Air Conditioning Load Prediction Based on Synergistic Feature Selection

Hua Meng, Hao Sun, Saishuai Dai, Tingting Xu, Yuting Yao, Yingjun Ruan2025 · DOI

TL;DR

Researchers have developed a "physics-informed" feature selection framework that significantly boosts the accuracy of cooling load predictions for metro stations. By combining mathematical metrics like Mutual Information with thermodynamic principles, the model achieves up to a 23% reduction in error compared to standard data-driven methods without increasing computational costs.

What Problem Does This Paper Address?

Predicting the air conditioning (AC) load of large buildings—especially underground metro stations—is notoriously difficult. These systems are influenced by a complex web of factors: outdoor weather, indoor passenger density, and the massive thermal inertia of the surrounding soil.

While modern machine learning models (like LSTMs) are great at finding patterns, they often struggle with "feature selection." Standard mathematical approaches often pick features that are statistically correlated but physically redundant (e.g., using too many historical load data points), while ignoring "weaker" statistical signals that are physically vital (e.g., indoor humidity). This leads to models that are either overfitted, computationally heavy, or physically nonsensical.

Methods

The authors propose a two-stage Synergistic Feature Selection framework:

  1. Preliminary Screening: They start with a pool of temporal features (historical loads, temperatures, and humidity). They use Pearson correlation to find associations but then apply a thermodynamic consistency screen. For example, if a feature suggests that rising temperatures decrease cooling demand during summer, it’s discarded as physically implausible.
  2. MI-Based Ranking & Synergistic Refinement: They use Mutual Information (MI)—a measure of how much information one variable provides about another—to rank the remaining 46 features.
  3. The "Synergy" Step: Instead of just picking the top-ranked MI features (which are usually just historical load values), they strategically incorporate features from different physical categories (Outdoor Temp, Indoor Temp, Humidity) even if they have lower MI scores. This reduces "intra-category redundancy" and improves the model's ability to map different physical drivers to the final load.

The framework was tested primarily using a Long Short-Term Memory (LSTM) network and validated with a Backpropagation Neural Network (BPNN).

Datasets

The study utilized a high-resolution dataset from a metro station in China’s "Hot Summer and Warm Winter" region.

  • Scope: Operational records from May 1 to August 31, 2020.
  • Granularity: Data collected at 2-minute intervals, totaling 88,560 samples.
  • Variables: Chilled/cooling water temperatures, flow rates, power consumption, indoor/outdoor air temperature, and relative humidity.

Key Insights and Ideas

  • More Math $\neq$ Better Physics: Relying solely on MI rankings often leads to selecting multiple historical load timestamps (e.g., $L_{t-1}, L_{t-2}, L_{t-3}$). The authors found that after the first few historical points, adding more creates "feature competition" that actually degrades performance.
  • Diversity Over Rank: A "weaker" feature from a new physical category (like outdoor temperature $T_{t,o}$) is often more valuable to the model than a "stronger" feature from a category already represented.
  • The "Sweet Spot": The optimal feature set wasn't the one with the highest total MI, but the one that balanced historical load data with key indoor and outdoor thermal indicators. Specifically, a combination of four historical load values plus outdoor and indoor temperature yielded the best results.

Key Results

The synergistic approach outperformed conventional MI-based selection across the board:

  • Accuracy: Achieved a 5.62% increase in the correlation coefficient ($R$).
  • Error Reduction: Reduced Mean Absolute Error (MAE) by 16.13%, Root Mean Square Error (RMSE) by 22.76%, and Mean Absolute Percentage Error (MAPE) by 16.44%.
  • Efficiency: The computational overhead was negligible, with training times increasing only slightly (from ~1048s to ~1203s), a small price to pay for a significant jump in reliability.

Limitations and Future Work

While the results are promising, the authors note a few areas for growth:

  • Offline vs. Online: The current model is designed for offline prediction. Further optimization is required to enable real-time, online load prediction for active HVAC control.
  • Generalizability: While the method is applicable to various building types and climates, the specific "optimal" feature set (the six temporal features identified here) is likely unique to metro stations in humid, subtropical climates.
  • Automation: Future work could involve automating the "synergistic" step, which currently requires some domain expertise to categorize features physically.

Summary updated

Abstract

This study proposes a physics-informed feature selection framework to enhance the accuracy and interpretability of data-driven cooling load prediction models.Using operational data from a metro station HVAC system, we first identify temporal features through Pearson correlation analysis, followed by thermodynamic consistency screening to eliminate physically implausible candidates.Mutual information (MI) is then employed to quantify feature importance, while our Synergistic approach strategically incorporates physically critical but lower-MI features to improve model accuracy.Tested with LSTM models, the proposed method achieves a 5.6% improvement in prediction accuracy and 16-23% error reductions compared to conventional MI-based selection, with negligible computational overhead.Further validation with BPNN and diverse datasets confirmed robustness.By bridging data-driven metrics and domain-specific physics, this approach offers a balanced trade-off between accuracy and interpretability for building energy prediction.
BibTeX
@misc{meng2025air,
  author = {Hua Meng and Hao Sun and Saishuai Dai and Tingting Xu and Yuting Yao and Yingjun Ruan},
  title = {{Air Conditioning Load Prediction Based on Synergistic Feature Selection}},
  year = {2025},
  doi = {10.46855/energy-proceedings-11893},
}
Figure