Feature optimization with metaheuristics for Artificial Neural Network-based chiller power prediction
Building energy efficiency is crucial for global sustainability efforts, with chillers representing major energy consumers in commercial buildings. Accurate prediction of chiller power consumption remains challenging due to complex operational parameters, with feature selection being critical for mo...
Saved in:
| Main Authors: | , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Elsevier LTD
2025
|
| Subjects: | |
| Online Access: | http://umpir.ump.edu.my/id/eprint/44263/1/Feature%20Optimization%20with%20Metaheuristics%20for%20Artificial%20Neural%20Network-based%20Chiller%20Power%20Prediction.pdf http://umpir.ump.edu.my/id/eprint/44263/ https://doi.org/10.1016/j.jobe.2025.112561 https://doi.org/10.1016/j.jobe.2025.112561 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Building energy efficiency is crucial for global sustainability efforts, with chillers representing major energy consumers in commercial buildings. Accurate prediction of chiller power consumption remains challenging due to complex operational parameters, with feature selection being critical for model performance. This study aims to develop an improved feature selection approach that enhances prediction accuracy while reducing computational complexity in chiller consumption forecasting. This paper presents a novel Evolutionary Mating Algorithm (EMA) hybridized with Artificial Neural Networks (ANN) for optimizing feature selection. The EMA-ANN approach was compared against other metaheuristic-ANN hybrid models using operational data from a commercial building's chiller system. EMA-ANN demonstrated superior prediction accuracy with the lowest Mean Absolute Error (0.2235), Root Mean Square Error (0.4150), and highest coefficient of determination (R² = 0.9689). The algorithm identified seven optimal features primarily comprising temperature and humidity parameters. The algorithm’s unique evolutionary mating mechanism with adaptive crossover rate (Cr = 0.85), enabled effective feature space exploration, resulting in a 38.3% reduction in RMSE and 6.0% improvement in R2 compared to models without feature selection. This research contributes a novel hybrid model, identifies key features for chiller power prediction, and establishes a benchmark for evaluating feature selection algorithms in building energy applications. |
|---|