[Paper Review] Transfer Learning in the Field of Renewable Energies -- A Transfer Learning Framework Providing Power Forecasts Throughout the Lifecycle of Wind Farms After Initial Connection to the Electrical Grid

Jens Schreiber|arXiv (Cornell University)|Jun 3, 2019

Domain Adaptation and Few-Shot Learning15 references5 citations

TL;DR

This paper proposes a transfer learning framework for wind power forecasting that enables accurate predictions throughout a wind farm's lifecycle—even with limited or no historical data—by leveraging knowledge from similar wind farms and weather models. The approach uses ensemble models, multi-task learning, and self-optimizing mechanisms to adapt forecasts dynamically, outperforming traditional physical models and enabling automated, scalable deployment in renewable energy systems.

ABSTRACT

In recent years, transfer learning gained particular interest in the field of vision and natural language processing. In the research field of vision, e.g., deep neural networks and transfer learning techniques achieve almost perfect classification scores within minutes. Nonetheless, these techniques are not yet widely applied in other domains. Therefore, this article identifies critical challenges and shows potential solutions for power forecasts in the field of renewable energies. It proposes a framework utilizing transfer learning techniques in wind power forecasts with limited or no historical data. On the one hand, this allows evaluating the applicability of transfer learning in the field of renewable energy. On the other hand, by developing automatic procedures, we assure that the proposed methods provide a framework that applies to domains in organic computing as well.

Motivation & Objective

• To address the challenge of limited or no historical power generation data in new or modified wind farms, which hinders effective machine learning-based forecasting.
• To develop an automated, self-optimizing transfer learning framework applicable to renewable energy forecasting across the entire lifecycle of wind farms.
• To enable continuous adaptation to new conditions such as night-time turbine shut-offs or changes in numerical weather prediction models.
• To extend transfer learning beyond vision and NLP into organic computing and energy systems by creating a reusable, adaptive framework.
• To improve forecast accuracy over traditional physical models by leveraging knowledge from similar wind farms and weather data sources.

Proposed method

• Utilizes a Cross-Source Generalized Ensemble (CSGE) model that combines predictions from multiple wind farms to improve generalization and adaptability.
• Employs multi-cross-task learning (MCL) with a spatial abstraction layer to learn generic representations of both NWP inputs and spatial relationships between wind farms.
• Applies transfer learning by fine-tuning models on new data, allowing the system to adapt to novelties such as night-time shut-offs or new weather models.
• Integrates probabilistic forecast analysis and sensitivity analysis to pre-select the most influential features (e.g., wind speed at 100m and 10m altitude) for model input.
• Uses automatic configuration and self-improvement procedures to reweight models based on performance and data availability, enabling continuous optimization.
• Employs a loss function-optimized CSGE that allows dynamic reconfiguration and adaptation to new data and conditions without full retraining.

Experimental results

Research questions

• RQ1 How can historical source domain data and forecast models be effectively transferred to a new wind farm with limited or no historical data?
• RQ2 What types of historical source domain data and forecast models are most suitable for transfer in wind power forecasting?
• RQ3 How can the framework automatically integrate increasing amounts of historical data into the model for a new wind farm over time?
• RQ4 How can the system self-configure and self-optimize as new data and conditions (e.g., night shut-offs) emerge?
• RQ5 Can multi-task learning with spatial and NWP abstraction layers enable robust, generalizable forecasting across diverse wind farms and weather models?

Key findings

• Wind speed at 100m altitude is the most influential feature in power forecasting across all evaluated terrains and machine learning models.
• Wind speed at 10m altitude ranks second in influence, with air pressure and wind speed variability at 100m also showing significant impact.
• The CSGE approach achieves state-of-the-art performance in both classification and regression tasks, outperforming standard ensemble methods on real-world datasets.
• The framework enables automatic adaptation to new conditions such as night-time turbine shut-offs through retraining and fine-tuning of the ensemble model.
• Multi-cross-task learning with spatial and NWP abstraction layers allows transfer of knowledge across different weather models and wind farm locations.
• Initial results confirm that transfer learning-based forecasting significantly outperforms traditional physical models in forecast accuracy, even with minimal data.