TLDR: OKG-LLM is a novel framework that significantly enhances global sea surface temperature (SST) prediction. It achieves this by constructing a unique Ocean Knowledge Graph (OKG) to represent diverse oceanographic knowledge and then effectively aligning and fusing this structured knowledge with fine-grained numerical SST data using large language models (LLMs). This approach addresses limitations of previous data-driven methods by integrating rich domain knowledge, leading to more accurate, robust, and efficient predictions across various forecasting lengths and LLM backbones.
Sea surface temperature (SST) prediction is a vital task in ocean science, impacting everything from weather forecasting and fisheries management to tracking storms. While current data-driven methods have shown success, they often miss out on using the vast amount of ocean knowledge gathered over decades. This oversight limits how accurate these predictions can be.
Recently, large language models (LLMs) have emerged, showing great potential for incorporating specialized knowledge into various tasks. However, applying LLMs to SST prediction has been challenging because it’s difficult to combine complex ocean domain knowledge with numerical data.
To tackle this, researchers have proposed a new framework called Ocean Knowledge Graph-enhanced LLM (OKG-LLM) for global SST prediction. This work introduces the first systematic effort to build an Ocean Knowledge Graph (OKG) specifically designed to represent diverse ocean knowledge relevant to SST prediction. The OKG captures both the unique characteristics of individual sea regions and the intricate connections between them.
How OKG-LLM Works
The OKG-LLM framework integrates structured ocean knowledge with detailed SST observation data. It consists of four main modules:
1. Time-series Encoding: This module processes raw SST data, normalizing it and segmenting it into patches to extract temporal features.
2. Knowledge Graph Encoding: This is where the Ocean Knowledge Graph (OKG) comes in. The OKG is built to include five critical types of entities related to SST variability: ocean currents, climatic zones, monsoon systems, geographic regions, and special oceanic areas. This module distills symbolic information from the OKG into low-dimensional representations, enriching the numerical data with structural and semantic knowledge. It uses both pre-trained entity embeddings to capture regional inter-correlations and a method to verbalize local graph structures into natural language prompts for fine-grained semantic representation.
3. LLM-Empowered Alignment: This crucial module aligns and merges the knowledge-based embeddings with the temporal embeddings. It uses a cross-attention mechanism to create unified, context-aware representations for each ocean region. These representations are then fed into a pre-trained LLM (like GPT-2 or Llama) to learn high-dimensional SST patterns.
4. Prediction Output Projection: Finally, a trainable transformer decoder refines the LLM’s output, modeling spatio-temporal dependencies, and a linear layer produces the final SST predictions.
Also Read:
- Empowering Service Robots with Lifelong Learning and Planning through Knowledge Graphs
- Efficient Anomaly Detection in Time Series Using TriP-LLM
Key Contributions and Performance
The development of OKG-LLM marks a significant step, being the first attempt to unify domain-specific ocean knowledge with observational data for SST forecasting. The creation of the Ocean Knowledge Graph (OKG) itself is a pioneering effort, providing a fine-grained, open-source knowledge graph to support ocean science. Extensive experiments on real-world global SST datasets show that OKG-LLM consistently outperforms nine state-of-the-art methods, demonstrating its effectiveness and robustness.
For instance, compared to TimeLLM, one of the best-performing LLM-based models, OKG-LLM shows a notable improvement in prediction accuracy across various prediction lengths. It also maintains its strong performance even with longer forecasting sequences, unlike many baseline methods that decline. Furthermore, OKG-LLM achieves this superior accuracy with manageable computational costs, striking a good balance between performance and efficiency.
A detailed study confirmed that each component of OKG-LLM—the Time Series Encoding, Knowledge Graph Encoding, and Fine-grained Alignment modules—is essential for its superior performance. The framework also proved to be adaptable, enhancing prediction performance regardless of the underlying LLM backbone used, including general models like GPT-2 and domain-specific ones like OceanGPT.
Visual comparisons of prediction errors across the globe show that OKG-LLM significantly reduces errors, especially in climate-sensitive areas like the El Niño-Southern Oscillation (ENSO) region. Predictions from OKG-LLM consistently align more closely with actual data in various oceanic regions, highlighting the benefit of integrating structured oceanographic knowledge.
This research demonstrates that combining domain knowledge with LLMs is a highly effective and robust approach for complex ocean prediction tasks. For more technical details, you can refer to the full research paper: OKG-LLM: Aligning Ocean Knowledge Graph with Observation Data via LLMs for Global Sea Surface Temperature Prediction.
