Adaptive spatiotemporal graph learning for multi-horizon probabilistic wind power forecasting Authors: Dong Hua, Geyu Huang, Peiyi Cui, Suisheng Liu, Bojun Deng, Yuanting Feng Published: 2026 ABSTRACT AbstractAccurate and reliable forecasting of wind power generation is a cornerstone for ensuring the operational flexibility, economic efficiency, and grid stability of modern power systems with high renewable penetration. While recent advances in deep learning have improved point and probabilistic forecasting, most existing methods still emphasize error minimization at isolated horizons or locations, neglecting the spatiotemporal dependencies, uncertainty calibration, and temporal coherence that are critical in real-world decision-making contexts. In this study, we propose a novel adaptive spatiotemporal graph neural network (ST-GNN) framework for multi-horizon probabilistic wind power forecasting, designed to jointly optimize deterministic accuracy, distributional sharpness, and temporal stability. The framework integrates site-specific meteorological inputs and SCADA data into a dynamically evolving graph structure, where edges are adaptively reweighted to capture changing inter-site correlations driven by wind climatology. A unified multi-objective loss function is formulated, combining mean error minimization with probabilistic metrics such as the Continuous Ranked Probability Score (CRPS) and a temporal smoothness regularizer that mitigates spurious fluctuations across consecutive horizons. Comprehensive case studies are conducted using two years of high-resolution data from a wind farm cluster consisting of twelve geographically diverse sites. Results demonstrate that the proposed model achieves up to 13% reduction in RMSE at 1–4 hour horizons and maintains improvements above 5% at 24-hour horizons when compared with state-of-the-art baselines. Probabilistic evaluations indicate a reduction in calibration error, with empirical coverage deviations generally within 3% across central quantiles and uncertainty estimates that adapt to different meteorological regimes. Moreover, the temporal rank-order consistency of predictions remains significantly higher than existing models, reducing dispatch-related instabilities and improving multi-site coordination. These findings suggest that the integration of spatiotemporal graph structures with probabilistic training objectives can provide a flexible and potentially generalizable paradigm.Subject terms: Climate sciences, Mathematics and computing Introduction The integration of large-scale wind power into modern power systems has emerged as both a technical opportunity and an operational challenge. As decarbonization targets tighten and renewable portfolio standards expand, wind generation is expected to play a dominant role in meeting global energy demand1,2. However, the inherently intermittent and non-stationary nature of wind resources introduces substantial uncertainty into short- and medium-term power system operations. Accurate forecasting of wind power is therefore a cornerstone of secure, economic, and reliable grid management, influencing a broad spectrum of decision-making processes including unit commitment, reserve scheduling, congestion management, and market bidding3–5. Traditional forecasting approaches, ranging from statistical time series models to deep learning architectures, have delivered steady progress in predictive accuracy. Yet, most existing methods treat each wind farm as an independent forecasting task or assume that spatial correlations can be adequately captured through simple feature concatenation, overlooking the dynamic and heterogeneous nature of spatiotemporal dependencies between geographically distributed sites6,7. In practice, wind farms do not operate in isolation. Their generation patterns are often correlated due to shared meteorological drivers, terrain-induced flow effects, and atmospheric phenomena such as wake propagation or weather front movement. These spatial dependencies evolve over time, influenced by changing wind directions, pressure gradients, and mesoscale climatic conditions. Temporal dependencies, on the other hand, arise from the autocorrelated nature of wind speed and power output, as well as from the persistence of weather regimes. Existing forecasting frameworks generally lack a unified mechanism to model such dynamic spatiotemporal interactions explicitly and adaptively, especially under multi-horizon prediction requirements where both accuracy and reliability need to be sustained across several hours ahead. Without capturing these dependencies, forecast outputs risk being spatially incoherent or temporally unstable, which can lead to inefficient or even risky operational decisions8–10. The objective of this research is to develop a novel forecasting methodology that explicitly models and leverages both the spatial and temporal relationships inherent in multi-site wind power data, while simultaneously providing calibrated probabilistic forecasts. This work proposes an adaptive spatiotemporal graph neural network (ST-GNN) framework capable of learning time-varying inter-farm relationships from heterogeneous data sources, integrating them with temporal sequence modeling to produce multi-horizon predictions. The model is designed to operate on a dynamically evolving graph where nodes represent wind farms and edges encode learned measures of spatial influence that adapt to meteorological regimes. By embedding physical insights into the graph construction process–such as gating edge strengths by wind direction and imposing constraints derived from turbine characteristics–the proposed framework remains both data-driven and physically interpretable. The proposed approach follows a modular pipeline. First, multi-source inputs are encoded, combining historical SCADA records, high-resolution numerical weather predictions, and geospatial attributes. A dynamic adjacency matrix is then computed to reflect the current and anticipated connectivity between sites, capturing both static factors like distance and dynamic factors like directional wind alignment. Graph neural network layers propagate information across this adaptive topology, enabling each site’s forecast to benefit from relevant upstream conditions at other locations. Temporal dependencies are modeled using recurrent neural units that operate on the graph-refined embeddings, ensuring that sequence patterns are preserved and multi-horizon consistency is achieved. The prediction head produces both point forecasts and probabilistic outputs, the latter expressed in quantiles and parametric distributional forms, ensuring that uncertainty is communicated alongside central estimates. The model is trained end-to-end under a unified objective that jointly minimizes deterministic errors, probabilistic scoring rules, and structural regularizers that enforce sparsity, smoothness, and temporal stability in the learned graph. What distinguishes this work from previous efforts is its combination of three elements rarely addressed together: a truly adaptive spatial graph that reconfigures with weather regimes, a temporal learning module that directly consumes these dynamic spatial encodings, and a dual-output probabilistic structure that offers