Temporal Features Boost HVAC Load Forecasting Accuracy

Researchers investigating short-term energy demand in smart buildings have demonstrated that incorporating temporal load-history information substantially improves the accuracy of heating, ventilation, and air conditioning (HVAC) load forecasting. Published in Energy Informatics, the study utilized real operational data from the PLEIAData building dataset to evaluate how different temporal feature representations influence predictive performance. The findings indicate that models relying solely on contemporaneous environmental variables exhibit limited predictive capability, whereas the inclusion of recent operational history allows for significantly more stable and accurate one-hour-ahead predictions.

The research focused on the Pleiades building complex located at the University of Murcia, Spain, specifically analyzing data from Building A. The dataset covers nearly a full year of operation, from January 1 to December 18, 2021, providing detailed measurements of indoor environmental conditions, outdoor weather variables, and HVAC system activity. By using hourly-resolution data, the study aligned with common practices in short-term building energy forecasting, where one-hour-ahead predictions are essential for real-time operation and demand-response strategies.

To assess the impact of feature engineering, the authors evaluated three forecasting models representing both machine learning and deep learning paradigms: Extreme Gradient Boosting (XGBoost), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU). These models were tested under progressively enriched temporal feature configurations, allowing for a controlled comparison of how different input representations affect performance. The experimental design isolated the effects of temporal representation from model-architecture differences by ensuring all models received identical temporally enriched inputs.

The feature engineering process began with baseline environmental variables, such as outdoor temperature and solar radiation. The authors then introduced autoregressive features, which include the most recent observed HVAC load, to capture the strong temporal persistence inherent in building energy systems. Further enrichment involved rolling statistical features, such as rolling means computed over the most recent three observations, which help smooth fluctuations and capture short-term trends in both load and input variables.

Results showed that baseline models using only environmental variables achieved an R² value of approximately 0.34, indicating limited predictive power. However, when autoregressive load information was added, performance improved dramatically, with R² values exceeding 0.70 for most configurations. The XGBoost model combined with the extended temporal feature representation achieved one of the best and most stable performances, reaching an R² value of approximately 0.799. This suggests that the primary driver of forecasting accuracy is the inclusion of recent load history rather than the complexity of the model architecture itself.

Interpretability analyses using permutation feature importance confirmed that the current HVAC load is the dominant predictor, substantially more influential than any other variable. While environmental factors like solar radiation and its rolling representation provide complementary information, their contribution remains secondary to the temporal load dynamics. This finding underscores the importance of thermal inertia and operational continuity in shaping near-future energy demand, as recent system states contain valuable predictive information for short-term forecasting tasks.

Robustness was further evaluated through an expanding-window walk-forward validation framework, which tests model generalization across multiple temporal segments. This approach revealed that performance remains relatively stable across different feature set complexities, with diminishing returns observed when expanding from a balanced feature set to a more extended one. The walk-forward validation results were lower than those from a single train-test split, reflecting a more realistic assessment of model generalization under varying temporal conditions and data distributions.

Ablation studies examined the impact of specific feature components, such as lag depth and rolling window size. Extending the lag structure beyond the current load provided only modest improvements, while larger rolling windows led to performance degradation due to excessive smoothing. The combination of rolling mean and standard deviation statistics yielded the best observed results, indicating that short-term variability information provides complementary predictive value. These findings suggest that carefully engineered temporal representations are more critical than increasing feature complexity or model depth.

The study acknowledges several limitations, including the focus on a single building and the lack of direct occupancy measurements. The transferability of the proposed framework across different buildings, occupancy patterns, and climatic conditions was not explicitly evaluated. Additionally, the recurrent neural networks were used with a sequence length of one, functioning primarily as nonlinear regressors over engineered features rather than fully exploiting long-term sequence-learning capabilities. Future research may explore cross-building validation and multi-horizon forecasting to address these gaps.

