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.
Object detection in autonomous driving scenes is one of the important research directions in computer vision. The researches focus on ensuring real-time and accurate object detection of objects by autonomous vehicles. Recently， a rapid development in deep learning technology had been witnessed， and its wide application in the field of autonomous driving had prompted substantial progress in this field. An analysis was conducted on the research status of object detection by YOLO （You Only Look Once） algorithms in the field of autonomous driving from the following four aspects. Firstly， the ideas and improvement methods of the single-stage YOLO series of detection algorithms were summarized， and the advantages and disadvantages of the YOLO series of algorithms were analyzed. Secondly， the YOLO algorithm-based object detection applications in autonomous driving scenes were introduced， the research status and applications for the detection and recognition of traffic vehicles， pedestrians， and traffic signals were expounded and summarized respectively. Additionally， the commonly used evaluation indicators in object detection， as well as the object detection datasets and automatic driving scene datasets， were summarized. Lastly， the problems and future development directions of object detection were discussed.
Decoding motor imagery EEG （ElectroEncephaloGraphy） signal is one of the crucial techniques for building Brain Computer Interface （BCI） system. Due to EEG signal’s high cost of acquisition， large inter-subject discrepancy， and characteristics of strong time variability and low signal-to-noise ratio， constructing cross-subject pattern recognition methods become the key problem of such study. To solve the existing problem， a cross-subject dynamic multi-domain adversarial learning method was proposed. Firstly， the covariance matrix alignment method was used to align the given EEG samples. Then， a global discriminator was adapted for marginal distribution of different domains， and multiple class-wise local discriminators were adapted to conditional distribution for each class. The self-adaptive adversarial factor for multi-domain discriminator was automatically learned during training iterations. Based on dynamic multi-domain adversarial learning strategy， the Dynamic Multi-Domain Adversarial Network （DMDAN） model could learn deep features with generalization ability between cross-subject domains. Experimental results on public BCI Competition IV 2A and 2B datasets show that， DMDAN model improves the ability of learning domain-invariant features， achieving 1.80 and 2.52 percentage points higher average classification accuracy on dataset 2A and dataset 2B compared with the existing adversarial learning method Deep Representation Domain Adaptation （DRDA）. It can be seen that DMDAN model improves the decoding performance of cross-subject motor imagery EEG signals， and has generalization ability on different datasets.
To improve the accuracy of watershed runoff volume prediction， and considering the lack of model transparency and physical interpretability of data-driven hydrological model， a new runoff forecast model named Graph Attention neTwork and Dual-stage Attention mechanism-based Long Short-Term Memory network （GAT-DALSTM） was proposed. Firstly， based on the hydrological data of watershed stations， graph neural network was introduced to extract the topology of watershed stations and generate the feature vectors. Secondly， according to the characteristics of hydrological time series data， a runoff forecast model based on dual-stage attention mechanism was established to predict the watershed runoff volume， and the reliability and transparency of the proposed model were verified by the model evaluation method based on attention coefficient heat map. On the Tunxi watershed dataset， the proposed model was compared with Graph Convolution Neural network （GCN） and Long Short-Term Memory network （LSTM） under each prediction step. Experimental results show that， the Nash-Sutcliffe efficiency coefficient of the proposed model is increased by 3.7% and 4.9% on average respectively， which verifies the accuracy of GAT-DALSTM runoff forecast model. By analyzing the heat map of attention coefficient from the perspectives of hydrology and application， the reliability and practicability of the proposed model were verified. The proposed model can provide technical support for improving the prediction accuracy and model transparency of watershed runoff volume.
When the Unmanned Aerial Vehicle （UAV） cluster attacks ground targets， it will be divided into two formations： a strike UAV cluster that attacks the targets and a auxiliary UAV cluster that pins down the enemy. When auxiliary UAVs choose the action strategy of aggressive attack or saving strength， the mission scenario is similar to a public goods game where the benefits to the cooperator are less than those to the betrayer. Based on this， a decision method for cooperative combat of UAV clusters based on deep reinforcement learning was proposed. First， by building a public goods game based UAV cluster combat model， the interest conflict problem between individual and group in cooperation of intelligent UAV clusters was simulated. Then， Muti-Agent Deep Deterministic Policy Gradient （MADDPG） algorithm was used to solve the most reasonable combat decision of the auxiliary UAV cluster to achieve cluster victory with minimum loss cost. Training and experiments were performed under conditions of different numbers of UAV. The results show that compared to the training effects of two algorithms — IDQN （Independent Deep Q-Network） and ID3QN （Imitative Dueling Double Deep Q-Network）， the proposed algorithm has the best convergence， its winning rate can reach 100% with four auxiliary UAVs， and it also significantly outperforms the comparison algorithms with other UAV numbers.
Accurate traffic flow prediction is very important in helping traffic management departments to take effective traffic control and guidance measures and travelers to plan routes reasonably. Aiming at the problem that the traditional deep learning models do not fully consider the spatial-temporal characteristics of traffic data， a CNN-LSTM prediction model based on attention mechanism， namely STCAL （Spatial-Temporal Convolutional Attention-LSTM network）， was established under the theoretical frameworks of Convolutional Neural Network （CNN） and Long Short-Term Memory （LSTM） unit and with the combination of the spatial-temporal characteristics of urban traffic flow. Firstly， the fine-grained grid division method was used to construct the spatial-temporal matrix of traffic flow. Secondly， CNN model was used as a spatial component to extract the spatial characteristics of urban traffic flow in different periods. Finally， the LSTM model based on attention mechanism was used as a dynamic time component to capture the temporal characteristics and trend variability of traffic flow， and the prediction of traffic flow was realized. Experimental results show that compared with Gated Recurrent Unit （GRU） and Spatio-Temporal Residual Network （ST-ResNet）， STCAL model has the Root Mean Square Error （RMSE） index reduced by 17.15% and 7.37% respectively， the Mean Absolute Error （MAE） index reduced by 22.75% and 9.14% respectively， and the coefficient of determination （R2） index increased by 11.27% and 2.37% respectively. At the same time， it is found that the proposed model has the prediction effect on weekdays with high regularity higher than that on weekends， and has the best prediction effect of morning peak on weekdays， showing that it can provide a basis for short-term urban regional traffic flow change monitoring.
Single Long Short-Term Memory （LSTM） network cannot effectively extract key information and cannot accurately fit data distribution in trajectory prediction. In order to solve the problems， a short-term trajectory prediction model of aircraft based on attention mechanism and Generative Adversarial Network （GAN） was proposed. Firstly， different weights were assigned to the trajectory by introducing attention mechanism， so that the influence of important features in the trajectory was able to be improved. Secondly， the trajectory sequence features were extracted by using LSTM， and the convergence net was used to gather all aircraft features within the time step. Finally， the characteristic of GAN optimizing continuously in adversarial game was used to optimize the model in order to improve the model accuracy. Compared with Social Generative Adversarial Network （SGAN）， the proposed model has the Average Displacement Error （ADE）， Final Displacement Error （FDE） and Maximum Displacement Error （MDE） reduced by 20.0%， 20.4% and 18.3% respectively on the dataset during climb phase. Experimental results show that the proposed model can predict future trajectories more accurately.
