federated Averaging Algorithm,FedAvg）聚合参数并反馈,对个性化层权重矩阵由本地客户端采用深度前馈神经网络（Deep Feedforward Neural Network,DFNN）进行训练优化,以消除数据异质性的影响,提高预测精度。最后,结合实际异构船舶航行油耗算例进行对比实验,结果表明,相比于其他模型,CatBoost联合个性化联邦学习预测方法的预测精度更高,对降低异构船舶航行油耗具有一定的指导意义。

关键词: 异构船舶航行油耗预测, 个性化联邦学习, 基于类别型特征的梯度提升, 联邦平均算法, 深度前馈神经网络