Forecasting Electric Load by Aggregating Meteorological and History-based Deep Learning Modules
Masoud Bashari, Ashkan Rahimi‐Kian
Abstract
Accurate day-ahead (or 24-hours ahead) electric load forecasting for power systems is crucial for system's optimal operations. In evolving smart distribution grids, the importance of precise electric load forecast in day-ahead is even more important for distributed energy management systems (DERMS) and demand response (DR) programs, which are used by the independent system operators (ISO) and power utilities (PU) for day-ahead system planning and optimal operations. This paper captures both dynamic members' interdependencies and the impact of meteorological factors on the load sequence. In this regard, Long Short-Term Memory (LSTM) is applied to use the historical load sequences to forecast the 24 hours ahead values of the system load. On the other hand, a Deep Feedforward Neural Network (DFNN) is applied to map the forecasted meteorological parameters to the upcoming 24-hourly values of the system load. Finally, based on the historical errors of these two engines, a Beta distribution generates a probabilistic weight for aggregating the forecasted values at each hour. The proposed forecasting model performs better than single engines based on Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), and Root Mean Square Error (RMSE) metric when applied to day-ahead load forecasting for the city of Toronto.