Forecasting electricity consumption is a major challenge for sustainable energy management, especially in cities with rapid population growth and accelerated urbanization, where energy demand typically exceeds production. Short-term load forecasting (STLF) emerges as a pivotal tool in the decision-making processes underlying the day-to-day operation and planning of modern power grids. It has become an indispensable component of efficient energy management systems. While existing literature has extensively explored STLF using conventional statistical methods such as ARIMA and advanced deep learning techniques like LSTM, CNN, and Transformer architectures, these studies often lack a comprehensive examination of these methodologies across multiple short-term forecast horizons. In this study, leveraging historical hourly electrical load data from Kinshasa, we employ popular statistical models such as ARMA alongside advanced deep learning architecture LSTM to identify the most effective model for horizons t+1 (very short-term), t+12, and t+24. Our comparative analysis shows that ARMA model has the best MAPE for t+1, but LSTM exhibit superior performance for longer forecast horizons. These findings suggest the necessity of employing a diverse array of models, each specifically tailored to a specific forecast horizon, to formulate an optimal forecasting policy crucial for resilient energy management in dynamic urban landscapes.
Multi-Horizon Short-Term Electrical Load Forecasting: A Comparative Analysis of Statistical Models and Deep Neural Networks
Paciello V.;
2024-01-01
Abstract
Forecasting electricity consumption is a major challenge for sustainable energy management, especially in cities with rapid population growth and accelerated urbanization, where energy demand typically exceeds production. Short-term load forecasting (STLF) emerges as a pivotal tool in the decision-making processes underlying the day-to-day operation and planning of modern power grids. It has become an indispensable component of efficient energy management systems. While existing literature has extensively explored STLF using conventional statistical methods such as ARIMA and advanced deep learning techniques like LSTM, CNN, and Transformer architectures, these studies often lack a comprehensive examination of these methodologies across multiple short-term forecast horizons. In this study, leveraging historical hourly electrical load data from Kinshasa, we employ popular statistical models such as ARMA alongside advanced deep learning architecture LSTM to identify the most effective model for horizons t+1 (very short-term), t+12, and t+24. Our comparative analysis shows that ARMA model has the best MAPE for t+1, but LSTM exhibit superior performance for longer forecast horizons. These findings suggest the necessity of employing a diverse array of models, each specifically tailored to a specific forecast horizon, to formulate an optimal forecasting policy crucial for resilient energy management in dynamic urban landscapes.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.