Load Forecasting of Integrated Energy System Based on Combination of Decomposition Algorithms and Meta-learning

Aiming at the problem of limited load forecasting accuracy due to high sensitivity of inter-load correlation and poor seasonal generalization in the regional integrated energy system (IES), a multivariate load combination forecasting method based on the combination of decomposition algorithms and meta-learning is proposed. First, the correlation between multivariate loads in different time periods is quantified based on the dynamic maximal information coefficient, and the characteristic input variables are constructed according to the dynamic correlation results. Then, the load sequence is divided into multiple sub-sequence units by window sliding and divided into multiple tasks by variational modal decomposition to avoid the forward-looking bias problem caused by the overall decomposition. Finally, the subsequence is forecasted using a bi-directional long short-term memory model and the gradient iterations are reduced by a model-agnostic meta-learning algorithm, which reconstructs the subsequence and then fuses the fully connected layers to output the prediction. Based on the IES dataset of Arizona State University Tempe Campus, USA, the proposed hybrid model is verified to have higher IES multivariate load forecasting accuracy. © 2024 Automation of Electric Power Systems Press. All rights reserved.