Spatio-Temporal Graph Neural Network

The foundational step involves representing wind farms and solar plants as nodes in a graph, connected by edges that encode spatial relationships. These edges are not merely geographic distances; they are weighted by correlation coefficients derived from historical generation data, topographic similarity, and meteorological connectivity. This creates a dynamic adjacency matrix where two geographically distant sites might be strongly connected if they are subject to the same weather front, enabling the model to learn how a cloud bank moving through one node will sequentially impact downstream nodes.

At each individual node, a Temporal Convolutional Network (TCN) or Long Short-Term Memory (LSTM) layer processes the historical time series of local generation, irradiance, and wind speed. Unlike standard recurrent networks, TCNs use dilated causal convolutions to capture hierarchical temporal patterns—from minute-by-minute ramp events to diurnal cycles—without suffering from vanishing gradients. This encoder compresses the raw temporal sequence into a latent feature vector that represents the current state of the asset, ready to be shared with its neighbors.

This is the core spatial learning mechanism. In each message-passing layer, every node aggregates the latent state vectors from its connected neighbors using a Graph Convolutional Network (GCN) or Graph Attention Network (GAT). A GAT variant is often preferred because it learns to assign dynamic attention weights to different neighbors, allowing a solar farm to automatically prioritize information from an upwind cloud sensor over a distant, uncorrelated site. This iterative propagation ensures that each node's representation is informed by the broader atmospheric system.

The fusion block interleaves temporal encoding and spatial message passing in a sandwich architecture. A typical block first applies a temporal convolution to capture local trends, followed by a graph convolution to exchange spatial information, and finally another temporal layer to process the newly received spatial context. Stacking multiple fusion blocks allows the network to learn multi-scale spatio-temporal features: early layers capture fast-moving cloud shadows affecting neighboring sites, while deeper layers model synoptic-scale weather systems propagating across entire regions.

The final layer transforms the enriched node embeddings into a probabilistic forecast rather than a single deterministic value. This is achieved by training the decoder with a Pinball Loss function to output specific quantiles (e.g., P10, P50, P90) of the future power distribution. For grid operators, this quantifies the uncertainty: a wide prediction interval signals a volatile weather regime requiring higher operating reserves, while a narrow interval indicates high confidence. The decoder can also output parameters of a parametric distribution, such as a Gaussian mixture, to capture complex, multi-modal forecast scenarios.

Instead of relying on a static, pre-computed adjacency matrix based on physical distance, advanced ST-GNNs learn the graph structure itself during training. An adaptive adjacency matrix is parameterized as a learnable embedding for each node, and the edge weight between any two nodes is computed as the similarity of their embeddings. This allows the model to discover latent spatial dependencies—such as two wind farms in different valleys that share a correlated response to a large-scale pressure gradient—that are not captured by static geographic or meteorological priors, significantly improving forecast accuracy.

The foundational spatial aggregation layer that operates directly on graph-structured data. A GCN updates a node's representation by averaging and transforming the feature vectors of its immediate neighbors.

• Message Passing: The core mechanism where nodes exchange information along edges.
• Adjacency Matrix: Defines the physical or correlational connectivity between renewable generation sites.
• Spectral vs. Spatial: GCNs can be defined via graph Fourier transforms or direct neighbor aggregation.

An architecture that integrates recurrent units like LSTMs or GRUs into the node update function. This allows the model to capture temporal dynamics at each sensor location while simultaneously aggregating spatial information from neighbors.

• Node-Level RNN: Each node in the graph maintains its own hidden temporal state.
• Diffusion Convolution: Often paired with recurrent units to handle multi-step spatial dependencies.
• Sequence-to-Sequence: Enables multi-step ahead forecasting of wind speed or solar irradiance.

Replaces fixed adjacency matrices with learned, dynamic attention weights. The model automatically discovers which neighboring wind farms or solar plants are most relevant for predicting a specific site's output at a given time.

• Graph Attention Network (GAT): Computes attention coefficients between node pairs.
• Dynamic Dependencies: Captures changing correlations due to weather patterns moving across a region.
• Multi-Head Attention: Stabilizes learning by allowing the model to attend to different spatial relationships simultaneously.

A parallelizable alternative to recurrent layers for capturing temporal patterns. TCNs use dilated causal convolutions to exponentially increase the receptive field, allowing the model to look far back in time without the vanishing gradient issues of RNNs.

• Causality: Ensures predictions only depend on past time steps, preventing data leakage.
• Dilation: Enables efficient modeling of long-range temporal dependencies in wind ramps.
• Residual Connections: Stabilize training of very deep temporal stacks.

The critical pre-processing step of defining the graph's topology. Nodes represent physical assets like turbines or pyranometers, while edges encode relationships.

• Distance-Based: Connect k-nearest neighbors based on geographic proximity.
• Correlation-Based: Connect sites with highly correlated historical generation profiles using Pearson correlation.
• Hybrid Graphs: Combine physical distance and data-driven similarity to create a robust adjacency structure.

Advanced architectures that capture cross-space-time correlations simultaneously, rather than processing space and time in separate sequential modules. This is crucial for modeling the propagation of weather fronts.

• Spatio-Temporal Graph: A unified graph where edges connect nodes across both spatial and temporal dimensions.
• Synchronized Convolution: Applies filters that slide across both the spatial graph and the time axis concurrently.
• Propagation Physics: Better captures the finite travel time of cloud shadows and wind gusts between sites.