Accurate forecasting of the Frequency Restoration Control Error (FRCE) is essential for enabling energy storage systems to optimize their operations on the markets. As renewable penetration increases and introduces greater variability in generation and demand, reliable short-term FRCE prediction - here using 15-minute measurement intervals - becomes critical for ensuring frequency stability, operational reliability, and efficient storage dispatch. This study investigates short-term FRCE forecasting across multiple horizons, including day-ahead and near real-time scenarios. Several machine learning models are developed and assessed: Linear Regression, XGBoost, LSTM networks, and a hybrid ensemble combining SVM and XGBoost. The models use temporal, meteorological, and system-related inputs, enriched with lagged features to capture daily and seasonal FRCE patterns. Using extensive real-world data from Italy as the case study, the models are evaluated under diverse system conditions, with special attention to periods of high variability. For day-ahead forecasts, the hybrid ensemble delivers the best performance, achieving an RMSE of 21.658 and an MAE of 16.587, and remains robust during rapid FRCE fluctuations. For near real-time forecasting (15 minutes ahead), the model anticipates extreme events and peak deviations effectively, reaching an RMSE of 16.335 and an MAE of 12.401, demonstrating strong accuracy even during high-demand, high-variability periods.
La previsione accurata del Frequency Restoration Control Error (FRCE) è essenziale per consentire ai sistemi di accumulo di energia di ottimizzare le proprie operazioni sui mercati. Con l’aumento della penetrazione delle energie rinnovabili e la conseguente maggiore variabilità nella generazione e nella domanda, una previsione affidabile del FRCE a breve termine - qui basata su intervalli di misurazione di 15 minuti - diventa fondamentale per garantire la stabilità della frequenza, l’affidabilità operativa e una gestione efficiente dello stoccaggio. Questo studio indaga la previsione a breve termine del FRCE su più orizzonti temporali, includendo scenari day-ahead e near real-time. Sono stati sviluppati e valutati diversi modelli di machine learning: Regressione Lineare, XGBoost, reti LSTM e un ensemble ibrido che combina SVM e XGBoost. I modelli utilizzano input temporali, meteorologici e relativi al sistema, arricchiti con feature laggate per catturare i pattern giornalieri e stagionali del FRCE. Utilizzando dati reali estesi dall’Italia come caso di studio, i modelli sono stati valutati in diverse condizioni di sistema, con particolare attenzione ai periodi di alta variabilità. Per le previsioni day-ahead, l’ensemble ibrido offre le migliori prestazioni, raggiungendo un RMSE di 21,658 e un MAE di 16,587, e si mantiene robusto durante rapide fluttuazioni del FRCE. Per la previsione near real-time (15 minuti in anticipo), il modello anticipa efficacemente eventi estremi e picchi di deviazione, raggiungendo un RMSE di 16,335 e un MAE di 12,401, dimostrando elevata accuratezza anche nei periodi di alta domanda e elevata variabilità.
Forecasting frequency restoration control error in day-ahead and near-real-time horizons with data-driven approach
BAKAEVA, ALINA
2024/2025
Abstract
Accurate forecasting of the Frequency Restoration Control Error (FRCE) is essential for enabling energy storage systems to optimize their operations on the markets. As renewable penetration increases and introduces greater variability in generation and demand, reliable short-term FRCE prediction - here using 15-minute measurement intervals - becomes critical for ensuring frequency stability, operational reliability, and efficient storage dispatch. This study investigates short-term FRCE forecasting across multiple horizons, including day-ahead and near real-time scenarios. Several machine learning models are developed and assessed: Linear Regression, XGBoost, LSTM networks, and a hybrid ensemble combining SVM and XGBoost. The models use temporal, meteorological, and system-related inputs, enriched with lagged features to capture daily and seasonal FRCE patterns. Using extensive real-world data from Italy as the case study, the models are evaluated under diverse system conditions, with special attention to periods of high variability. For day-ahead forecasts, the hybrid ensemble delivers the best performance, achieving an RMSE of 21.658 and an MAE of 16.587, and remains robust during rapid FRCE fluctuations. For near real-time forecasting (15 minutes ahead), the model anticipates extreme events and peak deviations effectively, reaching an RMSE of 16.335 and an MAE of 12.401, demonstrating strong accuracy even during high-demand, high-variability periods.| File | Dimensione | Formato | |
|---|---|---|---|
|
2025_12_Bakaeva_Thesis_01.pdf
accessibile in internet per tutti
Descrizione: Thesis text
Dimensione
3.17 MB
Formato
Adobe PDF
|
3.17 MB | Adobe PDF | Visualizza/Apri |
|
2025_12_Bakaeva_Executive Summary_02.pdf
accessibile in internet per tutti
Descrizione: Executive summary text
Dimensione
916.1 kB
Formato
Adobe PDF
|
916.1 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/10589/246984