Deep learning offers a new path for tropical cyclone forecasting， yet its black-box characteristic hinders its application in critical fields. Therefore， a review was conducted on the application research of explainable deep learning in tropical cyclone forecasting. First， the latest advances of deep learning research on tropical cyclone forecasting were classified systematically. Second， the advantages and disadvantages of traditional methods of forecasting were reviewed. Third， attention is focused on the applications of Convolutional Neural Network （CNN）， Recurrent Neural Network （RNN）， Long Short-Term Memory （LSTM） Network， and Transformer for forecasting tasks， and the frontier explorations of methods such as Generative Adversarial Network （GAN） were discussed. On this basis， the importance of explainable deep learning in tropical cyclone forecasting was analyzed in a deep way， the specific application of explainable deep learning to tropical cyclone forecasting was explored in three dimensions： feature importance interpretation， spatio-temporal mode interpretation， and uncertainty quantification， and its potential in enhancing model transparency and user confidence was pointed out. Finally， to address current challenges faced by explainable research， future research prospects of combining physical mechanisms， multi-granularity visualization， and causal inference were introduced.

In order to solve the problem that malicious clients in federated learning destroy the reliability of the global model by uploading malicious updates， a poisoning attack detection and defense algorithm for federated learning， FedDyna， was proposed. Firstly， an abnormal client detection scheme was designed to use the historical standard deviation of cosine similarity and Euclidean distance to detect abnormal updates preliminarily， and a multi-view model evaluation mechanism was combined to further detect suspicious clients. Then， an adaptive adjustment strategy was proposed to reduce the participation weights of abnormal clients gradually according to the weight adjustment factor until the malicious updates were removed from the model training process. The defense performance of FedDyna in different attack scenarios was evaluated on the EMNIST and CIFAR-10 datasets， and the algorithm was compared with the existing advanced defense algorithms. Experimental results show that， under the condition of a fixed attack frequency， a comparison of the effectiveness between the FedDyna algorithm and the Scope algorithm is conducted： When faced with three attack types， namely Projected Gradient Descent （PGD）， Model Replacement （MR）， and PGD+MR， FedDyna achieves the best results， reducing the Attack Success Rate （ASR） by 1.07 and 0.53， 1.49 and 1.45， 10.55 and 1.25 percentage points respectively； Under the EMNIST dataset subjected to Cosine Constraint Attack （CCA）， although FedDyna experiences a slight decrease in ASR， it still achieves the second-best results. Additionally， when evaluated against comparison algorithms in different attacker pools， FedDyna’s ASR performs optimally under most conditions and ranks second-best under the remaining conditions. Notably， in scenarios with varying attack intensities， FedDyna achieves an impressive average global Model Accuracy （MA） of up to 98.5%. It can be seen that FedDyna has shown significant robustness against poisoning attacks in different attack scenarios and can detect and eliminate poisoning models effectively.

To address the challenges of slow convergence and low model accuracy in non-independent and identically distributed （Non-IID） scenarios， this paper proposed a Data Augmentation scheme based on conditional Generative Adversarial Network in Federated learning （FDA-GAN）. First， a conditional generator for class selection was designed， adding an independent network module to each class and using the label as conditional information to more accurately extract specific features for each class. Second， a client selection strategy covering all classes was proposed. Based on the comprehensive reward of the clients， a client set containing as many classes as possible was selected for training， ensuring that the Generative Adversarial Network （GAN） could learn the complete class distribution. Finally， generated samples were used to augment the local datasets of the clients， optimizing the feature composition of the local data and reducing bias between clients. Experimental results show that under DIRichlet distributed （DIR） data partitioning， compared to CAP-GAN （Collaborated gAme Parallel learning based on GAN）， FDA-GAN improves the MNIST Score （MNIST inception Score） and Mode Score by 2.67 and 1.08 respectively， and reduces the FID （Fréchet Inception Distance） and MMD （Maximum Mean Discrepancy） scores by 55.12 and 2.56 respectively； in different Non-IID scenarios， the FedAvg （Federated Averaging） and FedProx （Federated Proximal） algorithms， when combined with FDA-GAN， converge within 50 communication rounds， with accuracy improvements of at least 30.36 percentage points. This demonstrates that FDA-GAN can improve the quality and diversity of generated samples， and when combined with baseline algorithms， it can significantly improve the accuracy and convergence speed of the federated model.

Traditional Federated Learning （FL） does not consider collaborative fairness， leading to a mismatch between the reward obtained by the client and its actual contribution. To address this issue， a Federated learning fairness algorithm based on Personalized Submodel and K-means clustering （FedPSK） was proposed. Firstly， the neurons in the neural network were clustered according to their activation patterns， and only the importance of the cluster center neurons after clustering was evaluated. And the scores of the cluster center neurons were used to represent the scores of other neurons in the cluster， which reduced the time consumption of neuron evaluation. Then， the number of neurons and their labeling included in the client submodel were selected through hierarchical selection method， and a submodel with a complete neural network structure was constructed for each client. Finally， collaborative fairness was achieved by distributing submodels to the clients. Experimental results on different datasets show that FedPSK improves the correlation coefficient of fairness measurement by 2.70% compared with FedSAC （Federated learning framework with dynamic Submodel Allocation for Collaborative fairness）. In terms of time overhead， FedPSK reduces by at least 84.12% compared with FedSAC. It can be seen that FedPSK improves the fairness of FL algorithm， and reduces the time overhead of the algorithm execution greatly， verifying the efficiency of the proposed algorithm.