Flight delay prediction is highly complex due to multiple influencing factors, including weather conditions, aircraft types, and airline strategies, with delays in preceding flights potentially causing cascading effects on subsequent flights. This paper proposes a flight delay prediction model based on an improved Spatio-Temporal Synchronous Graph Convolutional Network (STSGCN), incorporating collaborative methods and mechanism design to enhance the model's ability to process and predict using multi-source data. First, flight features are integrated with external features such as weather conditions, aircraft types, and airline codes to construct dynamic graph-structured data, enhancing the comprehensiveness of feature representation. Second, the model introduces a spatio-temporal convolutional network with attention mechanisms, collaboratively capturing complex dependencies in both temporal and spatial dimensions. Finally, flight chain features are integrated to enable a two-stage prediction of flight delays. Experimental results show that, compared with the existing model STSGCN, the proposed model reduces the average error in single flight delay prediction by 6 %, achieving 6.63 minutes, and has significant improvements in delay propagation pattern recognition and accuracy.

The aircraft scheduling process is often significantly affected by limited aircraft resources and the propagation of flight delays, resulting in considerable losses for airlines and passengers. This paper proposes an optimization model based on graph reinforcement learning to minimize aircraft resource utilization and reduce delay propagation to the greatest extent.First, a spatial temporal graph convolutional network is employed to analyze the propagation characteristics of flight delays, and a flight connection network is established based on the flight schedule and the propagation characteristics analyzed of flight delays. Next, the optimization objectives are set to minimize the used aircraft resource and the total propagation delays, with variable weights assigned to these objectives to meet the specific needs of different airlines. Finally, a graph reinforcement learning algorithm is developed for aircraft scheduling, and the graph reinforcement learning algorithm is compared with genetic algorithms and ant colony algorithms. Experimental results demonstrate that the graph reinforcement learning algorithm shows significant advantages both in terms of optimization performance and time efficiency. After optimizing the real-world aircraft scheduling plan, the number of aircraft used decreased by 2.21%, and propagation delays were reduced by 18.04%. The algorithm also achieved favorable results under different weight configurations, demonstrating strong adaptability.

Flight delay prediction is a key issue in the aviation domain, directly influencing airline operating efficiency and passenger pleasure. Current prediction methods still have some restrictions when handling complicated spatiotemporal dependencies and large-scale datasets because they are based mostly on classic statistical methods or deep learning models. A novel Cross-Attention Temporal Aggregation Large Language Model (CATA-LLM) is suggested in order to address these issues. In order to accomplish the best possible flight delay forecasting, this model uses a spatiotemporal cross-attention mechanism in conjunction with sophisticated large pre-trained language models. CATA-LLM efficiently fuses spatial and temporal embeddings by using multi-head cross-attention layers to gradually optimize the model parameters, increasing training stability, and predicting accuracy. Experimental results demonstrate that CATA-LLM significantly surpasses existing methods, reducing prediction errors by 42.9%.

Existing algorithms for the flight string scheduling problem often fall short due to their NP-hard nature and difficulties with local optima. To enhance efficiency, we propose a two-stage deep reinforcement learning column generation optimization algorithm (2-Stage DRLCG). This method lever-ages "offline training and online decision-making" to improve scheduling. In the first stage, we create a flight connection diagram based on airport connections and minimum transfer times. The second stage utilizes a deep Q network (DQN) to optimize column generation through a reduced cost (RC) approach. Experiments using data from a major Chinese airline demonstrate significant improvements in cost reduction and computational efficiency, highlighting the potential of deep reinforcement learning in flight scheduling.

Frequent flight delays have become a significant issue in global air traffic, adversely affecting passenger travel experiences and aviation system operational efficiency. To effectively capture the spatio-temporal dynamics of flight delays, this paper introduces a novel prediction model called the Layer-by-Layer Spatio-Temporal Graph Convolutional Network (LST-GCN). This model incrementally captures spatial-temporal dependencies through hierarchical graph convolution. It integrates temporal modeling with layer-by-layer graph convolutional techniques, utilizing Temporal Convolutional Networks (TCN) combined with the Transformer mechanism to extract temporal features of flight delays. Additionally, it employs Graph Convolutional Networks (GCN) and Graph Attention Mechanisms (GAT) to explore the topological relationships and dynamic influences between airports in an alternating layer-by-layer manner, constructing multiple graph adjacency structures. By incorporating specific node aggregation methods, this approach accurately captures the topological features and dynamic interactions within the airport network. The extracted spatio-temporal features are then fused into a unified high-dimensional representation, which is input into a bidirectional long short-term memory network (BiLSTM) to perform flight delay predictions. This study utilizes the 2023 annual flight punctuality data provided by the Bureau of Transportation Statistics (BTS) as a case study. Compared to the traditional Spatio-Temporal Synchronous Graph Convolutional Network (STSGCN), the proposed method achieves approximately a 3% reduction in error, demonstrating its effectiveness in capturing complex spatio-temporal dependencies.

