Presentation Details
| GenSolar: A Stochastic Generative Model for Minute-Resolution Solar Irradiance and Applications Marc Abou Anoma. Independent Researcher, Tokyo, Japan |
Abstract
We present a transformer-based neural network framework for stochastic generation of daily irradiance time series at minute resolution for both direct and diffuse irradiance components. We train the model on less than two years of irradiance data and show that the generated daily irradiance profiles realistically depict clear-sky conditions as well as cloudy and rainy events. We further show that the same pre-trained model can be reused for multiple downstream applications, including not only conditional short-term forecasting (via autoregressive rollout) but also generating a realistic synthetic irradiance dataset for training a control policy via reinforcement learning. Using data generated by the model, we train a simple policy-gradient-based single-axis solar-tracking controller that dynamically decides when to cease sun tracking and instead flatten the panels to maximize energy capture under diffuse-light conditions. We show that this control policy, trained on synthetic data, outperforms a classical single-axis sun-tracking strategy on real measured irradiance data.
No part of this publication may be reproduced, distributed, or transmitted in any form or by any means, including photocopying, recording, or other electronic or mechanical methods, without the prior written permission of the author.
No part of this publication may be reproduced, distributed, or transmitted in any form or by any means, including photocopying, recording, or other electronic or mechanical methods, without the prior written permission of the author.
Content Locked. Log into a registered attendee account to access this presentation.