<p>The rapid increase in industries and buildings, massive emission of gases, and decreasing forests due to urbanization create an extraordinary threat to climate health. Alleviating climate change caused by temperature increase is one of the biggest concerns of humankind. Accurate air temperature forecasting is crucial because of the variable temperature characteristics over different regions. Effective air temperature forecasting may create credibility for future planning to maintain the environmental sustainability of cities and climate health. In this research, the trend, seasonal, and residual characteristics of air temperature time series are utilized by decomposing the series to build effective deep learning models and then recomposing the outcomes to obtain the final air temperature forecasts, called the vipin-Decompose-deep-Recomposed (vDR) model. Furthermore, a novel optimal ensemble approach using a genetic algorithm (GA) has been proposed using vDR models called vDR-GAE. The performance of both proposed models is evaluated using recent 30-year univariate air temperature datasets from fifteen significant cities worldwide, with RMSE, MAPE, and <InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="12145_2025_1728_Article_IEq1.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="21" /> </InlineMediaObject> <EquationSource Format="TEX">$R^2$</EquationSource> <EquationSource Format="MATHML"><math> <msup> <mi>R</mi> <mn>2</mn> </msup> </math></EquationSource> </InlineEquation> used to measure performance. Quantitative measures and graphical analysis show that the proposed vDR models outperform traditional deep learning models where average improvement ranges from RMSE: 36.6%-46.9%, MAPE: 27.9%-51.23% and <InlineEquation ID="IEq2"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="12145_2025_1728_Article_IEq1.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="21" /> </InlineMediaObject> <EquationSource Format="TEX">$R^2$</EquationSource> <EquationSource Format="MATHML"><math> <msup> <mi>R</mi> <mn>2</mn> </msup> </math></EquationSource> </InlineEquation>:2.61%-2.89%. Moreover, the proposed vDR-GAE model surpasses the individual vDR models across all performance measures. AIC-BIC, Diebold Mariano, and Friedman ranking with Holm’s statistical tests are also applied to the predictive values and results, validating that the proposed models’ performances are compelling and distinct from traditional deep learning models.</p>

Decompose-deep-recompose models and genetic algorithm based optimal ensemble method (GAE) to enhance the air temperature forecasting of world’s major urban cities

期刊论文
发表日期：

2025年2月6日

查看全文

Vipin Kumar

摘要

The rapid increase in industries and buildings, massive emission of gases, and decreasing forests due to urbanization create an extraordinary threat to climate health. Alleviating climate change caused by temperature increase is one of the biggest concerns of humankind. Accurate air temperature forecasting is crucial because of the variable temperature characteristics over different regions. Effective air temperature forecasting may create credibility for future planning to maintain the environmental sustainability of cities and climate health. In this research, the trend, seasonal, and residual characteristics of air temperature time series are utilized by decomposing the series to build effective deep learning models and then recomposing the outcomes to obtain the final air temperature forecasts, called the vipin-Decompose-deep-Recomposed (vDR) model. Furthermore, a novel optimal ensemble approach using a genetic algorithm (GA) has been proposed using vDR models called vDR-GAE. The performance of both proposed models is evaluated using recent 30-year univariate air temperature datasets from fifteen significant cities worldwide, with RMSE, MAPE, and $R^2$ $R^{2}$ used to measure performance. Quantitative measures and graphical analysis show that the proposed vDR models outperform traditional deep learning models where average improvement ranges from RMSE: 36.6%-46.9%, MAPE: 27.9%-51.23% and $R^2$ $R^{2}$ :2.61%-2.89%. Moreover, the proposed vDR-GAE model surpasses the individual vDR models across all performance measures. AIC-BIC, Diebold Mariano, and Friedman ranking with Holm’s statistical tests are also applied to the predictive values and results, validating that the proposed models’ performances are compelling and distinct from traditional deep learning models.