Meta-PV

Official implementation of Meta-PV, a robust meta-learning framework for day-ahead photovoltaic power forecasting under limited observations.

Overview

Meta-PV addresses rapid adaptation to previously unseen PV sites by combining task-level meta-learning with inequality-aware task-loss regularization based on the Theil index.

Supported Methods

• Transfer Learning (TL)
• Model-Agnostic Meta-Learning (MAML)
• Task-Agnostic Meta-Learning (TAML) with task-loss variance regularization
• Task-Agnostic Meta-Learning (TAML) with softmax-based task reweighting
• Task-Agnostic Meta-Learning (TAML) with Theil index regularization

Supported Models

• MLP
• CNN
• Transformer
• GNN

Datasets

• NLR Solar Power Data for Integration Studies
• UK-PV dataset

Dependencies

Separate dependency files are provided for the NLR and UK-PV experiments. The NLR implementation uses torchmeta, whereas the GNN-based UK-PV experiments require a custom meta-learning implementation because the GNN architecture is not directly supported by torchmeta.

• requirements_nlr.txt
• requirements_ukpv.txt

Installation

Clone the repository:

git clone https://github.com/<username>/Meta-PV.git
cd Meta-PV