Foundation-Style Pretraining for Scalable and Transferable Distributed PV Forecasting

Dayin Chen; Xiaodan Shi; Xin Xu; Zhongyue Su; Shibo Zhu; Ruizhe Deng; Dongxiao Zhang; Yuntian Chen

cs.LGOct 20, 2026

Foundation-Style Pretraining for Scalable and Transferable Distributed PV Forecasting

Authors: Dayin Chen, Xiaodan Shi, Xin Xu, Zhongyue Su, Shibo Zhu, Ruizhe Deng, Dongxiao Zhang, Yuntian Chen,

Organizations: Department of Building Environment and Energy Engineering, The Hong Kong Polytechnic University, Hong Kong SAR, China · International Centre of Urban Energy Nexus, The Hong Kong Polytechnic University, Hung Hom, Kowloon, Hong Kong SAR, China · Ningbo Institute of Digital Twin, Eastern Institute of Technology, Ningbo, 315200, China · Department of Computer and Systems Sciences, Stockholm University, Stockholm, Sweden ·

Abstract

The rapid growth of distributed photovoltaic (PV) generation is increasing the demand for accurate forecasting across large numbers of geographically dispersed stations. Existing approaches are commonly developed for individual sites or specific forecasting horizons, requiring repeated training and maintenance as new stations are integrated and limiting direct reuse at previously unseen sites. To address this scalability gap, this paper presents a domain-specific foundation-style pretraining framework that learns a shared forecasting model from heterogeneous multi-site and multi-horizon PV data. NWP-assisted forecasting is formulated as a sequence-completion problem that unifies historical PV power, meteorological observations, future numerical weather prediction (NWP) forecasts, and time encodings. Three pretraining objectives are designed and systematically evaluated, including forecast-aligned reconstruction, generic reconstruction, and mixed reconstruction. The mixed objective combines the operational structure of PV forecasting with random sequence reconstruction and provides the most balanced performance across forecasting accuracy, degraded NWP conditions, and downstream adaptation. Experiments on 479 distributed PV stations use 340 stations for model development and 139 spatially held-out stations for zero-shot evaluation. Across horizons from 6 hours to 7 days, the multi-site pretrained MR model outperforms representative station-specific and shared baselines, as well as generalist time-series models, including a recent large-scale pretrained foundation model. The pretrained model can also be directly deployed at new PV stations and subsequently adapted to region-specific operational objectives when local data become available, providing a scalable forecasting paradigm for large distributed PV fleets.

Figures & tables

Figure 1: Overview of the proposed foundation-style pretraining framework for scalable distributed PV forecasting. A shared model is pretrained across multiple PV sites and forecasting horizons using a sequence-completion formulation and reconstruction-based objectives, enabling zero-shot transfer to unseen stations, robustness to degraded NWP availability, and downstream adaptation to region-specific operational objectives.

Category	Variables
PV power	PV power output (elec)
Meteorological observations (OBS)	Relative humidity (rh); 10-m wind components ( $u_{10}$ , $v_{10}$ ); 2-m air temperature ( $t_{2m}$ ); surface solar radiation downwards (ssrd); surface thermal radiation downwards (strd); top-of-atmosphere solar radiation (tsr); total cloud cover (tcc)
Numerical weather prediction (NWP)	Forecasted global horizontal irradiance (ghi_f); surface pressure (sp_f); relative humidity (rh_f); 2-m air temperature (t2m_f); 10-m wind direction (wd10_f); 10-m wind speed (ws10_f); total cloud cover (tcc_f)
Time encoding (TE)	Hour of day (sine, cosine); minute of hour (sine, cosine); day of year (sine, cosine)

Table 1: Variables included in the multivariate forecasting sequence, grouped into four categories. ERA5 weather reanalysis is used as a proxy for meteorological observations (OBS).

Figure 2: Illustration of rolling NWP alignment for the same target day. Solar irradiance forecasts for day $d$ are shown using NWP products issued from days $d-7$ to $d-1$ .

Figure 3: Overview of the multi-horizon sample construction, reconstruction-based pretraining objectives, and shared forecasting backbone. FA applies forecasting-aligned reconstruction, GR performs generic reconstruction with random masking, and MR combines the two masking strategies. The resulting masked sequences are processed by a shared PatchTSMixer-based backbone and reconstruction head, from which the future PV segment is extracted as the forecasting output. White blocks denote masked entries. Pretraining masks are used only during pretraining.

Figure 4: Spatial distribution and dataset split of the 479 distributed PV stations in Ningbo. Colors distinguish the 340 development stations from the 139 spatially held-out stations used for zero-shot evaluation, while marker size represents the installed PV capacity of each station.

		6 h		1 day		3 days		7 days
Method	#Params	MAE	RMSE	MAE	RMSE	MAE	RMSE	MAE	RMSE
Shared forecasting baseline
Shared Transformer	1.07M	0.124	0.161	0.122	0.159	0.125	0.163	0.131	0.170
Generalist time-series baselines
Chronos	1.08M	0.229	0.307	0.237	0.317	0.260	0.340	0.262	0.344
Chronos-MV	1.52M	0.130	0.196	0.161	0.237	0.208	0.292	0.226	0.309

Table 2: Comparison with shared and generalist time-series models on the common quality-controlled test set of the spatially held-out PV stations under different forecasting horizons. Lower values indicate better performance. Parameter counts are approximately matched for Shared Transformer, Chronos, and the proposed models, while TimesFM-3 is evaluated at its original pretrained scale. All models are evaluated without target-station adaptation.

