With the increasing penetration of renewable energy sources and electric vehicles (EVs), coordinated operation of EV aggregator (EVA) and distribution system operation (DSO) becomes a complex multistage and multidimensional stochastic problem. The motivation behind this paper is to develop a decentralized mechanism to offer a computationally efficient and almost optimal on-line policy for such problem under the framework of transactive energy control (TEC). First, a heterogeneous decomposition-based TEC is designed by utilizing the heterogeneous interactions between DSO and EVA. Then, a decentralized approximate dynamic programming-based algorithm is proposed to offer almost optimal dynamic TEC policies. A decentralized value function approximation approach with temporal difference learning is further employed for entities to learn how to utilize the flexibilities of their resources in response to the stochastic exogenous information. Case studies demonstrate the effectiveness of the proposed algorithm in terms of optimality, robustness, computation efficiency, and scalability.

To balance the unexpected power disturbances, an independent system operator (ISO) should assign the dynamic power regulation commands to all the regulation resources via an automatic generation control (AGC) dispatch. It can be described as a bi-objective Pareto optimization by considering the minimizations of total power deviation and total regulation mileage payment. In this work, a novel dropout deep neural network assisted transfer learning (DDNN-TL) is proposed to rapidly approximate the high-quality Pareto optimal solutions for AGC dispatch. Firstly, the training data is generated from the Pareto optimal solutions and fronts obtained by various multi-objective optimization algorithms according to the anticipated total power regulation commands. Secondly, the network parameters of DDNN will be updated via an off-line training with these data at each frequency regulation service period. Finally, based on an efficient transfer learning with a correction of infeasible solutions, DDNN-TL can directly approximate the high-quality Pareto optimal solutions for on-line decision of AGC dispatch. Case studies of DDNN-TL for bi-objective Pareto AGC dispatch are carried out on a two-area load frequency control model and Hainan power grid of China Southern Power Grid (CSG), which demonstrates its superior performance on optimization speed and stability compared with other multi-objective optimization algorithms.

Non-Intrusive Load Monitoring (NILM) has become a research boom due to its energy-saving benefits as the smart grid develops and the greenhouse effect intensifies. However, existing NILM models may perform poorly in the target domains collected from different sources, or those containing unseen appliance types, blocking the rollout of NILM. In contrast to the methods that update the trained model on the unseen data to handle domain shift or heterogeneous label space, this paper proposes a novel generalizable NILM method based on metric-based meta-learning and supervised pre-training. It beforehand learns the cross-domain meta-knowledge through meta-training to eliminate the need for parameter updates in the target domain. Moreover, this method is specifically designed for practical deployments at the edge where both labeled instances and computation resources are limited. The proposed method transforms the appliance current into a metric vector via lightweight 1D convolution and then uses the relation module to propagate the labels via similarity. The experimental results on three public datasets and the proposed dataset demonstrate that the proposed method can achieve high generalization performance in different datasets, user’s houses, appliance brands, and appliance types.

It is important to achieve an efficient home energy management system (HEMS) because of its role in promoting energy saving and emission reduction for end-users. Two critical issues in an efficient HEMS are identification of user behavior and energy management strategy. However, current HEMS methods usually assume perfect knowledge of user behavior or ignore the strong correlations of usage habits with different applications. This can lead to an insufficient description of behavior and suboptimal management strategy. To address these gaps, this paper proposes non-intrusive load monitoring (NILM) assisted graph reinforcement learning (GRL) for intelligent HEMS decision making. First, a behavior correlation graph incorporating NILM is introduced to represent the energy consumption behavior of users and a multi-label classification model is used to monitor the loads. Thus, efficient identification of user behavior and description of state transition can be achieved. Second, based on the online updating of the behavior correlation graph, a GRL model is proposed to extract information contained in the graph. Thus, reliable strategy under uncertainty of environment and behavior is available. Finally, the experimental results on several datasets verify the effectiveness of the proposed model.

In the presence of the increasing penetration of electric vehicles (EVs) and conflict of independent optimization objectives among each electric vehicle aggregator (EVA), real-time optimal scheduling (RTOS) of large-scale EVs based on dynamic non-cooperative game approach is proposed for optimal decision makings in a dynamic pricing market. First, real-time optimal scheduling framework is designed to describe the flow of energy and information. Then, equivalent model of large-scale EVs is formulated to address "curse of dimensionality" caused by a large number of decision variables. Then, the potential game theory is used to study the existence and uniqueness of the Nash equilibrium (NE) solution. Finally, a distributed approach based on alternating direction method of multipliers (ADMM) is designed to achieve the equilibrium. Case studies demonstrate that the proposed approach achieves peak load shifting and reduces cost of EVAs significantly. Furthermore, the proposed method obtains higher-quality solution compared with other methods and is more applicable for real-time optimal scheduling of large-scale EVs due to its high computation efficiency and privacy protection.