Abstract
As modern power grids integrate increasing levels of renewable energy, maintaining operational security becomes a complex challenge due to heightened uncertainty. This paper proposes the Sparse Weighted Oblique Decision Tree (SWODT) framework to extract interpretable and computationally efficient security rules. By utilizing oblique splits and Elastic Net regularization, the SWODT generates shallow, sparse trees that capture complex security boundaries. These rules are subsequently embedded into economic dispatch optimizations using the Big-M method. Experimental results on benchmark systems demonstrate that SWODT achieves superior accuracy and significantly reduces the computational burden on mixed-integer linear programming (MILP) solvers compared to traditional univariate approaches.
Introduction
The transition toward power systems dominated by wind and solar energy introduces significant operational volatility. Defining "security boundaries"—the safe operating limits of a grid during contingencies—is critical for reliability. Traditional Univariate Decision Trees (UDTs), such as CART, often struggle with this task; they require excessive depth to approximate complex boundaries, leading to poor interpretability and computational bottlenecks when integrated into optimization frameworks like Economic Dispatch (ED).
To address these limitations, this research introduces the Sparse Weighted Oblique Decision Tree (SWODT). Unlike UDTs that split data based on a single variable, oblique trees use linear combinations of features to create hyperplane splits, allowing for a more compact representation of the security region.
Methodology
The proposed framework consists of three primary stages: power system security simulation, sparse oblique decision tree training, and rules extraction for embedding into dispatch models.
Sparse Weighted Oblique Decision Tree (SWODT)
The core of the methodology lies in the SWODT algorithm, which incorporates several key innovations:
- Oblique Splits: Each node performs a split based on a hyperplane (a linear combination of multiple features), enabling the tree to capture diagonal boundaries with fewer levels.
- Weighted Soft-Splits: The authors employ a sigmoid-based soft split during training to ensure the objective function is differentiable, facilitating more efficient optimization.
- Elastic Net Regularization: To ensure the resulting rules are sparse and interpretable, the model applies both $L_1$ (Lasso) and $L_2$ (Ridge) penalties. This approach drives irrelevant weights to zero while managing the high correlation typical of power system data.
- Modified OWL-QN Algorithm: A specialized Orthant-Wise Limited-memory Quasi-Newton method is used to solve the non-smooth optimization problem created by the $L_1$ penalty.
Rules Extraction and Embedding
Once trained, the linear rules from the tree are extracted as sparse matrices. These are converted into linear constraints using the "Big-M" method, allowing the security boundaries to be directly integrated into MILP-based economic dispatch models.
Experimental Results
The SWODT method was validated using three benchmark systems: the IEEE 30-bus system, the RTS-GMLC system, and the NREL-118 bus system, all modified for high renewable penetration (36%–47%).
Accuracy and Interpretability
SWODT consistently outperformed standard oblique trees (WODT) in accuracy, particularly as tree depth increased, due to the effective prevention of overfitting by the regularization terms. Furthermore, SWODT produced trees with approximately half the number of leaf nodes compared to dense oblique trees, significantly enhancing human interpretability.
Computational Efficiency and Grid Security
Embedding SWODT rules into economic dispatch significantly improved grid reliability. In the tested systems, the percentage of secure operating states rose from a baseline of 53–76% to over 85–95%. Crucially, the sparsity of the rules reduced the computational time for the MILP solver. For the large-scale NREL-118 system, the solution time was reduced by 50% (from 0.1s to 0.05s), a vital improvement for real-time grid operations.
Conclusion
The Sparse Weighted Oblique Decision Tree offers a robust solution for power system security by bridging the gap between machine learning interpretability and optimization efficiency. By enforcing sparsity through Elastic Net regularization, the framework provides clear, linear security rules that enhance grid reliability while maintaining the high-speed performance required for modern economic dispatch.
Summary updated