A low frequency power-based event alignment method for trained NILM systems

Non-Intrusive Load Monitoring (NILM) Systems are one of the main proposals to characterize the household power consumption. This work proposes an alignment method based on the low frequency band power consumption which overrides the high frequency noise. To this end, a Gaussian kernel its used to smooth the power signal in order to obtain the right time where a device is turned on, and therefore making the device characterization task feasible. To develop this kind of systems several steps are required: First the devices within the system must be characterized with its high frequency fingerprint, second a data acquisition system must be implemented to obtain the household feeding voltage frequency components with a very wide spectrum from low to high frequencies. The last step will be the data disaggregation task. Unfortunately, the first step is not straightforward as there are no agreements for the manufacturers to provide their devices frequency fingerprint. Therefore, some training data must be obtained in order to characterize such devices. Unfortunately this task is realized by hand by creating a set of events where the devices are turned on and off along the monitoring period. It is very common for the data to have some misalignment with the monitoring data and therefore no device characterization can be done. Several problems are faced when trying to align the events data set mainly due to the high frequency noise which makes this a very hard task.