both quantile-based and distributional forecasts. This triad is further strengthened by a set of rigorously defined constraints that inject domain knowledge into the learning process, ensuring that the forecasts adhere to turbine operational limits, quantile ordering, and empirical coverage targets. The resulting model not only improves forecast accuracy in conventional metrics but also delivers spatially coherent and temporally reliable predictions–attributes essential for operational scenarios such as regional reserve allocation or coordinated dispatch. The methodological novelty extends beyond model architecture to the training process, which incorporates a unified loss function blending multiple objectives with physical and statistical regularization terms. This design ensures that improvements in one aspect of performance do not come at the expense of others, yielding a balanced and operationally robust forecasting tool. Additionally, the framework’s flexibility allows it to ingest new data sources, adapt to new wind farm configurations, and be deployed in markets or grid regions with varying meteorological characteristics. In summary, this paper makes four main contributions. First, it introduces an adaptive spatiotemporal graph neural network for wind power forecasting that learns and updates inter-farm dependencies in response to evolving meteorological conditions. Second, it embeds physical constraints and domain knowledge into both the model architecture and training process, ensuring operational feasibility and interpretability. Third, it integrates deterministic and probabilistic forecasting in a single end-to-end framework, providing both central estimates and well-calibrated uncertainty measures for multi-horizon predictions. Fourth, it establishes a comprehensive empirical evaluation framework on real-world multi-site datasets, demonstrating consistent improvements over state-of-the-art baselines in both accuracy and reliability, while maintaining stability in the face of high-variability events. Through this combination of adaptive spatial modeling, temporal reasoning, probabilistic output, and physical realism, the proposed work advances the frontier of wind power forecasting toward methods that are not only more accurate but also more aligned with the operational needs of future renewable-rich power systems. Literature review The forecasting of wind power has evolved through several distinct methodological eras, reflecting advances in data availability, computational capacity, and the growing operational importance of renewable generation. Early studies were dominated by statistical time series models such as autoregressive moving average (ARMA) and its variants, including ARIMA and seasonal ARIMA, which sought to capture autocorrelation structures in wind speed or power output. These approaches often relied on stationarity assumptions and linear relationships, limiting their ability to adapt to the non-linear and non-stationary nature of wind processes11–13. Efforts to extend these models incorporated exogenous variables such as temperature, pressure, and upstream wind measurements, leading to ARIMAX and transfer function frameworks. While these enhancements improved short-term accuracy, the linear modeling structure still constrained the capture of complex interactions between meteorological drivers and turbine performance14,15. The emergence of numerical weather prediction (NWP) models provided a richer set of predictors for wind power forecasting. Integrating NWP outputs with statistical post-processing techniques, such as model output statistics (MOS) or bias correction, became standard practice. Studies demonstrated that blending site-specific measurements with mesoscale weather forecasts improved performance for horizons beyond a few hours. However, such integrations often treated each wind farm independently, failing to account for spatial dependencies between geographically distributed sites. This omission is significant, given that weather systems, especially frontal passages and low-pressure systems, influence multiple wind farms in correlated patterns that evolve over time16–18. Machine learning methods began to reshape the field by offering greater flexibility in capturing non-linear relationships. Support vector regression (SVR), decision trees, and ensemble methods such as random forests and gradient boosting were widely explored. These approaches handled high-dimensional feature sets more effectively than classical statistical methods and showed improved performance when combined with meteorological inputs. Nevertheless, they typically relied on handcrafted features and static learning structures, which did not explicitly model temporal dynamics beyond short lags, nor did they integrate spatial correlations in a principled way19,20. Adaptive and regime-aware forecasting under non-stationarity Recent studies have increasingly recognized that wind power generation exhibits strong non-stationary behavior driven by evolving meteorological regimes, seasonal transitions, and mesoscale atmospheric dynamics. To address this challenge, a growing body of work has focused on adaptive and regime-aware forecasting frameworks that explicitly model time-varying system dynamics. One line of research leverages regime-switching mechanisms, such as hidden Markov models or clustering-based approaches, to partition wind patterns into distinct states and train specialized predictors for each regime. While these methods improve local accuracy, they often rely on predefined regime definitions and lack the flexibility to adapt continuously to evolving conditions. Another direction explores adaptive learning strategies, including online learning, attention mechanisms, and dynamic parameter updating, enabling models to adjust their behavior in response to changing data distributions. In particular, attention-based temporal models and adaptive weighting schemes have demonstrated improved responsiveness to non-stationary inputs, though many of these approaches focus primarily on temporal dynamics without explicitly capturing spatial interdependencies. More recently, regime-aware spatiotemporal models have emerged, integrating graph-based representations with adaptive mechanisms to account for evolving spatial relationships. These models attempt to reflect how wind propagation patterns and inter-site correlations shift under different meteorological conditions. However, existing approaches often employ either static graph structures or partially adaptive mechanisms, and typically lack a unified treatment of spatial adaptation, temporal consistency, and probabilistic uncertainty within a single framework. Compared with these recent developments, the proposed framework advances the state of the art by introducing a fully adaptive spatiotemporal graph structure that evolves continuously with meteorological conditions, while simultaneously incorporating temporal regularization and probabilistic calibration. This integrated design enables the model to capture non-stationary dynamics more effectively, maintain consistency across forecasting horizons, and produce reliable uncertainty estimates, addressing key limitations observed in prior adaptive and regime-aware forecasting