Abstract
In this paper a dynamic Artificial Neural Network (ANN) model called Meta-ANN is developed for forecasting the short-term grid load. The primary ingredient of the model is a base module which is an ANN trained over a large historical data set to learn the long-term trend and seasonality of grid load. To capture the nonstationary pattern of the grid load, an error-correction module based on the idea of meta-learning is integrated into the model. This module finetunes the base module according to most recent prediction errors. For each day of interest, Meta-ANN generates a new ANN model started from the base module by tracing the gradient of the prediction loss on recent observations weighted by learning rates with specific structures. The full Meta-ANN model is trained by jointly optimizing the base module and error-correction module via gradient descent algorithms. The implementation based on gradient descent algorithms is detailed with streamlined mathematical formulations. The proposed model is tested on the open-access data from Elia, a Belgian transmission system operator, for forecasting the daily mean load and load profile. The numerical study shows that Meta-ANN makes more accurate and robust prediction by effectively capturing the nonstationary pattern in grid loads.
•Meta-learning is used to build a dynamical ANN for short term load forecasting.•The model input is very concise and easily interpretable.•The Meta-ANN model and its implementation are detailed in streamlined math forms.•The proposed model is validated on open-access real data from Elia, a Belgium TSO.