22 January 2024 in IEEE Internet of Things Journal
Demand response (DR) programs aim to improve grid reliability, reduce customer costs, and manage building energy demand. However, coordinating autonomous demand response in grid-interactive buildings faces challenges due to uncertainties in system model parameters, complex energy demands, and interdependent operational constraints. Specifically, the energy management problem in grid-interactive buildings is formulated as a Partially Observable Markov Decision Process (POMDP), where each building has only local information and no knowledge of other buildings’ behavior. This makes achieving consistent coordinated energy management challenging. To address these issues, this paper proposes a novel approach called Sequential Update-based Multi-Agent Deep Reinforcement Learning (SUMADRL). This approach updates the policy of each building sequentially, considering the updates already made by previous buildings. This ensures that the joint policy update of buildings is always highly aligned with the overall energy optimization goal and improves coordinated energy management among buildings. Additionally, the sequential update scheme provides theoretical guarantees for achieving monotonically improving rewards and converging to Nash equilibrium, enhancing credibility and stability. Experimental results show that the proposed method significantly reduces energy costs, achieves efficient load shaping, and demonstrates superior effectiveness and stability in demand response compared to the Multi-Agent Deep Deterministic Policy Gradient (MADDPG) algorithm.
7 March 2024 in IEEE Transactions on Instrumentation and Measurement
Non-intrusive load monitoring (NILM) involves deploying monitoring devices at the customer’s bus entrance to decompose the total power captured into individual operational data for each device. Existing NILM load identification algorithms primarily adopt closed-set identification, where the same types of devices are used for both training and testing. However, a more realistic scenario is open-set identification, where device types in the training and test sets differ, encompassing not only devices from the training set but also unknown devices. To address this challenge, we propose a method leveraging reconstruction and supervised contrastive learning based on voltage-current (V-I) trajectories. Supervised contrastive learning improves the coherence of V-I trajectories within identical appliance classes, facilitating the model’s precise comprehension and emulation of distinctive appliance behavioral patterns. By introducing an autoencoder (AE) into the model, associating each appliance class with its unique subset of features. Known V-I trajectories activate associated features, preserving the topological characteristics of the trajectory. In contrast, unknown V-I trajectories lack reconstruction mappings, resulting in thoroughly mutated trajectories. Moreover, we innovate data augmentation for contrastive learning by introducing a denoising autoencoder (DAE). This approach introduces noise in V-I trajectories, the model learns to remove noise while preserving real appliance characteristics, generating more robust V-I trajectories. Experimental results on two publicly available datasets demonstrate the method’s outstanding performance.