Beijing Institute of Technology Press Co., Ltd
image:
Learning to Forecast for Better Scheduling: A Regret-Aware Framework for PV-BESS Optimization
view more
Learning to Forecast for Better Scheduling: A Regret-Aware Framework for PV-BESS Optimization
Credit: GREEN ENERGY AND INTELLIGENT TRANSPORTATION
Researchers have developed a decision-aware forecasting framework for photovoltaic-battery energy storage systems, or PV-BESS, that improves how solar generation is predicted for day-ahead scheduling. Instead of treating forecasting and operational decision-making as two separate steps, the new approach trains the forecasting model to account directly for downstream scheduling objectives such as economic return and output smoothness. The result, according to the study, is a system that not only predicts well, but also helps the storage system operate more profitably and more stably.
PV-BESS systems are becoming increasingly important as grids absorb more solar power. Pairing photovoltaics with battery storage can help offset the intermittency of solar generation, smooth power output, and support more reliable dispatch. But operating such systems well depends heavily on forecasting. If the forecast is inaccurate, the charging and discharging plan can be suboptimal, reducing arbitrage value and increasing fluctuations. In many existing systems, forecasting and optimization are still handled in a predict-then-optimize sequence, where the forecasting model is trained only to minimize statistical error and does not directly consider the practical impact of those errors on later scheduling decisions.
The authors of the new study argue that this separation creates a structural weakness. A forecast that looks good according to conventional error metrics does not necessarily produce the best operational decision. In other words, the model may be optimized for prediction accuracy in a narrow mathematical sense, while the energy system itself cares about a different outcome: revenue, smooth dispatch, and reduced operational regret. To address that mismatch, the researchers designed a decision-aware training strategy that couples day-ahead photovoltaic power forecasting with downstream scheduling objectives.
At the center of the framework is a surrogate decision model called the Regret Network, or R-Net. The purpose of R-Net is to estimate decision regret from simulated forecast-decision data and then feed that information back into model training in a differentiable form. This allows the forecasting system to learn not only from how close it is to the measured solar output, but also from how costly its forecasting errors become when translated into scheduling decisions. In effect, the forecasting model is taught to care about the real operational consequences of being wrong.
The forecasting backbone itself is based on a Transformer architecture enhanced with numerical weather prediction information. The researchers then optimized the model using a hybrid loss function called ReMix, which is designed to balance two targets at once: statistical forecasting accuracy and decision performance. The highlights of the paper also note that R-Net is built as a CNN-LSTM surrogate model, allowing decision loss to be approximated efficiently rather than requiring a full downstream optimization loop at every training step. Together, these elements create a learning framework in which forecasting and scheduling are more tightly aligned.
The study tested the approach on two real-world datasets, one from a centralized photovoltaic station in Daqing and another from a distributed system in Ningbo. According to the paper, the method reduced decision regret by as much as 19% and increased daily revenue by 3.8% while still maintaining high forecasting accuracy. These results matter because they suggest that the gains are not limited to abstract modeling improvements. Instead, they translate into measurable operational benefits, improving the economic value of PV-BESS scheduling while also contributing to more stable system behavior.
The broader significance of the study lies in its challenge to a long-standing workflow in energy AI. For years, many machine learning systems in energy have been developed as if prediction and decision-making were separate problems. This paper suggests that for applications like solar-plus-storage scheduling, that separation can be inefficient. A model trained to minimize decision regret may deliver more useful operational outcomes than one trained only to minimize forecast error. That insight could influence not only PV-BESS scheduling, but also a wider range of energy applications in which forecasts are ultimately used to drive control actions, market bids, or dispatch plans.
Further research will still be needed to understand how well the framework generalizes to additional market settings, battery configurations, and weather regimes. But the study provides a clear example of how decision-aware machine learning can improve both economic and operational performance in renewable energy systems. As solar deployment grows and storage becomes more tightly integrated into grid operations, methods that learn forecasting and decision-making together may become increasingly valuable tools for extracting more stability and value from clean energy infrastructure.
Reference
Author:
Dayin Chen a b d, Guotao Wang a b d, Xiaodan Shi e g, Mingkun Jiang h, Shibo Zhu a b d, Haoran Zhang f, Dongxiao Zhang c d, Yuntian Chen c d, Jinyue Yan a b
Title of original paper:
Learning to Forecast for Better Scheduling: A Regret-Aware Framework for PV-BESS Optimization
Article link:
https://www.sciencedirect.com/science/article/pii/S2773153725001355
Journal:
Green Energy and Intelligent Transportation
DOI:
10.1016/j.geits.2025.100385
Affiliations:
a Department of Building Environment and Energy Engineering, The Hong Kong Polytechnic University, Hung Hom, Kowloon, Hong Kong SAR, China
b International Centre of Urban Energy Nexus, The Hong Kong Polytechnic University, Hung Hom, Kowloon, Hong Kong SAR, China
c Zhejiang Key Laboratory of Industrial Intelligence and Digital Twin, Eastern Institute of Technology, 315200, Ningbo, China
d Ningbo Institute of Digital Twin, Eastern Institute of Technology, 315200, Ningbo, China
e School of Business, Society and Technology, Mälardalens University, Västerås, 72123, Sweden
f School of Urban Planning and Design, Peking University, Shenzhen, China
g Center for Spatial Information Science, the University of Tokyo, 5-1-5 Kashiwanoha, Kashiwa-shi, Chiba, 277-8568, Japan
h PV Industry Innovation Center, State Power Investment Corporation, 710061, Xi’an, Shaanxi, China
Disclaimer: AAAS and EurekAlert! are not responsible for the accuracy of news releases posted to EurekAlert! by contributing institutions or for the use of any information through the EurekAlert system.
Media Contact
Ning Xu
Beijing Institute of Technology Press Co., Ltd
xuning1907@foxmail.com
Beijing Institute of Technology Press Co., Ltd
Copyright © 2026 by the American Association for the Advancement of Science (AAAS)
Copyright © 2026 by the American Association for the Advancement of Science (AAAS)
