The urban heat island is a vastly documented climatological phenomenon, but when it comes to coastal cities, close to desert areas, its analysis becomes extremely challenging, given the high temporal variability and spatial heterogeneity. The strong dependency on the synoptic weather conditions, rather than on city-specific, constant features, hinders the identification of recurrent patterns, leading conventional predicting algorithms to fail. In this paper, an advanced artificial intelligence technique based on long short-term memory (LSTM) model is applied to gain insight and predict the highly fluctuating heat island intensity (UHII) in the city of Sydney, Australia, governed by the dualistic system of cool sea breeze from the ocean and hot western winds from the vast desert biome inlands. Hourly measurements of temperature, collected for a period of 18 years (1999-2017) from 8 different sites in a 50 km radius from the coastline, were used to train (80%) and test (20%) the model. Other inputs included date, time, and previously computed UHII, feedbacked to the model with an optimized time step of six hours. A second set of models integrated wind speed at the reference station to account for the sea breeze effect. The R2 ranged between 0.770 and 0.932 for the training dataset and between 0.841 and 0.924 for the testing dataset, with the best performance attained right in correspondence of the city hot spots. Unexpectedly, very little benefit (0.06-0.43%) was achieved by including the sea breeze among the input variables. Overall, this study is insightful of a rather rare climatological case at the watershed between maritime and desertic typicality. We proved that accurate UHII predictions can be achieved by learning from long-term air temperature records, provided that an appropriate predicting architecture is utilized.
Keywords: Desert winds, A.I. forecasting models; LSTM; Regional climate change; Sea breeze; Synoptic conditions; Urban heat island.
Copyright © 2019 Elsevier B.V. All rights reserved.