| Energy crisis has aroused more and more worries and concerns,and it's important to find a more efficient way to use electrical power.Non-intrusive load monitoring(NILM)aims at estimating the electricity consumption of individual appliances through the household's aggregate power consumption.Providing more detailed electricity consumption information can bring benefits to residents,power grids and policy makers.This thesis conducts research on three aspects:event detection,load identification and energy disaggregation,and proposes a set of state-of-the-art NILM algorithms.(1)Three mainstream event detection algorithms are compared and analyzed.First,a sliding window with margins is introduced.Window with margins pays more attention to the data at the edges of the sequence,and gets higher detection rate of change points.The fundamental active power is selected as load signature,and three event detection algorithms are tested on BLUED dataset,the Bayesian Information Criterion,the Hotelling T~2 test and the Cumulative Sum,among which the Hotelling T~2 test achieves the best performance.(2)A general,low-complexity convolutional neural network(CNN)model is designed for load identification.After an event is detected,the CNN model uses the current data to identify the specific load that generated this event.After merging and compressing the CNN model and the event detection algorithm,the overall algorithm is deployed on an embedded device based on STM32f4,and tested on a public dataset.The results verify the effectiveness and feasibility of the algorithm.(3)A deep neural network model based on self-attention mechanism is designed for energy disaggregation.The proposed model is tested on UK-DALE dataset and compared with basic benchmark models and state-of-the-art models.The results show that our proposed model outperforms the other models in various metrics,and has outstanding performance in disaggregating appliances with long operating cycle and complex,low-power appliances. |
PDF Full Text Request