Accuracy Comparison between Recurrent Neural Networks and Statistical Methods for Temperature Forecasting

Temperature is a critical weather factor that affects human beings in numerous ways. Therefore, accurate temperature forecasting is crucial in many fields. Numerous techniques for predicting temperature have been investigated in various studies. Auto-Regressive Integrated Moving Average (ARIMA) and Seasonal Auto-Regressive Integrated Moving Average (SARIMA) are two of the conventional methods that have been used for forecasting. In this research, a modern machine learning based approach, Recurrent Neural Networks (RNN), is compared with traditional methods in terms of temperature forecasting. A dataset gathered in Szeged, Hungary, was considered in this study, which contains hourly reported data from 2006 to 2016. The weekly average temperature was considered for building the models, which were evaluated for accuracy using Mean Squared Error (MSE). The results revealed that the RNN model gives more accurate predictions than the traditional time series approaches in temperature forecasting.