Enhancing photovoltaic power generation nowcasting with sky image analysis using multi-modal attention networks

Oluwatoyosi Bamisile; She Kun; Chiagoziem C. Ukwuoma; Dongsheng Cai; Chibueze D. Ukwuoma; Chinedu I. Otuka; Chidera O. Ukwuoma; Olusola Bamisile

doi:10.1016/j.solener.2025.114117

Enhancing photovoltaic power generation nowcasting with sky image analysis using multi-modal attention networks

Oluwatoyosi Bamisile (Lead / Corresponding author)
, She Kun (Lead / Corresponding author)
, Chiagoziem C. Ukwuoma
, Dongsheng Cai
, Chibueze D. Ukwuoma
, Chinedu I. Otuka
, Chidera O. Ukwuoma
, Olusola Bamisile

Energy Environment and Society

Research output: Contribution to journal › Article › peer-review

Abstract

The growing demand for renewable energy has heightened the importance of accurate photovoltaic (PV) power forecasting, especially under fluctuating weather conditions. Traditional and many deep learning-based models struggle to interpret the complex visual patterns in sky images, resulting in reduced prediction accuracy, particularly during periods of cloudiness or seasonal variability. In this study, we introduce a novel deep learning model that enhances solar nowcasting by dynamically identifying and emphasising the most informative features across multiple visual dimensions. The model introduces a compact multi-scale CNN backbone specifically tailored to the spectral–spatial characteristics of sky images, ensuring efficient feature extraction under real-time constraints. A custom lightweight attention mechanism is embedded to enhance cloud–irradiance saliency detection, delivering transformer-like selectivity while retaining low parameter and computational cost. These features are coupled with a regularised deep regression head that integrates high-level interactions, forming a unified architecture that advances photovoltaic nowcasting by balancing accuracy, interpretability, and efficiency. The study used the novel Sky Images and Photovoltaic Power Generation Dataset provided by Stanford University to evaluate the proposed model using evaluation metrics such as RMSE, MSE, MAE, MAPE, and R2. The proposed model achieves an overall RMSE of 2.259 and R2 of 0.913, with particularly strong performance on sunny days (RMSE of 0.461, R2 of 0.996) and consistent results under cloudy conditions (RMSE of 3.158, R2 of 0.824). Seasonal analysis reveals that the model maintains robust accuracy across different climatic conditions, with channel attention excelling during high-irradiance summer and autumn days, and spatial attention effectively capturing complex cloud structures in winter and spring. These outcomes depict the model’s ability to deliver more reliable short-term power forecasts by leveraging deeper visual understanding, ultimately contributing to more efficient solar energy management.

Original language	English
Article number	114117
Journal	Solar Energy
Volume	303
Early online date	7 Nov 2025
DOIs	https://doi.org/10.1016/j.solener.2025.114117
Publication status	Published - Jan 2026

Keywords

Attention Mechanism
Deep Learning
Photovoltaic Output Prediction
Sky Images
Solar Forecasting

ASJC Scopus subject areas

Renewable Energy, Sustainability and the Environment
General Materials Science

UN SDGs

This output contributes to the following UN Sustainable Development Goals (SDGs)

Access to Document

10.1016/j.solener.2025.114117

Cite this

@article{3751f5aabb1a464c842741666f514cc9,

title = "Enhancing photovoltaic power generation nowcasting with sky image analysis using multi-modal attention networks",

abstract = "The growing demand for renewable energy has heightened the importance of accurate photovoltaic (PV) power forecasting, especially under fluctuating weather conditions. Traditional and many deep learning-based models struggle to interpret the complex visual patterns in sky images, resulting in reduced prediction accuracy, particularly during periods of cloudiness or seasonal variability. In this study, we introduce a novel deep learning model that enhances solar nowcasting by dynamically identifying and emphasising the most informative features across multiple visual dimensions. The model introduces a compact multi-scale CNN backbone specifically tailored to the spectral–spatial characteristics of sky images, ensuring efficient feature extraction under real-time constraints. A custom lightweight attention mechanism is embedded to enhance cloud–irradiance saliency detection, delivering transformer-like selectivity while retaining low parameter and computational cost. These features are coupled with a regularised deep regression head that integrates high-level interactions, forming a unified architecture that advances photovoltaic nowcasting by balancing accuracy, interpretability, and efficiency. The study used the novel Sky Images and Photovoltaic Power Generation Dataset provided by Stanford University to evaluate the proposed model using evaluation metrics such as RMSE, MSE, MAE, MAPE, and R2. The proposed model achieves an overall RMSE of 2.259 and R2 of 0.913, with particularly strong performance on sunny days (RMSE of 0.461, R2 of 0.996) and consistent results under cloudy conditions (RMSE of 3.158, R2 of 0.824). Seasonal analysis reveals that the model maintains robust accuracy across different climatic conditions, with channel attention excelling during high-irradiance summer and autumn days, and spatial attention effectively capturing complex cloud structures in winter and spring. These outcomes depict the model{\textquoteright}s ability to deliver more reliable short-term power forecasts by leveraging deeper visual understanding, ultimately contributing to more efficient solar energy management.",

keywords = "Attention Mechanism, Deep Learning, Photovoltaic Output Prediction, Sky Images, Solar Forecasting",

author = "Oluwatoyosi Bamisile and She Kun and Ukwuoma, \{Chiagoziem C.\} and Dongsheng Cai and Ukwuoma, \{Chibueze D.\} and Otuka, \{Chinedu I.\} and Ukwuoma, \{Chidera O.\} and Olusola Bamisile",

note = "Publisher Copyright: {\textcopyright} 2025 International Solar Energy Society.",

year = "2026",

month = jan,

doi = "10.1016/j.solener.2025.114117",

language = "English",

volume = "303",

journal = "Solar Energy",

issn = "0038-092X",

publisher = "Elsevier",

}

TY - JOUR

T1 - Enhancing photovoltaic power generation nowcasting with sky image analysis using multi-modal attention networks

