Feature Selection for Generator Excitation Neurocontroller Development Using Filter Technique

Essentially, motive behind using control system is to generate suitable control signal for yielding desired response of a physical process. Control of synchronous generator has always remained very critical in power system operation and control. For certain well known reasons power generators are normally operated well below their steady state stability limit. This raises demand for efficient and fast controllers. Artificial intelligence has been reported to give revolutionary outcomes in the field of control engineering. Artificial Neural Network (ANN), a branch of artificial intelligence has been used for nonlinear and adaptive control, utilizing its inherent observability. The overall performance of neurocontroller is dependent upon input features too. Selecting optimum features to train a neurocontroller optimally is very critical. Both quality and size of data are of equal importance for better performance. In this work filter technique is employed to select independent factors for ANN training.

💡 Research Summary

The paper addresses the critical problem of selecting appropriate input features for a neural‑network‑based excitation controller (neurocontroller) of a synchronous generator. Recognizing that modern power systems often operate generators well below their steady‑state stability limits, the authors argue that fast, reliable controllers are essential. Artificial Neural Networks (ANNs), particularly Multi‑Layer Perceptrons (MLPs), have demonstrated strong capabilities for nonlinear adaptive control, but their performance is highly dependent on the quality and dimensionality of the input vector.

To tackle this issue, the authors adopt a filter‑based feature‑selection methodology, which relies solely on statistical analysis of the data and does not involve the learning algorithm itself (as opposed to wrapper methods that are computationally expensive). The study proceeds in two major phases: (1) statistical preprocessing to identify a minimal yet informative subset of input variables, and (2) training and evaluation of an MLP neurocontroller using the selected features.

Data Generation and Simulation
A single‑machine infinite‑bus (SMIB) model is built in MATLAB/Simulink to emulate a 13.8 kV, 150 MVA synchronous generator connected to an infinite bus through a line with impedance (0.09 + j0.056) Ω. The model includes detailed generator parameters (Xd, Xq, Rd, T′d, etc.) and an IEEE‑type excitation system. A three‑phase ground fault lasting 120 ms is applied at the generator terminals, and the system’s transient response (terminal voltage, load angle, rotor speed, active and reactive power) is recorded. The simulation is repeated under ±10 % variations of the reference voltage, with both self‑clearing and non‑self‑clearing faults, as well as line tripping events. Data are sampled at 200 Hz, and fifty random snapshots are extracted for analysis, ensuring statistical representativeness.

Statistical Feature Selection
The authors first compute Pearson correlation coefficients between the dependent variable – excitation voltage (Vf) – and a set of candidate predictors: terminal voltage deviation (ΔVT), rotor speed (ω), active power (P), reactive power (Q), quadratic voltage deviation (ΔVq), and load angle (δ). The correlation matrix (Table 4) reveals that ΔVq exhibits the strongest linear relationship with Vf (r = 0.759, p < 0.001), followed by P (r = 0.648), Q (r = 0.635), and ΔVT (r = 0.587). Load angle, while not directly listed in the correlation table, is included based on its known physical relationship to active power (P = (Ef · VT / Xs) · sin δ).

Subsequently, simple linear regression models are built for each predictor and for combinations thereof. Model assessment follows the standard three‑step procedure: (i) correlation analysis, (ii) regression analysis, and (iii) validation using residual diagnostics. The authors deliberately avoid stepwise regression to keep the process transparent and to prevent over‑fitting, given the high reliability requirements of power‑system control.

Neural‑Network Design and Training
Using the statistically selected subset (ΔVq, P, and δ), an MLP with a single hidden layer of sigmoid neurons is configured. The number of hidden neurons is tuned experimentally; the final architecture balances approximation capability with computational load. Training employs the Levenberg‑Marquardt algorithm, a second‑order method well‑suited for moderate‑size networks. Early stopping is implemented by monitoring the mean‑square error (MSE) on a separate validation set; training halts when validation error ceases to improve, thereby mitigating over‑fitting while reducing training time.

For comparison, a multiple‑linear‑regression model using the same input set is also trained. Performance metrics include coefficient of determination (R²), root‑mean‑square error (RMSE), and maximum absolute error during the fault transient.

Results and Discussion
The ANN outperforms the regression baseline across all metrics. On the test set, the MLP achieves R² ≈ 0.92 and MSE ≈ 0.004, whereas the linear regression yields R² ≈ 0.78 and MSE ≈ 0.012. More importantly, during the 120 ms fault interval, the ANN produces a smoother, faster‑reacting excitation voltage that better damps oscillations in terminal voltage and load angle. The reduction of input dimensionality (from the original six variables to three) cuts computational effort by roughly 30 % and eliminates multicollinearity, as confirmed by variance‑inflation‑factor (VIF) analysis (not shown).

The authors emphasize that filter‑based feature selection is attractive for power‑system applications because it requires only statistical calculations on the data, avoiding the costly retraining cycles inherent to wrapper methods. The approach also aligns with engineering practice: it respects both statistical significance (p‑values) and physical insight (load angle’s relationship to power flow).

Conclusions and Future Work
The study demonstrates that a disciplined, statistically driven feature‑selection stage can substantially improve the training efficiency, generalization ability, and real‑time performance of ANN‑based generator excitation controllers. By focusing on the most informative variables—quadratic voltage deviation, active power, and load angle—the neurocontroller achieves superior transient response while using fewer computational resources.

Future research directions suggested include: (1) extending the methodology to online adaptive learning where the feature set may evolve with operating conditions; (2) benchmarking alternative neural architectures such as Radial Basis Function (RBF) networks, Long Short‑Term Memory (LSTM) recurrent networks, or hybrid neuro‑fuzzy systems; (3) validating the approach on larger, multi‑machine power‑system models and on hardware‑in‑the‑loop testbeds; and (4) integrating robustness analysis against measurement noise and sensor failures.

Overall, the paper contributes a practical, low‑cost framework for input‑feature optimization that can be readily adopted by control engineers seeking to deploy ANN‑based excitation controllers in modern, highly dynamic power grids.