Most electric energy consumption in the building sector is provided by fossil fuels, leading to high greenhouse gas emissions. However, the increasing need for sustainable infrastructure has triggered a significant trend toward smart buildings, which enable optimal and efficient resource usage. In this context, accurate mid-term energy load forecasting is crucial for energy management. This study proposes a hybrid forecasting model obtained through the combination of machine learning (ML) and deep learning (DL) approaches designed to enhance forecasting accuracy at an hourly granularity. The hybrid two-layer architecture first investigates the model’s performance individually, such as decision tree (DT), random forest (RF), support vector regression (SVR), Extreme Gradient Boosting (XGBoost), FireNet, and long short-term memory (LSTM), and then combines them to leverage their complementary strengths in a two-layer hybrid design. The performance of these models is assessed on smart building energy datasets with weather data, and their accuracy is measured through performance metrics such as mean squared error (MSE), root mean squared error (RMSE), and R-squared ($R^2$). The collected results show that the XGBoost outperformed other ML models. However, the hybrid model obtained by combining FireNet and XGBoost models delivers the highest overall accuracy for the performance parameters. These findings highlight the effectiveness of hybrid models in terms of prediction accuracy. This research contributes to reliable energy forecasting and supports environmentally sustainable practices.

Intensive energy consumption in the building sector remains one of the primary contributors to climate change and global warming. Within Renewable Energy Communities (RECs), improving energy management is essential for promoting sustainability and reducing environmental impact. Accurate forecasting of energy consumption at the community level is a key tool in this effort. Traditionally, engineering-based methods grounded in thermodynamic principles have been employed, offering high accuracy under controlled conditions. However, their reliance on exhaustive building-level data and high computational costs limits their scalability in dynamic REC settings. In contrast, Artificial Intelligence (AI)-driven methods provide flexible and scalable alternatives by learning patterns from historical consumption and environmental data. This study investigates three Machine Learning (ML) models, Decision Tree (DT), Random Forest (RF), and CatBoost, and one Deep Learning (DL) model, Convolutional Neural Network (CNN), to forecast community electricity consumption using real smart meter data and local meteorological variables. The study focuses on a REC in Loureiro, Portugal, consisting of 172 residential users from whom 16 months of 15 min interval electricity consumption data were collected. Temporal features (hour of the day, day of the week, month) were combined with lag-based usage patterns, including features representing energy consumption at the corresponding time in the previous hour and on the previous day, to enhance model accuracy by leveraging short-term dependencies and daily repetition in usage behavior. Models were evaluated using Mean Squared Error (MSE), Mean Absolute Percentage Error (MAPE), and the Coefficient of Determination $R^2$. Among all models, CatBoost achieved the best performance, with an MSE of 0.1262, MAPE of 4.77%, and an $R^2$ of 0.9018. These results highlight the potential of ensemble learning approaches for improving energy demand forecasting in RECs, supporting smarter energy management and contributing to energy and environmental performance.

Renewable Energy Communities (RECs) are crucial in advancing decentralised and sustainable energy systems. Accurate forecasting of short-term electricity demand is essential to support REC operational planning, improve energy efficiency and reduce dependence on external supply. This study presents a practical and effective forecasting framework for hourly electric load in a residential REC in Loureiro, Portugal. The approach is based exclusively on exogenous variables like weather conditions and calendar-based features. Two tree-based machine learning models, Extreme Gradient Boosting (XGBoost) and Categorical Boosting (CatBoost), were applied to predict REC electric energy demand. The performance of the models was evaluated using standard regression metrics, including the Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and the coefficient of determination ( R ² ). Both models achieved high accuracy on unseen data, with CatBoost yielding significantly better results: MAE, RMSE, and R ² are equal to 1.21 kWh, 2.31 kWh, and 0.9718, respectively. An intraday 24-hour comparison confirmed the model’s ability to capture intraday consumption dynamics. These findings highlight the potential of exogenous feature-driven forecasting approaches to support efficient energy management in communityscale systems.

Residential batteries in energy communities usually run local rule-based controllers whose community-level effect is rarely checked against measured data. This study uses one year of smart-meter, PV and battery records from a real 20-household Irish community (24 kWp PV, 200 kWh storage) to test how much can be achieved with locally available data only. The existing controllers charged all batteries in the same night-tariff hours. With household flows netted hourly at the community level, this raised grid import by 14.5% and CO₂ emissions by 15.8% compared with no batteries, raised the maximum hourly import from 33 to 83 kW, and created a new morning peak. A lightweight framework, combining a drift-adaptive LightGBM load forecast and a mixed-integer dispatch with real-time correction, reduced grid import by 10.5–13.5%, emissions by 11.6–14.6%, and energy imported above a 14.1 kW reference threshold by 87% relative to the measured operation, with each scenario computed in seconds. A forecast-free controller acting on the same threshold reduced this exceedance at least as much, while the forecast-based optimisation lowered the import by a further 2.6%. LightGBM was 10.9% more accurate than a seven-day average, yet its dispatch did not differ significantly (Wilcoxon p = 0.53); correcting the systematic underestimation of daily peaks reduced the energy above the threshold by 14%. Sharing PV reduced imports by 9.7% and benefited mainly households without PV, whereas pooled storage mainly relieved the grid, with the effect increasing with better information.