Yongqian Liu
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Research Interests
Yongqian Liu has a total of 181 co-authors. Here are the top co-authors by number of shared publications:
- Jie Yan
State Key Lab of Alternate Electrical Power System with Renewable Energy Sources, School of New Energy, North China Electric Power University, Beijing, China
24 shared publications Shuang Han
18 shared publications- Shuang Han
North China Electric Power University
17 shared publications - Han Wang
State Key Lab of Alternate Electrical Power System with Renewable Energy Sources, School of New Energy, North China Electric Power University, Beijing, China
7 shared publications - Li Li
School of New Energy, North China Electric Power University, Beijing 102206, China
6 shared publications - Shuang Han
School of New Energy, North China Electric Power University, Beijing 102206, China
6 shared publications
Latest Publications
- 14 October 2026
- Published by Elsevier in Renewable Energy
- Vol. 278
Medium-term wind power forecasting based on the strategy of “independent encoding - joint decoding” considering multi-source data
- 14 October 2026
- Published by Elsevier in Renewable Energy
- Vol. 278
Medium-term wind power forecasting based on the strategy of “independent encoding - joint decoding” considering multi-source data
- 12 October 2026
- Published by Elsevier in Renewable Energy
- Vol. 278
A physics-guided dual-channel convolutional neural network wake model for a wind turbine over a hill
- 12 October 2026
- Published by Elsevier in Renewable Energy
- Vol. 278
A physics-guided dual-channel convolutional neural network wake model for a wind turbine over a hill
- Open Access
- 18 September 2026
- Published by Elsevier in SSRN Electronic Journal
Dynamic Adaptive Multimodal Early Fault Diagnosis Model of Offshore Wind Turbine Bearings via Energy-Decoupled Learning and Bilateral Acoustic–Vibration Coordination
Offshore wind turbine bearings are susceptible to degradation, while restricted offshore accessibility can delay maintenance and increase failure-related costs, highlighting the need for early fault detection and diagnosis. However, early bearing faults produce weak and stage-dependent acoustic and vibration responses, limiting the generalization of models trained on developed faults. To address these issues, a dynamic adaptive multimodal framework based on energy-decoupled learning and acoustic-vibration coordination, DAMEC-Net, is proposed. First, a Dynamic Adaptive Rectangular Convolution Network dynamically predicts convolution kernel sizes according to local structures at different spatial locations in vibration time-frequency images, enhancing the extraction of multiscale weak fault features. Second, an Energy-Decoupled Learning Constraint introduces energy attenuation perturbations and enforces feature consistency across energy levels, promoting evolution-robust intrinsic representations. Finally, a Bilateral Complementary Feature Coordination module performs bidirectional interaction between global acoustic and vibration semantics with adaptive fusion, exploiting cross-modal complementarity while mitigating semantic shift. Trained on faulty samples and tested on developing samples, DAMEC-Net achieves accuracies of 96.13% and 100% on two public bearing datasets. Ablation and visualization results further verify the effectiveness of the proposed modules in improving early fault diagnosis and cross-stage generalization, supporting DAMEC-Net as a candidate algorithmic module for multimodal offshore drivetrain condition monitoring.
- 18 September 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessDynamic Adaptive Multimodal Early Fault Diagnosis Model of Offshore Wind Turbine Bearings via Energy-Decoupled Learning and Bilateral Acoustic–Vibration Coordination
Offshore wind turbine bearings are susceptible to degradation, while restricted offshore accessibility can delay maintenance and increase failure-related costs, highlighting the need for early fault detection and diagnosis. However, early bearing faults produce weak and stage-dependent acoustic and vibration responses, limiting the generalization of models trained on developed faults. To address these issues, a dynamic adaptive multimodal framework based on energy-decoupled learning and acoustic-vibration coordination, DAMEC-Net, is proposed. First, a Dynamic Adaptive Rectangular Convolution Network dynamically predicts convolution kernel sizes according to local structures at different spatial locations in vibration time-frequency images, enhancing the extraction of multiscale weak fault features. Second, an Energy-Decoupled Learning Constraint introduces energy attenuation perturbations and enforces feature consistency across energy levels, promoting evolution-robust intrinsic representations. Finally, a Bilateral Complementary Feature Coordination module performs bidirectional interaction between global acoustic and vibration semantics with adaptive fusion, exploiting cross-modal complementarity while mitigating semantic shift. Trained on faulty samples and tested on developing samples, DAMEC-Net achieves accuracies of 96.13% and 100% on two public bearing datasets. Ablation and visualization results further verify the effectiveness of the proposed modules in improving early fault diagnosis and cross-stage generalization, supporting DAMEC-Net as a candidate algorithmic module for multimodal offshore drivetrain condition monitoring.
