High resolution temperature forecasting using functional time series decomposition and advanced predictive models

The air temperature (AT) is a significant factor in the environmental processes, human health, agriculture, and energy systems, so correct forecasting is very important. This paper explores the prediction of AT in a functional time series (FTS) model that models the high-frequency temperature data as smooth curves over the day that can model both deterministic seasonal cycles and other stochastic dynamics. Smoothing splines and Fourier-based functional representations are used to model and predict short-term variations in temperature by a functional autoregressive model, FAR(1). The performance of FAR(1) is systematically contrasted to those of classical statistical models such as ARIMA and VAR and machine-learning-based ones such as artificial neural networks (ANN) and autoregressive neural networks (ARNN). Accuracy of an error forecast is measured in terms of mean absolute error (MAE), mean absolute percentage error (MAPE) and root mean square error (RMSE). Empirical findings show the FAR(1) model records reduced forecasting errors and increased predictive stability both in months and in hours than the competition methods. Such results outline the utility of functional data analysis to utilize the smooth and periodic organization of temperature data. The research offers a practical and holistic model of predicting high-frequency AT, and it can be employed in the better decision-making in agricultural, energy and safety management among the population that operate during the climate variability.