Scholarlink

Abstract <p>This study investigates machine learning methods for approximating the temperature difference between urban and rural areas (urban heat island intensity) using examples from Moscow and St. Petersburg. Predictors consist of characteristics of large-scale meteorological conditions derived from long-term, regionally averaged observational data from rural weather stations and global ERA5 reanalysis data from 2012 to 2023. A key feature of meteorological data is the delayed dependencies between processes, where the value of a target variable is influenced by factors that act with a time lag. Two approaches were explored to account for these dependencies: explicit feature engineering to generate lag-related features for the CatBoost regression model, and application of the long short-term memory recurrent neural network (LSTM), for sequence modeling. The dependence of modeling results on the length of the lookback period was investigated. Experimental results showed that LSTM did not exceed the accuracy of CatBoost with expert-designed temporal features. The most informative data for modeling urban heat island corresponded to a lookback depth of 3 time steps (9-h history). The study revealed the critical importance of accounting for temporal dependencies in modeling urban heat islands.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Comparison of Machine Learning Methods for Accounting Lagged Relationships in Urban Heat Island Modeling

  • 期刊论文

  • 发表日期：

    2025年12月1日

查看全文
  • K. F. Nazmutdinov,
  • M. I. Varentsov

摘要

Abstract

This study investigates machine learning methods for approximating the temperature difference between urban and rural areas (urban heat island intensity) using examples from Moscow and St. Petersburg. Predictors consist of characteristics of large-scale meteorological conditions derived from long-term, regionally averaged observational data from rural weather stations and global ERA5 reanalysis data from 2012 to 2023. A key feature of meteorological data is the delayed dependencies between processes, where the value of a target variable is influenced by factors that act with a time lag. Two approaches were explored to account for these dependencies: explicit feature engineering to generate lag-related features for the CatBoost regression model, and application of the long short-term memory recurrent neural network (LSTM), for sequence modeling. The dependence of modeling results on the length of the lookback period was investigated. Experimental results showed that LSTM did not exceed the accuracy of CatBoost with expert-designed temporal features. The most informative data for modeling urban heat island corresponded to a lookback depth of 3 time steps (9-h history). The study revealed the critical importance of accounting for temporal dependencies in modeling urban heat islands.