A Data-Driven Bottom-Up Approach for Spatial and Temporal Electric Load Forecasting

Abstract

With the rapid urbanization, electrical infrastructure spreads to raw areas without existing loads. How to achieve accurate long-term load forecasts based on land use plans is a realistic problem. On the other hand, load forecasting (LF) should be extended to high spatial resolutions to guide middle- or low-voltage planning and time domain to consider the impacts of distribution generations and diversified users on multi-period system demands. A data-driven bottom-up spatial and temporal LF approach is proposed in this paper to solve these challenges. Land plots are treated as basic LF resolution to describe available multi-attribute data in smart grids and modern cities. Kernel density estimation and adaptive k-means are adopted to aggregate typical load densities and profiles of different land use types. Stacked auto-encoders are utilized to forecast the unknown plot load quantities. The neighbor plot loads are summed up to obtain the estimated loads of larger areas based on clustered load profiles. Case studies demonstrate that the proposed LF is more applicable than benchmark methods both in accuracy and application potential. The estimated hierarchical spatial and temporal results are of great significance to guide load balancing, power system planning, and user integration in different voltage levels.