methods. The advent of deep learning brought a new wave of architectures tailored to sequential data. Recurrent neural networks (RNNs), and particularly long short-term memory (LSTM) networks and gated recurrent units (GRUs), became popular for their ability to capture long-range temporal dependencies in wind speed and power series. Convolutional neural networks (CNNs) and temporal convolutional networks (TCNs) were also introduced, offering parallelizable and stable alternatives for sequence modeling. Studies comparing LSTM, GRU, and TCN approaches reported that deep architectures could outperform traditional models, particularly when trained on large volumes of high-resolution SCADA and meteorological data21,22. Still, most applications focused on single-site forecasting or treated spatial effects in an ad hoc manner, such as through concatenation of features from nearby sites without explicit modeling of their interdependencies. Spatiotemporal modeling has emerged as a promising direction for wind power forecasting, drawing inspiration from research in traffic flow prediction, climate modeling, and sensor networks. Graph-based approaches, in particular, have gained traction for their ability to represent irregular spatial topologies. Graph convolutional networks (GCNs) extend convolution operations to non-Euclidean domains, enabling the aggregation of information from neighboring nodes in a network. Initial applications in renewable forecasting have used static adjacency matrices based on geographical distance or correlation coefficients, applying GCN layers in combination with temporal models like LSTM to create spatiotemporal hybrid architectures. These models have shown that explicitly encoding the spatial structure can improve forecast coherence and stability across multiple sites. However, static graphs fail to capture the dynamic nature of wind farm relationships, which vary with changing meteorological conditions23,24. Adaptive graph learning represents a more recent evolution, in which the adjacency structure itself becomes a learnable and time-varying component of the model. In the context of wind forecasting, this allows the network to adjust spatial connectivity in response to evolving weather regimes, wind direction shifts, and localized turbulence25,26. Some studies have incorporated attention mechanisms into the graph structure, enabling the model to weight neighbors according to their current relevance rather than fixed proximity measures. While these adaptive mechanisms have demonstrated improvements in capturing spatiotemporal dependencies, many implementations have not integrated physical constraints or domain-specific priors, potentially limiting interpretability and operational acceptance. Probabilistic forecasting has developed in parallel as a means to quantify uncertainty in wind power predictions. Approaches such as quantile regression, Bayesian neural networks, and ensemble learning have been used to generate prediction intervals or full probabilistic distributions. Scoring rules like the continuous ranked probability score (CRPS) and pinball loss have provided rigorous evaluation metrics for these probabilistic outputs. Combining probabilistic forecasting with spatiotemporal models remains relatively underexplored, with most work focusing on either deterministic spatiotemporal architectures or probabilistic models that operate at single sites. Hybrid approaches that integrate physical modeling with machine learning have also gained interest. Coupling outputs from physical models like the Weather Research and Forecasting (WRF) system with data-driven post-processing allows the exploitation of both meteorological theory and statistical learning. Some works have incorporated turbine power curve physics directly into the loss function or network architecture, constraining predictions to feasible ranges. Others have integrated wake modeling or turbulence intensity estimates into feature engineering. These hybrid strategies aim to balance predictive performance with physical realism, a balance that is crucial for operational deployment. Regularization and constraint enforcement in deep learning-based forecasting have been recognized as important for achieving stable and interpretable models. Graph Laplacian regularization, sparsity constraints, and temporal smoothness penalties have been employed to promote coherence and avoid overfitting. In probabilistic settings, quantile crossing penalties and coverage calibration constraints have been used to ensure the validity of uncertainty estimates. However, the systematic integration of such constraints into a unified end-to-end framework, especially one that addresses both spatial and temporal aspects adaptively, has been rare. The literature on multi-horizon forecasting emphasizes the challenge of maintaining accuracy and calibration across increasing lead times. Sequence-to-sequence architectures, multi-output regression, and recursive prediction strategies have been compared, each with trade-offs in error propagation and temporal consistency. Recent work has explored multi-task learning setups in which forecasts for different horizons are learned jointly, sharing intermediate representations to improve generalization. Yet, in the domain of wind power, these strategies have often been implemented without an explicit graph-based spatiotemporal backbone or with only static spatial encodings24,27. In addition, recent advances have explored robust and adaptive forecasting under distribution shifts, which are particularly relevant in renewable energy systems where data distributions evolve due to changing weather regimes and system conditions. Zhao et al.28 proposed a causal self-supervised graph learning framework that enhances robustness against distribution shifts by leveraging invariant causal representations across varying conditions. This approach improves generalization by reducing reliance on spurious correlations in historical data. In a related line of work, Zhao et al.29 introduced a missing data-aware spatiotemporal graph network with a hierarchical downsampling–upsampling structure, which explicitly addresses data incompleteness and enhances robustness in real-world forecasting scenarios. Their method demonstrates improved resilience to both missing data and structural variability in load patterns. Furthermore, Zhao et al.30 developed a spatiotemporal pyramidal graph modeling approach for short-term load forecasting, which captures multi-scale temporal dependencies and hierarchical spatial features. This pyramidal design enables the model to better adapt to complex and evolving demand patterns across different temporal resolutions. While these studies provide important advances in handling distribution shifts, missing data, and multi-scale dynamics, they primarily focus on either robustness or representation learning aspects. In contrast, the proposed framework aims to complement these directions by jointly addressing adaptive spatial dependency learning, temporal consistency across forecasting horizons, and probabilistic uncertainty calibration within a unified model, thereby providing a more comprehensive solution for non-stationary wind power forecasting. In sum, the trajectory of wind power forecasting