Introduction
Short-term weather forecasting (3-18 hours) is often described as a "black art." While meteorologists have access to sophisticated thermodynamic models and statistical predictors, these tools are frequently insufficient for localized, high-impact events like severe convective storms. Experienced forecasters rely heavily on "local expertise"—heuristics regarding how specific geographical features or subtle qualitative changes in cloud cover influence storm development.
The primary challenge in automating this expertise lies in the volume and variety of data. A robust system must handle quantitative measurements (temperature, pressure), graphical data (contour maps), and qualitative human observations (station reports). Before the development of METEOR, much of this qualitative data remained unanalyzed because there was no efficient way to codify it for statistical models. This paper describes METEOR, a system designed to "understand" the state of the world by integrating diverse data types into a unified knowledge representation.
Methods and Knowledge Representation
METEOR integrates two primary AI formalisms to bridge the gap between raw data and heuristic expertise:
- Hybrid Representation: The system uses frames (implemented in Interlisp-D) to represent hierarchical domain knowledge and condition-action rules (via OPS4) for strategic reasoning and event matching.
- The "View" Scheme: A novel extension to frame-based inheritance was developed to handle the dual nature of meteorological concepts. For instance, a "Cloud" might belong to an "Observation" hierarchy (data at a point in time) and an "Event" hierarchy (a process over time). The view scheme allows a node to inherit different properties depending on the perspective through which it is accessed, preventing data duplication and property clashes.
- Automatic Rule Generation: To optimize performance, METEOR automatically generates production rules from frame descriptions on an as-needed basis. This allows the system to utilize fast matching algorithms (like the Rete algorithm) to recognize complex meteorological events defined within the frames.
- Spatial and Temporal Reasoning: The forecast area is represented as a 2D grid of "mapels" (map elements). Time-varying properties, such as storm location, are managed using "time frames," which function as add/delete lists of property values over a timeline.
System Architecture
The METEOR architecture is distributed to handle the distinct requirements of data processing and symbolic reasoning. Data collection and initial map generation are performed on a VAX 11/780, while the high-level AI reasoning engine resides on a Xerox 1100 Lisp Machine.
Key Insights and Results
A critical distinction in METEOR is between an observation (a data point at a specific time/place) and an event (a process inferred from observations spanning space and time). By parsing human remarks, METEOR can automate "look-out-the-window" strategies, such as identifying that cirrostratus clouds may decrease the likelihood of a storm—a heuristic grounded in thermodynamics but often ignored by purely numerical models.
The system was refined using historical data from the 1984 Alberta hail season. At launch, the knowledge base included 140 production rules and 300 frames, with the system dynamically generating upwards of 300 additional frames during a single forecast run to represent specific meteorological activities. The final output is a map outlining storm initiation regions and predicted intensity, accompanied by qualitative justifications for the forecast.
Conclusion and Future Work
METEOR demonstrates that integrating qualitative human observations with quantitative data can enhance the depth of automated weather forecasting. While the system successfully links procedural rules to deep frame-based knowledge, future work aims to improve causal justifications and manage the real-time complexity of incorporating increasingly granular qualitative data.
Summary updated