Modeling and Prediction of Meteorological Parameters Using the Arima and LSTM Methods: Sivas Province Case

Çetintaş A. O., Apaydın H.

Novel & Intelligent Digital Systems: Proceedings of the 3rd International Conference (NiDS 2023), Athens, Greece, 28 - 29 September 2023, vol.1, pp.250-261, (Full Text)

  • Publication Type: Conference Paper / Full Text
  • Volume: 1
  • Doi Number: 10.1007/978-3-031-44097-7_27
  • City: Athens
  • Country: Greece
  • Page Numbers: pp.250-261
  • Ankara University Affiliated: Yes

Abstract

The modeling of meteorological parameters sheds light on the determination of agricultural water needs and dry periods. Predicting precipitation and, therefore, droughts provides various benefits such as increased crop yield and the prevention of harvest losses. The power of prediction plays a crucial role in achieving improved water management, enhanced crop yield, risk mitigation, economic stability, and sustainable agriculture. It is crucial for rural communities that make a living through agricultural production, to benefit from these advantages. From past to present, meteorological parameters have been estimated with many statistical models and machine learning methods. This study aims to determine the prediction success of the statistical method ARIMA and artificial neural network model LSTM by using meteorological data of Zara, Susehri, Ulas, Kangal, Gemerek and Divrigi districts stations of Sivas province, Turkey. The study was conducted using daily data collected over a period of 10 years from all districts. While R2 ranged between 0.06–0.94 in the ARIMA models, R2 success was between 0.63–0.96 in the LSTM models. According to the results obtained, LSTM layer with 8–16-32 neurons, epoch value in the range of 100–300, learning rate of 5e−4 and 146e−5, between 1e−2 –1e−6 decay, Adam optimization, in which ReLu activation is used in each layer, in the estimation of meteorological parameters in the region. It has been determined that the LSTM method with batch size configuration in the range of 8–256 is the best alternative. Although the ARIMA model is a common model that has been used for many years, it is determined that the LSTM model are superior to the ARIMA model with the diversity and controllability of the variables, and more successful results can be obtained.