research reflects a gradual but incomplete shift from independent-site, point-based predictions toward integrated, spatiotemporal, and probabilistic modeling. While advances in deep learning, graph neural networks, and adaptive modeling have laid a foundation for capturing the complex dynamics of wind generation, the combination of dynamically learned spatial structures, temporally aware sequence modeling, and physically constrained probabilistic outputs remains a frontier that has only begun to be addressed in the literature. Mathematical modeling The system structure in Figure 1 illustrates the end-to-end adaptive spatiotemporal graph neural network framework, mapping data acquisition, dynamic graph learning, and temporal modeling into a coherent forecasting pipeline that produces both point and probabilistic wind power predictions. The multi-horizon objective function jointly represents deterministic prediction accuracy, probabilistic consistency, spatial structure regularization, and temporal stability within a graph-aware learning process. The first term combines absolute and squared error components across each node ι, time step τ, and horizon δ, with horizon-aware weights and error-balance parameters . The second term incorporates the probabilistic loss through forecasted quantiles and distributional parameters, including the mean, variance, and skewness. The remaining terms regularize the learned graph topology through the Laplacian , promote sparsity, limit abrupt changes in graph structure, and penalize unrealistic jumps between adjacent forecast horizons. The final term controls overall model complexity. The objective is optimized end-to-end over the model parameters , dynamic graphs , and node embeddings . Adaptive spatiotemporal graph neural network framework for multi-horizon wind power forecasting. To improve clarity and interpretability, the overall objective function can be decomposed into four functional components, each serving a distinct role in the learning process: (i) Deterministic accuracy term (): This component minimizes point-wise prediction errors across all nodes and forecasting horizons, combining absolute and squared loss terms to balance robustness and sensitivity. It ensures that the model achieves high accuracy in predicting central estimates. Deterministic accuracy term (): This component minimizes point-wise prediction errors across all nodes and forecasting horizons, combining absolute and squared loss terms to balance robustness and sensitivity. It ensures that the model achieves high accuracy in predicting central estimates. (ii) Probabilistic consistency term (): This term corresponds to , which integrates CRPS-based distributional matching, quantile regression loss, likelihood-based estimation, and calibration constraints. It enables the model to generate reliable and well-calibrated uncertainty estimates rather than only point forecasts. Probabilistic consistency term (): This term corresponds to , which integrates CRPS-based distributional matching, quantile regression loss, likelihood-based estimation, and calibration constraints. It enables the model to generate reliable and well-calibrated uncertainty estimates rather than only point forecasts. (iii) Spatial structure regularization (–): These terms regulate the learned graph structure. The Laplacian regularization () promotes smoothness of node embeddings over the graph, the sparsity constraint () prevents over-connected structures, and the temporal consistency penalty on adjacency matrices () avoids abrupt changes in spatial relationships across time. Spatial structure regularization (–): These terms regulate the learned graph structure. The Laplacian regularization () promotes smoothness of node embeddings over the graph, the sparsity constraint () prevents over-connected structures, and the temporal consistency penalty on adjacency matrices () avoids abrupt changes in spatial relationships across time. (iv) Temporal stability and model regularization (, ): The temporal smoothness term () penalizes unrealistic fluctuations between consecutive forecasting horizons, ensuring coherent multi-step predictions. The parameter regularization term () controls model complexity and prevents overfitting. Temporal stability and model regularization (, ): The temporal smoothness term () penalizes unrealistic fluctuations between consecutive forecasting horizons, ensuring coherent multi-step predictions. The parameter regularization term () controls model complexity and prevents overfitting. This decomposition highlights that the objective function is designed to jointly balance prediction accuracy, uncertainty reliability, spatial coherence, and temporal consistency, which are all critical for operational wind power forecasting. The hyperparameters associated with each component are selected based on a combination of domain knowledge and validation-based tuning. In practice, the coefficients are first initialized to comparable scales to ensure that no single component dominates the optimization process. Subsequently, a grid search is performed on the validation set to identify a balanced configuration that jointly optimizes RMSE and probabilistic calibration metrics. In particular, and are treated as the primary trade-off parameters between deterministic accuracy and probabilistic quality, while – are set to smaller values to act as structural regularizers rather than dominant objectives. The temporal smoothness coefficient is tuned to prevent excessive oscillations without oversmoothing dynamic changes. Similarly, the parameters – within are calibrated to balance distributional sharpness, calibration, and monotonicity constraints. Empirically, we observe that the model performance is stable across a reasonable range of these hyperparameters, as further demonstrated in the sensitivity analysis presented in the Results section. The probabilistic loss in Equation (2) provides the distributional component of the global objective. The first term is a continuous CRPS formulation that integrates over quantile levels , measuring both forecast error and distributional mismatch. The second term uses weighted quantile pinball penalties with weights and sharpness factor . The third term introduces a parametric negative log-likelihood for the predictive distribution defined by . The fourth term enforces quantile monotonicity by penalizing crossings between levels . The final term encourages empirical reliability through rolling-window coverage constraints that align predicted intervals with observed coverage frequencies. Together, these components support accurate, calibrated, and temporally consistent probabilistic forecasts. Forecast outputs must remain physically realistic–strictly bounded between zero and the turbine’s rated capacity at each site ι. This serves as a hard feasibility envelope across all time and horizons. Quantiles are required to be non-decreasing in level α; this constraint maintains the ordering structure of uncertainty and avoids logical inconsistencies in probabilistic forecasts. To avoid overly wide prediction intervals, this upper bound controls the spacing between any two quantiles, calibrated through historical spread specific to site ι. Abrupt forecast changes violate turbine