Efficient neural network representations for energy data analytics on embedded systems

Abstract

Electrical energy consumption data contain a great wealth of information. Currently, however, only limited insights are possible into the data that are collected by millions of smart meters everyday. A major reason for this limitation is the high computational and memory demand of corresponding analysis methods. This makes their execution on embedded processing systems virtually impossible. For example, Non-Intrusive Load Monitoring (NILM) methods strive to extract the individual power demand of an appliance from the aggregated power data of a household. Recent Non-Intrusive Load Monitoring (NILM) solutions are based on highly memoryintensive neural networks. This limits their application to powerful hardware systems and strongly hampers their wide practical adoption. In this work, we demonstrate how the application of neural network model compression techniques can be applied to make state-of-The-Art Non-Intrusive Load Monitoring (NILM) solutions operable on systems with limited resources. Through comparatively analyzing the impact of size reduction techniques on the models, we show how a balance between neural network size and NILM accuracy can be found. We moreover verify the operability on a real embedded hardware system by means of a practical evaluation. Model compression enables NILM methods to be executed on a vastly greater number of devices, at similar levels of performance as compared to unmodified algorithms.