Abstract
<title>Abstract</title> <p>The prediction of electricity price is still one of the most complex problems in the current energy markets considering the high volatility, nonlinear dependence, multiple seasonalities, and the increasing impact of external market factors. The day-to-day dynamics of high-frequency electricity price data are complex and difficult to model using traditional statistical and machine learning methods. To overcome these constraints, the functional data analysis (FDA) approach to modeling and forecasting electricity prices is developed where hourly electricity prices are modeled as continuous curves and the temporal structure of the data is used. In particular, we have used functional autoregressive models (FAR(p)) and their extension with functional exogenous variables (FARX(p)) to model the historical paths of prices and the effect of the electricity demand. The methodology is tested based on hourly prices of electricity from the Croatian Electricity Market and compared with the most widely used statistical and machine learning models. From the empirical results, it can be concluded that the FARX(p) model outperforms all other models both for hourly, weekly and monthly forecasting horizons by providing the lowest MAE, MAPE and RMSE values. The results reveal that the functional forecasting models are able to successfully capture the temporal dependence and exogenous market effects, and that these models yield significant gains in predictive accuracy and forecasting robustness. In addition to their statistical strength, the proposed models offer a useful tool for making decisions, allowing market participants to make more informed bids, better risk management, and better operational planning in competitive electricity markets. The findings show the significant promise of FDA-related approaches to improve next generation electricity price forecasting systems.</p>