According to research, US adults view weather forecasts almost 300 billion times a year. And the relevance of reliable forecasts is very high, as they are the ones that can predict dangerous weather events, such as blizzards, hurricanes, or flash floods 7–10 weeks before the event. According to preliminary estimates, the cost of these forecasts is $31.5 billion per year.
Artificial Intelligence can be used to improve the quality of forecasts as well as make them cheaper. That is what a new study by US scientists was aimed at. Using a convolutional neural network, the authors have developed a machine-learning-based weather forecasting system called Deep Learning Weather Prediction (DLWP). The model learns from past weather data, which differs from standard numerical weather prediction models that create mathematical representations of physical laws. DLWP can project future weather for 2–6 weeks for the entire globe.
The authors compared DLWP with classical modern numerical weather prediction models. The evaluation showed that standard forecasts work better for short forecast periods, e.g. 2–3 weeks. And the DLWP model was able to show excellent results for 4–6 weeks ahead.
Although the DLWP model cannot yet compete with existing models, the prospects are amazing. AI is more efficient than other approaches. DLWP can make an ensemble forecast in just 3 seconds, which consists of 320 independent runs of the model. The model has also shown that it can warn of hurricanes in 4–5 days, which can indeed save many lives.
But at the same time, scientists are convinced that classical forecasting methods should not be abandoned. People should use rules of thumb and pattern recognition methods not only as learning tools but also to guard against the loss of vital experience that meteorologists bring to severe weather situations or when models don’t fit the basics.
But artificial intelligence can be a reliable aid in making forecasts, especially in the long term.