A Multi-Scheme Ensemble Using Coopetitive Soft-Gating With Application to Power Forecasting for Renewable Energy Generation

André Gensler, Bernhard SickarXiv 2018 · DOI · arXiv

Abstract In this article, we propose a novel ensemble technique with a multi-scheme weighting based on a technique called coopetitive soft gating. This technique combines both ensemble member competition and cooperation in order to maximize the overall forecasting accuracy of the ensemble. The proposed algorithm combines the ideas of multiple ensemble paradigms (power forecasting model ensemble, weather forecasting model ensemble, and lagged ensemble) in a hierarchical structure. The technique is designed to be used in a flexible manner on single and multiple weather forecasting models, and for a variety of lead times. We compare the technique to other power forecasting models and ensemble techniques with a flexible number of weather forecasting models, which can have the same, or varying forecasting horizons. It is shown that the model is able to outperform those models on a number of publicly available data sets. The article closes with a discussion of properties of the proposed model which are relevant in its application.

1. Introduction

Renewable energy has become the backbone of the modern electrical grid, yet its inherent intermittency presents significant challenges for grid operators. Because wind and solar power outputs are entirely dependent on meteorological conditions, accurate power forecasting is essential for maintaining grid stability and minimizing financial risks associated with energy imbalances. While various ensemble techniques have been developed to combine different forecasts, many struggle to adapt dynamically to changing weather patterns or varying lead times.

In this paper, we introduce the Coopetitive Soft-Gating Ensemble (CSGE). By blending the concepts of competition and cooperation, this model dynamically adjusts the influence of specific sub-models based on the current weather context and the forecasting horizon. This approach ensures that the system can leverage the strengths of "specialist" models during specific weather events while maintaining a robust baseline through "generalist" cooperation.

2. Methodology

The core innovation of the CSGE is its hierarchical architecture, which processes information from Numerical Weather Prediction (NWP) models through multiple power forecasting models. The term "coopetitive" refers to the mechanism where ensemble members compete for influence while simultaneously cooperating to produce a single, optimized final output.

The hierarchical architecture of the Coopetitive Soft-Gating Ensemble (CSGE), illustrating the flow from Numerical Weather Prediction (NWP) models to power forecasting models and final weighted aggregation.

The weighting of each model is determined by three distinct "soft-gating" aspects:

  1. Global Weighting: A fixed term based on the model’s overall historical performance. Stronger models receive a baseline advantage.
  2. Local Soft-Gating: This assesses the "local" quality of a model based on the current weather situation. Using a k-Nearest Neighbor (k-NN) approach, the system identifies how models performed in similar historical weather conditions (e.g., specific wind speeds or directions) and assigns more weight to historical specialists for that scenario.
  3. Lead Time-Dependent Soft-Gating: Models are weighted based on the forecasting horizon ($k$). This allows the ensemble to prioritize persistence models for very short-term (intraday) forecasts and shift toward complex machine learning models for day-ahead horizons.

3. Soft-Gating Dynamics

The normalization and combination of these weights are controlled by a hyperparameter $\eta$, which defines the "softness" of the gating. This parameter essentially decides how aggressively the system should favor the best-performing models (competition) versus averaging them out (cooperation).

The influence of the gating parameter η on model weighting relative to RMSE quality. Higher values of η increase competition by favoring models with lower error, while η=0 results in equal weighting (pure cooperation).

As shown in the figure above, a high $\eta$ value creates a steep weighting curve that heavily rewards models with low Root Mean Square Error (RMSE). Conversely, lower $\eta$ values lead to a more uniform distribution of weights, ensuring that the ensemble remains robust even if the top-performing model's accuracy fluctuates.

4. Experimental Evaluation and Results

The CSGE model was evaluated using the EuropeWindFarm collection, a dataset comprising 45 onshore and offshore wind farms over two consecutive years. The model was compared against standard techniques such as Bagging, Boosting, and simple Multi-Model Ensembles (MME).

Key findings include:

  • Single Weather Model Performance: For day-ahead forecasts using one weather source, CSGE variants achieved a 7.53% improvement in forecasting skill over the baseline.
  • Multi-Model Ensemble Performance: When combining three different weather models, the CSGE-M (multi-scheme) approach outperformed all competitors, showing an 8.78% improvement in RMSE compared to the best non-ensemble model.
  • Intraday Accuracy: In short-term scenarios where a persistence model was included, the CSGE+P approach outperformed the persistence baseline by 49.13% in RMSE.

5. Discussion and Conclusion

The CSGE framework offers several advantages for real-world application. Its "specialist advantage" ensures that rare but critical weather events are handled by the most capable models. Furthermore, the use of soft-gating ensures smooth transitions between models, avoiding the jittery forecasts associated with "hard-gating" selection methods.

While the computational complexity of the k-NN local quality assessment is a consideration for scaling, the system's inherent robustness—allowing it to redistribute weights if a weather model fails—makes it a powerful tool for modern grid management. Future work will focus on integrating uncertainty assessments and model pruning techniques to optimize computational efficiency.

Summary updated

Abstract

In this article, we propose a novel ensemble technique with a multi-scheme weighting based on a technique called coopetitive soft gating. This technique combines both, ensemble member competition and cooperation, in order to maximize the overall forecasting accuracy of the ensemble. The proposed algorithm combines the ideas of multiple ensemble paradigms (power forecasting model ensemble, weather forecasting model ensemble, and lagged ensemble) in a hierarchical structure. The technique is designed to be used in a flexible manner on single and multiple weather forecasting models, and for a variety of lead times. We compare the technique to other power forecasting models and ensemble techniques with a flexible number of weather forecasting models, which can have the same, or varying forecasting horizons. It is shown that the model is able to outperform those models on a number of publicly available data sets. The article closes with a discussion of properties of the proposed model which are relevant in its application.
BibTeX
@misc{gensler2018multi,
  author = {André Gensler and Bernhard Sick},
  title = {{A Multi-Scheme Ensemble Using Coopetitive Soft-Gating With Application to Power Forecasting for Renewable Energy Generation}},
  year = {2018},
  eprint = {1803.06344},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1803.06344},
}
Figure