Data needs for load forecasting at different aggregation levels using LSTM networks

With the mass introduction of renewable energies and distributed energy resources to the power grid, load forecasting approaches are adopted to improve the balancing of supply and demand, optimize network planning, as well as produce economic benefits for the existing energy system. To investigate the impact of the aggregation level and the training data types on the forecasting performance, this paper proposes an hourly short-term load forecasting method for three aggregation levels based on the long short-term memory (LSTM) network and Gaussian process. Five input data combinations are defined to represent a variety of training data types, and the proposed model is evaluated using real power consumption data. Finally, the optimal forecasting model and input data combination are determined for each aggregation level by observing the error metrics.