New Algorithm Enhances Heating and Cooling Load Predictions

By Dr Silpaja Chandrasekar, PhDReviewed by Susha Cheriyedath, M.Sc.Jul 3 2024

In a paper published in the journal Energies, researchers explored predicting heating and cooling loads using machine learning (ML) algorithms. They focused on selecting the best regressor and tuning hyperparameters with a novel multi-objective plum tree algorithm (MOPTA).

They optimized five hyperparameters for each by comparing the extra trees, gradient boosting regressors (GBR), and random forest regressors (RFR) from sklearn. The algorithm achieved competitive root mean square errors (MSE) for heating and cooling, comparable to results from standard multi-objective algorithms and prior studies using the same dataset.

Background

Past work has explored ML for predicting heating and cooling loads using techniques like support vector regression (SVR), K-means, ensemble learning, and neural networks (NN), achieving superior performance metrics. Ensemble models and NN have demonstrated high accuracy, while optimization algorithms have improved SVM and RF methods. Some studies used gated recurrent units (GRU) to achieve low error values. However, there are often computational efficiency issues when these methods are scaled to handle big datasets or real-time applications.

Optimization Algorithms Evolution

The PTA draws inspiration from natural phenomena like flowering and fruiting cycles of plum trees. It operates by initializing flowers and plums in a search space defined by configurable parameters. Flowers represent potential solutions, and plums retain the best-performing solutions.

Through iterative phases—fruitiness, ripeness, and stores—the algorithm updates flower positions based on fitness evaluations and stochastic processes, ensuring exploration and exploitation of the search space. This iterative refinement continues until convergence, yielding an optimized solution for complex optimization problems.

PTA's efficacy in predicting heating and cooling loads was evaluated using the energy efficiency dataset. Employing algorithms like extra trees regressor (ETR), GBR, and RFR, PTA optimized hyperparameters to minimize RMSE, R-squared, mean absolute error (MAE), and mean absolute percentage error (MAPE) across 5-fold cross-validation. Standardization techniques ensured fairness in training and testing data treatment, enhancing algorithm performance in accurately predicting heating and cooling loads.

MOPTA extends PTA's capabilities by integrating a multi-objective fitness function. By converting plum positions into integers and applying them across algorithm selections and hyperparameters, modeling and optimization aim to minimize heating and cooling RMSE simultaneously. This novel approach enhances solution diversity and quality by using an external archive to manage Pareto-optimal solutions, thereby refining the algorithm's ability to handle complex optimization tasks effectively.

Adapting PTA to MOPTA involves strategic enhancements, including using an external archive to manage Pareto-optimal solutions. This adaptation refines the selection of ripe and unripe plums in each iteration, ensuring optimal solutions are consistently improved. Through these adaptations, MOPTA expands its utility across diverse optimization challenges, demonstrating its versatility and robustness in achieving high-quality solutions across multiple objectives.

Experimental Setup Summary

The experiments were conducted using Python version 3.12.3 and the sklearn library on a high-performance machine running Windows 11 Pro. All computations were central processing unit (CPU)--based to ensure consistency and reliability throughout the experiments. The energy efficiency dataset comprised 768 samples with eight attributes and two responses, focusing on 12 building shapes simulated in Ecotect. It was randomly split into five folds for cross-validation, ensuring robustness in evaluating predictive models for heating and cooling loads.

Hyperparameters were carefully configured across the experiments, guided by ranges inspired by previous research to optimize the performance of algorithms like GBR, RFR, and ETR. These configurations aimed to minimize metrics such as RMSE, R-squared, MAE, and MAPE across the folds, ensuring comprehensive model evaluation.

The MOPTA configuration parameters were tailored to enhance multi-objective optimization, integrating techniques to manage search spaces effectively. Results demonstrated MOPTA's superiority in minimizing RMSE for heating and cooling predictions compared to default parameter settings and benchmark algorithms like multi-objective grey wolf optimizer (MOGWO), multi-objective particle swarm optimization (MOPSO), and non-dominated sorting genetic algorithm II (NSGA-II). Additionally, computational load analysis highlighted MOPTA's efficiency despite its longer running time compared to simpler algorithms like GBR, RFR, and ETR, showcasing its viability in complex optimization tasks.

Conclusion

To sum up, the study introduced a novel approach that leveraged the MOPTA for predicting heating and cooling loads using the energy efficiency dataset from the University of California ML repository. The algorithm outperformed individual predictors such as GBR, RFR, and ETR. It competed against established multi-objective optimization methods such as MOGWO, MOPSO, and NSGA-II. Comparisons with previous research using similar algorithms demonstrated significant improvements in predictive accuracy.

Enhancing algorithm performance through hybridization techniques or incorporating strategies like Levy flights could further optimize results for future research directions. Additionally, exploring the algorithm's applicability to broader engineering contexts within energy efficiency prediction, especially in heavy industry buildings, incorporating more detailed building characteristics, was identified as a promising avenue for extending its utility and impact.