A neural Network for weather prediction
As a demonstration of the power of neural networks, we have programmed a neural network that will predict the weather in Oslo on a given day, based on the weather on the preceding five days. Through training on weather data from the Blindern weather station since 1950, the network has learned to recognize patterns in the way the weather develops. When given the weather data from five consecutive days, the network will use its experience with how the weather has behaved in the past to guess what the weather will be the following day. You can try it out by picking a date in the field below, and it will fetch weather data from the five preceding days from met.no, and feed it into the neural network to make a prediction. If you pick a past date, you will also be shown the observed weather on that date. The weather database unfortunately does not include today's weather or future weather predictions, so it cannot be used to predict the weather further into the future than today.
The weather is, of course, a very complex system, and reliably predicting the weather would also require knowledge of weather systems in nearby areas. It should therefore come as no surprise that there is a limit to how well you can predict the weather based on the last few days' weather alone. The average errors (in technical terms: the standard errors) of the predictions are as follows:
- Precipitiation: ±4.1 mm
- Avg. temperature: ± 2.1 °C
- Avg. cloud cover: ±24%
- Avg. wind speed: ±1.2 m/s
Note that the given temperatures are 24 hour averages, and include nighttime temperatures. They will therefore typically be lower than what you might expect from daytime temperatures.