	6 h		1 day		3 days		7 days
Method	MAE	RMSE	MAE	RMSE	MAE	RMSE	MAE	RMSE
Station-specific baselines
Pers-L	0.188	0.257	0.195	0.264	0.231	0.302	0.242	0.311
Pers-D	0.239	0.309	0.243	0.311	0.255	0.324	0.251	0.320
Ridge	0.105	0.141	0.107	0.143	0.116	0.154	0.131	0.172
LGBM	0.109	0.147	0.112	0.151	0.122	0.163	0.138	0.182

Table 3: Comparison with conventional station-specific baselines on 139 unseen PV stations under different forecasting horizons. Lower values indicate better performance. The proposed framework is evaluated in a zero-shot transfer setting without station-specific retraining.

Panel of Figure 5 — Figure 5: Station-level RMSE distributions of Ridge, Transformer, Informer, and MR across unseen PV stations under four forecasting horizons. Lower values indicate better predictive accuracy.

Panel of Figure 6 — Figure 6: Representative zero-shot forecasting trajectories on unseen PV stations under different prediction horizons. The four subplots compare the proposed variants with conventional station-specific baselines. While most methods capture the general generation trend at short horizons, larger discrepancies appear at longer horizons, particularly around rapid ramps and multi-peak patterns.

		6 h		1 day		3 days		7 days
Scenario	Method	MAE	RMSE	MAE	RMSE	MAE	RMSE	MAE	RMSE
Full forecast	FA	0.100	0.137	0.106	0.145	0.116	0.156	0.133	0.178
	MR	0.095	0.129	0.100	0.135	0.111	0.147	0.127	0.168
	GR	0.107	0.140	0.116	0.150	0.130	0.165	0.143	0.180
Partial NWP (30% masked)	FA	0.111	0.148	0.118	0.157	0.127	0.167	0.142	0.187
	MR	0.097	0.131	0.102	0.137	0.113	0.149	0.129	0.170

Table 4: Forecasting performance under degraded future NWP availability on 139 held-out PV stations. Results are averaged over all test stations. Lower values indicate better performance.

Panel of Figure 7 — Figure 7: Relative RMSE degradation under degraded future NWP availability. Degradation is computed relative to the full-forecast setting. Lower values indicate stronger robustness to reduced or unavailable future NWP inputs.

Horizon	Method	Region	Improved stations (%)	Mean $\Delta$ KPI (pp)	Mean KPI (%)
1 day	FA	Huabei	97.84	0.98	77.66
		Nanfang	99.28	7.44	52.52
		Shandong	90.65	0.20	85.74
	MR	Huabei	98.56	1.03	79.60
		Nanfang	100.00	8.67	56.86
		Shandong	99.28	0.29	86.81

Table 5: Downstream adaptation to region-specific operational objectives under the 1-day and 3-day forecasting settings. Results are averaged over 139 held-out PV stations. Bold values indicate the highest final Mean KPI for each forecasting horizon and regional criterion.

Solar Power Forecasting

Solar power forecasting predicts photovoltaic generation and irradiance from weather, imagery, plant data, and recent output to support reliable grid and energy decisions.

See all 45 papers

PARA-PV: Physics-Aware Retrieval-Augmented PV Prediction Based on Frozen Foundation Model and Distribution Shift Correction
Jul 9, 2026Hang Fan, Weican Liu, Ying Lu +3Time Series Forecasting Non-Stationary Time Series Forecasting
School of Economics and Management, North China Electric Power University, 102206, Beijing, China · School of Electrical and Electronic Engineering, Nanyang Technological University, 639798, 50 Nanyang Avenue, Singapore · Department of Electrical Engineering, Tsinghua University, 100084, Beijing, China
Constructed regional photovoltaic generation fields improve ultra-short-term distributed solar forecasting
Oct 12, 2026Dayin Chen, Xiaodan Shi, Xin Xu +5Solar Power Forecasting Spatiotemporal Forecasting
Zhejiang Key Laboratory of Industrial Intelligence and Digital Twin, Eastern Institute of Technology, Ningbo 315200, China. · Ningbo Key Laboratory of Advanced Manufacturing Simulation, Eastern Institute of Technology, Ningbo 315200, China · Department of Building Environment and Energy Engineering, and the International Centre of Urban Energy Nexus, The Hong Kong Polytechnic University, Hong Kong SAR, China. +1
Time series Foundation Models based on Physics-Informed Synthetic Histories for Cold-Start Photovoltaic Forecasting
Jun 5, 2026Lorenzo Longarini, Alessandro Rongoni, Simone Silenzi +2Zero-Shot Time Series Forecasting Time Series Forecasting
Sistemi 2000 srl, Civitanova Marche, Italy · Department of Engineering ”Enzo Ferrari”, University of Modena and Reggio Emilia, Modena, Italy · Department of Political Sciences, Communication and International Relations, University of Macerata, Macerata, Italy

Jul 9, 2026J/K move · Enter open · S save

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

Hang Fan, Weican Liu, Ying Lu, Dunnan Liu+2

School of Economics and Management, North China Electric Power University, 102206, Beijing, China · School of Electrical and Electronic Engineering, Nanyang Technological University, 639798, 50 Nanyang Avenue, Singapore · Department of Electrical Engineering, Tsinghua University, 100084, Beijing, China

Time Series Forecasting Non-Stationary Time Series Forecasting Solar Power Forecasting Physics-Informed ML Retrieval-Augmented Time Series Forecasting

Open paper AskPDF arXiv

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. 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 PV-specific 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 physics-constrained 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. Our code is available at https://github.com/weican1103/PARA-PV.

Abstract

Solar Power Forecasting

Explore similar work

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

Constructed regional photovoltaic generation fields improve ultra-short-term distributed solar forecasting

Time series Foundation Models based on Physics-Informed Synthetic Histories for Cold-Start Photovoltaic Forecasting