Prediction and selection in smart grids. A practical approach

Marc Frincu

Abstract: Smart grids generate massive amounts of data which can be used to enable an effective demand response. In demand response customers are asked by utilities to reduce their consumption during peak hours to avoid blackouts due to an excess in demand. However, the granularity and fluctuation of the consumption data makes consumption predictions challenging. In addition, external factors, such as weather and customer behavior need to be taken into account to further refine the prediction. Efficiently predicting the consumption for the next 4 to 24 hours ahead at 15 minute intervals is only the first step. The second step is to predict the curtailment potential of each customer based on historical events, a difficult task given the sparsity of these events and the amplitude of the data. Finally, based on the curtailment potential the set of customers needs to be selected from the total pool based also on the geographical distribution and the grid topology. In this talk we go briefly through each step, outlying some basic approaches and results obtained on a real smart grid installed on the USC campus.