Yaoran Chen

Accurate wind power forecasting is essential for energy management and yields substantial economic benefits for modern industrial power systems. To handle the complex non-stationary nature of wind power, recent advancements increasingly adopt multi-scale and direct forecasting paradigms. However, these methods face two challenges: (1) existing multi-scale models typically use scale-invariant feature fusion, which limits their adaptability to dynamic meteorological conditions like sudden wind ramps; (2) direct forecasting with conventional point-wise losses (e.g., Mean Squared Error) can be structurally misaligned with the conditional joint likelihood when future labels are autocorrelated. To address these challenges, we propose the Scale-Conditioned Structure-Regularized Direct Forecasting (S2DF) framework. S2DF introduces a Scale-Conditioned Aggregation (SCA) mechanism that extracts scale-specific summaries to adaptively fuse features based on instantaneous weather states. Additionally, it incorporates a Structure-Regularized Direct Forecasting (SRDF) objective that aligns predictions with leading orthogonal modes of the training labels, helping preserve dominant temporal structure. Experiments on four real-world datasets show that S2DF consistently improves predictive accuracy and structural consistency across representative multi-scale models.

Accurate and rapid wake prediction is essential for wind farm control and layout optimization, yet deep learning models struggle to process complex flow data on adaptive grids. We investigate sequence-based modeling of time-averaged wake fields, predicting them directly from non-uniform computational fluid dynamics (CFD) data. This formulation supports efficient, high-fidelity surrogate modeling for practical wind farm applications. This setting presents two challenges: (1) Uniform Grid Dependency: conventional pixel-based models require interpolation onto uniform grids, which dilutes information in high-gradient regions; and (2) Discontinuous Prediction: treating time-averaged wakes as static snapshots interrupts spatial causality along the streamwise direction. We propose the Triple Domainal Adaptive Fusion Transformer (Triformer), which directly predicts wakes from non-uniform data while explicitly modeling this causality. Triformer serializes wake fields into ordered sequences, preserving the adaptive resolution of the original data. It then combines three complementary modules: a Temporal Domain module for ordered upstream-to-downstream progression of the time-averaged wake, a Spectral Domain module for global periodic patterns, and a Feature Domain module for local correlations. These modules bridge gaps in spatial causality. Experiments on 840 CFD simulations demonstrate coefficients of determination above 0.99 and a 70% error reduction compared with baseline methods.

Accurately measuring a learner's cognitive state from observable behavioral data is a fundamental challenge in personalized education, a challenge that can be framed as a latent state estimation problem with inherent measurement uncertainty. Conventional student models produce deterministic point estimates, failing to capture the rich, potentially multimodal distribution of possible knowledge states or to quantify the uncertainty associated with each measurement. This paper introduces Generative-MPC (G-MPC), a framework based on conditional diffusion models that infers continuous, rich posterior distributions over knowledge states from sparse interaction data, explicitly quantifying measurement uncertainty through sampling variance. Two optional enhancements, an uncertainty penalty in the planning objective and adaptive weighting of deep knowledge tracing features (G-MPC-DKT), can be activated when needed. A hybrid variant (Hybrid-G-MPC) reduces computational cost via heuristic guidance. In simulation, the MPC planner uses known ground-truth dynamics to isolate cognitive state estimation effects from transition model error; in real-world evaluation, it relies on a transition model learned from student data. Experiments on a simulated environment and the real-world ASSISTments 2012–2013 dataset demonstrate consistent outperformance over state-of-the-art baselines. In simulation, G-MPC-DKT achieves the highest cumulative reward (419.5) and final knowledge level (0.428); a lightweight configuration reduces inference time by 94.3% while retaining 97.7% of the reward. On real-world data, G-MPC-DKT achieves a 24.0% improvement over the strongest baseline. Ablation studies confirm that the generative measurement approach alone accounts for the majority of the performance gain, validating the practical value of explicit uncertainty-aware state estimation in building deployable adaptive measurement systems.

