To confront the inaccuracy and imperfection of the environmental model, this article proposes a hybrid knowledge-data-driven reinforcement learning (KDD-RL) approach to solve the sequential optimal power flow problem during real-time operation. An improved soft actor-critic algorithm is proposed to train the control policy and formulate the sequential dispatch commands to the generators. To promote the safe exploration of the reinforcement learning algorithm, a hybrid knowledge-data-driven safety layer is developed to convert the unsafe actions into the safety region. Furthermore, a security-constrained linear projection model with an inactive constraint identification process is proposed to accelerate the computation efficiency of the safety layer. Numerical simulation results verify the superiority and scalability of the proposed approach in improving the decision-making efficiency and promoting the security operation of the power systems.

Increasing renewable energy resources introduces uncertainties into utility grids, calling for more frequent use of alternative current optimal power flow (AC-OPF) than before. Conventional AC-OPF solvers are too slow for real-time applications; to speed up the computation, deep neural network (DNN) based solvers are introduced. However, their results typically cannot enforce the security constraints of the utility grid. To overcome this difficulty, FRMNet is proposed in this paper, which combines DNN with feasibility restoration mapping (FRM). The DNN receives the load as input and outputs a partial solution. The following FRM maps the infeasible solution to the feasibility region to satisfy the AC-OPF constraints. The proposed FRMNet has four advantages compared to conventional and DNN-based solvers. First, the feasibility of overall solutions is guaranteed via FRM. Second, FRMNet is a self-supervised training method so that it avoids the laborious preparation of the optimal solutions for a training dataset using conventional solvers. Third, FRM is differentiable so that AC-OPF information encoded in gradients is transferred to DNN, which leads to a DNN with high feasibility rate. Fourth, the overall computation time of FRMNet is fast due to its DNN's high feasibility rate. Comprehensive experiments are conducted on IEEE standard cases. Results show that FRMNet's outcomes satisfy all constraints at a significantly faster average speed than conventional solvers. When FRM participates in training, the performance of DNN is stably improved and achieves better results than other DNN-based models. Even DNN alone achieves a high feasibility rate, a negligible degree of constraint violations and a small optimality gap, compared to the state-of-the-art DNN based solvers.

In recommender systems, the patterns of user behaviors (e.g., purchase, click) may vary greatly in different contexts (e.g., time and location). This is because user behavior is jointly determined by two types of factors: intrinsic factors , which reflect consistent user preference, and extrinsic factors , which reflect external incentives that may vary in different contexts. Differentiating between intrinsic and extrinsic factors helps learn user behaviors better. However, existing studies have only considered differentiating them from a single, pre-defined context (e.g., time or location), ignoring the fact that a user’s extrinsic factors may be influenced by the interplay of various contexts at the same time. In this article, we propose the intrinsic-extrinsic disentangled recommendation (IEDR) model, a generic framework that differentiates intrinsic from extrinsic factors considering various contexts simultaneously, enabling more accurate differentiation of factors and hence the improvement of recommendation accuracy. IEDR contains a context-invariant contrastive learning component to capture intrinsic factors, and a disentanglement component to extract extrinsic factors under the interplay of various contexts. The two components work together to achieve effective factor learning. Extensive experiments on real-world datasets demonstrate IEDR’s effectiveness in learning disentangled factors and significantly improving recommendation accuracy by up to 4% in NDCG.