Search

In this paper, the demonstration application project of non-intrusive load monitoring (NILM) technology is introduced, including the NILM-based smart meter, low voltage residential power information acquisition system of Jiangsu Province, data acquisition and display, etc. On this basis, considering the practical conditions, four technical characteristics of NILM system are expounded, which are processing data in different devices, pushing information with different time delays, interacting with users, improving performance continuously. A NILM research platform called cloud-terminal coordination system is built, and an unsupervised learning algorithm is explored, which provides a way of performance evolution for the system. This paper is expected to provide reference for researchers who committee to promoting the engineering application of NILM technology.

Basing on the hardware of smart electric power meter and aiming at the requirements of non-intrusive load identification application, the method of off-line learning and on-line classification based on support vector machine is studied. Two-sided cumulative sum based event detection method is employed to cut out the waveform windows between the start-up and shut-down of electric appliances, and then the characteristics of the electrical appliances are calculated. By using the feature data collected from electrical appliances samples, the parameters and support vectors of several binary classification decision functions are optimized through the MATLAB toolbox in an offline way. Then, a multiple classifier is constructed with several two class classifiers. Finally, the parameters of the multiple classifier are solidified into smart electric power meters to identify the type of electrical appliance load on-line. In order to verify the feasibility of the proposed method, the electrical appliance load classification test was carried out on the hardware platform of a single-phase electric power meter in the laboratory. The test results show that the accuracy rate of non-intrusive electrical appliance load identification is above 80%. The method proposed in this paper makes a large number of parameter optimization processes to be put on the computer, which is able to eliminate the online learning process and reduce the demand for calculation and storage resources of electric power meter, and has practical application value for the research of energy Internet on the user side.

Non-intrusive load monitoring (NILM) is an intelligent perception technology of electricity consumption information on the consumer side. Event detection, the fundamental and indispensable step of the NILM framework, has a direct impact on the accuracy of the results of load decomposition. This paper presents a novel NILM event detection approach based on Kalman filter and improved cumulative sum (CUSUM) algorithm. Firstly, this paper use Kalman filter to handle acquired data for sliding windows to make the power curve more smooth and suitable for next event detection step. An adaptive factor is introduced in the traditional CUSUM algorithm to improve the detection accuracy. The experiment results show that the proposed approach can not only effectively detect various input and cutoff events with different power level but also suits for the detection of the long transient events.

Event detection is an important foundation of non-intrusive load monitoring algorithm. In this paper, the common household appliance load events are classified, and a new triple-threshold event detection algorithm is proposed aimed at solving the problems of false detection and missing detection in the practical application. Firstly, a low power threshold is used to realize high-sensitive detection of the load events, and secondly the detected events are spliced according to the time threshold to get the complete events. Thirdly, the high threshold is used to discriminate the complete event set to filter out the disturbance caused by load fluctuation. Finally, the results are modified with a correction logic. The test results carried with static data show that, the algorithm proposed in this paper is more accurate for positioning the time of putting into and cutting off load, which is conducive to improve the accuracy of transient interval interception of load events, and has advantages in detecting slow rising load events. In addition, the algorithm proposed in this paper has a small amount of calculation, which can meet the requirements of application in the hardware of smart meter.

Focusing on the current impulse waveforms generated when the fixed or variable frequency air conditioners, vacuum cleaner, microwave oven and other electrical appliances are turned on, the characteristic parameters such as impulse amplitude, rising time, dropping amplitude, and falling time are defined, and a multi-feature Bayesian classification model is established. Considering that the classification algorithm works on the smart meter hardware, the computing and storage resources are limited, as a result that the model scale needs to be reduced as much as possible, a meant parameters Bayesian classification method is proposed. The current impulse waveform samples of different appliances are divided into several groups, and the mean value of the characteristic values of the samples is used as the parameters of the classification model. Under the same model capacity, the generalization error of the model is reduced. The simulation results show that the accuracy of Bayesian classifier constructed with meant value parameters is better than that of ordinary Bayesian classifier. In the laboratory, the impulse classification test is carried out on a single-phase smart meter, and the test results verify the feasibility of the proposed method.