The thesis analyzes the main approaches to Electricity Price Forecasting (EPF), ranging from traditional statistical models to machine learning techniques, highlighting their strengths and limitations. Based on this assessment, a predictive model for Italian Intraday Market prices is developed using the XGBoost algorithm, selected for its ability to capture non-linear relationships, ensure robustness to noise and outliers, and provide interpretable variable importance measures. The model is trained on an hourly dataset constructed by integrating data from GME, Terna, and ENTSO-E on electricity demand, renewable generation, gas prices, and market prices. The model structure is optimized through walk-forward validation and Bayesian hyperparameter search, ensuring temporal consistency. Performance metrics, evaluated via RMSE and R², indicate a good degree of accuracy and stability. The model is also applied in a forward-looking exercise to 2060, through the generation of synthetic variables consistent with different decarbonization scenarios. The results show the model’s ability to capture non-linear market dynamics and zonal price differences, as well as the growing impact of renewable penetration on intraday price levels. This suggests that reliable price forecasts can support more efficient and resilient system management, effectively contributing to the energy transition. Finally, the model is extended by incorporating non-energetic drivers, specifically 2-meter temperature, precipitation, and seismic activity. The econometric analysis indicates statistical significance for temperature and precipitation, while seismic activity does not appear systematically relevant. The inclusion of meteorological drivers yields only a modest numerical improvement in performance metrics, but enhances the parsimony and structural consistency of the model, indicating that a selective incorporation of weather-based variables may deepen understanding of intraday price dynamics in an increasingly flexible electricity system.
La tesi analizza i principali approcci di Electricity Price Forecasting (EPF), dai modelli statistici tradizionali alle tecniche di machine learning, evidenziandone punti di forza e limiti. Sulla base di tale analisi, viene sviluppato un modello di previsione dei prezzi del Mercato Intraday italiano fondato sull’algoritmo XGBoost, scelto per la sua capacità di gestire relazioni non lineari, garantire robustezza rispetto a rumore e outlier e fornire strumenti di interpretazione delle variabili. Il modello è addestrato su un dataset a frequenza oraria costruito integrando dati provenienti da GME, Terna ed ENTSO-E relativi a domanda, produzione rinnovabile, prezzi del gas e prezzi di mercato. L’ottimizzazione della struttura è effettuata tramite validazione walk-forward e ricerca bayesiana degli iperparametri, nel rispetto della sequenzialità temporale. Le performance, valutate attraverso RMSE e R², mostrano un buon livello di accuratezza e stabilità. Il modello viene inoltre applicato in un esercizio prospettico fino al 2060, mediante la generazione di variabili sintetiche coerenti con gli scenari di decarbonizzazione. I risultati evidenziano sia la capacità del modello di catturare le dinamiche non lineari del mercato e le variazioni zonali dei prezzi, sia l’impatto crescente della penetrazione rinnovabile sui livelli di prezzo intraday. Ne emerge come previsioni affidabili possano supportare una gestione più efficiente e resiliente del sistema elettrico, accompagnando in modo efficace la transizione energetica. Infine, il modello è esteso introducendo driver non puramente energetici, in particolare la temperatura a 2 metri dal suolo, le precipitazioni e l’attività sismica. L’analisi econometrica mostra significatività per temperatura e precipitazioni, mentre l’attività sismica non risulta sistematicamente rilevante. L’inclusione dei driver meteorologici comporta un miglioramento numericamente limitato delle metriche, ma accresce la parsimonia e la coerenza strutturale del modello, suggerendo che un’integrazione selettiva di variabili meteo possa approfondire la comprensione delle dinamiche intraday in un sistema elettrico sempre più flessibile.
Long-term price forecasting in the italian intraday electricity market using a machine learning model
Battista, Raffaele
2024/2025
Abstract
The thesis analyzes the main approaches to Electricity Price Forecasting (EPF), ranging from traditional statistical models to machine learning techniques, highlighting their strengths and limitations. Based on this assessment, a predictive model for Italian Intraday Market prices is developed using the XGBoost algorithm, selected for its ability to capture non-linear relationships, ensure robustness to noise and outliers, and provide interpretable variable importance measures. The model is trained on an hourly dataset constructed by integrating data from GME, Terna, and ENTSO-E on electricity demand, renewable generation, gas prices, and market prices. The model structure is optimized through walk-forward validation and Bayesian hyperparameter search, ensuring temporal consistency. Performance metrics, evaluated via RMSE and R², indicate a good degree of accuracy and stability. The model is also applied in a forward-looking exercise to 2060, through the generation of synthetic variables consistent with different decarbonization scenarios. The results show the model’s ability to capture non-linear market dynamics and zonal price differences, as well as the growing impact of renewable penetration on intraday price levels. This suggests that reliable price forecasts can support more efficient and resilient system management, effectively contributing to the energy transition. Finally, the model is extended by incorporating non-energetic drivers, specifically 2-meter temperature, precipitation, and seismic activity. The econometric analysis indicates statistical significance for temperature and precipitation, while seismic activity does not appear systematically relevant. The inclusion of meteorological drivers yields only a modest numerical improvement in performance metrics, but enhances the parsimony and structural consistency of the model, indicating that a selective incorporation of weather-based variables may deepen understanding of intraday price dynamics in an increasingly flexible electricity system.| File | Dimensione | Formato | |
|---|---|---|---|
|
2025_12_Battista.pdf
non accessibile
Descrizione: Testo della tesi
Dimensione
40.53 MB
Formato
Adobe PDF
|
40.53 MB | Adobe PDF | Visualizza/Apri |
|
2025_12_Battista_Executive Summary.pdf
non accessibile
Descrizione: Executive Summary della tesi
Dimensione
1.11 MB
Formato
Adobe PDF
|
1.11 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/246661