Sparse Matrix-Vector multiplication （SpMV） is an important numerical linear algebraic operation. The existing optimizations for SpMV suffer from issues such as incomplete consideration of preprocessing and communication time， lack of universality in storage structures. To address these issues， an adaptive optimization scheme for SpMV on heterogeneous platforms was proposed. In the proposed scheme， the Pearson correlation coefficients were utilized to determine highly correlated feature parameters， and two Gradient Boosting Decision Tree （GBDT） based algorithms eXtreme Gradient Boosting （XGBoost） and Light Gradient Boosting Machine （LightGBM） were employed to train prediction models to determine the optimal storage format for a certain sparse matrix. The use of grid searches to identify better model hyperparameters for model training resulted in both of those algorithms achieving more than 85% accuracy in selecting a more suitable storage structure. Furthermore， for sparse matrices with the HYBrid （HYB） storage format， the ELLPACK （ELL） and COOrdinate （COO） storage format parts in these metrices were computed on the GPU and CPU separately， establishing a CPU+GPU parallel hybrid computing mode. At the same time， hardware platforms were also selected for sparse matrices with small data sizes to improve computational speed. Experimental results demonstrate that the adaptive computing optimization achieves an average speedup of 1.4 compared to the Compressed Sparse Row （CSR） storage format in cuSPARSE library， and average speedup of 2.1 and 2.6 compared to the HYB and ELL storage formats， respectively.

Implementing diagonal Sparse Matrix Vector multiplication （SpMV） on Graphics Processing Unit （GPU） can make full use of the parallel computing capabilities of GPU and accelerate matrix vector multiplication. However， related mainstream algorithms have problems such as a large amount of zero-element filling data and low computational efficiency. In response to the above problems， a diagonal SpMV algorithm DIA-Dynamic （Diagonal-Dynamic） was proposed. Firstly， a new dynamic partition strategy was designed to divide the matrix into blocks according to different characteristics， which greatly reduced the zero-element filling while ensuring high computational efficiency of GPU， thereby removing redundant calculations. Then， a diagonal sparse matrix storage format BDIA （Block DIAgonal） was proposed to store block data， and the data layout was adjusted to improve memory access performance on GPU. Finally， based on the bottom of GPU， the conditional branch optimization was performed to reduce branch judgments， and dynamic shared memory was used to solve the problem of irregular access of vectors. Compared with the state-of-the-art Tile SpMV algorithm， DIA-Dynamic has the average acceleration ratio of 1.88； compared with the cutting-edge BRCSD （Diagonal Compressed Storage based on Row-Blocks）-Ⅱ algorithm， DIA-Dynamic has the average zero-element filling reduced by 43%， and the average acceleration ratio reaches 1.70. Experimental results show that DIA-Dynamic can effectively improve the computational efficiency of diagonal SpMV on GPU， shorten the computing time， improving the program performance.

With the advancement of technologies such as sensor networks and global positioning systems， the volume of meteorological data with both temporal and spatial characteristics has exploded， and the research on deep learning models for Spatiotemporal Sequence Forecasting （STSF） has developed rapidly. However， the traditional machine learning methods applied to weather forecasting for a long time have unsatisfactory effects in extracting the temporal correlations and spatial dependences of data， while the deep learning methods can extract features automatically through artificial neural networks to improve the accuracy of weather forecasting effectively， and have a very good effect in encoding long-term spatial information modeling. At the same time， the deep learning models driven by observational data and Numerical Weather Prediction （NWP） models based on physical theories are combined to build hybrid models with higher prediction accuracy and longer prediction time. Based on these， the application analysis and research progress of deep learning in the field of weather forecasting were reviewed. Firstly， the deep learning problems in the field of weather forecasting and the classical deep learning problems were compared and studied from three aspects： data format， problem model and evaluation metrics. Then， the development history and application status of deep learning in the field of weather forecasting were looked back， and the latest progress in combining deep learning technologies with NWP was summarized and analyzed. Finally， the future development directions and research focuses were prospected to provide a certain reference for future deep learning research in the field of weather forecasting.

Compared with ordinary networks， hypernetworks have complex tuple relationships， namely hyperedges. However， most existing network representation learning methods cannot capture the tuple relationships. To solve the above problem， a Heterogeneous hypernetwork Representation learning method with Hyperedge Constraint （HRHC） was proposed. Firstly， a method combining clique extension and star extension was introduced to transform the heterogeneous hypernetwork into the heterogeneous network. Then， the meta-path walk method that was aware of semantic relevance among the nodes was introduced to capture the semantic relationships among the heterogeneous nodes. Finally， the tuple relationships among the nodes were captured by means of the hyperedge constraint to obtain high-quality node representation vectors. Experimental results on three real-world datasets show that， for the link prediction task， the proposed method obtaines good results on drug， GPS and MovieLens datasets. For the hypernetwork reconstruction task， when the hyperedge reconstruction ratio is more than 0.6， the ACCuracy （ACC） of the proposed method is better than the suboptimal method Hyper2vec（biased 2nd order random walks in Hyper-networks）， and the average ACC of the proposed method outperforms the suboptimal method， that is heterogeneous hypernetwork representation learning method with hyperedge constraint based on incidence graph （HRHC-incidence graph） by 15.6 percentage points on GPS dataset.

The task of spatio-temporal sequence prediction has a wide range of applications in the fields such as transportation， meteorology and smart city. It is necessary to learn the spatio-temporal characteristics of different data with the combination of external factors such as precipitation and temperature when making station wind speed predictions， which is one of the main tasks in meteorological forecasting. The irregular distribution of meteorological stations and the inherent intermittency of the wind itself bring the challenge of achieving wind speed prediction with high accuracy. In order to consider the influence of multi-site spatial distribution on wind speed to obtain accurate and reliable prediction results， a Graph-based Dynamic Switch-Attention Network （Graph-DSAN） wind speed prediction model was proposed. Firstly， the distances between different sites were used to reconstruct the connection of them. Secondly， the process of local sampling was used to model adjacency matrices of different sampling sizes to achieve the aggregation and transmission of the information between neighbor nodes during the graph convolution process. Thirdly， the results of the graph convolution processed by Spatio-Temporal Position Encoding （STPE） were fed into the Dynamic Attention Encoder （DAE） and Switch-Attention Decoder （SAD） for dynamic attention computation to extract the spatio-temporal correlations. Finally， a multi-step prediction was formed by using autoregression. In experiments on wind speed prediction on 15 sites data in New York State， the designed model was compared with ConvLSTM， Graph Multi-Attention Network （GMAN）， Spatio-Temporal Graph Convolutional Network （STGCN）， Dynamic Switch-Attention Network （DSAN） and Spatial-Temporal Dynamic Network （STDN）. The results show that the Root Mean Square Error （RMSE） of 12 h prediction of Graph-DSAN model is reduced by 28.2%， 6.9%， 27.7%， 14.4% and 8.9% respectively， verifying the accuracy of Graph-DSAN in wind speed prediction.