- Open Access
- 31 July 2026
- Published by Elsevier in SSRN Electronic Journal
Ring-Based Feature-Level Knowledge Distillation for Federated Probabilistic Forecasting across Heterogeneous Wind Farms
As wind power transitions from a supplementary to a major source of electricity supply, weather driven power fluctuations pose growing risks to power-system balancing and dispatch. Probabilistic forecasting has therefore become essential for quantifying these operational risks. Collaborative modeling across wind farms can draw on complementary information from multiple sites to improve uncertainty characterization. Under data privacy constraints, however, collaboration generally relies on conventional federated learning methods that favor a single global distribution. Such methods may therefore fail to retain important site-specific characteristics. To address this limitation, a personalized federated framework for probabilistic wind power forecasting, termed RFKD-PFL (Ring-based Feature-level Knowledge Distillation for Personalized Federated Learning), is proposed. First, a ring based feature-level knowledge distillation mechanism replaces conventional round-wise parameter averaging. Transferable shared knowledge is extracted through progressive feature transfer among wind farms, reducing the loss of local information associated with parameter aggregation. Second, an adaptive distillation strategy accounts for differences in wind power output distributions across sites. The strategy dynamically adjusts the influence of shared knowledge on each local model, improving personalized learning across heterogeneous wind farms. Finally, a joint non-crossing constraint is introduced for multiple quantiles. A quantile-crossing penalty is combined with sorting based correction to maintain properly ordered interval bounds and improve the coherence of predicted probability distributions. Experiments were conducted on real-world data from eight wind farms in Gansu Province. Compared with conventional federated learning methods based on parameter aggregation, RFKD-PFL reduced normalized root mean square error (NRMSE) by 14.05–22.11%. It also reduced average quantile loss (Average QL) by 19.51–28.20%. These results show that RFKD-PFL produces more accurate and coherent probabilistic forecasts across heterogeneous wind farms.
- 31 July 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessRing-Based Feature-Level Knowledge Distillation for Federated Probabilistic Forecasting across Heterogeneous Wind Farms
As wind power transitions from a supplementary to a major source of electricity supply, weather driven power fluctuations pose growing risks to power-system balancing and dispatch. Probabilistic forecasting has therefore become essential for quantifying these operational risks. Collaborative modeling across wind farms can draw on complementary information from multiple sites to improve uncertainty characterization. Under data privacy constraints, however, collaboration generally relies on conventional federated learning methods that favor a single global distribution. Such methods may therefore fail to retain important site-specific characteristics. To address this limitation, a personalized federated framework for probabilistic wind power forecasting, termed RFKD-PFL (Ring-based Feature-level Knowledge Distillation for Personalized Federated Learning), is proposed. First, a ring based feature-level knowledge distillation mechanism replaces conventional round-wise parameter averaging. Transferable shared knowledge is extracted through progressive feature transfer among wind farms, reducing the loss of local information associated with parameter aggregation. Second, an adaptive distillation strategy accounts for differences in wind power output distributions across sites. The strategy dynamically adjusts the influence of shared knowledge on each local model, improving personalized learning across heterogeneous wind farms. Finally, a joint non-crossing constraint is introduced for multiple quantiles. A quantile-crossing penalty is combined with sorting based correction to maintain properly ordered interval bounds and improve the coherence of predicted probability distributions. Experiments were conducted on real-world data from eight wind farms in Gansu Province. Compared with conventional federated learning methods based on parameter aggregation, RFKD-PFL reduced normalized root mean square error (NRMSE) by 14.05–22.11%. It also reduced average quantile loss (Average QL) by 19.51–28.20%. These results show that RFKD-PFL produces more accurate and coherent probabilistic forecasts across heterogeneous wind farms.
