Learning a Generator Model from Terminal Bus Data

Abstract: In this work we investigate approaches to reconstruct generator models from measurements available at the generator terminal bus using ML techniques. The goal is to develop an emulator which is trained online and is capable of fast predictive computations. The training is illustrated on synthetic data generated based on available open-source dynamical generator model. Two ML techniques were developed and tested: (a) standard vector auto-regressive (VAR) model; and (b) novel customized long short-term memory (LSTM) deep learning model. Trade-offs in reconstruction ability between computationally light but linear AR model and powerful but computationally demanding LSTM model are established and analyzed.

• The paper demonstrates a generator-agnostic approach to reconstructing generator models from terminal bus data using machine learning techniques.
• The paper compares a VAR model, known for rapid and efficient prediction in linear dynamics, with a custom LSTM model that excels in handling non-stationary data.
• The paper shows that the LSTM-based emulator significantly reduces NRMSE under varying noise conditions, indicating its promise for real-world power grid applications.

Learning a Generator Model from Terminal Bus Data

Introduction

The paper "Learning a Generator Model from Terminal Bus Data" (1901.00781) explores the use of machine learning techniques to reconstruct generator models based solely on data available at the terminal bus of the generator. The research is centered around developing online trainable emulators capable of rapid predictive computations. Two key models are investigated: the vector auto-regressive (VAR) model, which offers simplicity and computational efficiency, and a custom long short-term memory (LSTM) model, which provides superior reconstruction capabilities at a higher computational cost. The study leverages synthetically generated data to demonstrate the efficacy of these models, emphasizing a generator-agnostic approach.

Technical Background

Vector Auto-Regressive Model

The VAR model is utilized for its proven effectiveness in handling stationary time series. In this context, it is extended to manage multivariate data through equations that incorporate both endogenous and exogenous variables. The paper outlines how conditional dependence within time-series data informs the selection of the model order $p$, where $p$ is critical for accurate prediction of generator dynamics.

Figure 1: Colormap interpretation of inverted correlation matrices for each pair of variables. Upper triangular part of each plot corresponds to degree of conditional dependence of first variable's present and second variable's past. Largest sequence of non-zero elements ending with zeros constitutes a measure for model order $p$.

Long Short-Term Memory Network

The LSTM model, noted for its efficacy in modeling long sequences, is enhanced using weight-drop techniques to mitigate overfitting. The customized model addresses the challenges of non-stationary data, directly applying to input variables without requiring transformations for statistical stationarity. This model leverages the architecture's ability to learn complex temporal dependencies and manage large fluctuations in data effectively.

Experimental Framework and Data

The research employs synthetic data generated via the OpenModelica tool and OpenIPSL package, simulating a typical power generator scenario. This data includes stochastic perturbations to model uncertainty, following dynamics outlined by a telegraph process. Such synthetic data provides a controlled environment to evaluate the robustness and scalability of the ML models.

Figure 2: Telegraph process.

Results and Discussion

The study demonstrates the VAR model's capability to rapidly adapt and predict generator states under varied noise conditions, albeit with limitations when dealing with strong non-linear dynamics. In contrast, the WD-LSTM model excels in handling both typical and adversarial data conditions, showcasing a significant reduction in normalized root mean squared error (NRMSE) across scenarios. Notably, the LSTM-based model exhibits resilience to high-order noise and parameter randomization, indicating its robustness and potential applicability in real-world scenarios.

Figure 3: WD-LSTM: Estimated density of the NRMSE distribution.

Figure 4: WD-LSTM: predicted values vs True values.

Conclusion

This work highlights the potential of adaptive machine learning frameworks in generator model reconstruction, showcasing a VAR model suitable for linear dynamics and an LSTM framework adept at capturing complex, nonlinear temporal dependencies. While the former provides a lightweight implementation, the latter affords comprehensive modeling power, meriting consideration for integration into current power grid predictive practices.

The paper concludes with a roadmap for future work encompassing physical model integration, application to real-world data, and expansion to other energy system components. These endeavors aim to refine the predictive toolkit available for power system management, enhancing operational reliability and efficiency.