General Lab Information

Energy Systems Division

Utility Applications

Development of a Dynamic, Granular, and Multi-Day Outage Forecasting Tool


Started in August 2021. Phase I completed and moving to Phase II.


The objective of this proposed study is to develop a data-driven dynamic, granular and multi-day power outage forecasting (up to 72 hours lead time) tool by leveraging Brookhaven Lab expertise in atmospheric science, power grid engineering, machine learning applications and probabilistic risk assessment.


We will incorporate meteorological data from long-term forecasts products and ancillary information (e.g., tree coverage maps, tree trimming schedules), historical grid status information from ConEdison, operational forecast model output, machine learning techniques, and probabilistic risk assessment towards the development of a computationally efficient, quantitative algorithm for predicting weather-related grid impacts. The planned tool can be used in conjunction with the expertise of ConEd meteorologists as a tool for informing storm preparation and restoration efforts.


  • M. Yue, M. Jensen, T. Zhao, S. Endo, A. Spaulding, and D. Cerrai, “Development of a Machine Learning-based, Dynamic, Granular Grid Outage Forecasting Algorithm,” submitted to Climate Conference, 2023.
  • T. Zhao, S. Endo, M. Yue, M. Jensen, A. Marschilok, B. Nugent, B. Cerruti, C. Spanos, “Machine Learning-based Dynamic Granular Electric Outage Forecasting,” submitted to IEEE PES General Meeting 2023.

A Feasibility Study on Machine Learning-Based Solution to Locate Faults in Underground Distribution Grids


Expected to start in January 2023


We proposed to pursue a machine learning (ML)-based solution for identifying granular fault information including what, where, and when a fault(s) occurs in the underground distribution system. In parallel to model development, hardware-based solutions to locate faults will also be benchmarked.


The proposed feasibility study includes (1) building a simulation model for generating training data with detailed models of protection devices, (2) developing an ML-based data fusion scheme, and (3) progressing towards a fault-locating model based on dynamic neural networks (DNNs).


The expected outcome is a more realistic simulation model for generating high-fidelity data, and a state-of-the-art data fusion scheme for multi-rate measurements. Preliminary results will be used in developing the proposal for ARPA-E’s funding opportunity. Ultimately a DNN-based fault-location model will be developed that is expected to further improve the accuracy.