- Open Access
- 21 July 2026
- Published by Elsevier in SSRN Electronic Journal
Dynamic Adaptive Multimodal Early Fault Diagnosis of Wind Turbine Bearings via Energy-Decoupled Learning and Bilateral Acoustic–Vibration Coordination
Early bearing fault diagnosis is vital to prevent minor defects from becoming severe failures, ensure wind turbine safety and reduce operation and maintenance costs. However, the weak impulse energy of early faults makes fault features easily submerged by noise, and their time-frequency structure changes significantly with fault developing, posing challenges for accurate early diagnosis. To address these issues, a dynamic adaptive multimodal framework based on energy-decoupled learning and acoustic–vibration coordination, DAMEC-Net, is proposed. First, a Dynamic Adaptive Rectangular Convolution Network dynamically predicts rectangular convolution kernel sizes according to local structures at different spatial locations in vibration time–frequency images, enhancing the extraction of multi-scale weak fault features at early stage. Second, an Energy-Decoupled Learning Constraint (EDLC) applies energy attenuation perturbations and enforces feature consistency across energy levels, reducing the interference of signal energy variations and guiding the model to learn intrinsic representations robust to fault evolution. Finally, a Bilateral Complementary Feature Coordination module exploits cross-modal complementarity through bidirectional interaction between the global semantics of the two modalities and adaptive gated fusion, while alleviating semantic shift. When trained on mid- and late-stage fault samples and tested on early-stage samples, the proposed method achieves accuracies of 96.13% and 100% on two public bearing datasets, respectively. Ablation experiments and feature visualization analyses confirm the contributions and synergy of the core modules, demonstrating that the proposed method can improve both the diagnostic performance and generalization capability for weak early bearing faults by mitigating evolution-induced feature distribution shift.
- 21 July 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessDynamic Adaptive Multimodal Early Fault Diagnosis of Wind Turbine Bearings via Energy-Decoupled Learning and Bilateral Acoustic–Vibration Coordination
Early bearing fault diagnosis is vital to prevent minor defects from becoming severe failures, ensure wind turbine safety and reduce operation and maintenance costs. However, the weak impulse energy of early faults makes fault features easily submerged by noise, and their time-frequency structure changes significantly with fault developing, posing challenges for accurate early diagnosis. To address these issues, a dynamic adaptive multimodal framework based on energy-decoupled learning and acoustic–vibration coordination, DAMEC-Net, is proposed. First, a Dynamic Adaptive Rectangular Convolution Network dynamically predicts rectangular convolution kernel sizes according to local structures at different spatial locations in vibration time–frequency images, enhancing the extraction of multi-scale weak fault features at early stage. Second, an Energy-Decoupled Learning Constraint (EDLC) applies energy attenuation perturbations and enforces feature consistency across energy levels, reducing the interference of signal energy variations and guiding the model to learn intrinsic representations robust to fault evolution. Finally, a Bilateral Complementary Feature Coordination module exploits cross-modal complementarity through bidirectional interaction between the global semantics of the two modalities and adaptive gated fusion, while alleviating semantic shift. When trained on mid- and late-stage fault samples and tested on early-stage samples, the proposed method achieves accuracies of 96.13% and 100% on two public bearing datasets, respectively. Ablation experiments and feature visualization analyses confirm the contributions and synergy of the core modules, demonstrating that the proposed method can improve both the diagnostic performance and generalization capability for weak early bearing faults by mitigating evolution-induced feature distribution shift.
- Open Access
- 21 July 2026
- Published by Elsevier in SSRN Electronic Journal
Weather-Regime-Aware Cross-Frequency Joint Probabilistic Forecasting of Wind--Solar Power for Supply-Target Attainment
High penetration of wind and photovoltaic (PV) generation makes it difficult to determine whether aggregate renewable output will attain prespecified supply targets under uncertain weather. Existing probabilistic forecasters commonly model source-specific marginal distributions and use NWP variables as core predictors of future power. However, NWP information is less often transformed into explicit weather regimes that condition internal feature routing. This study presents WCMA-Former, an end-to-end framework that predicts the conditional joint distribution of wind and PV power and maps it to supply-target attainment probabilities (STAP). Kumaraswamy marginals linked by a Frank Copula represent bounded outputs and wind--PV dependence. The cross-frequency adaptive decomposition and interaction (CF-ADI) module exchanges information across short-term, intraday, and longer-horizon branches, while an NWP-driven mixture-of-experts encoder--decoder routes representations according to four latent weather regimes. Evaluation on 15-min operational data from 30 wind farms and 30 PV stations in a Chinese provincial grid considered targets equal to 10-50% of installed wind--PV capacity. WCMA-Former achieved a mean Brier score of 0.0571 and probability estimation accuracy of 87.46%, exceeding the baselines by 1.18--8.96 percentage points.
