Missing data-aware robust electrical load forecasting based on hierarchical downsampling-upsampling spatiotemporal graph network

In real-world power systems, missing data is a frequent and inevitable issue due to sensor malfunctions and data transmission errors. This poses huge challenges for electrical load forecasting tasks since incomplete data can disrupt temporal patterns and obscure spatial dependencies. Traditional forecasting methods usually address this by imputing missing values before prediction, but such preprocessing can introduce bias and lead to error accumulation. To this end, this paper proposes a novel end-to-end missing data-oriented load forecasting framework based on the hierarchical downsampling-upsampling spatiotemporal graph network (HDU-STGNN). The proposed method can directly predict future load demand from partially observed inputs without requiring any data imputation process. Specifically, a temporal downsampling module is first developed to summarize observed portions of the input sequence into coarse-to-fine representations. This allows the model to capture long- and short-term load patterns while reducing sensitivity to irregularly missing entries. Then, a spatial coarsening and upsampling mechanism is proposed for effective information propagation across distant nodes. This allows the model to recover spatial dependencies even when local observations are unavailable. Finally, a multi-resolution attention fusion layer is utilized to adaptively reweight spatiotemporal features, which helps the model focus on reliable signals and suppress noise caused by incomplete data. Extensive experiments on real-world load datasets, covering three aggregation levels and missing rates up to 90 %, demonstrate the robustness of HDU-STGNN under varying degrees of data sparsity. In particular, the proposed model achieves an average reduction of 6.7 % in MAE and 9.9 % in MAPE compared with the strongest baseline across all settings, with improvements reaching up to 17.4 % under high missingness conditions.