Project Background

Adverse weather events (e.g., hurricanes and wildfires) are causing more and more widespread outages in the power grid as climate change continues, resulting in massive economic costs and even casualties in some cases. Being able to predict outages given a weather forecast is crucial for disaster response and resilience enhancement planning. Existing outage forecasting is mostly based on physical models, i.e., using actual topology with circuit, protection device, etc. and power flow models to simulate service degradation under disruptions. Although it is being used in some utilities, the results are far from being accurate and scalable due to enormous uncertainties and complicated outage mechanisms. It is imperative to develop an advanced forecasting tool for electricity sector to predict how outages evolves as weather events progress. Upon success, the tool is not only useful for disaster response, but also provides a quantitative evaluation framework for system resilience against certain types of extreme weather events.

Scientific Opportunities

In recent years, machine learning and artificial intelligence (ML/AI) have been extremely useful in developing prediction tools when large volume data are available. ML can replace sophisticated physical models with simplified and standardized functions (e.g., neural network or stochastic models) that maps inputs to outputs, achieving the same prediction purpose. The success of a ML work is highly dependent on data of quality and quantity. Thanks to the efforts in the past few years in collecting electricity customer outage announcement from publicly accessible utility websites, the project team has built a data platform with automatic outage data collection and accumulated high-resolution datasets for a few states on the east coast. In this work, we will leverage this electricity customer outage dataset, together with weather and terrain information that are publicly accessible to develop a ML/AI based outage prediction tool. This tool will combine stochastic models, which is good at describing random process such as outages, and deep neural network (DNN), which is good at prioritizing numerous input features, and build a hybrid temporal and spatial model to predict outage progress during an weather event. The results cannot only be used for prediction, but also can be used for resilience evaluation. For instance, the results can be used to determine which weather event the system is most vulnerable to and what kinds of weather variables (e.g., precipitation, wind speed, or temperature) are most impactful in an weather event, which could be used for resilience evaluation and enhancement. Therefore, this tool will help utility owners and grid operators to plan disaster response and system resilience enhancement.

Research Goal

This project will use big data and machine learning to develop data-driven prediction model for weather-induced customer power outages. We will expand existing electricity customer outage data platform to collect high-resolution and real-time power outage data from public resources, and then leverage advanced machine learning approaches to develop power outage forecasting models. The machine learning models will also offer causal-effect analysis to analyze the contributing factors of power outages, provide a quantitative resilience evaluation framework, and yield policy suggestions for enhancing power grid resilience.

Research Plan

Our framework for the modeling and analysis of weather-related outages in power grids is based on big data analytics and machine learning. First, we are developing a real-time outage data repository and geographically labeled outage data interface. The data repository collects real-time customer outage data with geographic labels (city, county or zip code). Second, we will perform data driven modeling of weather-related outages based on machine learning. We will leverage multiple public data sources, including the geographically labeled outage data and the numerical weather data from NOAA, as well as other environmental data, such as vegetation data, elevation data, and land cover data. Considering the spatial and temporal interdependencies in the power grid and weather-related power outages, we will use deep neural networks or statistical learning models to establish a spatio-temporal learning model to fit the real outage processes, and thus obtain the trained model as a data-driven model of weather-related outages. The data-driven model can be used for predicting weather-related outages, and thus give guidance for the emergency management functions of utility crews and governmental emergency agencies. Third, we will develop decision support framework for power system resilience evaluation and resilience enhancement against weather-related outages. The forecasting results will be used for quantitatively evaluating system resilience against extreme weather events. We will also perform sensitive analysis on the trained model and derive potential resilience enhancement policies for system operators.

Deliverables and Impacts

The deliverables include a nation-wide power outage database, an open-source software tool for near-real-time prediction of power outages and power system resilience evaluation, technical reports of dataset and software tool implementation, publications on peer-reviewed journals or conferences, and presentations at conferences and workshops.

Public presentations:

“Toward Simulation, Risk Assessment and Mitigation of Weather Hazards to Power Grids” (panel session), IEEE PES General Meeting (virtual meeting). 2020.

“Real Transmission and Distribution Outage Data Collection and Statistical Analysis to Model and Quantify Resilience” (panel session), IEEE PES General Meeting (virtual meeting). 2020.

“Modeling and Prediction of Weather induced Power Outages Using Neural Networks”, Poster presentation at the Physics-Informed Machine Learning (PIML) workshop. Santa Fe, NM, USA. 2020.

“A spatio-temporal analysis for power grid resilience to extreme weather”, Shixiang Zhu, Rui Yao, Yao Xie, Feng Qiu, to be submitted.

Team and contact

Argonne National Laboratory (lead)
Georgia Institute of Technology

Lab Lead Team Members

Dr. Rui Yao
Dr. Alinson Santos Xavier
Mr. Shixiang Zhu (Intern)
Mr. Ryan Newkirk (Intern)

Project PI: Dr. Feng Qiu
Principal Computational Scientist and Group Manager
fqiu@​anl.​gov