Machine Learning applied to Urban Heat Island Intensity prediction at the Metropolitan Region of Sao Paulo, Brazil

The urban heat island effect is characterized by the urban environment having increased temperatures when compared to the rural areas. This phenomenon negatively affects urban environments, increasing energy consumption and exacerbating health issues in vulnerable populations. To delve deeper into this effect, data spanning from 2009 to 2019 were collected from 30 meteorological stations managed by Centro de Gerenciamento de Emergências (CGE) of São Paulo city's government. This data was then combined with land use and occupation data from the Unidades Homogêneas de Uso e Ocupação do Solo Urbano (UHCT) database of the State of São Paulo. A cluster analysis utilizing PCA and k-means resulted in four distinct clusters: A, B, C, and D. Cluster A comprises meteorological stations located in the more central and urbanized areas of the São Paulo metropolitan region, exhibiting higher average temperatures and rainfall. On the other hand, Cluster D encapsulates stations positioned in peri-urban areas with lower average temperatures and heightened relative humidity. Feature engineering was subsequently performed on the time series data to mold it into a dataset apt for supervised regression models. This transformation involved the incorporation of lag variables, cyclical variables, and window statistics variables. Three regression models, namely Ridge, Gradient Boosting, and Multi-layer Perceptron, were trained and had their performance evaluated for predicting urban heat island intensity over a 6-hour window. Across all clusters, the Gradient Boosting model demonstrated the best performance, yielding Mean Squared Errors (MSE) of 3.81 for Cluster A, 3.64 for Cluster B, 2.81 for Cluster C, and 2.89 for Cluster D. However, it's crucial to note that all models faced challenges in effectively predicting air temperature during cold front events.