Machine Learning Approach to Event Detection for Load Monitoring

Abstract

Today’s Electricity market is highly interconnected. The actions of customers, transmission and distribution systems, generation systems, markets, and service providers are correlated. The dynamic stability of the grid depends on the dynamic characteristics of all these components and systems. The least understood component is the consumer load. While average power is metered for billing, the voltage or current relationships and dynamic characteristics of loads are not monitored. It is the purpose of this work to analyze high-bandwidth voltage and current measurements from an office building to characterize certain load features. The focus is on transient events under the premise that some features will be revealed when loads switch on and off, or change mode of operation. This approach involves certain tasks. First an event needs to be identified, and the duration of transient needs to be estimated. Second, features that might be useful for classifying transients must be determined. Third, classification is performed using machine learning Techniques. Signal Processing tools based on changes in current waveforms are used to detect events. Many features are examined in this work including voltage and current harmonics, power factor, active, reactive, and distortion power. Machine Learning techniques guide the classification task. These include self-organising maps and neural network. Analysis and results are presented based on actual load measurements.