Overall, the study highlights the central role of temporal feature design in short-term HVAC load forecasting. By demonstrating that robust temporal representations can significantly enhance predictive accuracy across diverse model architectures, the research provides valuable insights for developing intelligent energy-management systems. The findings support the adoption of leakage-aware evaluation and interpretable forecasting analysis to ensure reliable and effective building energy prediction frameworks in practical applications.

The study was conducted as a controlled comparative evaluation using real operational data from the Pleiades building complex at the University of Murcia, Spain. The dataset, introduced by Ibarra et al., provides detailed information on building operation, HVAC system activity, and environmental conditions measured in a real smart-building environment. The Pleiades complex consists of three adjacent buildings, A, B, and C, which exhibit similar construction characteristics and operational behavior. To avoid redundant modeling while preserving representative HVAC operational characteristics, the experiments were conducted using only Building A. The dataset covers the period between January 1 and December 18, 2021, representing nearly a full year of building operation. It integrates multiple sources of information, including indoor environmental conditions, outdoor weather variables, HVAC operational states, and building energy consumption. These measurements enable comprehensive analysis of building energy behavior and provide a valuable resource for data-driven research in smart building energy systems.

The forecasting problem was defined as one-hour-ahead HVAC load prediction, a task widely adopted in practical HVAC scheduling, building energy management, and demand-response applications. The choice of hourly resolution was motivated by the need to reduce high-frequency fluctuations and facilitate clearer analysis of short-term operational dynamics and temporal feature representations. The dataset provides measurements at multiple temporal resolutions, including both 10-minute and hourly aggregated data. In this study, the hourly-resolution dataset corresponding to Building A was selected. The processed version provided by the dataset authors was used, containing cleaned and synchronized measurements suitable for forecasting analysis. The forecasting target is the hourly HVAC energy consumption, denoted as dif_cons_real, which represents the energy demand associated with HVAC system operation. All variables listed in the dataset are treated as candidate predictors, while the final feature selection is performed based on statistical analyses and modeling considerations.

To ensure a comprehensive evaluation, Pearson correlation analysis was initially conducted on the full set of candidate variables. The correlation heatmap indicates that multiple variables exhibit moderate to strong linear relationships with HVAC energy consumption, confirming that building energy demand is influenced by a combination of environmental, operational, and control-related factors. In particular, device type, the number of active HVAC devices, outdoor temperature, indoor temperature, and the seasonal indicator show moderate positive correlations with HVAC load. In contrast, the HVAC control variable exhibits a moderate negative correlation, reflecting its regulatory role in system operation. In addition to their relationships with the target variable, the predictors themselves exhibit strong interdependencies. For example, outdoor temperature is strongly correlated with indoor temperature and vapor pressure deficit, while indoor temperature shows a strong negative correlation with the control variable and strong positive correlations with both seasonal indicators and dew point temperature. These relationships indicate that temperature-related and humidity-related variables capture overlapping physical information associated with the building’s thermal environment.

A key observation from the correlation analysis is the near-perfect positive correlation between the number of active HVAC devices and the system-level variable representing device type. This indicates that system configuration and operational activity are tightly coupled in the dataset. In practical terms, buildings equipped with similar device types tend to exhibit consistent activation patterns, leading to a strong statistical dependency between system configuration and operational intensity. This suggests that the device type variable may not provide entirely independent information but instead reflects underlying operational patterns already captured by the number of active devices. From a modeling perspective, this redundancy implies that including both variables may introduce multicollinearity without substantially improving predictive performance. This observation is consistent with the experimental results, where the inclusion of the device type variable leads to only marginal improvements compared to configurations already containing the number of active devices. Therefore, feature selection should consider not only the strength of correlation with the target variable but also the dependencies among predictors in order to construct a compact and informative feature set.

