| Abstract [eng] |
Since the electricity sector liberalization start electricity price forecasting has gradually become the central and integral process in the planning and operation of electric utilities, energy suppliers, system operators and other market participants. Irregular extreme price volatility, price dynamic change, observed multi-seasonality combined requires complex mathematical model implementation. The aim of this work is to apply 2 stage Markov hidden regime switching model with AR(2) modification for daily Lithuania electricity price prediction, regime identification and analyze the impacts of training dataset, model complexity, hourly data aggregation method on the final model results. To analyze the impact of the training dataset length on the final model results 6 different training period variations were tested. It was identified that the longer data training period helps to better identify the hidden regimes in the time series but reduces overall prediction accuracy. Using the shorter training datasets improves the prediction accuracy, but due to worse regime identification also decreases model stability and increases prediction error volatility. It was found that the more complex models, that uses more regressors shows better results both in the electricity price prediction and hidden regime identification. The best identified model uses Lithuania electricity consumption, generation, wind power generation, gas price, oil price, biofuel price and energy flow from neighbor countries as input variables for the Hidden Markov regime model. Finally, the hourly time series aggregation to daily period method impact on the results were tested. Daily average, daily median, working hours average, night hours average and maximal electricity price average methods were tested. It was found that using these aggregation models yields different time-series characteristics and vastly impacts regime identification. While daily average method can be used to model overall price dynamics, night hours and working hours aggregation methods can be used to better model these two periods of day individually. |
