PARA-PV Physics-Aware Retrieval-Augmented PV Prediction Based on Frozen Foundation Model and Distribution Shift Correction

Jul 16, 2026

Accurate photovoltaic (PV) power forecasting is critical for reliable grid dispatch and renewable integration, yet remains difficult due to weather variability, day–night transitions, regime-dependent dynamics, and strict physical constraints. Existing data-driven methods mainly increase model complexity, often underusing PV-specific physics and producing biased forecasts in peak, ramping, or low-power regimes. PARA-PV embeds physics knowledge throughout the pipeline via physics-aware retrieval of consistent historical patches and analog trajectories, frozen Chronos prior calibration with a residual adapter, and physics-aware distribution shift correction using gated mean-shift/scale, guided by regime-aware reweighted physics-constrained loss.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages36

Reading Time1 h 31 min

Introduction

Background and motivation

Related work

Method Overview

Physics-aware retrieval-augmented forecasting

Frozen foundation-model prior calibration

Physics-aware distribution shift correction

Physics-constrained regime-aware loss

Experiments and Results

arXiv :2607 .08079v 1 [ cs .AI] 9 Jul 2026 PARA-PV: Physics-Aware Retrieval-Augmented PV Prediction Based on Frozen Foundation Model and Distribution Shift Correction Hang Fana , Weican Liub,∗, Ying Lua , Dunnan Liua , Long Chengc , Wei Weid,∗∗ a School of Economics and Management, North China Electric Power University, 102206, Beijing, China b School of Electrical and Electronic Engineering, Nanyang Technological University, 639798, 50 Nanyang Avenue, Singapore c School of Control and Computer Engineering, North China Electric Power University, 102206, Beijing, China d Department of Electrical Engineering, Tsinghua University, 100084, Beijing, China Abstract Accurate photovoltaic (PV) power forecasting is essential for reliable grid dispatch and renewable energy integration, yet it remains challenging because PV generation is jointly shaped by weather variability, day-night transitions, regime-dependent dynamics, and strict physical constraints. Existing data-driven methods typically improve temporal representation learning by increasing model complexity, but they insufficiently exploit PV-specific physical knowledge and may produce biased forecasts under peak, ramping, or low-power conditions. To address these limitations, we propose PARA-PV, a Physics-Aware Retrieval-Augmented framework that embeds physical knowledge throughout the forecasting process. The framework first encodes multivariate PV observations into patch-level representations and, through a physics-aware retrieval-augmented learner, retrieves historical patches and analog trajectories that are consistent with the current window in temporal shape, power level, PV operating state, and intra-day period; this yields a physically grounded base forecast. To supplement local memory with broader temporal knowledge, the base forecast is then calibrated against a frozen Chronos time-series foundation-model prior through a lightweight residual adapter, so that general temporal regularities are adapted to PVspecific dynamics without overriding the physically grounded prediction. Because residual conditional distribution shifts persist when weather and diurnal regimes change, a physics-aware distribution shift correction module subsequently adjusts the preliminary forecast using power, weather, timestamp, and day/night conditions, applying gated mean-shift and scale corrections selectively. Finally, a physicsconstrained loss function partitions the samples into peak, ramping, night-time, and regular regimes and adaptively reweights their error contributions, preventing the dominant regular regime from suppressing learning of operationally critical states. Experiments on two geographically distinct PV farms demonstrate that PARA-PV consistently outperforms seven representative baselines across multiple forecasting horizons, achieving lower point-forecast errors and sharper prediction intervals with fewer than 1M trainable parameters. Our code is available at [https://github.com/weican1103/PARA-PV](https://github.com/weican1103/PARA-PV). Keywords: Artificial intelligence, Foundation models, Physics-aware learning, PV power forecasting, Retrieval-augmented prediction. ∗ W. Liu is the Corresponding author ([Email: weican001@e.ntu.edu.sg](Email: weican001@e.ntu.edu.sg)). ∗∗ W. Wei is the Co-corresponding author ([Email: wei-wei04@mails.tsinghua.edu.cn](Email: wei-wei04@mails.tsinghua.edu.cn)). 1. Introduction 1.1. Background and motivation With the continuous growth of global energy demand, the development of reliable and low-carbon energy systems has become an urgent priority. Among renewable energy sources, solar energy offers clear advantages in terms of cleanliness, scalability, and wide geographical availability. Driven by advances in PV technology and policy support, PV generation has expanded rapidly and is becoming an increasingly important component of modern power systems [1] . Accurate PV power forecasting is essential for grid integration of s

FAQs

What makes PV power forecasting challenging in this work?

PV generation is shaped by weather variability and diurnal transitions while also following regime-dependent dynamics and strict physical constraints such as night-time zero output, capacity limits, and peak/ramping behavior.

How does PARA-PV produce a physically grounded base forecast?

It encodes multivariate PV observations into patch-level representations and uses a physics-aware retrieval-augmented learner to retrieve historical patches and analog trajectories matching the current temporal shape, power level, PV operating state, and intra-day period.

How is distribution shift handled when weather and day/night regimes change?

A physics-aware distribution shift correction module adjusts the preliminary forecast using power, weather, timestamp, and day/night conditions, applying gated mean-shift and scale corrections selectively.

Jul 16, 2026

Upload by Theodora