AI can provide more accurate weather forecasts

An AI tool developed and tested as part of a degree project at the University of Gävle can provide more accurate weather forecasts.

Humans have attempted to predict the weather for thousands of years. The oldest documented weather forecasts date back to Mesopotamia, a culture that emerged around 5000 years ago. Over time, the methods have become increasingly sophisticated, but the factors influencing the weather are numerous and complex, making it a challenge to predict the weather even today.

The use of AI for weather forecasting is a relatively new phenomenon. Today, there are several AI models that use historical data to determine how the weather has behaved over time, providing statistics that can be used to make forecasts.

Mohamad Safia and Rodi Abbas, computer science students, conducted their degree project work under the supervision of Mohammad Aslani, researcher of computer engineering. They developed and evaluated four different machine-learning models of weather forecasting. Artificial Neural Networks (ANN) proved to be the most effective among them. 13 years of weather data was used to train the models to achieve high accuracy based on Swedish cities.

“Developing AI with machine learning is a challenging and time-consuming process. It involves large amounts of data where the relationships between humidity, temperature, and other parameters form patterns over time. The AI tool learns to recognise these, which can then be used for weather forecasts,” Rodi Abbas says.

According to the degree project, their AI tool can make more precise assessments of historical weather conditions compared to the traditional methods currently in use. This suggests that the tool has the potential to improve weather forecasts. The researchers used the American weather data and forecasting provider Weatherstack to gather historical data and compare the accuracy of weather forecasts.

“When we tested our ANN model for weather forecasts over a two-month period and compared it to Weatherstack’s own forecasts, our model was 47 percent more accurate in its forecasts,” Mohamad Safia says.

A more extensive study is needed to definitively determine how much forecasts can improve with AI-based software for weather predictions, but, clearly, the potential is significant.

“With machine learning, you can rapidly analyze vast amounts of data to discern patterns. However, climate change has brought about challenges in weather prediction. Therefore, further research, data collection, and analysis are required to identify the climate changes, and adjust weather forecasts accordingly ” says Mohamad Safia.

The degree project has been approved and published in the scientific journal Atmosphere. External link.