EXPLORING THE APPLICATION OF ARTIFICIAL NEURAL NETWORKS FOR IMPROVED ACCURACY IN ENVIRONMENTAL TEMPERATURE PREDICTION

Abstract

Artificial neural networks are becoming increasingly popular tools in researching and forecasting weather conditions. The use of these networks for predicting ambient temperature in the short term holds great potential in fields where accurate and rapid forecasts are critically important. Weather warnings are considered key informational products as they help protect lives and property from dangers associated with extreme weather conditions. The reliability and timeliness of the received information are of significant importance, not just the fact of the warning itself. The aim of the research was to increase the accuracy of temperature forecasting and select the most effective neural network model for addressing the temperature prediction task. Weather parameters for the research were collected from the climate data center and divided into three datasets (training, testing, and validation). Neural networks were trained and tested. A recurrent neural network (RNN) and a combination of neural networks (convolutional and fully connected) were chosen as promising approaches to increase forecast accuracy. Using these models with the datasets, forecasts of future temperatures were made. The accuracy of these forecasts was verified using quality metrics such as mean absolute error (MAE), mean squared error (MSE), and mean absolute percentage error (MAPE). It was demonstrated that the proposed models have errors of 15.46% and 14.22% for forecasting temperature with the recurrent neural network and its combination, respectively. The results confirm that proposed models have the potential for successful application in temperature forecasting.