Forecasting Intraday Power Output by a Set of PV Systems using Recurrent Neural Networks and Physical Covariates

Overview

• Accurate forecasts of power output from photovoltaic (PV) systems are crucial for efficient energy grid operations.
• The researchers describe a neural autoregressive model designed for intraday PV power forecasting.
• The model incorporates a physical PV performance model as input features.
• The model addresses multiple geographically distributed PV sites using a single neural network.
• The approach uses a scale-free approach to model seasonal effects.
• The model was compared to alternative architectures and showed a 15.72% improvement over the baseline physical model.

Plain English Explanation

Solar power is becoming an increasingly important part of our energy grid, but accurately predicting how much power solar panels will generate moment-to-moment is challenging. This is a problem because grid operators need to know how much power will be available to properly balance supply and demand.

The researchers in this paper developed a neural network model that can forecast the power output of solar panels throughout the day. Their key innovation was to use a physical model of how solar panels work as one of the inputs to the neural network. This helped the model learn the underlying relationship between weather conditions and power generation.

Additionally, the researchers used a technique called "scale-free modeling" to better capture the seasonal patterns in solar power generation. This allowed the model to work across multiple solar sites without needing to be retrained for each location.

When tested, the researchers' model outperformed a baseline physical model by 15.72%. This shows that the combination of physical modeling and scale-free neural networks can lead to significant improvements in solar power forecasting.

Technical Explanation

The core of the researchers' approach is a neural autoregressive model for intraday PV power forecasting. This model builds upon a physical, deterministic PV performance model, using its output as input features for the neural network.

Importantly, the researchers used a single neural network to address multiple geographically distributed PV sites. They achieved this by embedding information about the specific PV site as covariates in the model. This allowed the model to learn patterns across sites without needing to be retrained for each location.

The researchers also adopted a scale-free modeling approach, which relies on the explicit modeling of seasonal effects. This helps the model generalize better and capture the cyclical nature of solar power generation.

Additionally, the researchers repurposed a model architecture initially developed for the retail sector, and introduced a novel truncated Gaussian output distribution. This output distribution is well-suited for modeling the bounded nature of PV power generation.

Through an ablation study and comparison to alternative architectures, the researchers demonstrated that the components of their proposed model variant work synergistically to achieve a 15.72% improvement in skill score over the baseline physical model.

Critical Analysis

The researchers acknowledge several limitations and areas for further research in their work. For example, they note that their model does not explicitly account for the impact of cloud cover, which can significantly affect PV power generation. Incorporating more granular weather data or additional physical modeling of cloud dynamics could potentially improve the model's accuracy.

Moreover, the researchers' evaluation was limited to a single geographical region. Testing the model's performance on a more diverse set of PV sites would help validate its scalability and generalization capabilities.

Another potential area for improvement is the incorporation of additional data sources, such as historical PV power generation data or information about the specific PV system configurations. This could further enhance the model's ability to capture the complex relationships between environmental factors and power output.

Finally, the researchers' use of a truncated Gaussian output distribution, while novel, may not be the optimal choice for all PV power forecasting scenarios. Exploring alternative output distributions or even more flexible modeling approaches, such as deep generative models, could lead to further performance gains.

Conclusion

This research demonstrates the potential of combining physical modeling and scale-free neural networks for accurate intraday forecasting of PV power generation. By leveraging a deterministic PV performance model as input features and adopting a single-model approach for multiple sites, the researchers were able to achieve a significant improvement over a baseline physical model.

The insights from this work have important implications for the integration of solar power into energy distribution grids. Accurate and reliable forecasts can help grid operators better manage the intermittent nature of solar generation, leading to more efficient and stable energy supply. As solar power continues to play a growing role in our energy mix, advancements in PV power forecasting will become increasingly crucial for the transition to a sustainable energy future.