Knowledge tracing (KT) is a core component of intelligent education systems, which aims to predict students’ future performance from their historical learning behaviors. Despite the remarkable progress brought by deep learning, existing models often struggle to effectively capture the dynamic evolution of students’ knowledge states, particularly the intertwined processes of knowledge-acquisition and -forgetting. To address this challenge, we propose the D ynamic K nowledge P erception for Temporal K nowledge T racing (DKPKT) model. DKPKT integrates two complementary modules: the interaction perception module employs a dynamic attention mechanism to extract key interaction features that signal shifts in knowledge states, and the knowledge perception module refines these representations by incorporating knowledge-forgetting and knowledge-acquisition factors, which enable more accurate modeling of knowledge dynamics over time. We evaluate DKPKT against several mainstream baseline models on three real-world educational datasets. Experimental results demonstrate that DKPKT achieves superior predictive performance, validating its effectiveness in modeling knowledge states and enhancing the accuracy of student performance prediction.

Accurate regional lightning-occurrence prediction is important for operational weather-risk management, but its development is challenged by the severe class imbalance of grid-hour lightning samples. This study proposes a repeated random undersampling (RUS) stacking ensemble that combines heterogeneous machine-learning models and produces probabilistic lightning-occurrence forecasts using atmospheric variables from the European Centre for Medium-Range Weather Forecasts (ECMWF) fifth-generation reanalysis (ERA5), together with spatial and temporal predictors. The effects of the undersampling ratio, ensemble size, and positive-class weighting were systematically evaluated, and a configuration with a RUS ratio of 1:15, 20 ensemble members, and a positive-class weight of 2 was selected. Using an operational threshold of 0.591 selected exclusively on the 2024 validation set, the final model achieved an area under the receiver operating characteristic curve (ROC-AUC) of 0.945, an area under the precision–recall curve (PR-AUC) of 0.235, a probability of detection (POD) of 0.521, and an F1-score of 0.302 on the independent 2025 test set. Physical-variable-group ablation and aggregated TreeSHAP (tree-based Shapley additive explanations) analyses were further conducted to interpret the predictions. Removing the spatiotemporal predictors produced the largest reduction in performance, followed by removing cloud and microphysical variables. Both the Light Gradient Boosting Machine (LightGBM) and categorical boosting (CatBoost) models consistently identified longitude, latitude, total-column cloud ice water, and convective available potential energy as the leading predictors, while higher cloud-ice-water content and stronger convective-instability indices generally shifted model outputs towards lightning occurrence. Direct transfer of the ERA5-trained ensemble to corresponding ECMWF forecast fields without retraining retained useful predictive skill and substantially outperformed the operational ECMWF lightning product over the collocated evaluation samples, although performance degradation indicated a cross-dataset distribution shift. These results demonstrate the value of combining imbalance-aware ensemble learning with physically interpretable predictors for regional lightning-risk forecasting, while the strong influence of geographic variables indicates that external validation and local recalibration or retraining are required before application to other regions.

Lightning poses localized risks to signaling, communication, and traction-power systems along high-speed railways, yet most nowcasting products are generated on regular grids rather than operational line segments. This study developed a grid-to-segment machine-learning framework for next-hour lightning warning along the Jinan West–Tai’an corridor. Ground-based lightning observations, hourly ERA5 fields, temporal variables, and engineered historical lightning features and spatial-neighborhood features were organized on a 0.25° grid. Data from 2014 to 2018 were used for training, 2019 for validation, and 2020 for independent testing. Grid probabilities were converted into warnings for six railway segments using 10 km buffers and maximum-probability aggregation. In 2020, the full extreme gradient-boosting (XGBoost) model, a tree-based ensemble-learning algorithm, achieved grid-level probability of detection (POD), false-alarm ratio (FAR), and critical success index (CSI) values of 0.55, 0.50, and 0.36; segment-level verification yielded 0.60, 0.39, and 0.43. To examine transfer to forecast-driven application, the trained model and threshold were fixed, and ERA5 meteorological inputs were replaced by short-lead ECMWF HRES forecasts for July 2025. POD decreased from 0.85 to 0.80 and CSI from 0.60 to 0.57, while FAR remained nearly unchanged. Under a predefined non-zero rule, HRES litota1 achieved 0.62, 0.71, and 0.25. ML-HRES therefore showed higher CSI and lower FAR than the direct litota1 baseline.