Recently Published Articles Scientific Reports Four different artificial intelligence (AI) models were proposed to predict energy consumption in educational facilities. The performance of the models, including decision tree, K-nearest neighbor (KNN), gradient boosting, and long short-term memory (LSTM) network, were evaluated and compared during the training and testing phases.
background
The building and construction sector accounts for more than one-third of global energy demand, with education facilities in particular contributing around 37% of carbon dioxide emissions associated with energy use, making these buildings in need of energy planning and forecasting solutions.
Energy usage in school buildings depends on many factors, including location, size, number of occupants, age, and the level of air conditioning. Accurate prediction of energy usage is therefore essential for the effective functioning of modern power grids. Furthermore, energy usage forecasting can aid in the development of effective demand-side management plans, intelligent control systems, and fault detection and diagnosis methods.
However, accurate energy consumption forecasting is difficult due to many unpredictable conditions and noisy data. As a result, currently used methods often produce inaccurate predictions. In this paper, we developed and validated various AI models to estimate the energy consumption of school buildings using real data.
Method
The model development process involved multiple stages starting with identifying factors influencing energy consumption based on literature review and experts. Actual energy consumption data from 352 educational facilities was filtered to create an outlier-free dataset.
The actual energy consumption data and related input parameters were then split into separate subsets for training, validation, and testing. Overall, 11 input variables and one output variable were selected from the refined dataset.
Descriptive statistics were conducted to obtain a concise summary that would make the collected data easier to understand and interpret. Additionally, various visualization tools were used to identify potentially hidden patterns, trends, and relationships in the data. Additionally, scatter plot matrices were created to explore the relationships and interactions between multiple variables.
Correlation statistics were used to investigate the relationship between the input variables and the target value (annual energy consumption of educational facilities). Furthermore, a parallel coordinate plot was developed to provide a comprehensive visual representation of the multivariate data.
In the second phase of the study, appropriate machine learning algorithms were selected and trained to estimate the annual energy usage of educational facilities. The accuracy of these algorithms was determined by parameters such as coefficient of determination (COD), mean absolute error (MAE), mean absolute percentage error (MAPE), and root mean square error (RMSE). Finally, the performance of each model was evaluated using the test dataset.
Results and discussion
A thorough data analysis revealed interesting relationships between the input parameters and the annual energy usage of educational facilities. For example, the scatter plot matrix suggested that higher values in the categories of “Number of Students”, “Number of Staff” and “Number of Classrooms” were likely to lead to higher energy usage.
However, the influence of the building's physical characteristics was less significant. Factors such as “building age” and “annual consumption” also did not have a direct effect on energy use. Furthermore, no strong direct relationship was found between “number of staff” and energy consumption.
On the other hand, “AC Capacity” showed a high correlation with energy consumption. Moreover, the Pearson correlation coefficient showed negative values for urban schools, indicating that certain urban factors may have led to more efficient energy use. The developed AI models performed differently based on their unique characteristics. Tree-based models such as decision trees and gradient boosting handled nonlinear relationships effectively, while KNN relied on local data patterns and LSTM was better at capturing temporal dynamics.
The decision tree model performed well on the training data, with a relatively low RMSE of 20,716.25 and MAE of 10,764.99, indicating good fit. Moreover, the prediction error was about 3.58% of the actual value, which is suitable for practical applications. On the other hand, the KNN model showed significantly higher errors and a perfect COD of 0.934134, possibly indicating overfitting. The RMSE of KNN further increased during the testing phase.
Gradient Boosting performed better than the other models in the training phase with a significantly lower RMSE and a near perfect COD; however, the model's generalization ability was not as good in the testing phase. Overall, the LSTM network performed well on both the training and testing data.
Conclusion and future prospects
Overall, the researchers were successful in developing an AI model trained on real data to predict energy consumption in educational facilities. Factors influencing energy use in such buildings were identified from the literature.
The results indicate that optimizing energy consumption through advanced AI tools can help educational facilities encourage students to engage in sustainability thinking in their everyday environments. Of note, this study included schools in hot climates; therefore, the results may not apply to other school types and climates. The researchers suggest conducting an expanded study to include different school structures and climates in the future.
Journal Reference
Tariq, R., Mohammed, A., Alshibani, A., Ramírez-Montoya, M.S. (2024). Complex artificial intelligence models for energy sustainability in educational facilities. Scientific Reports14(1), 15020. DOI: 10.1038/s41598-024-65727-5, https://www.nature.com/articles/s41598-024-65727-5
