Dual-level domain-adaptive spatiotemporal learning for ultra-short-term distributed PV power forecasting in distribution grids
Accurate ultra-short-term forecasting of distributed photovoltaic (DPV) power generation under domain shift is essential for reliable distribution-grid operation, distribution-network balancing, and local scheduling. However, existing forecasting models often degrade when transferred to new PV sites characterized by limited labeled data, heterogeneous system configurations, and variable weather conditions. To address these challenges, this paper proposes a dual-level domain-adaptive spatiotemporal learning framework for robust DPV forecasting in data scarce target domains. The framework jointly models dynamic inter-site dependencies, multi-scale temporal patterns, and domain-invariant representations. Specifically, a Dynamic Graph Attention (DGA) module learns time-varying spatial dependencies among PV sites directly from node features, eliminating the need for a predefined graph topology and improving robustness to nonstationary and partially observed inter-site interactions. The learned spatial embeddings are processed by a Multi-Scale Temporal Fusion (MSTF) block with four parallel temporal scales, k=1,2,3,6, to capture rapid ramps, intra-hour fluctuations, and diurnal trends. A lightweight stacked LSTM then refines short-horizon temporal dependencies and preserves forecasting continuity. Cross-domain transferability is achieved through a gradient-reversal-based adversarial mechanism that reduces source-target discrepancies. The adaptation is performed at two complementary representation levels: node-level alignment regularizes the DGA-derived site embeddings, whereas graph-level alignment constrains the LSTM-refined spatiotemporal representations. This dual-level formulation promotes the transfer of both local site-level patterns and global system-level dynamics across heterogeneous DPV domains. Experiments on two real-world DPV datasets demonstrate improved multi-horizon accuracy, enhanced peak-event tracking, and reduced source-target feature discrepancy, indicating the effectiveness of the proposed framework for transferable DPV forecasting in distribution grids. Copyright © 2026. Published by Elsevier Ltd.
Библиографическая ссылка
Ullah, Shafqat, Liguo, Wang, Khan, Aftab Alam, Dreglea, Aliona, Sami, Irfan, Sidorov D.N. Dual-level domain-adaptive spatiotemporal learning for ultra-short-term distributed PV power forecasting in distribution grids // Sustainable Energy, Grids and Networks. Vol.48. ID:102548. 2026. DOI: 10.1016/j.segan.2026.102548
Скопировать
WOS
SCOPUS
