Open source generative ML model unveiled for energy system planning

Researchers from the US Department of Energy’s (DoE’s) National Renewable Energy Laboratory (NREL) have developed an open source, publicly available generative machine learning model to simulate future energy-climate impacts to assist energy system planning.

Known as the Super-Resolution for Renewable Energy Resource Data with Climate Change Impacts (Sup3rCC), the open source model uses generative machine learning to produce downscaled future climate data sets that are available to the public at no cost.

According to NREL in a post, downscaled climate data is necessary to understand the impacts of climate change on local wind and solar resources and energy demand.

There are a multitude of existing downscaling methods, but they all have trade-offs in resolution, computational costs and physical constraints in space and time.

Developed by an NREL team of data scientist Grant Buster, senior software engineer Brandon Benton, researcher in applied mathematics Andrew Glaws and computational researcher Ryan King, Sup3rCC represents a new field of generative machine learning methods that can produce physically realistic high-resolution data 40 times faster than traditional dynamical downscaling methods.

“Sup3rCC will change the way we study and plan future energy systems,” said Dan Bilello, director of the Strategic Energy Analysis Center at NREL. “The tool produces foundational climate data that can be plugged into energy system models and provide much-needed insights for decision makers who are responsible for keeping the lights on.”

Have you read:
Planning key focus for Britain’s future National Energy System Operator
NREL to test energy management system at clean tech laboratory

Overcoming system research siloes

According to the lab, energy system research and climate research have traditionally been siloed for several reasons.

The resolution of traditional global climate models is too coarse across both time and space for most energy system models and enhancing resolution is computationally expensive.

Global climate models also do not always generate or save outputs that are required to model renewable energy generation and existing publicly available global climate model data sets are not commonly connected to the data pipelines and software used in energy system research.

Due to these challenges, most energy system planners have relied on historical high resolution wind, solar and temperature data to model electricity generation and demand.

But ignoring future climate conditions can be risky when it comes to planning a reliable energy system.

“Climate science is a complex field with massive amounts of data, huge uncertainties and not a lot of resources on how the information can or should be applied to other fields of study,” Buster said. “At NREL, we aim to bring the energy and climate modelling communities together to effectively and appropriately use climate information to guide energy system design and operation.”

Sup3rCC was created through a partnership between energy analysts and computational scientists at NREL to better incorporate multi-decadal changes in climate and meteorological variability in energy systems modelling.

“This work bridges the gap between energy system and climate research communities to significantly advance the developing field of energy-climate research,” added Bilello.

Also of interest:
ChatGrid to assist power grid operators with generative AI grid visualisation
Generative AI for smart grid modelling

Utilising generative AI

Sup3rCC overcomes computational challenges of traditional dynamical downscaling techniques by leveraging the power of recent advances in a generative machine learning technique called generative adversarial networks.

The model learns physical characteristics of nature and the atmosphere by studying NREL’s historical high resolution data sets, including the National Solar Radiation Database and the Wind Integration National Dataset Toolkit.

The model then injects physically realistic small-scale information that it has learned from the data sets into the coarse future outputs from global climate models.

As a result, Sup3rCC generates detailed temperature, humidity, wind speed and solar irradiance data based on the latest climate projections.

Sup3rCC outputs can then be used to study future renewable energy power generation, changes in energy demand and impacts to power system operations. The initial Sup3rCC data set includes data from 2015 to 2059 for the contiguous United States, and additional data sets will be released in the coming years.

“Our super-resolution work is unique in that we enhance the spatial and temporal resolution simultaneously and inject far more information than ever before,” King said. “Sup3rCC preserves the large-scale trajectories of climate simulations, while endowing them with realistic small-scale features that are crucial for accurate renewable energy resource assessments and load forecasting.”

Sup3rCC increases the spatial resolution of global climate models by 25 times in each horizontal direction and the temporal resolution by 24 times — representing a 15,000-fold increase in the total amount of data.

According to NREL, the model can do this process 40 times faster than traditional dynamical downscaling models so energy system planners and operators can get straight to planning at large scales.

The model is hoped to allow researchers at NREL and beyond to investigate weather events like future heat waves and the interplay between the electrical grid and renewable energy generation.

“Our approach dramatically reduces the computational cost of generating high spatial and temporal resolution data by several orders of magnitude,” added King.

“This allows us to consider changes in renewable resources and electrical demand in a multitude of future climate scenarios across multiple decades, which is critical for planning future energy systems.”

Outputs from Sup3rCC are compatible with NREL’s Renewable Energy Potential (reV) Model to study wind and solar generation and are interoperate with a suite of NREL modelling tools.

Users can access Sup3rCC data on Amazon Web Services and run reV in the cloud from their own desktop to see how wind and solar generation, capacity and system cost change under different climate scenarios.