Accurate and rapid wake prediction is crucial for wind farm control and layout optimization, yet current deep learning approaches face limitations in processing complex flow data from adaptive grids. In this work, we investigate sequence-based wake modeling—learning to predict flow fields directly from non-uniform CFD data. We identify two fundamental challenges in this domain: (1) Uniform Grid Dependency, where the rigid input requirements of conventional pixel-based models force the interpolation of adaptive grids, causing information dilution in high-gradient regions; (2) Discontinuous Prediction, where treating wake fields as static snapshots interrupts the propagation of spatial causality along the streamwise direction. To handle these challenges, we propose Triple Domainal Adaptive Fusion Transformer (Triformer), a framework enabling direct wake prediction from non-uniform data while explicitly modeling spatial causality. Specifically, to handle Uniform Grid Dependency, Triformer introduces a spatial serialization strategy that reformulates wake fields as ordered sequences to preserve the adaptive resolution of the original data. To handle Discontinuous Prediction, Triformer integrates a Temporal Domain module to model upstream-to-downstream dependencies, a Spectral Domain module to capture global periodic patterns, and a Feature Domain module to extract local correlations, thereby bridging the gaps in spatial causality. Experiments across 840 CFD simulations validate the utility of Triformer, achieving coefficients of determination exceeding 0.99 and reducing error by 70% compared to baseline methods.

The massive global expansion of wind energy demands extensive land use, increasingly creating conflicts with biodiversity conservation, agriculture, and local communities. Traditional wind farm layout optimization primarily focuses on maximizing power output while overlooking land-use efficiency. To address this tension, this study proposes a multi-objective co-design framework that integrates wake steering to concurrently optimize annual energy production (AEP) and capacity density, implemented via nested and sequential optimization workflows. The framework utilizes the NSGA-II algorithm coupled with the Generalized Serial-Refine (GSR) method, an advanced deterministic solver that optimizes yaw angles for wake steering. One-outer-iteration tests demonstrate that GSR matches the solution quality of the global-search Genetic Algorithm while requiring 3.4 times less computational cost, thereby making nested co-design tractable. Moreover, the sequential and nested workflows follow distinct evolutionary trajectories, leading to qualitatively different Pareto frontiers. By decoupling layout from control, the sequential approach converges to static‑optimum layouts at significantly lower computational cost, making it a pragmatic choice for engineering deployment. Nonetheless, the nested strategy remains indispensable for uniquely identifying control-complementary layouts in highly constrained sites. Notably, results from a Hong Kong offshore scenario show that the co-design framework attains a 10.7%-68.2% land-use reduction while preserving over 99.2\% of the maximum achievable AEP. By unlocking the land-sparing potential of wake steering, this study provides a novel paradigm to reconcile the wind capacity expansion with land conservation, thereby facilitating a more sustainable global energy transition.

Accurate and rapid wake prediction is crucial for wind farm control and layout optimization. However, conventional deep learning approaches face two challenges: (1) a structural mismatch between adaptive Computational Fluid Dynamics (CFD) grids and pixel-based convolutional models requiring uniform grid interpolation, causing information dilution in high-gradient regions, and (2) the lack of physical inductive bias to capture the anisotropic spatial causality inherent in wake flows. Unlike graph-based approaches that often treat spatial dependencies isotropically, this study introduces a sequence modeling paradigm that directly processes wake data on non-uniform CFD grids. By reformulating wake fields as spatially-ordered sequences via physics-informed sorting, this approach preserves adaptive resolution while strictly aligning the model with the upstream-to-downstream transport physics and vertical shear structures. We develop the Triple Domainal Adaptive Fusion Transformer (Triformer), extending the Transformer with physics-aware mechanisms: a Spectral Domain module capturing periodic patterns via Fast Fourier Transform (FFT) filtering, a Temporal Domain module modeling spatial evolution via Gated Recurrent Unit (GRU), and a Feature Domain module extracting local dependencies through multi-head attention. Validation using 840 CFD simulations demonstrates Triformer achieves coefficients of determination exceeding 0.99, reducing error by 70% compared to standard Transformers. Extension to multi-turbine scenarios enables accurate power prediction within seconds—versus hours for CFD simulations. The framework offers a generalizable solution for sequential data with strong periodicity and long-range dependencies.