In order to accurately predict the flight transit time at airports of different sizes, this paper proposes a two-stage flight transit time dynamic prediction model based on ensemble learning and Transformer to jointly and dynamically predict the flight transit time of multiple airports. First, considering the continuity of flight operation, the flight information is associated, and the key features are extracted from the historical information of the flight to construct a multi-airport flight transit data set; secondly, the parameter transfer and connection of the flight transit time are carried out; thirdly, the Bayesian optimization algorithm is introduced to optimize the hyperparameters of the LightGBM and XGBoost models, and the optimal hyperparameters are used to make a preliminary prediction of the flight transit time in the first stage; finally, combined with the real-time delay data and the actual operation of the previous flight, the Transformer model is used to dynamically adjust the initial flight prediction time in the second stage to obtain the final prediction result. The experimental results show that the two-stage prediction results reduce the MAE by 2.42 and 2.49 minutes compared with the one-stage MAE, which proves the effectiveness of this method, and it is better than the random forest, support vector machine and neural network models in terms of prediction error.

This study proposes an Information-guided Pheromone Puzzle Algorithm (IPPA) to address the Traveling Salesman Problem (TSP) with enhanced scalability, convergence speed, and solution quality. The algorithm integrates spectral clustering, dynamic pheromone adjustment, and a sliding window 2-opt local optimization strategy to improve both global search ability and local refinement. To evaluate its performance, extensive experiments are conducted on eight Euclidean TSP benchmark datasets, covering city sizes from 50 to 700. Comparative experiments are performed against a classical Genetic Algorithm (GA), a GA with affinity propagation (AP) for the Clustered Traveling Salesman Problem (CTSP), and five representative Ant Colony Optimization (ACO) algorithms. Results demonstrate that IPPA consistently achieves superior path quality and faster convergence. Across all Euclidean benchmark datasets, IPPA reduces tour length by 0.18% to 3.70% compared to the best performing baseline algorithm on each instance, while maintaining lower computational cost. These results validate the effectiveness and efficiency of IPPA for solving both small-scale and large-scale TSP instances.

As cloud storage systems evolve, "fail-slow" are gaining attention, during which drive I/O threads experience delayed responses and system performance is consistently lower than expected. This paper introduces Patch-sfp, a practical framework for detecting fail-slow in cloud storage systems. It employs a PatchTST-based prediction model, LPM, to anticipate future drive latency trends, and develops a latency threshold design mechanism to accurately identify fail-slow events.Given the diverse loads of the drives, a dynamic chunking strategy has been designed to dynamically chunk data from drives under different working environments, enhancing the ability of LPM to capture dependencies over time. Moreover, a MoH-Attention mechanism is introduced to improve the ability of LPM to learn complex relationships between drive data. Experimental results show that Patch-sfp effectively detects fail-slow in cloud storage systems, outperforming existing fail-slow detection methods in terms of performance, adaptability, and robustness.

In order to make the root cause location model more fully capture the long-term dependencies and complex temporal characteristics of metric data, and to add directed edges in the failure dependency graph to more clearly indicate the dependencies between metrics, a root cause localization model CFMRCL that integrates causal discovery algorithms, Recurrent Neural Networks (RNN), and Variational Autoencoders (VAE) is proposed. Use the causal discovery algorithm to construct a causal failure dependency graph (CFDG) to represent the dependence between metrics, and add directed edges to represent the causal relationship between failures, thereby more accurately reflecting the interconnection and propagation direction of failures. Secondly, for the feature extraction of metric data, RNN is used to better capture and model the long-term dependencies and temporal characteristics in the time series, and VAE is combined to learn the potential distribution of the data, thereby more comprehensively mining important features in metric data, improving the expressive ability and data modeling ability of feature extraction. The collaboration between each module improves the accuracy of CFMRCL in locating root causes. The proposed CFMRCL model is compared with the existing DejaVu model and other baseline models. Experiments on four data sets show that the CFMRCL model has greatly improved accuracy and efficiency in locating the root cause of failures. The CFMRCL model has a maximum accuracy of 97.87% in locating the top five root causes. The accuracy is up to 8% higher than the DejaVu model. The efficiency is on average 16% higher than the DejaVu model and 70% higher than the baseline model.

