Accurate and robust wind power forecasting plays a crucial role in ensuring the safety and stability of the power system. Hybrid spatiotemporal forecasting models based on graph convolutional networks have received widespread attention due to their advantages in spatial feature extraction. However, these methods are susceptible to the quality of the generated graph due to data noise and missing issues, resulting in suboptimal performance. In this paper, we propose a hybrid deep learning model based on spatiotemporal graph contrastive learning to address the above issues. Specifically, the model's encoder combines an adaptive graph convolutional network with LSTM to capture fine-grained spatiotemporal dependencies. To enhance the robustness of the encoder against data noise, we apply feature-level and topology-level data augmentation techniques to the model's input and design two contrastive learning auxiliary tasks from the temporal and spatial dimensions, respectively. Furthermore, to capture more comprehensive spatial correlations, we construct an adaptive graph by fusing the static graph with a learnable parameter matrix. Extensive experimental results on two real-world datasets demonstrate that our proposed model significantly outperforms other state-of-the-art methods.

Building extraction from remote sensing images (RSIs) requires exploring multiscale boundary detailed information and extracting it completely, which is challenging but indispensable. However, existing solutions tend to augment feature information solely through multiscale fusion and apply attention mechanisms to focus on feature relationships within a single layer while ignoring the multiscale information, which affects segmentation results. Therefore, enhancing the capability of the network to adaptively capture multiscale information and capture the global relationship of features remains a pivotal challenge in overcoming the aforementioned hurdles. To address the preceding challenge, we propose a Multiscale Direction Context-aware network with Global Attention (MDCGA-Net), employing a classic encoder–decoder architecture enhanced with direction information and global attention flow. Specifically, in the encoder part, the multiscale layer is used to extract contextual information from the interlayer. In addition, the multiscale direction context-aware module is adopted to adaptively acquire multiscale information. In the decoder part, we propose a global attention gate module to capture discriminative features. Furthermore, we construct an operation of attention feature flow to obtain the global relationship among the different features with long-range dependencies, which guarantees the integrity of results. Finally, we have performed comprehensive experiments on three public datasets to showcase the efficacy and efficiency of MDCGA-Net in building extraction.

Graph representation learning has become a mainstream method for processing network structured data, and most graph representation learning methods rely heavily on labeling information for downstream tasks. Since labeled information is rare in the real world, adopting self-supervised learning to solve the graph neural network problem is a significant challenge. Currently, existing graph neural network approaches attempt to maximize mutual information for self-supervised learning, which leads to a large amount of redundant information in the graph representation and thus affects the performance of downstream tasks. Therefore, the self-supervised graph information bottleneck (SGIB) proposed in this paper uses the symmetry and asymmetry of graphs to establish comparative learning and introduces the information bottleneck theory as a loss training model. This model extracts the common features of both views and the independent features of each view by maximizing the mutual information estimation between the local high-level representation of one view and the global summary vector of the other view. It also removes redundant information not relevant to the target task by minimizing the mutual information between the local high-level representations of the two views. Based on the extensive experimental results of three public datasets and two large-scale datasets, it has been shown that the SGIB model can learn higher quality node representations and that several classical network analysis experiments such as node classification and node clustering can be improved compared to existing models in an unsupervised environment. In addition, an in-depth network experiment is designed for in-depth analysis, and the results show that the SGIB model can also alleviate the over-smoothing problem to a certain extent. Therefore, we can infer from different network analysis experiments that it would be an effective improvement of the performance of downstream tasks through introducing information bottleneck theory to remove redundant information.

Road extraction from remote sensing images (RSIs) is a key research area in smart city development. While deep learning techniques have demonstrated remarkable effectiveness in this domain, existing approaches exhibit limitations: convolutional neural network (CNN)-based methods struggle to capture global contextual information for long-range road networks, vision transformer (ViT)-based methods fail to adequately extract multi-scale local features, and hybrid CNN-ViT architectures overlook the synergistic guidance between local and global features. To address these challenges, we propose LG-Umer, a UNet-like network that integrates Local-Global features with a novel attention mechanism, combining the complementary strengths of CNNs and ViTs within an encoder-decoder framework. Specifically, the encoder employs a Multi-scale Strip Deformational (MSD) module, which utilizes deformable convolutions to adaptively extract topological structures and variable-shaped local road features. In the decoder, a Multi-stage Gate Unit (MGU) module is introduced, incorporating a novel attention mechanism to model long-range dependencies by leveraging local features as attention operators for global feature refinement. Extensive experiments on three public benchmarks demonstrate the superiority of LG-Umer. It achieves IoU scores of 70.4%, 71.2% and 68.7% on the Massachusetts Road, DeepGlobe Road, and CHN6-CUG datasets, respectively, surpassing recent state-ofthe-art (SOTA) methods by 1.2%, 0.9%, and 1.1%. These results validate the effectiveness of our approach in balancing local detail preservation and global contextual modeling for road extraction tasks.

Efficient and accurate wind power forecasting is beneficial for the rational distribution and stable supply of electricity. Existing deep learning-based methods are generally focused on extracting effective spatiotemporal features from historical wind power data. However, the current prediction accuracy is often constrained by three aspects: temporal feature extraction, spatial feature extraction, and spatiotemporal fusion scheme. Specifically, the temporal feature extractor fails to efficiently capture long-term temporal dependencies, the spatial feature extractor does not account for comprehensive spatial relationships among wind turbines, and there is a lack of effective spatiotemporal fusion strategies. In this work, we design a spatiotemporal fusion architecture based on Mamba and graph convolutional network, which further enhances the accuracy of wind power forecasting. Our model, which is based on an encoder–decoder framework, primarily consists of three components: temporal block, spatial block, and spatiotemporal fusion block. Specifically, a temporal block is capable of efficiently extracting long-term temporal dependencies from wind power data. Spatial block takes into account the inherent and dynamic spatial relationships among wind turbines. The spatiotemporal fusion block employs a progressive fusion strategy that effectively integrates spatiotemporal features. Multi-step predictions are conducted on a Spatial Dynamic Wind Power Forecasting dataset and a private dataset, and experimental results show that our model outperforms other models in predictive performance.