dynamics. This constraint places a bound on temporal variation between consecutive horizons to enforce smooth multi-step transitions. Over a rolling window of size W, this ensures that empirical coverage of prediction intervals meets the desired nominal levels π, with tolerance \varepsilon safeguarding against finite-sample noise. Each wind farm i is allowed to attend only to a subset of others at time τ, enforcing graph sparsity via the degree budget Γ and promoting interpretability in learned spatial dependencies. Long-distance relationships are truncated when geospatial separation exceeds , reflecting the physical limitations of wind correlation over large distances. This normalizes each row of the adjacency matrix into a stochastic vector–ensuring numerical stability and interpretability of GNN attention weights as convex combinations. Adjacency strength is learned as a function of meteorological alignment , filtered by wind direction threshold Θ; thus encoding upwind–downwind influence using physics-informed gating. This bounds the Frobenius norm difference of adjacency matrices between adjacent time steps, preventing the graph structure from erratically oscillating in response to short-term noise. Once wind speed exceeds the rated threshold , power output saturates. This constraint models the nonlinear plateau behavior characteristic of turbine power curves. Below the cut-in wind speed , turbines do not generate. This enforces zero forecast output under no-wind conditions, anchoring the model to physical ground truth. Methodology All available data sources–historical SCADA , high-resolution weather forecasts , and site-specific geospatial embeddings –are concatenated and passed through a learnable encoder to produce dense feature vectors per node and time. Adjacency strength between sites i and j is adaptively computed based on geodesic distance , directional alignment , and wind field similarity, with learned weights , , and ; all passed through a sigmoid to yield normalized graph connectivity. Each node ι receives messages by aggregating contributions from its neighbors j, with each message transformed via a trainable function , preserving spatial relational context. Spatiotemporal node embeddings are recursively updated over graph layers using residual propagation. Linear transformations and extract meaningful representations as graph signals traverse the network. Once spatial dependencies are encoded, a GRU cell captures temporal evolution across forecasting steps. Hidden states roll forward conditioned on graph-aware embeddings . At each forecasting horizon δ, the point prediction is obtained by projecting GRU states through a regression head . Uncertainty is simultaneously captured by a quantile head , producing predicted quantiles across levels , parameterized by a shared temporal representation. The model also outputs parametric distributional forecasts for skewed or multi-modal behavior using a third head that maps temporal encodings to mean , scale , and skewness . The training objective integrates point-wise and probabilistic losses for each node, horizon, and time step, tuned via scalar weights and , as defined earlier. Finally, model parameters are updated via projected gradient descent, where denotes the projection operator that enforces all constraint sets from the mathematical modeling section. Results The empirical evaluation of the proposed adaptive spatiotemporal graph neural network (ST-GNN) framework is conducted using a comprehensive dataset collected from a real-world wind farm cluster located in an inland–coastal transitional region of eastern China. The cluster comprises 12 utility-scale wind farms, each equipped with between 20 and 36 turbines rated at 2.5 MW, yielding an aggregate installed capacity of approximately 810 MW. Supervisory Control and Data Acquisition (SCADA) data were acquired at a 10-minute resolution over a continuous 24-month period spanning January 2022 to December 2023, resulting in 105,120 timestamped records per site. Each record contains turbine-level and farm-level variables, including active power output (MW), rotor speed (rpm), nacelle wind speed (m/s), wind direction (degrees), and ambient temperature (C). In addition to SCADA data, high-resolution Numerical Weather Prediction (NWP) forecasts from the European Centre for Medium-Range Weather Forecasts (ECMWF) Integrated Forecasting System were retrieved at a 3 km spatial resolution and 1-hour temporal resolution, providing meteorological predictors such as multi-height wind speed (10 m, 80 m, 100 m), air pressure (hPa), relative humidity (%), and temperature (C). Geospatial features, including farm altitude (m), terrain roughness (dimensionless), and relative position vectors between wind farms, were extracted from a digital elevation model with 30 m resolution. Prior to model training, the dataset underwent a rigorous preprocessing pipeline. Missing or corrupted SCADA entries, accounting for approximately 0.8% of total observations, were imputed using a k-nearest neighbor approach constrained by physical plausibility limits derived from turbine specifications. NWP forecasts were temporally aligned to SCADA timestamps via cubic spline interpolation and spatially collocated to each farm centroid. All continuous variables were normalized using z-score scaling computed from the training set, while categorical features, such as wind sector classifications, were one-hot encoded. To assess generalization across varying meteorological regimes, the dataset was split chronologically into 70% for training (January 2022 to August 2023), 15% for validation (September 2023 to November 2023), and 15% for testing (December 2023 to February 2024). The test set was deliberately chosen to include both winter monsoon and transitional season conditions, which are known to exhibit rapid wind variability and directional shifts. The forecasting task targeted lead times of 1 to 24 hours ahead in hourly increments, yielding 24 separate prediction horizons per input timestamp. The computational experiments were executed on a high-performance workstation equipped with two NVIDIA A100 GPUs (40 GB each), dual AMD EPYC 7742 CPUs clocked at 2.25 GHz, and 512 GB of DDR4 RAM, running Ubuntu 22.04 LTS. All models were implemented in Python 3.10, leveraging PyTorch 2.0 for neural network components and DGL 1.1 for graph construction and message passing operations. Data preprocessing and feature engineering pipelines were managed through Pandas 2.1 and NumPy 1.26, with visualization support from Matplotlib 3.8 and Seaborn 0.13. Training was performed using the Adam optimizer with an initial learning rate of , decayed by a factor of 0.7 upon validation loss plateau, and a batch size of 256 time steps per GPU. Early stopping with a patience of 15 epochs was employed to prevent overfitting. On average, the proposed ST-GNN model required approximately 9 hours to converge for the full 24-horizon joint training, while baseline models completed within 3–6 hours. All experiments were repeated over three random initialization seeds to ensure robustness, and results are reported as averages across these runs. Baseline models and benchmark setup To ensure a rigorous and transparent evaluation, the proposed ST-GNN