- 21 July 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessWeather-Regime-Aware Cross-Frequency Joint Probabilistic Forecasting of Wind--Solar Power for Supply-Target Attainment
High penetration of wind and photovoltaic (PV) generation makes it difficult to determine whether aggregate renewable output will attain prespecified supply targets under uncertain weather. Existing probabilistic forecasters commonly model source-specific marginal distributions and use NWP variables as core predictors of future power. However, NWP information is less often transformed into explicit weather regimes that condition internal feature routing. This study presents WCMA-Former, an end-to-end framework that predicts the conditional joint distribution of wind and PV power and maps it to supply-target attainment probabilities (STAP). Kumaraswamy marginals linked by a Frank Copula represent bounded outputs and wind--PV dependence. The cross-frequency adaptive decomposition and interaction (CF-ADI) module exchanges information across short-term, intraday, and longer-horizon branches, while an NWP-driven mixture-of-experts encoder--decoder routes representations according to four latent weather regimes. Evaluation on 15-min operational data from 30 wind farms and 30 PV stations in a Chinese provincial grid considered targets equal to 10-50% of installed wind--PV capacity. WCMA-Former achieved a mean Brier score of 0.0571 and probability estimation accuracy of 87.46%, exceeding the baselines by 1.18--8.96 percentage points.
- 1 July 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Sustainable Energy
- Vol. 17 (3) , 2016-2027
A Novel Federated Learning Framework With Adaptive Parameter Updater for Wind Power Forecasting
- 1 July 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Sustainable Energy
- Vol. 17 (3) , 2016-2027
A Novel Federated Learning Framework With Adaptive Parameter Updater for Wind Power Forecasting
- Open Access
- 12 June 2026
- Published by Elsevier in SSRN Electronic Journal
Dynamic Wake Model for Wind Farm Based on Flow Field Decomposition and Physics-Guided Neural Networks
Fast prediction of dynamic wakes in wind farms is an important basis for operational control. Existing wind farm wake models often have difficulty describing the instantaneous characteristics of flow fields. To address this problem, this paper proposes a dynamic wake model based on flow field decomposition and physics-guided neural networks. The model decomposes the flow field into the background flow field, the turbine wake velocity deficit, and the interference from upstream turbines. Based on Taylor’s frozen hypothesis and an analytical wake model, a feature transformation module is constructed. This module converts time-series inflow, thrust coefficient, and yaw angle into two-dimensional flow field features with clear physical meanings. On this basis, a background branch, a wake branch, and a position-aware flow field refinement network are established. They are used to describe the dynamic evolution of the background flow field, extract dynamic wake velocity deficit features, and represent the influence of upstream turbines, respectively. Furthermore, fast prediction of the dynamic flow field in the wind farm is achieved through single-turbine region prediction and multi-turbine region stitching. Compared with high-fidelity large-eddy simulation results, the mean absolute percentage errors of the predicted flow fields are 5.17% and 6.29% under changes in operating conditions and wind farm scale, respectively. The results show that the proposed model can accurately and rapidly predict dynamic flow fields in multi-turbine wind farms. It also has a certain degree of scenario transferability.
- 12 June 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessDynamic Wake Model for Wind Farm Based on Flow Field Decomposition and Physics-Guided Neural Networks
Fast prediction of dynamic wakes in wind farms is an important basis for operational control. Existing wind farm wake models often have difficulty describing the instantaneous characteristics of flow fields. To address this problem, this paper proposes a dynamic wake model based on flow field decomposition and physics-guided neural networks. The model decomposes the flow field into the background flow field, the turbine wake velocity deficit, and the interference from upstream turbines. Based on Taylor’s frozen hypothesis and an analytical wake model, a feature transformation module is constructed. This module converts time-series inflow, thrust coefficient, and yaw angle into two-dimensional flow field features with clear physical meanings. On this basis, a background branch, a wake branch, and a position-aware flow field refinement network are established. They are used to describe the dynamic evolution of the background flow field, extract dynamic wake velocity deficit features, and represent the influence of upstream turbines, respectively. Furthermore, fast prediction of the dynamic flow field in the wind farm is achieved through single-turbine region prediction and multi-turbine region stitching. Compared with high-fidelity large-eddy simulation results, the mean absolute percentage errors of the predicted flow fields are 5.17% and 6.29% under changes in operating conditions and wind farm scale, respectively. The results show that the proposed model can accurately and rapidly predict dynamic flow fields in multi-turbine wind farms. It also has a certain degree of scenario transferability.
- 1 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Industrial Informatics
- Vol. 22 (7) , 5963-5974
An Efficient Wind Farm Condition Monitoring Method Balancing Complexity and Performance
The predictive maintenance based on condition monitoring (CM) and early fault warning is critical for ensuring the safe operation of wind turbines and reducing wind farm maintenance costs. However, current CM and fault detection methods focus on individual wind turbines rather than the entire wind farm. Furthermore, the growing complexity of models leads to higher computational requirements, which complicates edge deployment. To effectively address the challenges of the tradeoff between model scale and monitoring performance, we propose a clustering and lightweight wind turbine CM method. First, a clustering method incorporating temporal features and data correlations is proposed to group wind turbines and identify central representatives. Second, a lightweight neural circuit policy-driven spiking neural network is developed. The normal behavior models are trained on the central wind turbines and generalized across wind turbines in the same cluster. Finally, we quantitatively compare the proposed method with three baseline models using both healthy and faulty wind turbines to validate its excellent performance in CM at a wind farm comprising 80 wind turbines. The results demonstrate that the proposed method achieves CM and early fault detection for the entire wind farm using only three models. Additionally, the proposed method maintains remarkable stability and accuracy, despite being one-tenth the size or smaller compared to baseline models. The average root mean square error decreases by 9.68% during the healthy operation. The average warning time is 35.31 h before the failure occurrence while minimizing false alarms.
