A comparison of data-driven methods in prediction of weather patterns in central Croatia

The prediction of future weather patterns has recently become a very important research area due to the ongoing climate change process which causes extreme weather events and rapidly changing weather patterns. In this paper we compare the prediction accuracy of eight data-driven methods, which had been developed for time series prediction, on future weather patterns in central Croatia. The evaluated methods are Seasonal naïve, AutoRegressive Integrated Moving Average (ARIMA), Error-Trend-Seasonality (ETS), Exponential smoothing state space model with Box-Cox transformation (TBATS), Dynamic Harmonic Regression (DHR), Neural Network AutoRegression (NNAR), Support Vector Regression (SVR) and Long Short-Term Memory (LSTM). In our experimental evaluation, we use a historical data from 1961 to 2017 that contains temperature, air pressure and precipitation values for eight weather stations in central Croatia, and indices from two atmospheric oscillations, namely North Atlantic Oscillation (NAO) and Arctic Oscillation (AO). The results of our evaluation show that SVR is the best method, and that DHR and NNAR methods are also better than the other evaluated methods, as far as the accuracy of prediction is concerned. Among DHR and NNAR methods, DHR method is better for the prediction of temperature and air pressure, while NNAR method is better for the prediction of precipitation. Additionally, our evaluation shows that SVR, DHR and NNAR methods achieve a better prediction accuracy when oscillation indices are included as additional predictors.