Evaluation of machine learning models for predicting the System Marginal Price of an Electricity System - Spanish SMP Day - Ahead forecasting
Αξιολόγηση μοντέλων μηχανικής μάθησης για την πρόβλεψη της οριακής τιμής συστήματος ηλεκτρικής ενέργειας - Περίπτωση Ισπανίας
Master Thesis
Author
Ψάλτης, Αθανάσιος
Psaltis, Athanasios
Date
2023View/ Open
Keywords
Machine learning ; SMP ; Spanish energy system ; Sarima ; FBProphet ; LightGBM ; Time series forecastingAbstract
Electricity market price prediction plays a crucial role in ensuring the efficient operation of power systems, helping market participants make informed choices and promoting sustainable energy consumption practices. This thesis offers an in-depth examination of three distinct time series forecasting models: AutoRegressive Integrated Moving Average (ARIMA), Facebook Prophet, and Light Gradient Boosting Machine (LightGBM). These models are applied to forecast Short-Term Market Prices (SMP) in Spanish energy system. The research starts by providing an overview of the electricity market in Spain, emphasizing the
significance of precise price predictions for market players, grid operators, and policymakers. Additionally, the thesis underscores the significance of data analysis, manipulation, and preprocessing in preparing the SMP dataset for modeling. This involves techniques like data decomposition, stationarity testing, and Fourier transform to uncover underlying patterns and improve the quality of input data. The thesis provides a thorough review of each forecasting model, explaining their fundamental principles, strengths, and limitations. It then delves into the data preprocessing phase, illustrating how data decomposition techniques such as Seasonal Decomposition of Time Series (STL) can be used to separate trend, seasonality, and residual components. Stationarity tests like the Augmented Dickey-Fuller (ADF) test are employed to ensure the data is suitable for modeling, ensuring
consistent statistical properties over time. A comparative analysis is conducted using historical SMP data to assess the performance of ARIMA, Facebook Prophet, and LightGBM models in terms of forecasting accuracy, computational efficiency, and
robustness. Key metrics like Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE) are used to evaluate the predictive abilities of the models. The comparative analysis of the ARIMA, Facebook Prophet, and LightGBM models
extends to assessing their ability to capture the decomposed components, seasonality, and volatility patterns within SMP data. Furthermore, the research explores the incorporation of Fourier transform to identify and model cyclical patterns in the SMP data, enhancing the models' capability to capture underlying periodic behaviors. This demonstrates the adaptability of the models to various data manipulation techniques. In conclusion, this thesis offers a comprehensive approach to electricity market price forecasting by integrating data analysis and manipulation techniques with ARIMA, Facebook Prophet, and LightGBM models for SMP price prediction in Spain. The results empower market participants and researchers to choose suitable forecasting strategies based on data characteristics, facilitating more informed decisions in the dynamic electricity market environment, while considering the inherent seasonality, stationarity, and cyclicality of
the data.