framework is compared against a diverse set of representative baseline models spanning statistical, machine learning, and deep learning paradigms. These baselines are selected to reflect commonly adopted approaches in wind power forecasting and to provide a comprehensive performance benchmark. Specifically, the following models are implemented: ARIMA: A classical autoregressive integrated moving average model, capturing linear temporal dependencies in wind power series. ARIMA: A classical autoregressive integrated moving average model, capturing linear temporal dependencies in wind power series. SVR: Support vector regression with radial basis function (RBF) kernel, representing a traditional machine learning approach for nonlinear regression. SVR: Support vector regression with radial basis function (RBF) kernel, representing a traditional machine learning approach for nonlinear regression. Random Forest (RF): An ensemble learning method based on decision trees, capable of modeling nonlinear relationships with reduced overfitting. Random Forest (RF): An ensemble learning method based on decision trees, capable of modeling nonlinear relationships with reduced overfitting. LSTM: A long short-term memory network designed to capture temporal dependencies in sequential data, widely used in wind forecasting tasks. LSTM: A long short-term memory network designed to capture temporal dependencies in sequential data, widely used in wind forecasting tasks. CNN-LSTM: A hybrid deep learning model combining convolutional layers for feature extraction and LSTM layers for temporal modeling. CNN-LSTM: A hybrid deep learning model combining convolutional layers for feature extraction and LSTM layers for temporal modeling. Static GCN-LSTM: A spatiotemporal model using a graph convolutional network with a fixed adjacency matrix based on geographical distance, followed by LSTM for temporal modeling. Static GCN-LSTM: A spatiotemporal model using a graph convolutional network with a fixed adjacency matrix based on geographical distance, followed by LSTM for temporal modeling. Probabilistic LSTM (Quantile-LSTM): An extension of LSTM trained using quantile loss to generate probabilistic forecasts without explicit spatial modeling. Probabilistic LSTM (Quantile-LSTM): An extension of LSTM trained using quantile loss to generate probabilistic forecasts without explicit spatial modeling. All baseline models are trained using the same input features (SCADA, NWP, and geospatial data where applicable), identical training-validation-test splits, and consistent optimization settings to ensure fairness. Hyperparameters for each model are tuned via grid search on the validation set. For probabilistic baselines, prediction intervals are generated either through quantile regression (for Quantile-LSTM) or residual-based bootstrapping (for deterministic models), ensuring a consistent comparison with the proposed probabilistic framework. Table 1 summarizes the benchmark models and contrasts their modeling type, spatial representation, and probabilistic output capability with the proposed ST-GNN framework. Table 1Summary of baseline models used for comparison.ModelTypeSpatial ModelingProbabilistic OutputARIMAStatistical××SVRMachine Learning××Random ForestMachine Learning××LSTMDeep Learning××CNN-LSTMDeep Learning××GCN-LSTM (Static)Spatiotemporal DL\checkmark (Static)×Quantile-LSTMDeep Learning×\checkmarkProposed ST-GNNSpatiotemporal DL\checkmark (Adaptive)\checkmark Summary of baseline models used for comparison. Figure 2 presents the geographic configuration of the 12 wind farms within the inland–coastal transitional zone of eastern China, situated between latitudes 34.02N and 37.96N and longitudes 118.03E to 121.95E. The minimum geodesic distance between two neighboring farms is 18.4 km, while the farthest inland–coastal pair are separated by 315.7 km. In the displayed network, only site pairs with separations less than 200 km are connected, revealing a dense subgraph among inland farms and sparser long-distance connections toward coastal sites. Elevation across the network ranges from sea level to 286 m, which can influence wind speed distributions through orographic acceleration or sheltering effects. The spatial distribution highlights contrasting exposure: coastal farms are more directly influenced by marine air masses and sea–land breeze circulations, while inland farms are more responsive to continental boundary layer dynamics and topographically induced flow patterns. This figure establishes the physical backbone of the adjacency matrix used in the early stages of the model before adaptive graph learning reshapes it according to meteorological regimes. Spatial distribution and baseline connectivity of the wind farm cluster. Figure 3 visualizes the seasonal wind climatology for each site, constructed from 500 representative directional and speed observations per farm. Several inland farms display a pronounced northwesterly sector dominance (–), contributing over 26% of total wind occurrences and characterized by median speeds between and . In contrast, certain coastal farms exhibit bimodal regimes, with a secondary southeastern sector (–) contributing up to 18% of observations, often associated with monsoon-related wind reversals. Wind speeds above are recorded in multiple sites during winter months, corresponding to strong synoptic pressure gradients. Seasonal persistence is evident: for example, coastal sites consistently register northeast-dominant regimes during DJF, while inland sites display a clockwise seasonal shift, transitioning toward more southerly sectors in JJA. These differences in wind direction distribution are important for characterizing inter-farm correlations and directional dependency structures, both of which directly inform the adaptive reweighting of edges in the proposed spatiotemporal graph. Seasonal wind roses for individual wind farms. Figure 4 examines the reliability of the predictive distributions produced by the ST-GNN model, focusing on its probabilistic calibration and empirical coverage consistency. The left panel shows reliability curves for the predicted quantile coverage (x-axis) versus the empirically observed coverage (y-axis) using 11 evenly spaced quantile bins. The ST-GNN trace adheres closely to the reference line, indicating near-perfect calibration across most quantile levels. Slight over-coverage is visible at extreme ends (e.g., predicted 90% corresponds to observed ), but the overall deviation never exceeds 2–3%. In contrast, the baseline model systematically underestimates uncertainty, with empirical coverage falling 4–6% below predicted values in the central regions and up to 9% near the upper quantiles. This discrepancy suggests that the baseline model suffers from overconfident predictions, failing to adequately represent underlying uncertainty in volatile wind regimes. The right panel tracks coverage error (i.e., absolute deviation between predicted and actual coverage) across all 24 forecast horizons. ST-GNN maintains low and stable coverage errors, with values below 0.035 for h łe 12 and gradually rising to just over 0.05 at h = 24. By comparison, the baseline model shows a much steeper rise in coverage error beyond h = 8, reaching nearly 0.08 at long-range