Closed-Loop Reinforcement Learning for Short-Term Load Forecasting over a REST API Framework

Julien Guité-Vinet, Alexandre Blondin Massé, Éric Beaudry, Arnaud Zinflou

Keywords: short-term load forecasting, reinforcement learning, closed-loop design, model lifecycle, REST API

Abstract:

Time series forecasting is a crucial task in various domains because it allows for the forecast of future values based on historical data. This is essential for making informed decisions in fields such as finance, energy, retail, meteorology, and many others. Accurate forecasting can
help optimize operations, reduce costs, and improve planning and resource allocation. During the lifecycle of a model, it is important to manage the lifecycle of models, including training, validation, and hyperparameter tuning. In this paper, we propose a novel approach to short-term load forecasting using reinforcement learning in a closed-loop manner to optimize the lifecycle of models. We evaluate the proposed approach using a real-world energy dataset and a REST API framework. The results demonstrate an improvement of 0.77% in the baseline MAPE score while minimizing the need for manual intervention in managing the lifecycle of forecasting models. This highlights the system’s ability to autonomously adapt to changing conditions and optimize model performance over time.

Submission Number: 7