Spatial-temporal load forecasting using AMI data

Jin Xu, Meng Yue, Dimitrios Katramatos, Shinjae Yoo

Abstract: One of the critical requirements in power grid operation and planning is the ability to accurately forecast expected load. This allows for a heightened enhancement in grid operations, energy management, and planning. Load forecasting is historically based on aggregated spatial and temporal consumption data; with the deployment of Advanced Metering Infrastructure (AMI) systems, it can be achieved not only at a system level but also down to the consumer level. With this new increase in data, novel approaches and methods to load forecasting at a refined level can be explored. In this paper, a novel k-nearest Vector Autoregressive framework with exogenous input is proposed to spatial-temporally model household-level electricity demand from very short-term (15 min) to mid-term (2 weeks).We processed smart meter time series and geographical data from thousands of residential and commercial households. Our systematic experimental results showed an average of 27.3% RMSE and 31.6% MAPE improvement over the baseline model on a comprehensive 4-month dataset.