TL;DR
This paper introduces an adaptive Graph Neural Network (GNN) based hybrid model for long-term photovoltaic (PV) power forecasting. By dynamically learning relationships between PV generation and environmental factors, the model significantly improves prediction accuracy, especially for longer horizons, which is crucial for grid stability and renewable energy integration.
What Problem Does This Paper Address?
Accurate forecasting of photovoltaic (PV) power generation is essential for integrating renewable energy into the grid, ensuring real-time grid balancing, and optimizing energy storage. However, PV output is highly sensitive to environmental factors like weather, making precise prediction a significant challenge. Existing methods often struggle to effectively capture the complex, spatiotemporal interdependencies between PV generation and these environmental variables, particularly for long-term forecasts. This paper aims to bridge this gap by developing a model that can dynamically learn and leverage these intricate relationships.
Methods
The core of this work is a hybrid deep learning model featuring an Adaptive Graph Neural Network (GNN). The GNN is designed to capture latent relationships between PV power generation and various environmental variables. Key components include:
- Data-Driven Directed Graph Structure: Instead of relying on predefined relationships, the GNN learns an adjacency matrix that represents the connections between variables. This allows for a dynamic and adaptive understanding of interdependencies.
- TCN and Variable Interaction Layers: The architecture alternates Temporal Convolutional Network (TCN) layers for capturing temporal patterns with layers that model variable interactions. This dual approach is designed to better model the spatiotemporal coupling of data.
- Adaptive Graph Construction: A graph construction layer generates an initial adjacency matrix, which is then refined through adaptive learning. This process aims to capture hidden relationships among nodes (variables) in multivariate time series.
- Forecasting Module: A lightweight Multi-Layer Perceptron (MLP) network processes the GNN-enhanced data. It employs both continuous and interval sampling strategies to extract long-term trends and fine-grained local variations, enabling multi-horizon predictions.
Datasets
The study utilized PV power plant datasets collected via a SCADA system from three solar power plants located in North, Central, and Northwest China. These plants have installed capacities ranging from 30 MW to 200 MW. The dataset comprises power generation measurements sampled at 15-minute intervals over a 24-month period. The data was partitioned into training (70%), testing (20%), and validation (10%) sets. All three sites shared identical feature sets, including:
- Total solar irradiance (TSI)
- Direct normal irradiance (DNI)
- Global horizontal irradiance (GHI)
- Air temperature (AT)
- Air pressure (AP)
- Relative humidity (RH)
- Photovoltaic power generation (P)
Key Insights and Ideas
- The Power of Dynamic Relationships: The study highlights that adaptively learning the relationships between PV generation and environmental factors, rather than using static or predefined connections, significantly boosts forecasting accuracy. This is particularly true for longer prediction horizons where these relationships can become more complex.
- Spatiotemporal Synergy is Key: Models that simultaneously capture both spatial (inter-variable) and temporal dependencies consistently outperform those focusing on only one dimension. This underscores the interconnected nature of PV power generation and its influencing factors.
- Directed Graphs for Unidirectional Influence: For longer-term forecasting, where environmental factors influence PV output but not vice-versa, a directed graph structure proved more effective than an undirected one. This allows the model to better represent the causal flow of influence.
- Optimal Variable Selection Matters: The number of input variables, controlled by a correlation coefficient 'k', is critical. Too few variables lead to underutilization of information, while too many can introduce noise and degrade performance, especially for longer horizons. An optimal 'k' (found to be 4 in this study) balances these factors.
Key Results
The proposed adaptive GNN-based hybrid model demonstrated superior performance in long-term PV power forecasting:
- Accuracy Improvements: At a 384-step horizon, the model achieved average improvements of 2.19% in MSE and 1.57% in MAE compared to the best baseline. For the 768-step horizon, these improvements were 2.81% in MSE and 2.47% in MAE.
- Robustness: The model consistently ranked among the top performers across all prediction horizons and sites, showing strong robustness.
- Long-Term Forecasting Prowess: Visualizations of predicted vs. true power curves for the 768-step horizon showed the proposed model's exceptional ability to capture peaks, troughs, and fluctuations, aligning more closely with actual power variations than baseline models.
- Computational Efficiency: While the proposed model had a relatively larger parameter count and memory footprint, its computational complexity (MACs) and iteration time per training step were competitive, ranking favorably against many advanced models.
Limitations and Future Work
The authors acknowledge a key limitation: the dynamic adjacency matrix employed by the adaptive GNN, while beneficial, might lack sufficient stability. This could potentially limit its ability to fully capture and explain the complex relationships between PV power generation and environmental factors.
Future work will focus on:
- Improving Interpretability: Further exploring the interpretability of graph-based models to better understand the learned relationships.
- Optimizing PV Installation: Leveraging these insights to facilitate optimal PV installation planning by comprehensively incorporating environmental information, aiming to maximize power generation efficiency.
Summary updated