An Improved Steady- and Transient-State Mixed Non-Intrusive Load Monitoring Using Viterbi Algorithm

Non-intrusive load monitoring (NILM) is the most commonly used method to achieve load state identification, which is an important technology to realize power grid panoramic perception and support the carbon peaking and carbon neutrality goals. Depending on the detection object, NILM can be divided into two categories, i.e., steady-state signature based NILM and transient-state signature based NILM. Various studies have been conducted on both fields respectively, but the internal relationships between different states are rarely discussed. To explore the potential of associating the sequential states in load disaggregation problem, this paper thoroughly investigates the sequential logics between the load states in adjacent steady-state process and transient-state process, and makes use of them to improve the NILM performance. Firstly, considering the switching states of appliances, steady-state signature based NILM is solved by discrete particle swarm optimization (DPSO) algorithm. Secondly, transient-state signature based NILM is addressed by dynamic time warping (DTW) approach, to deal with the complex event characteristics. Then, a probability evaluation system is applied for decision-making, where multiple identification results with high confidence are selected to construct the candidate sets of independent identification results. Lastly, the candidate sets of adjacent steady-state and transient-state are associated together, and Viterbi algorithm is utilized to establish the probabilistic sequential model and optimize the load identification results. The proposed method is analyzed and validated on load consumption data from both low voltage network simulator (LVNS) and UK-DALE dataset. Four metrics, including F1 measure (F1), root mean square error (RMSE), mean absolute error (MAE), and normalized mean square error (NMSE), are utilized to evaluate the load disaggregation performance. The results show that the proposed method can effectively improve the overall load identification accuracy, especially for these appliances with high rated power. In the simulation-based case, the disaggregation performance for the incandescent light bulb rated at 40W is unsatisfactory by the proposed method, due to the similarity of the rated power and frequent fluctuating power values. However, such results for low-power appliances do not affect the overall performance, while the global NILM is greatly enhanced by integrating the different states together. In the measurement dataset-based case, all appliances, including the 85 W fridge, are accurately identified by the proposed method, along with a significant improvement compared with independent steady-state signature based NILM. Besides, the disaggregation results for dishwasher, the appliance with complex operation states, are remarkably promoted by the proposed method, indicating the superiority and applicability of this study. Finally, two conventional NILM approaches, i.e., Combinatorial optimization (CO) and factor hidden Markov model (FHMM), are also tested and compared in the cases. Although CO and FHMM outperform DPSO under traditional framework, by integrating the proposed strategy, the DPSO based NILM is largely enhanced from all metrics, generating a better NILM solution. The following conclusions can be drawn from the verification analysis: (1) By utilizing the sequential logics between adjacent load states, the load disaggregation performance could be largely enhanced compared with independent steady-state signature based NILM, as well as transient-state signature based NILM. Bedside, the proposed framework has a good compatibility with diverse independent NILM approaches. (2) Compared with CO and FHMM, the proposed model has definite physical significance. In addition, the proposed method shows remarkable enhancement in load disaggregation, even if the basic steady-state signature based NILM algorithm performs poorly. (3) The proposed model and method are flexible in time scalability, which is promising in load perception and prediction under the dynamic rolling mechanism. © 2023 Chinese Machine Press. All rights reserved.