A novel decompose-cluster-feedback algorithm for load forecasting with hierarchical structure

In load forecasting fields, electricity demand with hierarchical structure is very popular where there are some differences among investigated load series because of geography or customers’ habits. Common methods usually ignore their differences and introduce some complex models to improve forecasting performance. Therefore, appropriately dealing with the diverged series is necessary to achieve accurate predictions in hierarchical load forecasting. In this paper, we propose an iterative decompose-cluster-feedback algorithm, which is modified from CLC method, to further improve the performance of forecasts at the total level of hierarchy. Compared with CLC, this algorithm applies empirical mode decomposition (EMD) to decompose load series into sub-series with various amplitude–frequency characteristics, which can avoid directly operating on load series. Specifically, the divergence can have detrimental effects on forecasts if ignored. Finally, we test the proposed algorithm with three real tasks of load forecasting with hierarchical structure, and the experimental results show that the performance of our algorithm is at least 43% better than a SVR-BU method, 52% better than a TD-MLP and a TD-LSTM-SDE method, and 32% better than several methods belonging to middle-out method.