Despite the tightening of energy performance standards for buildings in various countries and the increased use of efficient and renewable energy technologies, it is clear that the sector needs to change more rapidly to meet the Net Zero Emissions (NZE) scenario by 2050. One of the problems that have been analyzed intensively in recent years is that buildings in operation use much more energy than they were designed to. This problem, known as the energy performance gap, is found in many countries and buildings and is often attributed to the poor management of building energy systems. The application of Artificial Intelligence (AI) to Building Energy Management Systems (BEMS) has untapped potential to address this problem and lead to more sustainable buildings. This paper reviews different AI-based models that have been proposed for different applications and different buildings with the intention to reduce energy consumption. It compares the performance of the different AI-based models evaluated in the reviewed papers by presenting the accuracy and error rates of model performance and identifies where the greatest potential for energy savings could be achieved, and to what extent. The review showed that offices have the greatest potential for energy savings (up to 37%) when they employ AI models for HVAC control and optimization. In residential and educational buildings, the lower intelligence of the existing BEMS results in smaller energy savings (up to 23% and 21%, respectively).

Integrating Artificial Intelligence (AI) into heating, ventilation, and air conditioning (HVAC) systems is a promising approach that helps enhance energy efficiency in buildings, which leads to cost savings and provides environmental benefits. However, the effective performance of the AI models depends not only on the model design but also on the data quality, reliability, size, availability, and management. This paper analyses recent studies that apply AI models, specifically Deep Learning and Hybrid models, to achieve energy efficiency in HVAC systems in buildings from a data perspective, examining various aspects of data management. This analysis aims to provide insights into data-related challenges in AIdriven HVAC systems and propose strategies to overcome them, ensuring more accurate, efficient, and reliable models. The findings reveal that combining multiple data types can enhance model performance and generalizability. The findings also indicate that data quality is overlooked by researchers in many studies, where only 31 % of the analysed papers discussed quality issues, reflecting that it is not yet a standard practice in this field. Additionally, this analysis highlights the scarcity of reliable and audited data. Therefore, and in response to this issue, this paper recommends accessible and reliable data resources that can be employed in AI applications for HVAC systems in buildings.

Integrating AI in HVAC systems is a promising approach that helps enhance energy efficiency in buildings, which leads to cost savings and provides environmental benefits. However, the effective performance of these AI models, especially in HVAC systems, depends not only on the model design but also on the data's quality, reliability, size, availability, and management. Data plays an important role in determining the accuracy and reliability of the AI model's performance. This paper analyses recent studies that apply AI models to achieve energy efficiency in HVAC systems from a data perspective, examining various aspects of data management in Deep Learning and Hybrid models applied to HVAC in buildings, such as data availability, the different data sources, type, quality issues, and data splitting methods. Through this analysis, the paper aims to provide insights into data-related challenges and recommend ways to overcome and mitigate them to develop AI models that perform more effectively. The paper highlights the importance of developing better data-handling practices to have more accurate, efficient, and reliable AI models in HVAC systems. The findings reveal that combining multiple data types can enhance model performance and generalizability. Moreover, the analysis concludes that the main data type for residential buildings is simulated data rather than real-world data; this could be due to privacy concerns. Meanwhile, commercial buildings have commonly utilized more structured and reliable dataset sources, enabling more precise modelling. The findings also indicate that data quality is overlooked by researchers in many studies, where only 31 % of the analysed papers discussed quality issues, reflecting that it is not yet a standard practice in this field. Additionally, this analysis addresses the scarcity of reliable and audited data. Therefore, and in response to this issue, this paper recommends accessible and reliable data resources that can be employed in AI applications for HVAC systems in buildings.

Feature engineering has become a common step in AI-based HVAC forecasting, often involving variables calculated from raw building management system (BMS) measurements, such as temperature differences, setpoint tracking deviations, airflow balance indicators, rolling statistics, and temporal or seasonal descriptors. Accurate short-term forecasting can provide a baseline of expected operation for anomaly and fault detection and can support control optimization and operator decision making. However, real-world deployment is complicated due to differences in BMS sensor availability and data quality, as well as the preprocessing and maintenance burden associated with complex feature sets. The actual contribution of these features to the performance of AI forecasting remains underexplored, particularly for short-term prediction of air handling unit (AHU) operation. This study evaluates the impact of features on short-term AHU forecasting using three deep learning (DL) architectures: Temporal Convolutional Networks (TCNs), Long Short-Term Memory (LSTM) networks, and a hybrid CNN–LSTM model. An actual operational AHU dataset from a BMS was used to predict key operational variables, including supply and extract air temperatures, supply and extract fan operating signals, and supply air temperature setpoint-tracking error. Fan signal balance was additionally evaluated as a derived indicator calculated from the two predicted fan signals. Four input configurations were evaluated: (i) full (74 inputs), containing raw BMS measurements, short-cycle temporal variables, engineered and dynamic features, and annual-calendar information; (ii) no annual calendar (68 inputs), identical to full but excluding annual-calendar variables; (iii) raw + short-cycle temporal (20 inputs); and (iv) raw-only (12 inputs). The models used a 60-min input history to forecast the following 30-min at one-minute resolution. Persistence and Ridge models were included as reference baselines. All models were trained and tested on identical data splits and forecasting horizons to ensure a fair comparison. Each DL experiment was repeated across five independent runs, and performance was evaluated using MAE, RMSE, and R². The TCN showed the strongest overall DL performance. Raw-only achieved the highest mean R² in 11 of 15 architecture–target comparisons using just 12 inputs. The best mean DL R² ranged from 0.916 for the fan signals to 0.993 for extract air temperature. Annual-calendar features improved the TCN results but provided no consistent benefit for the LSTM or CNN–LSTM. Ridge slightly outperformed the best DL configurations for temperature-related targets, reflecting the strong short-term continuity of these signals. These findings show that recent raw BMS measurements contain most of the information needed for accurate 30-min AHU forecasting, while explicit seasonal and engineered features provide limited additional value. The resulting simpler models may in the future be used as forecasting components in predictive control and fault detection systems. However, their control and energy-saving benefits must be tested separately.