Step-adaptive multimodal fusion network with multi-scale cloud feature learning for ultra-short-term solar irradiance forecasting
Step-adaptive multimodal fusion network with multi-scale cloud feature learning for ultra-short-term solar irradiance forecasting
Authors: Jingxin Zhang, Xiaoqin Wang Date: 2026-06-04 Paper ID: openalex:2606.06102
Summary
This paper addresses the challenges of ultra-short-term solar irradiance forecasting by proposing a multimodal fusion network that integrates ground-based cloud images with meteorological time-series data. The framework utilizes InceptionNeXt for multi-scale spatial feature extraction and a novel step-adaptive low-frequency compensation unit to adjust global information dynamically based on the forecast horizon. Combined with a TempAttnLSTM for capturing temporal dependencies, the method provides improved prediction accuracy compared to existing state-of-the-art models on NREL and practical datasets.
Key Contributions
- Proposes a multi-source fusion model using InceptionNeXt for multi-scale spatial cloud feature extraction.
- Introduces a step-adaptive low-frequency compensation unit to dynamically modulate global information across different forecast horizons.
- Demonstrates superior performance over state-of-the-art methods for ultra-short-term solar irradiance forecasting on NREL and real-world photovoltaic station data.
Key Concepts
- step-adaptive-low-frequency-compensation: A mechanism that dynamically modulates global frequency components conditioned on the specific forecast step in time-series prediction.
Archivist Review
I approved the ‘Step-adaptive Low-Frequency Compensation’ concept as it represents a novel and reusable inductive bias for time-series forecasting that specifically addresses the horizon-dependent nature of spectral features. The NREL dataset was approved as a core evaluation benchmark. Other proposed elements were rejected as being either too specific to this implementation (e.g., InceptionNeXt/TempAttnLSTM integration) or insufficiently novel as standalone contributions.
Approved Concepts
- Step-adaptive Low-Frequency Compensation: It addresses the failure of fixed frequency compensation strategies in ultra-short-term forecasting by allowing modulation tuned to the specific prediction horizon.
Datasets
Links
Metadata & Links
- url
- https://arxiv.org/abs/2606.06102
- paper_id
- 2606.06102
- paper_source
- openalex
- domain
- time-series
- tags
-
- architectures
-
- datasets
- NREL dataset
- concept_slugs
- step-adaptive-low-frequency-compensation
- dataset_slugs
- nrel-dataset
- skill
- TimeSeriesSkill
- created_at
- 2026-06-07T06:32:27Z