Using high-speed computer to tackle the computational demand, numerical weather prediction (NWP) is the technique used to forecast weather by solving a set of equations within a numerical model that describes the evolution of meteorological variables representing the atmospheric state. These variables include temperature, wind, pressure and moisture content.

In the model, the overall atmospheric state at any given instant is represented by the values of the variables at systematically arranged points set up within a three-dimensional grid. The larger the set of grid points, the higher the computational demand, the finer the model resolution and the more details in the future state of the atmosphere can be described.

Approximations and assumptions are made in the governing equations and representation of the physical processes. To solve the set of governing equations, initial conditions have to be properly represented and set up using the latest information. In practice, the initial state of the atmosphere is analyzed by taking a previous short-range model forecast, ingesting the latest meteorological observations to update the situation, and then starting a new forecast cycle.

At the Hong Kong Observatory, the Operational Regional Spectral Model (ORSM) is run at horizontal resolutions of 20 km (inner domain) and 60 km (outer domain) to provide 42-hour and 72-hour forecasts respectively. Graphical products from model outputs are generated to facilitate interpretation and other applications.

The HPC consists of 16 central processing units (CPU), 8 gigabytes of random access memory (RAM) and around 450 gigabytes of disk capacity. Its peak processing speed of 19.2 billion floating point operations per seconds (GFLOPS) is 5 000 times faster and its memory size 1 000 times bigger than the one installed at the Observatory in the late 1980s. In those days, constrained by available computer speed and memory resources, NWP model could only simulate a limited region of the atmosphere using a 100 km by 100 km grid (Fig. 2) and calculation was limited to once per day. Improving the model resolution to 20 km by 20 km has definitely been a significant advance (Fig. 3).

Forecast Applications :
Since the introduction of the HPC and ORSM, hourly rainfall information derived from weather radars and raingauges is routinely ingested into the model for more effective simulation of rain systems (Fig. 4), allowing forecasters to make useful reference of the quantitative precipitation forecasts and precipitation trends provided by objective model guidance. More detailed description of the motion and intensity changes of tropical cyclones has also been made possible (Fig. 5). Temperature distribution on a regional basis within Hong Kong, particularly during cold surges in winter, can also be meaningfully derived from the model output.

Outlook :
Numerical weather prediction has its inherent limitations. Inevitably, data will always contain errors; and in practice, observations are never quite dense enough to be truly representative of the real atmosphere. Subtle differences in the initial conditions can lead to a substantial divergence in the final computational results.

For small scale systems such as convective cells or thunderstorms with dimensions under 10 km that affect Hong Kong every summer, the current model resolution is still not quite fine enough (Fig. 6). The next target is to go for even higher resolution model with the aim of improving predictions in terms of rainstorm location, intensity and timing. This will also involve the ingestion of more observational data on the finer scale for input into such models.