Session: 01-03: AI for Energy Sustainability III

Paper Number: 138679

138679 - Time Series Scenario Generation for Stochastic Optimization and Uncertainty Quantification of Energy Systems: Classical Models, Deep Learning, and Beyond 

Abstract: 

Renewable energy sources such as wind and solar power play an essential role in reducing carbon emissions resulting from generating electricity, heat, and other energy products. However, these same renewable resources are intermittent and cannot reliably meet energy demand without either significant over-installation and curtailment, large amounts of energy storage, or complementary, dispatchable generators. The variability of renewable energy sources adds volatility to the supply side of the supply-demand relationship, which manifests as volatility in wholesale electricity price experienced by all generators in an electricity market.

Uncertainty in electricity price is driven by stochasticity in the weather, fuel prices, electricity demand, and generator availability, among many other possible drivers. Large, infrequent spikes in wholesale electricity price occur at the confluence of extreme behavior in these driving uncertainties and are a key statistical feature of the electricity price. The uncertainty introduced by these driving factors must be properly quantified to anticipate the risk and expected returns on investment for generators participating in energy markets and for grid operators to assess the robustness of the system to meet demand. After fitting a statistical time series model to the historical data of interest, Monte Carlo samples drawn from these models augment the historical data on which they were trained, and a distribution of performance metric, such as net present value of a system of interest, can be obtained. However, the real-world data of electricity markets pose a difficult modeling problem which traditional time series modeling approaches, such as ARIMA-based models, are ill-equipped to handle.

The presented work applies state-of-the-art generative deep learning models to generate synthetic time series data to enable robust optimization of energy systems in evolving electricity markets. We demonstrate the application of generative adversarial network (GAN) models, including convolutional neural network (CNN) and long short-term memory (LSTM) GAN models, for this scenario generation on wind power, solar power, and electricity demand multivariate time series from several US electricity markets. Each of these regional markets has unique statistical characteristics. We also introduce continuous-time generative models using neural stochastic differential equations (NSDEs) for this application. To our knowledge, this use of NSDEs for scenario generation of energy sector relevant time series is novel.

NSDEs are a promising category of models which use neural networks to learn both the deterministic and stochastic behaviors of an observed stochastic process. As a result of their mathematical formulation, NSDEs are generally more interpretable than black-box models, offer a natural way to generate synthetic time series of varying length and time resolution, allow for direct inclusion of explicit physics-driven and statistical components to the model, and provide flexibility in the time spacing, frequency, and completeness of training data.

We find that all generative deep learning models tested here produce more realistic synthetic data than the traditional statistical time series models, such as ARIMA-based models, currently used in this application. Further, we show that the newly introduced NSDE model provides equivalent performance to the other deep learning models while being more parsimonious, flexible, and interpretable. The AI-driven models presented here will enable more accurate uncertainty quantification, risk analysis, and robust optimization of integrated energy systems and future energy markets and electricity grids.

Presenting Author: Jacob Bryan Utah State University

Presenting Author Biography: Jacob Bryan is pursuing a PhD in Mechanical Engineering at Utah State University under the supervision of Dr. Hailei Wang. Jacob's research focuses on statistical time series modeling for scenario-based stochastic analysis and optimization of integrated energy systems. He has worked closely with Idaho National Laboratory over the last two years to implement these models into their FORCE framework for energy systems analysis. He aims to graduate in Summer 2024.

Authors:

Jacob Bryan Utah State University
Hailei Wang Utah State University