Microplastic pollution in riverine ecosystems poses critical environmental challenges, yet current modeling approaches inadequately capture the spatial heterogeneity and topological complexity of fluvial systems. This study develops an innovative spatiotemporal graph convolutional network (ST-GCN) framework that integrates hydrological connectivity, flow parameters, and microplastic characteristics for simultaneous migration pathway identification and pollution source tracing. This model constructs multi-scale graph representations encoding system structure and transport dynamics, implements spatial-temporal convolution layers with adaptive attention mechanisms, and employs a backpropagation-based algorithm for inverse source identification. Validation using 18 months of field observations from 45 monitoring nodes across a 127 km river reach demonstrates 87.3% pathway prediction accuracy and 94.3% source localization accuracy (R² = 0.841, p < 0.001), representing substantial improvements over conventional advection–diffusion models. The framework successfully identified three pollution sources during a real contamination incident within 6 h of detection, enabling rapid regulatory intervention. This research advances environmental modeling by demonstrating that graph neural networks effectively capture transport processes in networked hydrological systems, providing practical tools for watershed management and evidence-based pollution control decision-making.

Microplastic pollution in riverine systems poses critical environmental challenges, yet predictive modeling remains constrained by data scarcity and the computational limitations of traditional numerical approaches. This study develops a physics-informed neural network (PINN) framework that integrates advection–diffusion equations and turbulence modeling approaches with deep learning architectures to stimulate three-dimensional microplastic transport dynamics. The methodology embeds governing partial differential equations as soft constraints, enabling predictions under sparse observational conditions (requiring approximately three times fewer observation points than conventional numerical models), while maintaining physical consistency. Applied to a representative 15 km Yangtze River reach with 12 months of monitoring data, the model achieves improved performance with a root mean square error of 0.82 particles/m³ and a Nash–Sutcliffe efficiency exceeding 0.88, representing a 34% accuracy improvement over conventional finite volume methods. The framework successfully captures size-dependent transport behavior, identifies three primary accumulation hotspots exhibiting 3–5 times elevated concentrations, and quantifies nonlinear flux–discharge relationships with 6–8-fold amplification during high-flow events. This physics-constrained approach provides practical findings for pollution management and establishes an adaptable computational framework for environmental transport modeling in data-limited scenarios across diverse riverine systems.

The growing share of intermittent renewable electricity has increased the need for long-duration storage in industrial energy systems. Meanwhile, many industrial processes still release recoverable low-grade waste heat. Introducing this heat into pumped thermal energy storage (PTES) can improve thermal integration, but industrial waste heat is often unsteady, and its temperature and mass flow fluctuations may disturb the charging process. This study investigates an air-based reverse Brayton PTES system assisted by an industrial hot-water waste heat stream of approximately 100 °C. A dynamic model was developed in Simulink/Simscape. The shaft speed is fixed at 3000 rpm, and a PID controller regulates the molten-salt flow rate to maintain the thermal storage temperature. The results show that increasing the waste heat temperature from 95 °C to 105 °C mainly changes the charging-side heat distribution. The waste heat utilization power increases from 36.0 MW to 37.9 MW, while the regenerator power decreases from 126.8 MW to 122.0 MW. The thermal storage power increases slightly from 117.0 MW to 119.0 MW, with the mechanical input fixed at 81.0 MW. The influence of waste heat temperature is concentrated near the low-temperature heat exchanger, regenerator, and turbine outlet. Under dynamic disturbances, faster temperature ramps increase short-term deviations, but the PID-based molten-salt flow regulation keeps the storage temperature close to 550 °C, indicating that the proposed control strategy can suppress moderate thermal disturbances during charging. When waste heat temperature and mass flow rate vary together, same-direction changes strengthen the disturbance, whereas opposite-direction changes partly offset it. These results clarify the disturbance propagation mechanism of fluctuating industrial waste heat in the PTES charging loop and provide a basis for the dynamic design and temperature-control strategy of waste-heat-assisted PTES systems.

Growing demand for composite talent in intelligent construction has laid bare persistent gaps between traditional simulation-based teaching and the integrative competencies industry now expects. This paper presents the design, implementation, and empirical evaluation of a large language model (LLM)-driven virtual simulation teaching platform operating within an industry–education fusion framework. The platform adopts a four-layer microservice architecture integrating a domain-adapted LLM inference engine with retrieval-augmented generation and low-rank adaptation fine-tuning, a building information modeling (BIM)-driven three-dimensional scene assembler, a real-time physics simulation backend, and Bayesian knowledge tracing for adaptive difficulty calibration. Seven simulation modules covering representative intelligent construction workflows were developed and populated with authentic enterprise project data. A controlled experiment with 128 undergraduate participants over eight weeks compared the proposed platform against a conventional virtual simulation system. Results showed that the LLM-driven platform produced large effect sizes across knowledge mastery (Cohen’s d = 1.08), practical skill proficiency (d = 1.04), and integrative competency (d = 0.99), with the most pronounced advantages in open-ended tasks requiring cross-trade coordination and professional judgment. An analytic hierarchy process combined with fuzzy comprehensive evaluation (AHP-FCE) registered a 21.6% improvement in overall industry–education fusion effectiveness. Lower-performing students benefited disproportionately, with gains 58.2% larger than those of higher-performing peers. Taken together, these results indicate that embedding LLMs into virtual simulation environments can meaningfully advance adaptive engineering education; single-institution sampling, an eight-week horizon, and a residual hallucination risk temper the conclusions and mark clear priorities for follow-up work.