A framework for short-term activity-aware load forecasting

Abstract

In this paper, we present a framework for implementing short-term load forecasting, in which statistical time series prediction methods and machine learning-based regression methods, can be configured to benchmark their performance against each other on given data of smart meters and other related exogenous variables. Besides the prediction methods, forecasting performance also depends on the quality of training data. This is addressed by two characteristics of our framework on data collection and preprocessing. The first one is to introduce a human activity variable as an additional load influencing factor which reflects anomalous load patterns by aperiodic human activity. The second characteristic is to wavelet transform training data during the preprocessing stage to better extract redundant information from meter data. To investigate the feasibility of the proposed framework, a preliminary case study for predicting daily power consumption of several individual smart meters, using real-world data, is presented. The results indicate that, in general, the aggregation level of meter data and activity data matters.