This paper aims to identify incipient and subtle shifts in multivariate time series (MTS) before the occurrence of anomalies (or failures), as well as to localize a group of monitoring indexes accounting for the impending anomalies. This approach allows operators to take actions in advance to prevent possible anomalies, effectively enhancing the reliability of complicated systems. Specifically, we propose Graph Transformer, an unsupervised prediction model that comprehensively learns the complex patterns of MTS to predict whether and why the monitored object will suffer anomalies in the coming time (prediction windows). This model simultaneously learns the multi-grained temporal synchronous and asynchronous inter-dimension correlations in MTS through Transformer and Graph Attention Networks, and employs a multi-task adversarial training strategy to amplify the difference between normal and abnormal. Additionally, we propose a novel anomaly criterion and an interpretation method, both of which incorporate the reconstruction errors and the global errors in inter-dimension correlations to improve the performance of the proposed model in prediction and interpretation. On five real-world data sets, Graph Transformer achieves excellent prediction and interpretation performance in various prediction windows, and outperforms the state-of-the-art models. Furthermore, our proposed model exhibits strong performance in MTS anomaly detection, surpassing state-of-the-art models in this domain.

Hard drive failures persist in data centers, leading to data loss. To address these failures and enhance storage capacity, new hard drives are introduced to replace the failed ones. Therefore, data centers host hard drives from various suppliers, as well as different models from the same supplier. In the face of new hard drives, traditional models fail to deliver satisfactory prediction results. By estimating the remaining useful life(RUL) of a hard drive, operators can predict when a hard drive will fail and replace it at the optimal time, ensuring maximum utilization and reducing operating costs. Previous variant models based on LSTM and Transformer are slow to train, consume significant resources, and exhibit high RMSE. Therefore, this paper proposes a new MLP-LightGBM model for hard drive failure prediction with transfer learning performance, which not only achieves RUL prediction but also performs RUL prediction on new hard drives. The model consists of Time-processing and LightGBM. We embed the time series features of the hard drives into feature columns through a series of operations, including normalization, flipping, and MLP learning on the hard drive data. Subsequently, we use LightGBM to predict the RUL of the hard drives. We conducted experiments using ten years of datasets provided by the official Backblaze website, and the results indicate that our experimental outcomes are significantly superior to existing RUL prediction methods. For the first time, we apply transfer learning to predict RUL of hard drives and demonstrate that our model achieves accurate results in transfer learning.

Reed-Solomon coding has been widely adopted to protect data storage against failures in storage systems. Recently, proactive fault tolerance is coming to offer an added protection for data storage coping with the increased device failures. Reliability is critical for storage systems. Due to complex system states and fault tolerance patterns, it is very intricate to analyze the reliability of proactive-tolerance Reed-Solomon storage systems. In this paper, a reliability framework is proposed, which combines event-driven simulation with mathematical model to analyze the reliability of proactive-tolerance Reed-Solomon storage systems. Monte Carlo simulation model simulates storage systems operation according to the failure parameters of different subsystems generated from statistical data. A mathematical model is designed to analyze the probability of data loss under certain concurrent failures. The proposed framework models the impact of various device failures (including permanent device failures, transient failures and correlated failures), network bandwidth, and both accurate predictions and false alarms of proactive fault tolerance on the reliability. The proposed reliability framework can be adapted to different Reed-Solomon codes and system configurations, offering versatility in its application and assist system designers in optimizing trade-offs and comparing schemes, which is beneficial for system design and operation.

