Storm-Induced Power Grid Damage Forecasting Method for Solving Low Probability Event Data

Abstract

The data obtained from storm-induced damage to power grids possesses an inherent skewness distribution, which impedes the development of the damage forecasting model. An inaccurate damage forecasting model may fail to accurately forecast the damages and hinder the planning, preventive measures, and restorative actions for a storm event. This study investigates the challenges that must be overcome to yield an accurate model and proposes a machine learning-based damage forecasting method. A robust forecasting model was developed by identifying the key explanatory variables using the G-mean values. The method combines the application of the weighted extreme learning machine (ELM) and long short-term memory model (LSTM) to forecast power grid damage in response to storm events. The weighted ELM is used to classify the grid state for a storm in advance and the LSTM is subsequently used to forecast the number of grid damage cases. The actual storm event data were used to verify the efficacy of the proposed method using the root mean square error. The results demonstrate that the proposed method outperforms the regular forecasting method as it is more robust and accurate.

Keywords

• Extreme weather events
• machine learning
• power grid resilience
• imbalanced data
• predictive analytics