Hierarchical reconciliation of convolutional gated recurrent units for unified forecasting of branched and aggregated district heating loads

Independent hierarchical estimations for the branched and aggregated streams of district heating loads often conflict with each other, leading to significant uncertainties in the daily operations of heating plants. To address this issue, a hierarchical reconciliation of convolutional gated recurrent units (HR-CGRU) is proposed to simultaneously forecast both branched and aggregated heating loads. Specifically, the trend, seasonal, and residual components of each branch are decomposed using singular spectrum analysis. For each time series within the hierarchical mainline-branch-component system, a temporal convolutional network is employed for feature extraction, followed by a bidirectional gated recurrent unit with a multi-head attention mechanism for forecast modeling. To integrate time series information across the hierarchy, a novel hierarchical reconciler is developed to unify the branched and aggregated forecasts using a joint top-down mapping and bottom-up fusion strategy. The proposed HR-CGRU is evaluated using industrial heating load data collected from a combined heat and power plant in Quzhou City of China. Comparative methods are also applied to test various forecasting scenarios. The results demonstrate the superiority of the present HR-CGRU over existing methods, highlighting its effectiveness in forecasting district heating loads.