Forecasting is essential for ensuring grid stability given the unpredictable nature of renewables like solar power and wind energy. Deep learning methods that have been deployed to date can generate inaccurate forecasts in certain situations and fail to satisfy physical assumptions and constraints. In this paper, we propose a framework to enhance the accuracy and reliability of renewable energy forecast models. The RAG framework is augmented with safety guardrails to ensure that the forecasts are robust, accurate, and interpretable. Specifically, the framework retrieves historical weather-generation patterns relevant to the current state to increase the context-awareness of the time series models. We further deploy the guardrails to constrain predictions and avoid impossible values. Through experiments done on benchmark data from the National Renewable Energy Laboratory and Global Energy Forecasting Competition, we prove that our approach reduces prediction errors and constraint violations compared to the baselines.
Speaker(s): Rakesh More,
Agenda:
Renewable Energy Forecasting, Retrieval-Augmented Generation, Explainable AI, Smart Grid, Guardrails, Time-Series Prediction, AI Safety
Virtual: https://events.vtools.ieee.org/m/561841