Session: 02-01: AI for Energy Sustainability I

Paper Number: 156091

156091 - Evaluating Neural Network-Based Energy Forecasting With Mixed Variables and Limited Data 

Abstract: 

Large buildings are currently one of the top electrical energy consumers, accounting for 30% of global final energy consumption and contributing to 26% of energy-related emissions. Accurate electricity demand forecasting supports energy management processes, enabling efficient energy consumption in large buildings. These forecasts can also be used as a proxy to generate energy efficiency policies, which are essential in the context of climate change.

Different models have been applied to forecast electricity demand across various building types, including statistical methods and artificial intelligence (AI) models, particularly deep neural network architectures. Neural Networks models discover non-linear relationships between input data (historical energy data and exogenous data) and output data (forecasted consumption), thereby allowing energy consumption to be predicted based on previous energy use patterns and external factors influencing demand. External variables, such as meteorological data, enhance forecast accuracy by leveraging their correlation with electricity demand. Time-related variables also help the model, because it accounts for both seasonal and daily consumption trends, adjusting differences in demands on weekends, holidays, and across seasons.

This study proposes a deep neural network model to forecast electricity demand on a university campus using multivariate time series data, including electricity demand, meteorological, and time-related variables. The dataset was provided by Escuela Superior Politécnica del Litoral (ESPOL) a university campus in Guayaquil, Ecuador, and incorporates three years of electricity demand and meteorological data. A university campus, as a large building with large and variable energy demand, is an ideal case for electricity forecasting. In this case it is appropriate to include exogenous variables like meteorological data and specially time-related variables which can help the model align with the campus usage patterns, affected by academic schedules, workdays, weekends, holidays, and vacation periods. These inputs enable the model to adapt effectively to the distinct and fluctuating energy demands of a university context. Four forecasting horizons (1 day, 2 days, 7 days, and 30 days) were chosen to address various planning needs: immediate adjustments in case of demand fluctuations, as well as longer-term resource and budget planning. These horizons also help to examine how the performance of deep neural networks changes as the prediction time window increases.

In this study, electricity demand forecasting was carried out on a university campus on the four selected horizons. We evaluated four models: SARIMAX (a statistical model widely applied in electrical demand forecasting due to its ability to incorporate exogenous variables), along with three deep learning models (Recurrent Neural Network (RNN), Gated Recurrent Unit (GRU), and Long Short-Term Memory (LSTM)). The results illustrate the efficacy of deep neural networks in forecasting the campus's electricity demand, with consistently high performance across all evaluated time horizons. Among these networks, the LSTM demonstrates superior performance, particularly over the 7-day and 30-day horizons, indicating its ability to maintain accuracy as the prediction timeframe extends, due for their ability to learn long-term dependencies.

This study also seeks to answer a related question: how much historical data is necessary to train a well-performing model across these forecasting horizons? Findings indicate that using a year of data with all time-related variables achieves the best performance metrics, with no significant improvements observed when extending the training set beyond one year. However, it is still possible to achieve a well-performing model with a shorter dataset, by omitting time-related variables such as month and week, providing reliable forecasts with just three months of data. This approach is particularly feasible for buildings with limited historical data and where rapid deployment is desirable, outperforming many traditional statistical methods.

This study underscores the practical value of deep neural networks, particularly LSTM, for electricity demand forecasting in large buildings like university campuses. Future research should explore the potential of deep neural networks and other emerging architectures to achieve further improvements with even shorter training periods or alternative exogenous variables. This proposed methodology has broad applicability, adaptable for various types of buildings, contributing to more sustainable energy management practices.

Presenting Author: Jose Cordova-Garcia ESPOL University

Presenting Author Biography: Jose Cordova-Garcia is an Associate Professor with the Faculty of Electrical Engineering and Computer Science at ESPOL University. He received his PhD degree from the State University of New York at Stony Brook. Currently he leads the Artificial Intelligence and Analytics for Sustainability initiative. His research interests include differentiable optimization, cyber-physical systems, machine learning for optimization, and data issues in remote sensing. He was a recipient of a Fulbright Scholarship from the Department of State during his M.S. Studies.