forecasts. This suggests the value of incorporating probabilistic regularizers—such as the Continuous Ranked Probability Score (CRPS) and quantile-consistency constraints—into the objective function formalized in Equation (2). The use of deep spatiotemporal representations, coupled with dynamic graph learning, enables the ST-GNN to better model variance, temporal drift, and nonlinear dependencies, all of which are essential for achieving reliable uncertainty quantification. Taken together, the results in this figure show that the proposed model improves predictive sharpness while maintaining statistically meaningful uncertainty calibration over long forecasting horizons. Calibration and interval coverage performance of probabilistic forecasts. Figure 5 showcases the quality and coherence of the predicted probabilistic distributions generated by the ST-GNN model over three representative sites during a 72-hour operational window. Each panel overlays the forecast median, central 50% interval (between the 25th and 75th quantiles), outer 80% interval (10th to 90th quantiles), and the actual observed wind power series. One of the most salient observations is the absence of any quantile crossings—neither within individual prediction horizons nor across consecutive timesteps—demonstrating that the model’s quantile regression outputs remain ordered and coherent. This behavior reflects the successful implicit enforcement of monotonicity constraints and rank consistency in the training process. In terms of amplitude, the interquantile spread adjusts dynamically based on temporal volatility. During the first 24-hour block, all three sites exhibit relatively narrow uncertainty bands, with central IQR widths ranging from 4.5 to . This is consistent with meteorologically stable conditions, where boundary-layer wind regimes are less variable and easier to model. However, in the subsequent segment (hours 24–48), the IQR bands visibly expand—especially in the third site—reaching widths up to , corresponding to the passage of a fast-moving synoptic front that was identified in the meteorological inputs. Notably, the 10–90% prediction band captures nearly all significant variations in the actual time series: more than 89% of observed values remain within the full 80% interval across the entire window, with only brief excursions at major ramps. Such probabilistic integrity is crucial for downstream applications involving reserve scheduling or market bidding, where credible and well-calibrated uncertainty bounds reduce operational risk. Moreover, the median forecasts themselves track the actual time series with high fidelity across slow ramps and oscillatory transitions, which suggests that the point estimates are not merely central tendencies but structurally informed modes of the underlying conditional distributions. This result underscores the importance of coupling temporal smoothness and spatial awareness in the ST-GNN model, as achieved through the dual loss structure combining CRPS-based quantile regression and graph-informed regularization. Collectively, this figure validates that the model not only avoids logical failures such as quantile crossover, but also learns to express uncertainty in a physically meaningful, regime-adaptive manner. Quantile coherence and spread behavior over 72-hour prediction windows. Figure 6 presents a snapshot of the learned spatiotemporal graph at a representative timestamp during a high-variability period. Nodes correspond to the 12 wind farm sites, plotted in a circular layout, and edges are visualized with width and opacity scaled by their corresponding Wind Flow Influence Index (WFII). WFII represents the normalized attention-weighted strength of directional influence between sites as learned by the dynamic graph mechanism in the ST-GNN architecture. Unlike fixed adjacency matrices based on geographic proximity, the edge strengths here are emergent properties driven by meteorological regimes, cross-site wind correlations, and directional lag effects captured during training. Several systematic patterns are observed. First, the graph exhibits strong directional clustering: edges concentrated in certain sectors (e.g., northeast to southwest) become significantly more prominent, suggesting prevailing directional wind dependencies during the snapshot. In this case, a synoptic northeasterly flow regime dominates, and farms situated in the upstream quadrant show intensified outbound influence values. For instance, three inland farms positioned along the 30–90 arc exhibit high outbound WFII scores–evidenced by thick, opaque connections radiating toward the southern half of the network. This aligns with the physical reality that upwind farms in coherent flow paths carry predictive signal for downwind variability. Second, the graph is not fully connected in practice despite being rendered over a complete topology. Many edge weights fall below the visual threshold (set around 0.1 in normalized scale), meaning that during this regime, the model effectively sparsifies the graph, pruning weak or irrelevant relationships. This adaptive sparsification behavior allows the model to shift between more centralized topologies during convective events and more distributed structures during stable regimes. The interpretability of WFII values supports physically grounded analysis of the learned graph, particularly when tracking how individual sites transition between influencing and receiving roles across weather conditions. Snapshot of dynamic graph structure with edge-weighted wind flow influence index (WFII). Figure 7 illustrates the mean forecast bias across four meteorological regimes: stable, frontal, convective, and transition. The bias is computed as the average difference between the predicted and observed wind power values over all sites and horizons within each regime. In stable conditions, the ST-GNN model shows a minimal bias of approximately -0.12 MW, well within operational tolerance and indicative of strong calibration in low-variability environments. Under frontal regimes, the mean bias increases slightly to +0.85 MW, reflecting the model’s tendency toward overestimation during rapid wind ramps caused by synoptic weather systems. Convective regimes yield an average bias of around -0.48 MW, suggesting occasional under-prediction of short-lived gust events. The transition regime exhibits a bias near +0.22 MW, balancing between slight overestimation and underestimation episodes. The standard deviation bars, ranging from 0.4 MW in stable conditions to 0.7 MW in convective periods, highlight the influence of regime volatility on forecast dispersion. The color coding in deep blue conveys precision and reliability, while the white edge delineation preserves visual clarity. Forecast bias by regime. Figure 8 quantifies the percentage reduction in root mean squared error (RMSE) achieved by the ST-GNN model compared with the best-performing baseline, across forecast horizons from 1 to 24 hours. The improvement is most pronounced at short horizons, where the model achieves reductions of 12.8% at h = 1 and 12.3% at h = 2, reflecting the benefit of fine-grained spatiotemporal feature extraction in near-term prediction. The advantage gradually diminishes with increasing horizon