Modern weather predictions are primarily performed with numerical weather prediction models, which are programs that run on computers. To find ways to improve the forecast, it is important to first understand the sources of forecast errors.
It is a well-known fact that the accuracy of forecasts decreases with time. A major cause for the increase in prediction errors with time is the basic nature of atmospheric flow. Due to the nonlinear nature of fluid flow, small errors in the values representing fluid state can amplify with time, causing predictions to deviate from the actual evolution. Since the initial state we can obtain for the prediction is never perfect, the errors in the initial conditions will eventually contaminate the prediction to render it useless.
To make the mathematical equations governing atmospheric flow manageable by digital computers, they have to be discretized. As a result of the process, the computer models are approximations of the real atmosphere. It implies that the computer model involves omissions that contribute to the loss of accuracy of the prediction.
Another source of error is our imperfect knowledge of the different processes at work that change the atmospheric state. The numerical weather prediction models are essentially representations of our current understanding of these processes, and that knowledge is still fairly limited and incomplete.
A relatively straightforward means to decrease forecast errors in numerical weather prediction models is to increase their resolution, thereby decreasing errors resulting from the discretization. Increasing the resolution, however, implies an increase in the amount of information to be processed, and demands more memory and processing time on the computer. For a forecast to be a forecast, the results need to be obtained within a time shorter than the real time. That constrains the maximum resolution allowable for a particular model. The weather models have always been pushing the envelope of computing power. The availability of greater computing power will not only accommodate higher resolutions but also more sophisticated treatment of atmospheric phenomena.
The initial conditions for weather prediction models come from the vast meteorological observation network that collects weather data. Errors can result from defects and malfunctions of instruments or human errors. Improvements in instrumentation techniques and reporting procedures can help minimize these errors. The other problem is that the network of observations is highly non-uniform with many gaps. In remote areas and over most of the ocean surface observations are severely limited. Investments in expanding the network will definitely pay off. The use of satellites, which has a uniform global coverage, will also help in filling in the gaps.
Constant research and progress are being made in our understanding of the processes that produce weather phenomena and incorporating them into the weather prediction models. There are still major challenges in many areas and our knowledge will never be complete, but new advances in many areas are quickly improving our skills in making forecasts.