AI-Powered Equivalent Modeling of Mixed Wind Farms

By Dr Silpaja Chandrasekar, PhDReviewed by Susha Cheriyedath, M.Sc.Oct 26 2023

In a paper published in the journal Energies, researchers presented a novel approach for equivalently modeling mixed wind farms (WFs) after considering various wind turbine types, wind conditions, and voltage dips. They analyzed the electromechanical transient performance of mixed WFs, established an equivalent node model with appropriate variables, and used multiple artificial neural networks (ANNs) to capture the complex relationships between these variables. Their approach was tested through dynamic simulations in MATrix LABoratory (MATLAB), showing that it effectively represented the external characteristics of mixed WFs in diverse wind conditions and voltage dip scenarios.

Wind Farm Challenges and AI Advancements

China has seen a surge in large-scale WFs, driven by a transition to cleaner energy sources. However, this shift has introduced challenges in power system stability due to the variable nature of wind power. Modeling these complex systems is hindered by the "curse of dimensionality," and traditional approaches differ for various wind turbine types. While clustering methods have evolved, mixed wind farms pose unaddressed questions about coupling characteristics and the need for new clustering indicators to represent dynamic complexities effectively.

Previous research has advanced artificial intelligence (AI) in power systems, primarily in pattern recognition, prediction, control, optimization, and data classification. However, data limitations, especially in the context of wind farm dynamics, are a challenge. Recent attempts to use machine learning techniques like Long Short-Term Memory Neural Networks (LSTM) and Support Vector Machines (SVM) have faced issues, indicating a need for innovative methods in modeling WFs, particularly in mixed wind farm scenarios.

Advanced Modeling for Mixed WFs

The proposed approach differs from coherence-based WF modeling by treating the entire WF as a black box, leveraging AI to map input-output relationships. It focuses on the input and output characteristics of WFs without delving into their internal topology. This method uses ANNs to approximate the output of nonlinear systems by adjusting connection weights as model parameters. The modeling time depends solely on the ANN and its learning algorithm, independent of the weight type (WT) types or their number. However, a significant challenge lies in balancing the model's accuracy and generalization ability and obtaining training datasets, particularly for small-sample learning.

For this equivalent modeling, inputs like wind speed and voltage dip are crucial for WF dynamic performance. The refined method utilizes individual ANNs to predict output values at different time points, and these predictions are subsequently combined to create the complete dynamic curve. This sequential approach suits small-sample data, mitigating errors from constant input assumptions over extended periods.

Regarding mixed WFs, where various WT types coexist, the focus is on situations where WT parameters and technologies differ. The study examines scenarios with Doubly Fed Induction Generator (DFIG), Permanent Magnet Synchronous Generator (PMSG), and Squirrel-Cage Induction Generator (SCIG) WT types. Each type exhibits unique dynamic responses due to distinct control strategies. Factors influencing the external characteristics of mixed WFs include WT type, wind speed, wind direction, and fault voltage dip. These external characteristics manifest as variable and time-attenuating waveforms. The interplay between multiple WT types results in smoother power fluctuations than individual WTs, enabling the treatment of the mixed WF for identifying its external characteristics.

Equivalent Node Modeling for Mixed Wind Farm

Equivalent Node Model:  Power system dynamic equivalent modeling employs two methods: coherent theory retains the original structure. Estimation equivalent modeling treats it as a black box, offering flexibility for the identical node model in mixed WFs.

This model focuses on power nodes, characterizing external power features with active and reactive power at the Point of Common Coupling (POC) outputs, influenced by factors like WT type, wind conditions, and fault voltage dip.

Experimental Design and Data Collection: The estimation equivalent modeling relies on sample data. It divides wind direction into 16 azimuths and wind speed into 21 segments. This estimation considers four fault voltage dip scenarios. These factors resulted in 1344 input samples. Sampling is variable in fault periods and initial fault removal stages to capture the most dynamic changes.

BP-Based Equivalent Modeling: Researchers actively select the back-propagation (BP) neural network for modeling. It uses 22 input variables, including wind speeds and fault voltage dip, and predicts the active and reactive power of the mixed wind farm. The BP network employs an error correction learning algorithm to improve accuracy.

Model Performance Evaluation

Implementing the proposed equivalent node modeling method in MATLAB, the setup includes a hidden layer comprising 45 neurons, a learning rate 0.05, and an additional momentum factor of 0.7. Evaluation metrics comprise the mean absolute percentage error (MAPE) for active power and the mean absolute error (MAE) for reactive power. Researchers created 20 new testing data sets to assess the model's generalization, calculating active and reactive power errors at different stages of the electromechanical transient period.

Due to solid nonlinearity, the dynamic power errors are relatively large during the early fault and fault recovery phases. However, the overall errors are relatively concentrated, indicating good model stability. The model's performance under the first-ranking test data is within an acceptable range, except for an active power error exceeding 20% at 0.3 seconds. Considering all time points, the equivalent node model's error is 4.884%, representing the mixed WF's external characteristics. The model's dynamic response closely matches the detailed model's fluctuation process, particularly for reactive power.

Conclusion and Future Directions

To sum up, this paper presents a refined equivalent modeling approach for mixed WFs using multiple BP neural networks to capture dynamic power responses. It offers strong applicability, considering factors like WT type, wind conditions, and fault voltage dip. Simulations on an 11th Gen Intel Core i7 computer showed promise. Future research could address the challenges associated with measurement errors and noise in field data by combining small sample learning and filtering techniques to enhance accuracy and practical utility.