Improved Deep Belief Network and Model Interpretation Method for Power System Transient Stability Assessment

Abstract

The real-time transient stability assessment (TSA) and emergency control are effective measures to suppress accident expansion, prevent system instability, and avoid large-scale power outages in the event of power system failure. However, real-time assessment is extremely demanding on computing speed, and the traditional method is not competent. In this paper, an improved deep belief network (DBN) is proposed for the fast assessment of transient stability, which considers the structural characteristics of power system in the construction of loss function. Deep learning has been effective in many fields, but usually is considered as a black-box model. From the perspective of machine learning interpretation, this paper proposes a local linear interpreter (LLI) model, and tries to give a reasonable interpretation of the relationship between the system features and the assessment result, and illustrates the conversion process from the input feature space to the high-dimension representation space. The proposed method is tested on an IEEE new England test system and demonstrated on a regional power system in China. The result demonstrates that the proposed method has rapidity, high accuracy and good interpretability in transient stability assessment.

Artificial Intelligence Techniques for Power System Transient Stability Assessment

“…Mahato et al in [61] introduced a particular Bi-LSTM attention mechanism to the TSA problem, which featured LSTM layers. Wu et al proposed a deep belief network in [75]. Xia et al used an embedding algorithm in combination with deep learning for the transient stability assessment [76].…”

A Critical Review of Data-Driven Transient Stability Assessment of Power Systems: Principles, Prospects and Challenges

“…First, it is tough and expensive to obtain large-scale, balanced data with accurate labels in real-world applications [70,71]. Then, existing data-driven TSA methods act as a black box with poor interpretability [72,73], which also limits their application in actual power systems. Finally, most of data-driven TSA methods generally lack the adaptability to topological changes.…”

Sag Source Location and Type Recognition via Attention-based Independently Recurrent Neural Network

“…Deep learning, as a representative of data-driven method, has the following favorable properties: ① it enables a machine or model to automatically learn deep and multi-aspect features from data [18]; ② feature learning can be achieved without knowledge of the power system model or human expert experiences [19]. Thus, deep learning provides a powerful solution to the above problems.…”