However, it is important to emphasize that Pearson correlation captures only linear dependencies between variables. As a result, variables that appear weakly correlated with the target may still contain predictive information through nonlinear relationships or interaction effects. To address this limitation, a complementary nonlinear feature relevance analysis based on mutual information is performed. Unlike Pearson correlation, mutual information captures both linear and nonlinear dependencies between variables, making it suitable for identifying complex relationships and interaction effects that may not be revealed through linear analysis. In this study, mutual information scores were computed between candidate variables at time t and the HVAC load at the next time step, allowing the evaluation of predictive relevance in a one-step-ahead forecasting framework. The results show that the current HVAC load exhibits the highest mutual information with the target. This is expected, as the HVAC system exhibits strong temporal persistence due to building thermal inertia and operational continuity. Nevertheless, this result quantitatively confirms that short-term HVAC load dynamics are primarily governed by recent system states.

Importantly, beyond the autoregressive component, several additional variables demonstrate significant nonlinear relevance. In particular, the HVAC control parameter and indoor temperature show strong mutual information scores, indicating that control signals and internal thermal conditions play a critical role in shaping load evolution. Furthermore, the number of active HVAC devices, the system-level variable, and outdoor temperature also exhibit meaningful contributions. These findings provide an important extension to the correlation analysis: while some variables appeared less prominent or ambiguous under linear correlation, they exhibit higher relevance under mutual information. This suggests the presence of nonlinear dependencies and interaction effects, particularly between control signals, system characteristics, and environmental conditions. In contrast, seasonal indicators and weather-derived variables, including dew point temperature, vapor pressure deficit, and solar radiation, exhibit lower but non-negligible mutual information scores. These variables are therefore interpreted as providing complementary contextual information rather than acting as primary drivers of short-term load variability.

Overall, the mutual information analysis demonstrates that HVAC load prediction is influenced by a combination of autoregressive, environmental, operational, and control-related variables. However, even when nonlinear dependencies are taken into account, the predictive structure remains strongly dominated by recent load history. This finding supports the central premise of this study: temporal dependencies encoded through recent load observations play a more significant role than individual exogenous variables in short-term HVAC load forecasting. The structured and hierarchical feature selection strategy was adopted to construct baseline feature sets with varying levels of complexity. The correlation analysis revealed several groups of highly correlated variables representing similar physical phenomena. Notably, strong pairwise relationships were observed between indoor temperature and HVAC control signals, operational and system-level variables, and temperature-related weather indicators. These relationships indicate that multiple variables capture overlapping information about the building’s thermal state, system operation, and environmental conditions.

To account for this redundancy while preserving physical interpretability, three nested baseline feature sets were defined using different selection criteria. In the compact feature set, only one representative variable was selected from each highly correlated group, prioritizing variables with clear physical interpretation and consistent relevance across both correlation and mutual information analyses. This results in a minimal and interpretable feature set capturing the dominant thermal, operational, and control dynamics. In the balanced feature set, only near-perfect correlations were treated as exclusion criteria. Moderately correlated variables were retained when they represent distinct physical mechanisms, such as environmental conditions, control signals, and seasonal effects. This set aims to capture additional contextual information without introducing excessive redundancy. In the extended feature set, correlation-based constraints were relaxed, and variables identified as relevant through mutual information analysis were included. This configuration represents an upper-bound scenario in terms of feature richness, allowing the evaluation of whether additional variables provide meaningful predictive improvements.

Subject of Research: Impact of temporal feature engineering on short-term HVAC load forecasting accuracy

Article Title: Improving short-term HVAC load forecasting via temporal feature engineering: insights from the PLEIAData building dataset

Article References: Ermis, S. (2026). Improving short-term HVAC load forecasting via temporal feature engineering: insights from the PLEIAData building dataset. Energy Informatics, 9(1), Article 69. https://doi.org/10.1186/s42162-026-00670-5

Image Credits: AI Generated

DOI: 10.1186/s42162-026-00670-5

Keywords: HVAC load forecasting, temporal feature engineering, PLEIAData, XGBoost, LSTM, GRU, building energy consumption, machine learning, smart buildings, energy management, Improving, short-term