CSFS: More Efficient Feature Selection for Wind and Solar Power Forecasting

Introduction

Wind and solar power are becoming central to modern energy systems, but their output is tightly coupled to environmental conditions such as wind speed, irradiance and temperature. Unlike conventional generation, renewable production cannot be controlled with the same consistency. This makes reliable prediction essential for grid operation, energy trading, storage planning and the broader integration of clean power.

The arXiv paper “Improving Wind and Solar Power Prediction with Efficient Wrapper-based Feature Selection: An Empirical Study” focuses on a less visible but important part of the forecasting pipeline: choosing which input variables should be used in the first place.

Key points

• Two practical renewable forecasting tasks: The paper covers wind turbine power curve modeling and photovoltaic power prediction. For the wind domain, the authors conducted their own structured literature review; for photovoltaics, they synthesize findings from an existing survey on frequently selected input features.
• Feature selection remains underdeveloped: Although monitoring systems and environmental data sources can provide many variables, the study finds that feature selection is often limited or unsystematic. This can increase computational burden and introduce redundant or weakly useful inputs.
• A new method: CSFS: The authors propose Cluster-based Sequential Feature Selection, a clustering-based wrapper method designed to be model-agnostic. In other words, it is not tied to a specific forecasting model and can be inserted into different renewable energy prediction workflows.
• Performance with lower cost: The paper compares CSFS against established approaches, including wrapper-based sequential feature selection, filter-based methods and the embedded feature importance of Random Forests. The reported results show that wrapper-based methods generally produce better-performing feature subsets. CSFS achieves predictive performance comparable to standard SFS while reducing computational cost by 21% on average.
• Open-source implementation: The authors state that an implementation of CSFS is available on GitHub, supporting reproducibility and reuse.

Why it matters

The main contribution of this work is not a new forecasting model, but a more disciplined way to prepare the inputs that models depend on. As wind farms and solar plants collect more sensor readings, weather variables and operational signals, the number of candidate features can grow quickly. Feeding every available variable into a model may appear convenient, but it can make training slower, pipelines harder to maintain and results less robust.

CSFS is positioned as a practical preprocessing step: it aims to preserve the benefits of wrapper-based feature selection while reducing part of the computational overhead. For operators, this could support leaner prediction systems under constrained compute budgets. For researchers, it highlights the need to report not only model architectures and error metrics, but also how input variables are chosen.

The findings should still be tested across more sites, climates and forecasting horizons. Renewable energy data is highly context-dependent, and a feature subset that works well in one setting may not transfer directly to another. Even so, the study points to a useful lesson: before building ever more complex models, it may be worth choosing the data more intelligently.

Source: arXiv