نوع مقاله : مقاله پژوهشی

نویسندگان

1 دانشجوی کارشناسی ارشد، گروه مهندسی آب، دانشکده کشاورزی، دانشگاه تبریز، تبریز، ایران

2 گروه مهندسی آب، دانشکده کشاورزی، دانشگاه تبریز، تبریز، ایران

چکیده

دمای نقطه شبنم در زمینه‌های مختلف از جمله علوم هواشناسی جهت پیش‌بینی‌های مربوط به آب و هوا دارای اهمیت فراوانی می‌باشد. لذا ارائه مدل‌های مناسب جهت پیش‌بینی دقیق مقدار این متغیر هواشناسی برای استفاده عملی مهندسین کشاورزی و ایستگاه‌های مجاوری که در آن‌ها امکان اندازه‌گیری این دما وجود ندارد، ضروری می‌باشد. در این راستا، در پژوهش حاضر توانایی چهار مدل داده‌محور درخت گرادیان تقویتی، مدل درختی M5P ، جنگل تصادفی و جنگل تصادفی بهینه‌شده با الگوریتم ژنتیک در تخمین دمای نقطه شبنم روزانه مورد ارزیابی قرار گرفت. برای این منظور از داده‌های هواشناسی روزانه دو ایستگاه اردبیل و پارس‌آباد در بازه زمانی 1384 تا 1399 استفاده شد. پارامترهای هواشناسی مورد استفاده شامل حداقل، حداکثر و میانگین دما، رطوبت نسبی، ساعت آفتابی و سرعت باد بوده که در 10 ترکیب متفاوت به‌عنوان متغیرهای ورودی برای هر یک از مدل‌های مذکور در نظر گرفته شدند. در نهایت، جهت ارزیابی صحت نتایج و دقت عملکرد مدل‌ها، نتایج به‌دست‌آمده با استفاده از پارامترهای آماری با یکدیگر مقایسه شده و مدل‌ برتر انتخاب شد. مقایسه نتایج به‌دست آمده برای هر دو ایستگاه نشان داد که مدل‌ M5P-8با دارا بودن جذر میانگین مربعات خطای ◦C 54/0 و ضریب ویلموت برابر با 998/0 در ایستگاه اردبیل و مدل M5P-6 با جذر میانگین مربعات خطای ◦C 29/0 و ضریب ویلموت برابر با 00/1 در ایستگاه پارس‌آباد به‌عنوان برترین مدل‌ها معرفی شدند.

کلیدواژه‌ها

موضوعات

عنوان مقاله [English]

Evaluation of the Capabilities of Gradient Boosted Tree and Optimized Random Forest in Estimating Daily Dew Point Temperature

نویسندگان [English]

  • Mohsen Osouli 1
  • Fatemeh Mikaeili 1
  • Saeed Samadianfard 2

1 M.Sc. Student, Department of Water Engineering, Faculty of Agriculture, Tabriz University, Tabriz, Iran

چکیده [English]

Introduction

Dew point temperature (DPT) is the temperature at which the moisture or vapor in the air is converted to liquid water at a constant pressure due to the high concentration of water molecules. Accurate estimation of this temperature plays an important role in estimating frost, snow, rain, humidity and other meteorological parameters. Long-term changes in dew point temperature can be very effective in regulating the energy adjacent to the earth's surface and the water balance and greenhouse gases. DPT is not measured and recorded relative to common meteorological parameters such as rainfall and relative humidity at many meteorological stations, but is mainly estimated using regression methods, which in most cases estimate data are not reliable. Today, intelligent systems based on data mining have been considered in modeling hydrological and meteorological processes.

Methods

In the present study, daily meteorological data of Ardabil and Parsabad stations including mean temperature, minimum temperature, maximum temperature, relative humidity, sunshine hours, wind speed and dew point temperature were used for a period of 16 years (2005-2011). Based on this, 10 combined scenarios of meteorological variables were defined as the input of the studied models. The methods used in the present study include gradient boosted tree, M5P tree, random forest and random forest optimized by genetic algorithm. The results obtained from these models were compared using statistical indices such as coefficient of determination (R2), root mean square error (RMSE) and willmott’s index (WI) and the best models were selected.

Results

The results showed that at Ardabil station, M5P-8 and M5P-10 models with R2 = 0.994, WI = 0.998 and the RMSE = 0.54 ◦C had the best performance compared to the different patterns defined for each of the mentioned models. In the second place, the RF-GA-6 model showed remarkable performance compared to other defined patterns and methods. Pattern No. 6 with the input parameters of temperature and relative humidity in all methods except M5P, has provided the most accurate results. In Ardabil station, model RF-1 with R2 = 0.706, WI = 0.91 and the RMSE = 3.79 ◦C showed the weakest performance among the other methods. At Parsabad station, the M5P model in scenarios 6, 8, 9 and 10 with the same results with R2 = 0.999, WI = 1.00 and RMSE = 0.29 ◦C had the best prediction and the performance of the GA-RF-6 model was ranked second. The GBT-1 model with a R2 = 0.908, WI = 0.78 and RMSE = 4.38 ◦C presented the weakest results. In general, the M5P tree model in both stations was introduced as the superior model in estimating dew point temperature values and the GBT model with poor results was considered as an inefficient method in predicting daily dew point temperature values.

Conclusion

The overall conclusion showed that: 1) In Ardabil station, model M5P in the eighth scenario includes input data of relative humidity, wind speed, minimum, maximum and average temperature and in Pars Abad station, the model in the sixth scenario includes input data of relative humidity, minimum, maximum and average temperature, the most desirable results offered. 2) Comparing the results between the models, the M5P, GA-RF, RF and GBT models were ranked from more to less accuracy, respectively. 3) Parameters of temperature and relative humidity were introduced as the dominant variables in predicting dew point temperature. 4) By comparing the results of RF with GA-RF models, it was concluded that the genetic algorithm improved the performance of random forest and had a positive effect on all models.

کلیدواژه‌ها [English]

  • Ardabil
  • Intelligence models
  • Meteorological variables
  • Statistical evaluation