Realization of a Prototype for the Implementation of Non-Intrusive Load Monitoring

The Non-Intrusive Load Monitoring is a technique that can be used to recognize appliances that are being used in a house from their electrical characteristics, such as the profile of the current that they absorb. This technique can be employed to potentially achieve an appliance by appliance complete disaggregation of the loads connected to a power grid and therefore of the ongoing energy consumption. The Non-Intrusive Load Monitoring can have different applications, going from the partitioning of the electrical bill to the smart management of local power sources, like solar panels. The purpose of the proposed study is to realize a prototype which is capable of detecting the activation of an appliance, called “event”, from the waveform profile of the current absorbed from the power grid, and of classifying it. Moreover, our prototype is capable of communicating via Wi-Fi with a remote cloud server and possibly with other devices. The employed recognition mechanism is a Feed-Forward Neural Network (FFNN) whose classification is based on the spectrum measured after a certain time from the activation of an appliance, in its steady-state. In particular, we explored two different implementations of the described prototype, both realized with the same hardware resources: one in which the appliance classification is performed in a remote cloud server and the other one in which it is carried out in edge. Either way, the prototype receives as input the waveform of the current that is absorbed from the power grid and detects the eventual activation of an appliance. Afterwards, while in the first solution we transmit the waveform itself to the cloud to perform the remote frequency analysis and classification; in the second one we do those same operations at the local hardware level. At the end, in both cases, the classification outcomes are transmitted via Wi-Fi, through an MQTT broker, to multiple other devices, such as smartphones and computers. The results obtained with these two solutions present a comparable accuracy in the appliances recognition (~90%), but they strongly differ in the in the organization of the computational cost among the local hardware resources and the remote cloud, with the consequent need, in one case, of an appliance classification service in cloud and, in the other case, of a hardware powerful enough to perform the recognition. The optimization of this computational distribution, as well as a study of the cost-effectiveness trade-off for the employment of hardware with different levels of performances, can be the foundations of further studies to be conducted on these subjects.