Resilient Event Detection Algorithm for Non-Intrusive Load Monitoring Under Non-Ideal Conditions Using Reinforcement Learning

Event detection is critical in a non-intrusive load monitoring (NILM) solution. NILM is essential for the implementation of some demand-side management (DSM) techniques. This article proposes an event detection algorithm based on reinforcement learning for NILM purposes (RLNILM). The proposed method employs a number of simpler traditional event detection algorithms, e.g., LLR voting, or SWDC, to train the RLNILM agent through a feedback system that separates the RLNILM agent from directly accessing consumers' data. The performance of the proposed RLNILM method is validated using the real-world data from the iAWE dataset under ideal and non-ideal conditions. Four test scenario groups include cases where the frequency, input electric signals, or access to crucial grid information varies significantly. In all scenarios, the RLNILM agent outperforms traditional event detection algorithms used in the feedback system. The results show not only the proposed architecture increases the cyber security of the customers connected to NILM services, but also it improves the performance of real-time event detection algorithms in non-ideal grid conditions.