The uncertainty of ground truth values at the pixel scale obtained via upscaling directly affects the credibility of remote sensing product validation. This article provides an in-depth analysis of the sources of uncertainty in ground truth evapotranspiration (ET) at the pixel scale. This uncertainty is quantitatively evaluated, and the methods for its control are discussed. The results indicate that the uncertainty from upscaling methods is highest, followed by that from auxiliary data, with that from instrument measurements being the smallest. The relative accuracy of the ground truth ET at the pixel scale for the LAS1–LAS4 (LAS5–LAS7) regions is 89.15%–90.16% (81.56%–82.58%). The accuracy for homogeneous surfaces is relatively high at approximately 90%–93%, whereas for moderately and highly heterogeneous surfaces, it is lower, varying from approximately 81% to 92%. To control uncertainty, precise instrument calibration, strategic positioning, the use of diverse constraints, and robust modeling are recommended to increase measurement accuracy and prediction reliability. This uncertainty study includes the analyses of different sources of uncertainty and quantitatively evaluates the uncertainty of ground truth values over different heterogeneous underlying surfaces. The results can be used to objectively evaluate the accuracy of remote sensing products, thus advancing studies of the uncertainty of ground truth values at the pixel scale and enhancing the scientific, reliable, and systematic validation of remote sensing products. This approach can greatly promote the validation of remote sensing ET products over heterogeneous surfaces.

The spatio-temporal distribution of biocrusts can be used to monitor regional water resources in desert ecosystems. However, a lack of biocrust products from remotely sensed images with fine spatial resolution (FSR) limits scientific research in this area. To address this issue, we establish an estimation model for biocrusts (EMBC) in three steps for FSR images and map the large-scale fractional biocrust coverage (FBC) in deserts using Sentinel-2 images with a spatial resolution of 10 m. Firstly, we develop a fraction biocrust cover index (FBCI) based on radiative transfer theory. Next, a multi-angle calculation equation involvingFBCIis established and the parameters for a pixel dichotomy model are solved by an inverse method using a linear kernel-driven model. Finally, this pixel dichotomy model withFBCIis used to calculateFBC. We validate the model using field measurements and compare the validation results with those estimated by a random forest model and a backpropagation neural network model. This comparison demonstrates that the value ofFBCestimated by EMBC is highly consistent with field measurements (root mean square error (RMSE) = 0.0774, systematic deviation = -4.05%). Furthermore, the values of FBC estimated with EMBC and the two other models show a high level of consistency in terms of spatial distribution (RMSEFBCin a desert and is an important technique for monitoring drought in an arid environment.

With the increased use of high-spatial-resolution (HSR) images for vegetation monitoring in arid areas, more details of the low vegetation coverage and interference from the land “background” are captured in the corresponding images. From computational time and accuracy, the multi-angle method (MAM) in the pixel dichotomy model is a potential algorithm to apply in arid areas, but MAM needs the multi-angle vegetation index (VI) as the driver parameters. However, most HSR images are obtained in nadir mode, and the multi-angle information of reflectance is difficult to obtain, which limits the estimation of multi-angle VI from HSR images. To address this issue, this study used a “graphical method” to modify the radiation influence caused by the canopy structure and land “background.” We developed an inversion method of the linear kernel-driven model (KDM) and designed a random sampling method to estimate multi-angle VI from HSR images. Then, we proposed a new pixel dichotomy coupled linear KDM (PDKDM), validated using simulated, field-measured, and reference data. The results showed that the FVC in arid areas estimated by PDKDM was highly consistent with “true” data, with root-mean-square error (RMSE) < 0.062, RMSE < 1.125, and RMSE < 0.027 for comparison with simulated, field-measured and reference data, respectively. PDKDM addressed the issue with the previous MAMs to estimate FVC from HSR images in arid areas. This study provides a useful algorithm with high computational efficiency for producing HSR FVCs in arid areas.

Traditional machine learning methods for water body extraction need complex spectral analysis and feature selection which rely on wealth of prior knowledge. They are time-consuming and hard to satisfy our request for accuracy, automation level and a wide range of application. We present a novel deep learning framework for water body extraction from Landsat imagery considering both its spectral and spatial information. The framework is a hybrid of convolutional neural networks (CNN) and logistic regression (LR) classifier. CNN, one of the deep learning methods, has acquired great achievements on various visual-related tasks. CNN can hierarchically extract deep features from raw images directly, and distill the spectral–spatial regularities of input data, thus improving the classification performance. Experimental results based on three Landsat imagery datasets show that our proposed model achieves better performance than support vector machine (SVM) and artificial neural network (ANN).

Previous research has shown that the near infrared spectrum of leaves of plants starts to change in intravital trees or other vegetation, but there has been little related research on leaves fallen from vegetation and we design a experiment to study this aspect. Because of different water losing rates of leaves, different leaves present different changes. As to the change of the red edge and related parameters, we think the main possible reason is that the difference in life-form characteristics and vitality of species. This provides the basis for applying remote sensing to monitor the biological changes of vegetation stressed by a drought season. Whether other growth seasons will also present similar changes needs to be studied further.