随着可再生能源在发电侧所占比例的不断提高，在日前调度过程中，通过提升准确预测发、用电量的精度来降低与实际调度结果的偏差变得更加困难。因此，有必要根据电网实时状态和超短期预测技术进行前瞻调度，以修正机组运行计划。在优化调度方面，基于传统模型驱动方法的应用往往受到电力系统规模的限制，无法满足实时调度的要求。而数据驱动方法在面对大规模系统时可以提供强大的在线决策支持能力，且在面对可再生能源带来的系统不确定性增加的问题时，数据驱动方法也具有很强的适应能力。鉴于此，本文开展了基于数据驱动的可适应性电网前瞻优化调度技术方面的研究。首先，本文提出了一种基于深度强化学习的电网多时段滚动前瞻经济调度模型。其中，在传统多时段滚动前瞻调度模型基础上，通过引入深度强化学习的算法来训练前瞻调度决策智能体，建立了包含状态输入与动作输出的深度强化学习模型，并对奖励函数进行了设计，给出了不同奖励函数的权重设置方法。所构建的调度模型为决策智能体的具体训练以及多运行场景下的迁移奠定了理论基础。其次，本文提出了多时段滚动前瞻调度模型智能体的训练方式。其中，通过设计系统不平衡量分配环节，将模型驱动的方法引入到智能体的深度强化学习算法的训练过程中；与此同时，在深度确定性策略算法的基础上引入了神经网络的参数预训练环节来进一步提高智能体训练的效果，并基于IEEE30节点的算例进行正常运行场景下前瞻调度模型的训练效果测试。最后，在考虑了电网实际运行场景的多元化特性下，本文提出了多运行场景下基于迁移强化学习的前瞻调度智能体的并行计算框架。其中，给出了运行场景数据集的构建、特征数据集的提取与选择以及如何根据特征数据集进行运行场景的聚类的过程；此外，通过引入基于模型的迁移学习方法，可指导迁移智能体在新运行环境中进行高效探索，并基于SG126节点算例系统测试所提架构的有效性。本文研究内容以提升电网经济调度的智能化水平为出发点，充分利用了数据驱动方法的智能性、自适应性与快速性特点，有效地提升了调度模型应对短时不确定场景的适应能力，可为我国电力系统调控优化决策领域工作提供有益借鉴。

With the increasing proportion of renewable energy in the power source side, it is more difficult to reduce the deviation from the actual dispatch results by improving the accuracy of predicted power generation and load demand in the day-ahead dispatching process. Therefore, it is necessary to carry out look-ahead dispatching to correct the operation plan according to the real-time state of the power grid and ultra-short-term prediction technique. Regarding the optimal dispatching, the applications of traditional model-driven methods are often limited by the scale of power system and cannot meet the requirements of real-time dispatching. The data-driven method, on the other hand, can provide powerful online decision support capabilities in the face of large-scale systems, and are also highly adaptable when facing the problem of the increased uncertainty caused by renewable energy sources. In view of this, this thesis conducts research on the adaptive look-ahead dispatching technology based on data-driven method.Firstly, a multi-period rolling look-ahead economic dispatch model of power grid based on deep reinforcement learning algorithm is proposed. Specially, on the basis of the traditional multi-period rolling look-ahead dispatching model, by introducing the deep reinforcement learning algorithm to train the look-ahead economic dispatching decision-making agent, a deep reinforcement learning model including state input and action output is established, and the corresponding reward functions are designed, in which the weight setting methods pertinent to different reward functions are given. The constructed dispatching model lays a theoretical basis for the specific training and transfer of decision-making agents under multi-operation scenarios.Secondly, a training mode of multi-period rolling look-ahead dispatching model agent is proposed. Specially, the model-driven method is introduced into the training process of the deep reinforcement learning algorithm of the agent by designing the system imbalance allocation link. At the same time, based on the deep deterministic policy algorithm, the parameter pre-training link of neural network is introduced to further improve the effect of agent training, and the training effect of the look-ahead dispatching model under the normal operation scenario is tested using the IEEE 30-bus system.Finally, considering the diversified characteristics of the actual operation scenarios of the power grid, this thesis proposes a parallel computing framework of the look-ahead dispatching agent based on transfer reinforcement learning in multiple operation scenarios. Specially, the construction of the running scene data set, the extraction and selection method of the corresponding feature data set, and how to cluster the running scene according to the feature data set are given. In addition, by introducing the model-based transfer learning method, the transfer agent can be guided to explore efficiently in the new environment, and the effectiveness of the proposed framework is tested based on the SG126-bus system.The research content of this thesis takes improving the intelligent level of power grid economic dispatching as the starting point, makes full use of the intelligence, adaptability and rapidity of data-driven methods, and effectively improves the adaptability of dispatching model to short-term uncertain scenarios, which can provide useful reference for the work in the field of economic dispatch of power system in China.