PVformer: end-to-end transferable day-ahead PV forecasting framework for new PV stations with few-shot historical data

Summarise with VIRAT
Authors: Weican Liu
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Accurate day-ahead photovoltaic (PV) power forecasting is critical for stable grid operation, yet newly established stations often suffer from poor prediction performance due to limited historical data. To address this few-shot challenge, this paper proposes PVformer, an accurate end-to-end transferable forecasting framework. First, a multi-stage data preprocessing module is designed to enhance data quality and usability. This module utilizes the Maximum Mean Discrepancy (MMD) algorithm to select optimal pre-training datasets, applies the HI algorithm for outlier correction, and employs ReliefF for efficient feature selection. Second, we construct a customized end-to-end deep learning model comprising two independent encoders and a fusion decoder. This architecture effectively captures spatiotemporal characteristics and minimizes the error accumulation often found in non-end-to-end structures. The framework leverages a pre-training and fine-tuning strategy to transfer knowledge from data-rich source domains to the target station. Furthermore, a novel time-frequency domain fusion loss function is introduced. By combining penalty-weighted Mean Squared Error (MSE) with Fast Fourier Transform (FFT)-based frequency penalties, the model achieves robust dual-domain optimization. Experimental results demonstrate that PVformer significantly improves forecasting accuracy for newly established PV stations.

Fields of Study & Subjects

Solar Radiation and Photovoltaics Energy Load and Power Forecasting Photovoltaic System Optimization Techniques

Publication details

Journal / source
DR-NTU (Nanyang Technological University)
Publication date
1 January 2026
Language
EN
Document type
Thesis