AU - Bamisile, Oluwatoyosi

AU - Kun, She

AU - Ukwuoma, Chiagoziem C.

AU - Cai, Dongsheng

AU - Ukwuoma, Chibueze D.

AU - Otuka, Chinedu I.

AU - Ukwuoma, Chidera O.

AU - Bamisile, Olusola

PY - 2026/1

Y1 - 2026/1

N2 - The growing demand for renewable energy has heightened the importance of accurate photovoltaic (PV) power forecasting, especially under fluctuating weather conditions. Traditional and many deep learning-based models struggle to interpret the complex visual patterns in sky images, resulting in reduced prediction accuracy, particularly during periods of cloudiness or seasonal variability. In this study, we introduce a novel deep learning model that enhances solar nowcasting by dynamically identifying and emphasising the most informative features across multiple visual dimensions. The model introduces a compact multi-scale CNN backbone specifically tailored to the spectral–spatial characteristics of sky images, ensuring efficient feature extraction under real-time constraints. A custom lightweight attention mechanism is embedded to enhance cloud–irradiance saliency detection, delivering transformer-like selectivity while retaining low parameter and computational cost. These features are coupled with a regularised deep regression head that integrates high-level interactions, forming a unified architecture that advances photovoltaic nowcasting by balancing accuracy, interpretability, and efficiency. The study used the novel Sky Images and Photovoltaic Power Generation Dataset provided by Stanford University to evaluate the proposed model using evaluation metrics such as RMSE, MSE, MAE, MAPE, and R2. The proposed model achieves an overall RMSE of 2.259 and R2 of 0.913, with particularly strong performance on sunny days (RMSE of 0.461, R2 of 0.996) and consistent results under cloudy conditions (RMSE of 3.158, R2 of 0.824). Seasonal analysis reveals that the model maintains robust accuracy across different climatic conditions, with channel attention excelling during high-irradiance summer and autumn days, and spatial attention effectively capturing complex cloud structures in winter and spring. These outcomes depict the model’s ability to deliver more reliable short-term power forecasts by leveraging deeper visual understanding, ultimately contributing to more efficient solar energy management.

AB - The growing demand for renewable energy has heightened the importance of accurate photovoltaic (PV) power forecasting, especially under fluctuating weather conditions. Traditional and many deep learning-based models struggle to interpret the complex visual patterns in sky images, resulting in reduced prediction accuracy, particularly during periods of cloudiness or seasonal variability. In this study, we introduce a novel deep learning model that enhances solar nowcasting by dynamically identifying and emphasising the most informative features across multiple visual dimensions. The model introduces a compact multi-scale CNN backbone specifically tailored to the spectral–spatial characteristics of sky images, ensuring efficient feature extraction under real-time constraints. A custom lightweight attention mechanism is embedded to enhance cloud–irradiance saliency detection, delivering transformer-like selectivity while retaining low parameter and computational cost. These features are coupled with a regularised deep regression head that integrates high-level interactions, forming a unified architecture that advances photovoltaic nowcasting by balancing accuracy, interpretability, and efficiency. The study used the novel Sky Images and Photovoltaic Power Generation Dataset provided by Stanford University to evaluate the proposed model using evaluation metrics such as RMSE, MSE, MAE, MAPE, and R2. The proposed model achieves an overall RMSE of 2.259 and R2 of 0.913, with particularly strong performance on sunny days (RMSE of 0.461, R2 of 0.996) and consistent results under cloudy conditions (RMSE of 3.158, R2 of 0.824). Seasonal analysis reveals that the model maintains robust accuracy across different climatic conditions, with channel attention excelling during high-irradiance summer and autumn days, and spatial attention effectively capturing complex cloud structures in winter and spring. These outcomes depict the model’s ability to deliver more reliable short-term power forecasts by leveraging deeper visual understanding, ultimately contributing to more efficient solar energy management.

KW - Attention Mechanism

KW - Deep Learning

KW - Photovoltaic Output Prediction

KW - Sky Images

KW - Solar Forecasting

UR - https://www.scopus.com/pages/publications/105021088066

U2 - 10.1016/j.solener.2025.114117

DO - 10.1016/j.solener.2025.114117

M3 - Article

AN - SCOPUS:105021088066

SN - 0038-092X

VL - 303

JO - Solar Energy

JF - Solar Energy

M1 - 114117

ER -

Enhancing photovoltaic power generation nowcasting with sky image analysis using multi-modal attention networks

Abstract

Keywords

ASJC Scopus subject areas

UN SDGs

Access to Document

Other files and links

Fingerprint

Cite this