Spatiotemporal Graph Neural Network for Performance Prediction of Photovoltaic Power Systems

Ahmad Maroof Karimi, Yinghui Wu, Mehmet Koyutürk, Roger H. FrenchAAAI 2021 · OpenReview

Abstract

Accurate forecasting of photovoltaic (PV) power performance is essential for maintaining electrical grid reliability as solar deployment continues to rise. However, the complex non-linear variability of PV output makes forecasting a significant challenge. This paper presents a spatiotemporal graph neural network (st-GNN) that leverages the spatial and temporal coherence among power plants. By treating a network of 316 PV systems as a graph, the model utilizes regional environmental similarities to improve predictions. Experimental results demonstrate that this spatiotemporal approach significantly outperforms traditional temporal-only models, particularly as the forecast horizon extends.

Introduction

The rapid integration of photovoltaic systems into global energy grids introduces substantial variability due to environmental factors like cloud cover and temperature. Effective grid management requires precise power output predictions to mitigate uncertainty. Traditional forecasting methods often treat PV systems as isolated entities, ignoring the valuable information shared between geographically proximate plants. If one plant experiences a drop in output due to a passing cloud, its neighbors are likely to follow. This research addresses this gap by modeling PV systems as a connected network, capturing both the temporal dynamics of individual sites and the spatial correlations across the region.

Methodology: The st-GNN Architecture

The core of the proposed solution is a spatiotemporal graph neural network (st-GNN) designed to process the multi-dimensional dependencies of solar power data.

Model Structure

The architecture consists of multiple spatiotemporal (ST) blocks. Each block features a spatial graph convolution layer "sandwiched" between two temporal convolutional layers. This design allows the network to first extract temporal features, then propagate those features across the spatial graph to capture regional trends, and finally refine the temporal output.

The st-GNN architecture featuring sequential spatiotemporal blocks with temporal and spatial convolution layers.

Graph Construction and Sparsity

The network of PV plants is represented as a graph where each plant is a node. Edges between nodes are defined by geographical proximity, calculated using a Gaussian kernel and the Haversine formula. A critical finding of the study was the importance of graph sparsity. By applying a threshold ($\epsilon$), the researchers could filter out distant, "noisy" connections, focusing the model on meaningful regional clusters.

Visual representation of graph connectivity: (a) isolated nodes, (b) a fully connected graph, and (c) a sparse graph optimized for regional coherence.

Local Solar Noon Alignment

To improve model performance, the authors introduced a preprocessing step: aligning all time series data to "local solar noon." Because plants at different longitudes experience peak sunlight at different clock times, standardizing the data to the sun's position removes artificial asynchronicity, allowing the neural network to focus on the underlying physical patterns of solar generation.

Key Results

The st-GNN was evaluated using a dataset of 316 PV systems across the United States over a two-year period. The results indicate a substantial improvement in accuracy over 1-D temporal convolution models:

  • Accuracy Boost: For a 120-minute forecast, the st-GNN achieved a Mean Absolute Percentage Error (MAPE) of 11.01%, a significant reduction from the 18.98% error seen in temporal-only models.
  • Forecast Horizon: The benefit of the spatiotemporal approach becomes more pronounced as the prediction window increases. While short-term "nowcasts" (15 minutes) show competitive results, the st-GNN's ability to "see" approaching weather patterns through neighboring nodes gives it a clear advantage for 1-to-2-hour horizons.
  • Scalability: The model's performance improved when the number of systems in the network increased from 77 to 316, suggesting that denser sensor networks provide richer spatial information for the GNN to exploit.

Conclusion

By integrating spatial relationships into the forecasting pipeline, the st-GNN provides a more robust and accurate tool for PV power prediction. The combination of graph convolutions, sparsity optimization, and solar noon alignment allows the model to capture the complex physical realities of solar energy production. Future work will explore the integration of additional meteorological variables and the development of dynamic graph structures to account for real-time weather movements like wind-driven cloud patterns.

Summary updated

Abstract

In recent years, a large number of photovoltaic (PV) systems have been added to the electrical grid as well as installed as off-grid systems. The trend suggests that the deployment of PV systems will continue to rise in the future. Thus, accurate forecasting of PV performance is critical for the reliability of PV systems. Due to the complex non-linear variability in power output of the PV systems, forecasting PV power is a non-trivial task. This variability affects the stability and planning of a power system network, and accurate forecasting of the performance of the PV system can reduce the uncertainty caused during PV operation. In this work, we leverage spatial and temporal coherence among the power plants for PV power forecasting. Our approach is motivated by the observation that power plants in a region undergo similar environmental exposure. Thus, one power plant’s performance can help improve the forecast of other power plants' power values in the region. We utilize the relationship between PV plants to build a spatiotemporal graph neural network (st-GNN) and train machine learning models to forecast the PV power. The computational experiments on large-scale data from a network of 316 systems show that spatiotemporal forecasting of PV power performs significantly better than a model that only applies temporal convolution to isolated systems or nodes. Furthermore, the longer the future forecast time, the difference between the spatiotemporal forecasting and the isolated system forecast when only temporal convolution is applied increases further.
BibTeX
@inproceedings{karimi2021spatiotemporal,
  author = {Ahmad Maroof Karimi and Yinghui Wu and Mehmet Koyutürk and Roger H. French},
  title = {{Spatiotemporal Graph Neural Network for Performance Prediction of Photovoltaic Power Systems}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2021},
  url = {https://openreview.net/forum?id=pTUB2KfexK},
}
Figure