研究成果
Research Achievements

Existing approaches for long-term forecasting of PM_2.5 typically focus either on time-domain or frequency-domain features in isolation, neglecting their complementary interactions. This limitation restricts their capacity to effectively capture long-term trends. Moreover, the absence of explicit modeling of multi-scale correlations among influencing factors under complex environmental conditions may undermine both the stability and accuracy of model predictions. To overcome these limitations, we introduce AirMamba, a novel deep learning framework designed to enhance long-term PM_2.5 forecasting by integrating multi-scale correlation analysis with time-frequency interactions. Specifically, a multi-scale inter-variable correlations extractor module is developed to capture the complex interdependencies among variables across diverse temporal scales. The framework leverages the Maximum Overlap Discrete Wavelet Transform (MODWT) to decompose time series data into multi-scale high-frequency and low-frequency components, thereby facilitating a comprehensive time-frequency analysis. An enhanced bidirectional Mamba structure is then employed to model both long- and short-term dependencies within the time series, informed

Over-smoothing is a persistent challenge in Graph Neural Networks (GNNs), where node embeddings become indistinguishable as network depth increases, fundamentally limiting their effectiveness on tasks requiring fine-grained distinctions. This issue arises from the reliance on diffusion-based propagation mechanisms, which suppress high-frequency information essential for preserving feature diversity. To mitigate this, we propose a wave-driven GNN framework that redefines feature propagation through the wave equation. Unlike diffusion, the wave equation incorporates second-order dynamics, balancing smoothing and oscillatory behavior to retain high-frequency components while ensuring effective information flow. To enhance the stability and convergence of wave equation discretization on graphs, an energy-based mechanism inspired by kinetic and potential energy dynamics is introduced, balancing temporal evolution and structural alignment to stabilize propagation. Extensive experiments on benchmark datasets, including Cora, Citeseer, and PubMed, as well as real-world graphs, demonstrate that the proposed framework achieves state-of-the-art performance, effectively mitigating over-smoothing and enabling deeper, more expressive architectures.

Tropical cyclones (TCs) are highly destructive weather phenomena that cause extensive human and economic losses in affected regions. Accurate prediction of tropical cyclone intensity (TCI) is crucial for disaster preparedness and mitigation. Traditional TCI forecasting methods fail to extract nonlinear features and suffer from high computation costs. In recent years, deep learning methods have been increasingly used to address this challenge. However, current approaches often underutilize meteorological variables and satellite cloud imagery, and fail to capture correlations between multimodal data. In this article, we propose TCIque, a sequence-to-sequence model specifically designed for TCI forecasting. TCIque is designed to integrate multimodal data and retrieve correlational features between them based on the Wide and Deep concept. The "Wide" component leverages domain knowledge to extract statistical features, while the "Deep" component captures nonlinear correlations and spatio-temporal dynamics based on self-attention mechanisms. This unique combination allows the model to fully utilize diverse data sources, such as meteorological variables, satellite imagery, and expert-driven features, ensuring robust feature fusion. Furthermore, a predictive encoder–decoder architecture associated with the self-attention mechanism is employed to address the challenge of long-term dependency decay. Experimental results demonstrate that the TCIque model outperforms existing methods, achieving more accurate performance in TCI prediction by 60.9%, 51.6%, 39.2%, and 1.8% compared to the best performance of baselines, which includes ConvLSTM, PredRNN, TC-Pred, SCSTque, SAF-Net, TCI-Net, Tint, and Pred_3d at 6h, 12h, 18h, and 24h.

The security of Graph Neural Networks (GNNs) is crucial for ensuring the reliability and protection of the systems they are integrated within real-world applications. However, current approaches lack the ability to prevent GNNs from learning high-risk information, including edges, nodes, convolutions, etc. In this paper, we propose a secure GNN learning framework called Reinforcement Learning-based Secure Training Algorithm. We first introduce a model conversion technique that transforms the training process of GNNs into a verifiable Markov Decision Process model. To maintain the security of model we employ Deep Q-Learning algorithm to prevent high-risk information messages. Additionally, to verify whether the strategy derived from Deep Q-Learning algorithm meets safety requirements, we design a model transformation algorithm that converts MDPs into probabilistic verification models, thereby ensuring our method’s security through formal verification tools. The effectiveness and feasibility of our proposed method are demonstrated by achieving a 6.4% improvement in average accuracy on open-source datasets under adversarial attack graphs.

