This thesis aims to perform a comparative analysis of different electrical load modelling techniques applied to the Politecnico di Milano microgrid. The main objective is to identify a reliable and robust forecasting model, suitable for implementation in the University's microgrid, to improve the STLF with forecasting horizons of h+1 and h+24. Accurate demand forecasting would allow for optimised grid management and, if necessary, enable the system to operate independently from the Milan municipal grid during peak consumption periods, thus ensuring greater self-reliability. Through an extensive literature review, the most promising models were identified: SARIMAX, among the statistical models, and Random Forest (RF) and NARX, among the machine learning techniques. The study was developed based on the collection and analysis of quarter-hourly consumption data from different substations within the microgrid, complemented with exogenous meteorological and calendar variables to provide a better context for the time series profile.The assessment of the forecasting performance was carried out through a quantitative analysis, including the evaluation of the error metrics, the evaluation of the importance of the predictor variables, the sensitivity analysis on the optimised model structures and the residual analysis, as well as a qualitative analysis based on the comparison of the forecast curves. The results indicate that for h+1 forecasts, no model stands out with a clear advantage over the others. However, when the forecasting horizon is extended to h+24, the NARX model is superior to RF, as it is more effective in capturing the time dynamics of the series.
Il presente lavoro di tesi si propone di condurre un'analisi comparativa di diverse tecniche di modellazione dei carichi elettrici, applicate alla microgrid del Politecnico di Milano. L'obiettivo principale è individuare un modello predittivo affidabile e robusto, implementabile nel contesto della rete universitaria per migliorare le previsioni di STLF con orizzonti previsionali di h+1 e h+24. Un’accurata previsione della domanda consentirebbe di ottimizzare la gestione della rete ed eventualmente isolarla dalla rete urbana di Milano, nei momenti di massima criticità dei consumi, garantendone maggiore indipendenza. Attraverso un’approfondita analisi della letteratura, sono stati individuati i modelli più promettenti: SARIMAX, tra i modelli statistici, e Random Forest (RF) e NARX, appartenenti alle tecniche di apprendimento automatico. Lo sviluppo del lavoro si è basato su una raccolta e analisi di dati quartorari relativi ai consumi delle diverse cabine della rete, correlati da variabili esogene meteorologiche e di calendario, al fine di contestualizzare al meglio il profilo della serie storica. La valutazione delle performance previsionali è stata condotta attraverso un'analisi quantitativa, che ha incluso la misurazione delle metriche di errore, la valutazione della rilevanza delle variabili predittive, un’analisi di sensitività sulla struttura ottimizzata dei modelli e lo studio dei residui, nonché attraverso un’analisi qualitativa, basata sul confronto delle curve previsionali. I risultati hanno evidenziato che, per le previsioni h+1, nessun modello emerge con un chiaro vantaggio sugli altri. Tuttavia, ampliando l’orizzonte previsionale a (h+24), il modello NARX mostra una superiorità rispetto al RF, dimostrandosi più efficace nella cattura delle dinamiche temporali della serie.
Comparative analysis of predictive models for microgrid electricity demand estimation
Pollenghi, Andrea;Pichierri, Pier Giuseppe
2023/2024
Abstract
This thesis aims to perform a comparative analysis of different electrical load modelling techniques applied to the Politecnico di Milano microgrid. The main objective is to identify a reliable and robust forecasting model, suitable for implementation in the University's microgrid, to improve the STLF with forecasting horizons of h+1 and h+24. Accurate demand forecasting would allow for optimised grid management and, if necessary, enable the system to operate independently from the Milan municipal grid during peak consumption periods, thus ensuring greater self-reliability. Through an extensive literature review, the most promising models were identified: SARIMAX, among the statistical models, and Random Forest (RF) and NARX, among the machine learning techniques. The study was developed based on the collection and analysis of quarter-hourly consumption data from different substations within the microgrid, complemented with exogenous meteorological and calendar variables to provide a better context for the time series profile.The assessment of the forecasting performance was carried out through a quantitative analysis, including the evaluation of the error metrics, the evaluation of the importance of the predictor variables, the sensitivity analysis on the optimised model structures and the residual analysis, as well as a qualitative analysis based on the comparison of the forecast curves. The results indicate that for h+1 forecasts, no model stands out with a clear advantage over the others. However, when the forecasting horizon is extended to h+24, the NARX model is superior to RF, as it is more effective in capturing the time dynamics of the series.File | Dimensione | Formato | |
---|---|---|---|
2025_04_Pichierri_Pollenghi_Tesi.pdf
solo utenti autorizzati a partire dal 12/03/2026
Descrizione: Testo della Tesi
Dimensione
71.11 MB
Formato
Adobe PDF
|
71.11 MB | Adobe PDF | Visualizza/Apri |
2025_04_Pichierri_Pollenghi_Executive_Summary.pdf
solo utenti autorizzati a partire dal 12/03/2026
Descrizione: Testo dell'Executive Summary
Dimensione
706.04 kB
Formato
Adobe PDF
|
706.04 kB | 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/10589/236349