TL;DR
This paper introduces a novel robust state estimation (SE) method for active distribution networks (ADNs) that specifically addresses the challenges posed by bad data in both measurements and network topology parameters. By employing a two-pronged approach—a bad measurement data processing model and a robust estimation model—the authors significantly improve SE accuracy and robustness, even under adverse conditions, while also enhancing computational efficiency.
What Problem Does This Paper Address?
Modern active distribution networks (ADNs) are increasingly integrating renewable distributed generators (DGs), leading to more complex network structures and dynamic operational requirements. State estimation (SE) is crucial for understanding the real-time state of these networks (e.g., voltage and current at various nodes), providing essential information for operation and control. However, traditional SE methods struggle with two major issues:
- Bad Measurement Data: Faulty sensors, communication noise, or equipment malfunctions can introduce "bad data" into the measurements. This inaccurate data can significantly degrade SE accuracy and make traditional methods unreliable.
- Bad Topology Parameters: Changes in the ADN topology (e.g., due to switching operations or faults) can lead to outdated or incorrect topology information. This can cause errors in the Jacobian matrix calculations, a core component of many SE algorithms, further impacting accuracy.
These challenges are amplified in ADNs due to the intermittent nature of renewable DGs and the dynamic variations in load, making robust and accurate SE a critical research area.
Methods
The proposed method comprises two main components:
- Bad Measurement Data Processing Model:
- Pseudo-Measurement Generation: This model uses a Gated Recurrent Unit (GRU) neural network to adaptively learn from historical measurement data. It generates high-precision "pseudo-measurements" that replace the identified bad measurement data. This approach leverages the temporal dependencies in the data to create more accurate replacements than traditional linear interpolation or simpler ANN methods.
- Adaptive Learning: The GRU's update and reset gates allow it to dynamically adjust its memory of past states, enabling it to better capture complex, non-linear relationships in the historical data.
- Robust Estimation Model:
- Neural Network-Based Pre-Estimation (DSENN): This component uses a Dual-Output SE Neural Network (DSENN) that does not require topology parameters as input. This is a key innovation, as it directly addresses the problem of inaccurate topology data. The DSENN takes measurement data and outputs estimated voltage magnitudes and phase angles.
- Linear State Estimation: The outputs from the DSENN are then combined with actual micro-PMU (µPMU) measurements and fed into a linear SE algorithm. This two-stage approach (pre-estimation followed by linear estimation) enhances overall SE accuracy and robustness.
The combination of these methods aims to overcome the limitations of traditional SE techniques that are sensitive to both measurement and topology errors.
Datasets
- Load Data: The paper utilizes daily load data from Belgium [33] as a baseline for generating simulation data.
- Simulation Data: Actual power data is calculated using Matpower software [34] based on the load data and network parameters. This data is then used to simulate SCADA and µPMU measurements by adding Gaussian noise.
- Bad Data Simulation: To simulate bad data, the network topology is randomly altered, and 15% of the generated measurements are corrupted by adding noise with a specified mean and standard deviation.
The modified IEEE 33-node system is used as the testbed for simulations.
Key Insights and Ideas
- Decoupling Topology from Pre-Estimation: The DSENN's ability to perform pre-estimation without requiring topology parameters is a significant breakthrough. This directly mitigates the impact of inaccurate topology data, a major weakness in traditional methods.
- Adaptive Pseudo-Measurements: Using GRUs for pseudo-measurement generation allows for more sophisticated learning of historical data patterns compared to simpler interpolation or ANN methods, leading to more accurate replacements for bad data.
- Two-Stage Robustness: The combination of a neural network pre-estimator and a subsequent linear estimator provides a layered defense against various types of data corruption, enhancing overall system robustness.
- Computational Efficiency: By leveraging neural networks for pre-estimation, the proposed method avoids the iterative Jacobian matrix calculations common in traditional methods, leading to significantly faster computation times.
Key Results
- Improved Pseudo-Measurement Accuracy: The proposed GRU-based method achieved lower RMSE and MAE for pseudo-measurements compared to linear interpolation and ANN methods, demonstrating its superior ability to generate accurate replacements for bad data. For instance, the RMSE for pseudo-measurements was 2.35 × 10⁻² and MAE was 7.7 × 10⁻³ for the proposed method, outperforming others.
- Enhanced SE Accuracy: The overall SE accuracy, measured by MAE, was significantly improved. With the proposed method, MAEs were as low as 4.68 × 10⁻² (voltage magnitude) and 5.49 × 10⁻³ (voltage phase angle), outperforming existing methods.
- Robustness to Bad Data: The proposed method showed significantly lower SE errors compared to WLS and WLAV methods, especially as the proportion of bad data increased. This indicates strong robustness against both measurement and topology errors.
- Computational Efficiency: The average estimation time for the proposed method was drastically reduced to 0.14 ms, compared to 2.437 ms for WLS and 603.99 ms for WLAV, highlighting its computational advantage.
Limitations and Future Work
- Limitations:
- The paper acknowledges that while the DSENN does not require topology input for pre-estimation, the subsequent linear estimation stage might still be affected by bad topology parameters. However, the impact is stated to be significantly less than in traditional methods.
- Data Availability: The specific datasets used for training and testing are not publicly available due to privacy concerns.
- Future Work:
- The authors suggest future research directions that include considering the interpretability of the SE model and the correlation between measurements to further enhance SE accuracy.
Summary updated