A framework for wind field forecasting from sparse observations via integrated tensor completion and prediction

Accurate short-term wind speed predictions are crucial for future sustainable power systems, including renewable energy integration and grid stability. Existing approaches mainly focus on single-point or multi-site predictions at existing monitoring stations, with their performance heavily reliant on historical data. Consequently, these approaches are limited to locations with prior observations and fail to generate spatially continuous wind fields across regions. To address this limitation, this paper proposes a novel Tensor Completion-Convolutional Spatio-Temporal Prediction Network (TC-ConvSTPNet) that enables high-resolution, region-wide wind speed prediction. Firstly, a tensor completion network reconstructs high-dimensional wind speed tensors from sparse observational data. Then, a hybrid network integrating ConvLSTM2D and Conv2D is developed to model spatio-temporal dependencies for tensor prediction. Additionally, physical consistencies such as non-negativity and local conservation are ensured during the tensor completion and prediction. Finally, the proposed method is validated using real-world datasets. The results demonstrate that TC-ConvSTPNet outperforms conventional neural networks at monitored sites while accurately capturing wind speed trends even at unobserved locations. Besides its core capability of spatially continuous wind field forecasting from sparse inputs, the method facilitates data recovery under sensor outages or failures, as well as enhanced wind resource assessment in regions with limited historical observations. © 2026 Elsevier Ltd.