length due to the accumulation of meteorological uncertainty, settling at 5.8% at h = 24. This tapered profile is operationally meaningful because it reveals that the proposed architecture is particularly well-suited for short-term dispatch and real-time operational adjustments, where high precision yields the largest cost savings. The steady decline in improvement without abrupt drops indicates that the multi-horizon training objective maintains stable performance across the evaluated horizon range. It is worth noting that the reported RMSE improvements are measured against the best-performing baseline among the aforementioned models, ensuring that the comparison reflects a strong and competitive benchmark rather than a weak reference. Horizon-wise RMSE reduction. Hyperparameter sensitivity analysis Given the multi-objective structure of the proposed framework, which involves several weighting coefficients (e.g., – and –), it is essential to evaluate the robustness of model performance with respect to these hyperparameters. This subsection presents a systematic sensitivity analysis focusing on the most influential components of the objective function. We focus on three representative groups of hyperparameters: (i) the balance between deterministic and probabilistic learning ( vs. ), (ii) the strength of temporal smoothness regularization (), and (iii) the probabilistic calibration weights within (specifically and , corresponding to CRPS and negative log-likelihood terms, respectively). Each parameter is varied within a predefined range while keeping others fixed at their nominal values, and model performance is evaluated using RMSE and coverage error metrics. The results indicate that the model exhibits stable performance across a reasonably wide range of hyperparameter values. Specifically, increasing (probabilistic loss weight) improves uncertainty calibration but leads to a slight degradation in point prediction accuracy when excessively large, suggesting a trade-off between sharpness and accuracy. Conversely, overly small results in under-dispersed prediction intervals. For the temporal smoothness coefficient , moderate values effectively suppress unrealistic fluctuations across forecasting horizons, while excessively large values introduce over-smoothing, causing delayed response to rapid wind ramps. This suggests the necessity of balancing temporal consistency with responsiveness. Regarding probabilistic parameters, (CRPS weight) plays a dominant role in improving overall distributional quality, while (likelihood term) enhances tail behavior modeling. The interaction between these terms ensures both global calibration and local distributional fidelity. Overall, the proposed framework demonstrates robust behavior with respect to hyperparameter variations, and no abrupt performance degradation is observed within the tested ranges. This indicates that the model does not rely on overly fine-tuned parameter settings, supporting its practical applicability. Conclusion This study proposes a unified adaptive spatiotemporal graph neural network (ST-GNN) framework for multi-horizon probabilistic wind power forecasting. By integrating dynamic graph learning, temporal sequence modeling, and probabilistic prediction within a single end-to-end architecture, the proposed method addresses key limitations of existing approaches, particularly the lack of adaptive spatial modeling, temporal inconsistency across forecasting horizons, and insufficient uncertainty calibration. The main contribution of this work lies in three aspects. First, an adaptive graph construction mechanism is developed to capture time-varying inter-site dependencies driven by meteorological conditions, enabling the model to dynamically adjust spatial relationships rather than relying on static connectivity assumptions. Second, a unified multi-objective learning framework is formulated, jointly optimizing point prediction accuracy, probabilistic reliability, and temporal smoothness, which ensures both numerical precision and operational consistency across multiple horizons. Third, the integration of physics-informed constraints and probabilistic calibration techniques improves the interpretability and practical reliability of the forecasting outputs, making them more suitable for real-world grid operation and decision-making. Extensive experiments conducted on real-world multi-site wind farm data demonstrate that the proposed framework consistently outperforms a range of benchmark models across both deterministic and probabilistic metrics. The results show significant improvements in short-term prediction accuracy, stable performance across extended horizons, and well-calibrated uncertainty estimates, suggesting the effectiveness and robustness of the proposed approach. Despite these promising results, several directions for future research remain. First, the current framework can be extended to incorporate additional exogenous information sources, such as satellite observations or large-scale atmospheric circulation indicators, to further enhance predictive capability under extreme weather conditions. Second, the scalability of the dynamic graph learning mechanism could be explored for larger regional or national wind farm networks with higher spatial complexity. Third, future work may investigate the integration of advanced architectures, such as transformer-based temporal encoders or foundation models, to further improve long-horizon forecasting performance. Finally, the deployment of the proposed model in real-time operational environments, including its interaction with power system dispatch and market mechanisms, remains an important avenue for practical validation. Overall, this work provides a systematic and extensible framework for spatiotemporal probabilistic wind power forecasting, bridging the gap between data-driven learning, physical interpretability, and operational applicability in modern renewable energy systems. Author contributions Dong Hua: Conceptualization, problem formulation, methodology development, spatiotemporal modeling, and manuscript writing (original draft preparation). Geyu Huang: Data preprocessing, feature engineering, integration of SCADA and NWP data, and experimental result analysis. Peiyi Cui: Supervision, project administration, industrial application guidance, validation of modeling assumptions, and critical revision of the manuscript. Suisheng Liu: Case study design, scenario construction, benchmarking experiments, and performance evaluation. Bojun Deng: Algorithm implementation, neural network training, hyperparameter tuning, and visualization of figures and results. Yuanting Feng: Data curation, engineering support, interpretation of practical implications, and assistance in manuscript revision. All authors discussed the results, contributed to the interpretation of the findings, and approved the final version of the manuscript. Funding None. Data availability The datasets generated and/or analyzed during the current study are available from the corresponding author on reasonable request. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement The datasets generated and/or analyzed during the current study are available from the corresponding author on reasonable request.