- 1 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Industrial Informatics
- Vol. 22 (7) , 5963-5974
An Efficient Wind Farm Condition Monitoring Method Balancing Complexity and Performance
The predictive maintenance based on condition monitoring (CM) and early fault warning is critical for ensuring the safe operation of wind turbines and reducing wind farm maintenance costs. However, current CM and fault detection methods focus on individual wind turbines rather than the entire wind farm. Furthermore, the growing complexity of models leads to higher computational requirements, which complicates edge deployment. To effectively address the challenges of the tradeoff between model scale and monitoring performance, we propose a clustering and lightweight wind turbine CM method. First, a clustering method incorporating temporal features and data correlations is proposed to group wind turbines and identify central representatives. Second, a lightweight neural circuit policy-driven spiking neural network is developed. The normal behavior models are trained on the central wind turbines and generalized across wind turbines in the same cluster. Finally, we quantitatively compare the proposed method with three baseline models using both healthy and faulty wind turbines to validate its excellent performance in CM at a wind farm comprising 80 wind turbines. The results demonstrate that the proposed method achieves CM and early fault detection for the entire wind farm using only three models. Additionally, the proposed method maintains remarkable stability and accuracy, despite being one-tenth the size or smaller compared to baseline models. The average root mean square error decreases by 9.68% during the healthy operation. The average warning time is 35.31 h before the failure occurrence while minimizing false alarms.
- Open Access
- 23 March 2026
- Published by Elsevier in SSRN Electronic Journal
A High-Fidelity Wake Model for Wind Turbines in Complex Terrain via Feature Transformation and Data-Driven Fusion
With the rapid development of wind energy technology, complex terrain has become an important scenario for wind farm deployment. Under complex terrain conditions, wind turbine wake structures exhibit stronger nonlinear characteristics. Existing wake models therefore struggle to balance prediction accuracy and computational efficiency. To address this issue, this study proposes a high-fidelity wake prediction model for wind turbines in complex terrain based on feature transformation and data-driven fusion. First, a high-fidelity wake dataset covering different terrain slopes and operating conditions is constructed using large-eddy simulation combined with the immersed boundary method. A feature transformation module is then developed to convert inflow wind profile, thrust coefficient, and terrain information into a velocity deficit field and a terrain-induced background flow field with clear physical meaning. This process enhances the representation of key flow features. On this basis, a data-driven module is established to extract and fuse features from the two intermediate flow fields. As a result, fast and accurate prediction of wind turbine wakes in complex terrain can be achieved. In the generalization tests, the proposed model yields a mean absolute percentage error of only 4.20% compared with large eddy simulation results. In addition, compared with the BPA model coupled with the Coanda effect, the proposed model improves prediction accuracy by 83.00%.
- 23 March 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessA High-Fidelity Wake Model for Wind Turbines in Complex Terrain via Feature Transformation and Data-Driven Fusion
With the rapid development of wind energy technology, complex terrain has become an important scenario for wind farm deployment. Under complex terrain conditions, wind turbine wake structures exhibit stronger nonlinear characteristics. Existing wake models therefore struggle to balance prediction accuracy and computational efficiency. To address this issue, this study proposes a high-fidelity wake prediction model for wind turbines in complex terrain based on feature transformation and data-driven fusion. First, a high-fidelity wake dataset covering different terrain slopes and operating conditions is constructed using large-eddy simulation combined with the immersed boundary method. A feature transformation module is then developed to convert inflow wind profile, thrust coefficient, and terrain information into a velocity deficit field and a terrain-induced background flow field with clear physical meaning. This process enhances the representation of key flow features. On this basis, a data-driven module is established to extract and fuse features from the two intermediate flow fields. As a result, fast and accurate prediction of wind turbine wakes in complex terrain can be achieved. In the generalization tests, the proposed model yields a mean absolute percentage error of only 4.20% compared with large eddy simulation results. In addition, compared with the BPA model coupled with the Coanda effect, the proposed model improves prediction accuracy by 83.00%.