Evaluation of Uncertainty and Error of LSTM-Based Day-Ahead Load Forecasting Models

• Published in: 2021 IEEE PES Innovative Smart Grid Technologies Europe (ISGT Europe) pp. 1 - 6
• Main Authors: Zhu, S., Zou, M., Djokic, S. Z.
• Format: Conference Proceeding
• Language: English
• Published: IEEE 18-10-2021
• Subjects: Ensemble forecasting model; error forecasting; Europe; Load forecasting; long short-term memory model; Predictive models; Substations; Supply and demand; Uncertainty; Weather forecasting
• DOI: 10.1109/ISGTEurope52324.2021.9640191

Load forecasting is becoming increasingly important for planning and operational studies of electricity networks, which feature much higher levels of interactions between supply and demand sides, resulting in much larger variations of power flows. This paper evaluates uncertainty and error in a stacked bidirectional variant of a long short-term memory (SB-LSTM) model, which is applied for a day-ahead load forecasting. First, the paper analyses importance of the correct setting of hyperparameters of SB-LSTM model. Then, four different SB-LSTM forecasting models with four different data/window lengths are used to assess the "model uncertainty", i.e., variations in the forecasted demands due to multiple implementations of a specific SB-LSTM model on the same input data set. Afterwards, the four base SB-LSTM models are combined in a homogenous ensemble forecasting model, which dynamically integrates base learners and produces final predictions from their inputs. Finally, the fifth SB-LSTM model is built and trained using hindcasted errors of one base model to forecast its error on the test data. Input data for all SB-LSTM models are actual demands recorded in one Scottish MV substation, together with the corresponding meteorological and calendar data.