Analysing environmental risks with AI
Previous methods for predicting environmental risks are based on computationally intensive and sometimes lengthy simulations. The AI methods developed in the project enable locally precise prediction models for cities that integrate both complex urban development and local environmental data. In addition, the effects of various climate scenarios can be visualised ‘on the doorstep’.
The data for the environmental forecast was generated using simulations. Firstly, abstract simulation data on heat, water and wind was generated for selected training areas as time series on various spatial scales. The results were then used as training data for the development of an efficient, fine-meshed forecasting model for the entire city of Freiburg based on an artificial neural network (ANN).
Artificial intelligence methods are used for the identification of...
... heat spots in cities
Based on the prediction by the KNN and a 3D city model, spatially high-resolution data of the frequency for determining the heat load (universal thermal climate index, UTCI) under the current climatic conditions in the city were created. To calculate the same fine-meshed results with physical model simulations for the entire city of Freiburg would require many hours of high-performance computing. The network was then further developed so that the heat load could be predicted not only at the local level, but also for future decades in relation to the various climate projections.
...floodplains and critical structures
In the area of water, possible consequences (maximum local flooding depths and maximum local flow velocity) of heavy rainfall events for different intensities, pre-humidities and certain annualities in selected training areas were shown by means of simulation. In order to better recognise critical points and to place defence measures in a sensible manner, time sequences were created. These determine where flooding occurs first in the surrounding area, which can then be used, for example, as a warning sign in the municipality or for neighbouring municipalities. This allows critical structures and areas to be identified for hazard defence purposes.
...critical infrastructure against strong wind events
In the area of wind, flow simulations were carried out in building and vegetation-resolving physical models for selected, typical urban structures in different flow situations (so-called Large Eddy Simulations, LES). With the aim of analysing and systematising interactions between flow and roughness elements (houses, trees) in the selected, typical urban structures, the building-resolving flow calculation was improved. The simulation data served as training data for the development of new AI methods. City-wide wind simulations are hardly possible with physical models due to the computing power required. With the AI prediction algorithms developed, approximate results can be realised in real time and simulation data can be integrated on different spatial scales. The dynamic wind model provides the basis for a hazard analysis with regard to strong winds for the entire area of the city of Freiburg.
Publications
Simulation, Analysis & Prediction: Using AI to predict the urban climate
Modelling Tmrt in a complex urban environment using a convolutional encoder-decoder network
Simulation and Multiscaling Wind
Modelling mean radiant temperature in complex urban areas using a convolutional network approach
* Image source Img. 1: Briegel F, Wehrle J, Schindler D, Christen A, 2024: High-resolution multi-scaling of outdoor human thermal comfort and its intra-urban variability based on machine learning. Geoscientific Model Development, 17, 1667-1688.
** Image source Img. 3: Original paper in the journal Urban Climate (Wehrle et al., Introducing new morphometric parameters to improve urban canopy air flow modelling: A CFD to Machine-Learning study in real urban environments. Under review).