Spatio-temporal graph neural networks for multi-site PV power forecasting (2107.13875v2)

Abstract: Accurate forecasting of solar power generation with fine temporal and spatial resolution is vital for the operation of the power grid. However, state-of-the-art approaches that combine machine learning with numerical weather predictions (NWP) have coarse resolution. In this paper, we take a graph signal processing perspective and model multi-site photovoltaic (PV) production time series as signals on a graph to capture their spatio-temporal dependencies and achieve higher spatial and temporal resolution forecasts. We present two novel graph neural network models for deterministic multi-site PV forecasting dubbed the graph-convolutional long short term memory (GCLSTM) and the graph-convolutional transformer (GCTrafo) models. These methods rely solely on production data and exploit the intuition that PV systems provide a dense network of virtual weather stations. The proposed methods were evaluated in two data sets for an entire year: 1) production data from 304 real PV systems, and 2) simulated production of 1000 PV systems, both distributed over Switzerland. The proposed models outperform state-of-the-art multi-site forecasting methods for prediction horizons of six hours ahead. Furthermore, the proposed models outperform state-of-the-art single-site methods with NWP as inputs on horizons up to four hours ahead.

• The paper introduces two novel graph-based models (GCLSTM and GCTrafo) that leverage PV data to capture spatio-temporal dependencies.
• It integrates graph convolutions within LSTM and transformer architectures, yielding superior short-term forecasting performance.
• The study demonstrates improved precision and scalability for grid management by eliminating reliance on complex meteorological data.

Spatio-Temporal Graph Neural Networks for Multi-site PV Power Forecasting

The paper "Spatio-temporal graph neural networks for multi-site PV power forecasting" introduces two novel graph-based neural network models that aim to improve the accuracy of solar power generation forecasts across multiple photovoltaic (PV) sites. Traditional methods for solar power forecasting, which typically rely on machine learning combined with numerical weather predictions (NWP), lack fine temporal and spatial resolution. This paper proposes graph neural network approaches to leverage spatio-temporal dependencies inherent in PV sites, using PV production data alone to achieve higher forecasting precision without relying on costly and complex meteorological data.

Graph Signal Processing Perspective

The authors adopt a graph signal processing (GSP) approach, treating PV production data from multiple sites as signals on a graph. The graph structure naturally represents spatial relationships between PV installations, allowing the models to exploit spatial patterns and time series dependencies. This approach assumes that PV installations form a network of virtual weather stations, capturing atmospheric dynamics due to cloud cover influence on solar energy output. The models introduced are Graph-Convolutional Long Short Term Memory (GCLSTM) and Graph-Convolutional Transformer (GCTrafo), which incorporate graph convolutional layers within their architectures to enhance forecasting capabilities.

Models and Methodologies

Graph-Convolutional Long Short Term Memory (GCLSTM):

The GCLSTM model uses an encoder-decoder mechanism where the encoder, a GCLSTM network, processes past observed data, while the decoder generates forecasts for future time steps. By integrating graph convolution operations in the LSTM units, the model effectively captures complex spatio-temporal dependencies.

Graph-Convolutional Transformer (GCTrafo):

Inspired by the transformer architecture, GCTrafo employs attention mechanisms to maintain information across longer sequences, addressing the issue of fading memory in recurrent neural networks. Attention layers preceded by graph convolutions help embed node features within their spatial and temporal context, improving long-range forecasting performance.

Evaluation

The models were evaluated against state-of-the-art techniques using real and synthetic datasets representing PV installations distributed across Switzerland. The datasets used had high temporal resolution, allowing for a detailed comparative analysis between traditional and proposed methods. Key findings indicate that GCLSTM and GCTrafo outperform existing multi-site methods, particularly for shorter time horizons. Notably, GCLSTM demonstrates superior accuracy for forecasts up to four hours ahead, while GCTrafo shows promise for extended predictions.

Implications and Future Directions

The paper highlights the potential of graph neural networks in enhancing PV power forecasting without reliance on external meteorological data, which is often costly and complex to manage. The scalability of these models suggests significant applicability for grid management systems dealing with numerous PV sites. However, the authors acknowledge the limitations at longer time horizons, where competing models incorporating NWP inputs show reduced error rates.

The robustness and adaptability of the models to incomplete datasets and new PV site additions remain areas for exploration. Future research may involve probabilistic extensions of the models, offering a more comprehensive approach to uncertainty in PV production forecasts. As renewable energy systems integrate more deeply with power grids, advancements inspired by this work will be crucial for efficient energy management and increased solar energy penetration.

In conclusion, the research presents a compelling case for incorporating graph-based methodologies in solar forecasting, complementing existing machine learning and NWP techniques while pushing the boundaries of what's possible using PV production data alone.