ISSN 1000-0526
CN 11-2282/P
CN 11-2282/P
Experimental Study on Refined Forecasting of Urban Temperature in Large Cities Based on Ensemble Correction with Multiple Machine Learning Methods
- Article
- Figures
- Metrics
- Preview PDF
- Reference
- Related
- Cited by
- Materials
Abstract:
To improve the refined forecasting capability for large urban temperatures, this study employed 2m temperature forecast products from ECMWF and observational 2m temperature data from critical locations in Xi"an during the period 2023-2024, to assess the performance of ECMWF model temperature forecasts. Optimal feature factors were identified using subjective prior knowledge and lag correlation analysis. Subsequently, three machine learning models(XGBoost, LightGBM, CatBoost) were applied to separately correct biases in the 2m temperature forecasts at critical locations in Xi"an. To further improve the accuracy of temperature forecasts, these individual models were then integrated using a machine learning stacking approach. The results show that:(1) Overall, the ECMWF 2m temperature forecasts at critical points in Xi"an align well with the actual trends, but the temperatures forecast levels tended to be lower than the observed temperatures. The 2m temperature forecast error exhibits a diurnal variation characteristic, being smaller during the day and larger at night. During significant cooling and precipitation weather events, the 2m temperature forecasts at key points in the urban area of Xi"an exhibit a notable cold bias.(2)The three machine learning temperature forecasting models effectively reduced the forecast bias of 2m temperatures at key points in the urban area of Xi"an as predicted by ECMWF. During nighttime, the root mean square error (RMSE) reduction rate reached 30% to 46%. Bayesian optimization was employed to fine-tune the hyperparameters of the three individual temperature forecasting models. After parameter tuning, the RMSE of the three models on the test set across all time periods decreased by 0.037°C, 0.021°C, and 0.024°C, respectively, compared to before the tuning.(3)By integrating prior knowledge, seven physical quantities closely related to 2m temperature were introduced. Then, using lag correlation analysis, different upper-air region characteristic factors closely associated with each forecast lead time within 24h period were constructed. After optimizing the feature factors for the three individual models, the RMSE on the test set across all time periods decreased by 0.259°C, 0.243°C, and 0.272°C, respectively, compared to before the optimization.(4)The comparison between stacking ensemble and weighted ensemble methods indicated that the former performs better for 2m temperature forecasting. Specifically, the RMSE of the stacking ensemble is reduced by 0.045°C compared to the weighted ensemble, while the accuracy of forecasts within 2°C was improved by 0.021.At each forecast lead time, the stacking ensemble approach demonstrates improved forecast quality compared to individual model forecasts. During significant cooling and precipitation weather events, the stacking ensemble approach reduced temperature forecasts RMSE by 7% to 20% compared to individual model forecasts. Overall, the temperature forecasting model for key points in Xi"an, based on ensemble correction using multiple machine learning methods, effectively reduces the ECMWF model temperature forecast error and further enhances the quality of large urban temperature forecast.
Keywords:
Project Supported:
Clc Number:
Mobile website
