Study On Nonintrusive Load Monitoring Technology Based On HMM And Its Variants

Load monitoring technology is of great significance for energy conservation and emission reduction,planning,operation,and management of power system,and the implementation of smart grid.Nonintrusive load monitoring(NILM)technology is an efficient method for load monitoring.It has advantages of easy installation,low cost,high reliability,and ease of rapid promotion and so on over intrusive load monitoring(ILM).In recent years,methods based on hidden Markov model(HMM)and its variants have become one of the hot research topics in NILM field.By using HMM or its variants to model the aggregate load and appliances,the NILM problem can be described as the decoding problem of HMM or its variants.Once the NILM problem is described as the decoding problem of HMM,it can be solved by the Viterbi algorithm.Aim at the NILM application,the Viterbi algorithm is modified based on detecting load events.Since appliance states remain the same between two consecutive load events,the number of state sequences that need to be traversed by the Viterbi algorithm is reduced by considering that the appliance states only change at the load event times.Hence the computation complexity of the Viterbi algorithm is reduced.Compared with the standard Viterbi algorithm,the proposed modified Viterbi algorithm can reduce the computation time and be applicable for the scenarios where the number of appliances and the number of appliance states are much more.Difference factorial hidden Markov model and hidden semi-Markov model are combined as a new variant of HMM,which is referred to as difference factorial hidden semi-Markov model(DFHSMM).A method for NILM based on DFHSMM is proposed.The proposed method takes into account both the steady-state power signatures and state duration signatures and can discern appliances with similar power changes.It also can eliminate the impact of unmodeled appliances and be applicable for the scenarios where there exist unmodeled appliances.Moreover,the proposed method has high computational efficiency.It takes advantages of load events to greatly reduce the solution space of the established load monitoring model based on DFHSMM,groups the appliances into clusters according to their power signatures,and solves the DFHSMM-based load monitoring model to estimate appliance states for each appliance cluster respectively.