随着智能电网中分布式能源主体的不断涌现，传统机器学习算法开始受到隐私和效率层面的约束，无法用于有效分析能源数据的内在规律。而联邦学习作为一种分布式、隐私保护式的合作式机器学习框架，为该问题提供了一种全新的解决思路。本文系统回顾了联邦学习领域的研究进展，并据此提出改进联邦学习的若干框架，使其更好地应用于智能电网的跨主体联合负荷预测与优化调度问题。 首先，鉴于此前联邦学习尚未被系统性地引入智能电网，本文立足于珠海市智慧城市课题的多源用能数据，将横向、纵向的联邦学习框架同时引入智能电网的区域负荷预测问题。特别地，本文为该框架设计了一套基于Diffie-Hellman密钥交换协议的预加密技术，在保障多方数据隐私的同时提升联邦加密效率。最终的联邦负荷预测框架能够高效准确地实现跨样本或跨特征数据主体的联合训练与预测。 其次，在联邦学习预测问题中，能源数据往往同时分散在不同地区中以及每个地区的不同利益主体中，即数据在样本维度和特征维度上被同时割裂。已有的联邦学习框架难以直接解决智能电网中的这种多元数据割裂问题。有鉴于此，本文创新性地提出了一种基于XGBoost算法的二维联邦学习框架，以有机结合横向联邦学习和纵向联邦学习，实现二维数据割裂下的合作式联邦预测。该框架下的能源主体信息公平性和计算效率也都得到了保障。进一步地，为了更好地为各地能源主体打造一套个性化的预测模型，本文提出了一种基于梯度提升树的联邦多任务学习框架。它将联邦学习框架与多任务学习框架有机结合，使各地能源主体在合作建立共性预测模型的基础上，拓展各自的个性化预测模型。该框架可切实提升各主体预测的准确性和泛化性，并充分保障训练过程中的数据隐私和计算效率。 最后，本文提出了使用联邦学习提供优质暖启动解、辅助多能主体分布式优化的联合经济调度框架。该框架通过联邦学习和交替方向乘子法构建了一套完整的跨主体分布式计算体系，使并行计算和隐私保护的思想得以贯穿在多能系统的预测优化过程中。联邦学习提供的暖启动解能在保证结果最优性、可行性的前提下，极大提升分布式优化的收敛性，进而提升计算效率。 上述联邦学习框架的有效性通过定性/定量的理论分析和实际数值算例得到了验证。

With the increasing number of distributed entities within the smart grid, centralized machine learning is no longer suitable for analyzing the underlying energy patterns due to escalating concerns on privacy and efficiency. Fortunately, federated learning, a distributed and privacy-preserving machine learning framework, has the potentials to alleviate such concerns through collaborative training. This dissertation systematically reviews state-of-the-art literatures on federated learning, meanwhile proposing research directions that improves federated learning for the cross-entity collaborative load forecasting and optimal scheduling of smart grids. First, in light of the fact that federated learning was not systematically introduced into the smart grid, this dissertation studies the multi-source energy consumption dataset of the Zhuhai smart city project. It then introduces both horizontal and vertical federated learning frameworks into the district load forecasting problem of smart grids based on the dataset. Particularly, a pre-encryption mechanism based on the Diffie-Hellman Key Exchange Protocol is designed for the frameworks, which increases the federated encryption efficiency while preserving data privacy. The final frameworks can effectively coordinate collaborative training and prediction in smart grids, where data are separated in the sample dimension or the feature dimension by the distributed entities. It is also noticed that in smart grid prediction problems, data are often scattered both by districts and by stakeholders in each district, indicating a data separation in both the sample and feature dimensions. This is a practical smart grid setting that current federated learning frameworks cannot address. To this end, we propose a novel hybrid federated learning framework based on XGBoost, which seamlessly combines horizontal and vertical federated learning, to address the two-dimensional data separation. Information fairness and computational efficiency issues are also properly addressed in the framework. Furthermore, in order to better personalize the models for each district and entity, a federated multi-task learning framework based on gradient-boosted decision trees is proposed. It seamlessly integrates federated learning into multi-task learning, such that personalized local models can be built on top of the collaborative global models. The framework significantly increases prediction accuracy and generalizability, while maintaining the privacy and efficiency of federated learning. Finally, this dissertation proposes a novel framework where federated learning assists in the distributed optimization of multi-energy systems by providing high-quality warm starts. The framework leverages federated learning and the Alternating Direction Method of Multipliers algorithm to construct a consistent distributive pipeline, such that parallel computing and privacy preservation can be achieved throughout the prediction and optimization processes across the collaborating energy systems. The warm starts generated by federated learning can significantly increase the convergence speed and thereby computational efficiency of distributed optimization, meanwhile preserving optimality and feasibility of the solution. The effectiveness of the proposed federated learning frameworks is confirmed through qualitative/quantitative theoretical analyses and carefully-designed case studies.