With the growing interest in data-driven methods, Graph Neural Networks (GNNs) have demonstrated strong performance in PM_2.5 forecasting as a deep learning architecture. However, GNN-based methods typically construct the graph based solely on the distance between stations, and few methods introduce geographical factors that significantly affect the spatial dispersion of PM_2.5, leading to performance bottlenecks. Additionally, these methods often fail to process the dynamic wind–field data comprehensively, resulting in inaccurate PM_2.5 dispersion graph construction. These shortcomings greatly limit the interpretability of GNN models in forecasting air pollution. To address these issues, we propose a deep learning method that combines Graph Convolution Network (GCN) with Long Short-Term Memory (LSTM), leveraging geographical information within a dynamic graph. The model captures spatial dependencies between PM_2.5 monitoring stations using a dynamic directional graph derived from the wind–field data and a static graph to represent inherent geographical relationships. The combination of GCN and LSTM enables the extraction of both spatial and temporal correlations. The results of experiments suggest that our proposed model, which offers great interpretability, outperforms state-of-the-art methods, especially in 24, 30, and 36 hours forecasts.

Wind downscaling is crucial for refining coarse-scale wind estimates, improving local-scale predictions, and supporting various applications like risk assessment and planning. Dynamic downscaling models demand extensive computational resources and time, leading to a shift toward more efficient statistical downscaling, whereas it often overlooks inter-variable and inter-station spatial correlations. Addressing this, we propose TerraWind, a deep learning-based downscaling method for complex terrain regions. TerraWind enhances accuracy by incorporating topographic factors and inter-station linkages, capturing wind field interactions with terrain at multiple scales. Experimental results in Eastern China demonstrate that TerraWind reduces wind speed Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) by an average of 42.6% and 33.3%, respectively, compared to three interpolation methods (bicubic, bilinear, and Inverse Distance Weighting). Furthermore, TerraWind achieves an average reduction of 35.3% in wind speed MAE and 25.6% in wind speed RMSE compared to four deep learning models (Wind-Topo, DeepCAMS, RCM-emulator, and Uformer).

Social relationships offer crucial supplementary information for recommendations by leveraging users’ social connections to gain insights into their preferences. However, prevalent social recommendation methods often grapple with the issues of sparsity and noise, which curtail their effectiveness. In addition, these methods overlook the intricacies of user interactions within social networks, which could provide invaluable information. Addressing their deficiencies, this article introduces a novel sociological-theory-based multitopic self-supervised recommendation method (SMSR). This method integrates user attitude information into the construction of social relationships and utilizes dynamic routing to identify and categorize topics, thereby mitigating the impact of social noise on recommendation accuracy. Furthermore, we reveal sophisticated higher order user relations within these topics by using motifs. By combining the light graph convolutional network with balance theory, SMSR efficiently aggregates information from diverse social relations to gain its outstanding performance. Moreover, we have devised and integrated four self-supervised signals, inspired by social theory and derived from heterogeneous graph analysis, to more effectively exploit the rich structural and semantic information inherent in social relationship graphs. Empirical results from extensive experiments on publicly available datasets underscore SMSR’s superiority over the state of the art.

Satellite cloud imagery is pivotal for meteorologists in characterizing weather patterns, detecting climate anomaly regions, and predicting rain effects. The task of satellite cloud image forecasting is crucial, and while deep learning models have shown promise in predicting spatio-temporal data, traditional methods face challenges with extracting long-term spatio-temporal features and high computation costs. To address these issues, we propose the Re-parameterized Sequence-to-Sequence Satellite Cloud Imagery Prediction Network (Rep-SSCIPN). Rep-SSCIPN utilizes Rep-convolution layers to reduce inference-time cost and memory consumption, enhancing efficiency by converting re-parameterized blocks into a single convolution layer during inference. The sequence normalization attention mechanism in Rep-SSCIPN highlights crucial feature sequences and establishes their inter-dependencies. We validate our novel method using a real-world satellite cloud image dataset from the meteorological satellite “Himawari.” Experimental results showcase significant improvements in prediction accuracy and reconstruction quality compared to ConvLSTM, PredRNN, FCLSTM, LMC, SimVP and SCSTque models. The efficiency gains make Rep-SSCIPN a promising advancement for satellite cloud image prediction.

Online payment has become an influential method of transaction. While improving convenience, this method of payment also brings great financial risks. Prevalidating transactions can be effective in reducing the number of frauds. For this reason, recommendation algorithms have been introduced to measure the credibility of transactions by predicting users’ ratings of items. However, most algorithms deal with the relationships in social networks without distinction, mixing positive and negative information into the recommender system, which brings huge noise. And, they only generate a single interest representation for each user to measure the similarity between users and spread interest, ignoring the diversity of user interests. Moreover, they did not consider the propagation of different interests would be different. In this article, we propose a multiinterest and social interest-field framework (MISIF) for social recommendations in financial security, which introduces capsule networks into social recommendation and extends the traditional single-interest representation to user multiinterest embedding by dynamic routing (DR) and other methods to improve the expressiveness of user embedding. After that, we construct social interest fields to integrate social interests based on multiinterest embedding, which alleviates the noise in social networks and user data sparsity problems. Finally, we aggregate user multiinterest embedding and additional information through neural networks to obtain the final prediction scores. Experiments with three publicly available datasets show that our proposed MISIF framework outperforms the state-of-the-art social recommendation methods.