Source-linked AI summary
Systematic review of deep learning and machine learning for building energy
Ardabili Sina, Leila Abdolalizadeh, Csaba Mako, Bernat Torok, Mosavi Amir
TL;DR
Building-energy forecasting is important to energy management in smart cities, and the review examines machine- and deep-learning methods across building-energy demand, consumption, and load prediction. It uses a systematic PRISMA-based review and taxonomy to compare these methods, finding that robustness rankings vary by forecasting task and model family.
Problem
Building energy forecasting supports energy management in smart cities, and the study examines methods for predicting building-energy demand and consumption.
Source: §1 · p. 2Method
Results
For demand forecasting, hybrid and ensemble methods rank high in robustness; for consumption, DL, hybrid, and ensemble models rank highest; for load, hybrid and ensemble models rank higher while DL and SVM rank good.
Source: §8 · p. 41Takeaways & Limitations
The review reports different robustness tiers across demand, consumption, and load forecasting, rather than one shared ranking for all three tasks.
Source: §8 · p. 41The authors note computational cost associated with complex data in building optimization and energy management, including multistage and nonlinear behavior.
Source: §8 · p. 42
Abstract
from arXiv · showhide
The building energy (BE) management has an essential role in urban sustainability and smart cities. Recently, the novel data science and data-driven technologies have shown significant progress in analyzing the energy consumption and energy demand data sets for a smarter energy management. The machine learning (ML) and deep learning (DL) methods and applications, in particular, have been promising for the advancement of the accurate and high-performance energy models. The present study provides a comprehensive review of ML and DL-based techniques applied for handling BE systems, and it further evaluates the performance of these techniques. Through a systematic review and a comprehensive taxonomy, the advances of ML and DL-based techniques are carefully investigated, and the promising models are introduced. According to the results obtained for energy demand forecasting, the hybrid and ensemble methods are located in high robustness range, SVM-based methods are located in good robustness limitation, ANN-based methods are located in medium robustness limitation and linear regression models are located in low robustness limitations. On the other hand, for energy consumption forecasting, DL-based, hybrid, and ensemble-based models provided the highest robustness score. ANN, SVM, and single ML models provided good and medium robustness and LR-based models provided the lower robustness score. In addition, for energy load forecasting, LR-based models provided the lower robustness score. The hybrid and ensemble-based models provided a higher robustness score. The DL-based and SVM-based techniques provided a good robustness score and ANN-based techniques provided a medium robustness score.
1 Introduction
Building energy forecasting matters for smart-city sustainability, but existing surveys leave a gap in systematic review and model-accuracy evaluation. This study addresses that gap with a PRISMA-based review of machine- and deep-learning applications across building-energy demand, consumption, and load prediction.
1 Introduction
- Buildings account for about 40% of total energy consumption, making building-energy demand and consumption forecasting relevant to smart-city sustainability.The review also includes energy-load prediction within its scope.Source: §1 · p. 2
- Previous surveys covered related topics, but the authors identify a lack of a standard PRISMA-based review and in-depth model-accuracy evaluation for building energy.They also point to limitations in the concepts and application communities covered by earlier studies.Source: §1 · p. 3
- The study presents a comprehensive systematic review that compares machine- and deep-learning models across building-energy applications and summarizes their advantages and disadvantages.Its stated scope includes energy demand, consumption, and load prediction.Source: §1 · pp. 2, 4
2 Methodology
The review organizes its study pool by forecasting application and model family, and uses PRISMA stages to select the articles for analysis. Its final database contains 52 articles for qualitative and quantitative comparisons.
2 Methodology
- The review groups building-energy studies by demand, consumption, and load forecasting, with model-family subsections and separate treatment of deep-learning methods.Its taxonomy covers ANN, SVM, hybrid, ensemble, RNN, LSTM, and CNN approaches, alongside evaluation parameters and results.Source: §2 · p. 4
- The PRISMA search identified 350 articles in Web of Science and Scopus, then retained 70 after duplicate screening and title-and-abstract review.Full-text eligibility review selected 52 articles for the final review database.Source: §2 · pp. 4–5
- The final database contains 52 articles for qualitative and quantitative comparisons.Figure 2 presents the PRISMA flow chart used to form the database.Source: §2 · p. 5, Figure 2 · p. 5
3 Demand prediction in the BE sector
This section reviews machine-learning approaches to building-energy demand prediction and compares their performance using accuracy, reliability, and sustainability criteria.
3 Demand prediction in the BE sector
- The reviewed demand-prediction studies cover ANN, SVM, hybrid, and ensemble methods, with forecasting and optimization as common application purposes.The section introduces the reviewed studies as building-energy demand prediction work.Source: §3 · pp. 5, 8
3.1 ANN based studies
- ANN studies apply the method to building-energy forecasting, demand and power estimation, and cooling-demand optimization, while noting the need for robustness against overfitting.The section describes ANN training, testing, and validation, and reports that later work compared ANN with other methods to seek accurate models.Source: §3.1 · pp. 8–10
3.2 Support vector based studies
- Support-vector studies use supervised learning for classification and regression; reviewed comparisons often favor SVM over ANN, though one reports weaker performance than decision trees.The section describes hard- and soft-margin formulations and notes sensitivity to training-data selection in one forecasting study.Source: §3.2 · pp. 10–11
3.3 Hybrid based studies
- Hybrid methods combine complementary approaches, often pairing a prediction method with an optimization method; ANFIS combines fuzzy rules with ANN architecture.The reviewed hybrid studies include building-energy demand prediction and related prediction tasks.Source: §3.3 · pp. 12–13
3.4 Ensemble based studies
- Ensemble methods combine training algorithms to address weaknesses and improve training accuracy, with reviewed studies reporting improved reliability or performance over single methods.The section identifies boosting and bagging as common ensemble methods and describes building-energy demand applications.Source: §3.4 · p. 13
3.5 Criteria for evaluations
- The review evaluates the discussed techniques by their statistical performance in prediction, detection, optimization, and monitoring tasks.The evaluation criteria include MSE, RMSE, MAE, MAPE, accuracy, reliability, and sustainability.Source: §3.5 · pp. 13–14
3.6 Results
- The review compares building-energy demand prediction techniques using accuracy, reliability, and sustainability indices.Accuracy is derived from training-phase performance, reliability from testing-phase performance, and sustainability from reliability, accuracy, processing time, and other factors.Source: §3.6 · p. 15, Table 3 · p. 15
4 BE consumption prediction
BE consumption is measured in watt-hours over a specified period and can be calculated from building energy load and consumption hours. The section introduces ML-based forecasting studies of this measure.
4 BE consumption prediction
- BE consumption is measured in watt-hours over a specified period and can be calculated by multiplying BE load by consumption hours.Forecasting this measure is presented as useful for sustainable building-sector instruments and policy makers.Source: §4 · p. 18
4.1 ANN based studies
The review compares ML and DL approaches to building energy consumption prediction, covering ANN, support-vector, hybrid, and ensemble methods. Its robustness assessment places DL-based, hybrid, and ensemble models highest, while ANN, SVM, single ML, and LR models receive lower ratings.
4.1 ANN based studies
- ANN studies forecast building energy consumption with controller-assisted, user-preference, and simulation-based approaches, while another used monthly historical data and RMSPE analysis.The review notes that ANN accuracy may not cover datasets with different dimensions, motivating consideration of more accurate methods.Source: §4.1 · pp. 19–20
4.2 Support vector based methods
- In IoT-derived building data, SVM and ANN outperformed linear regression across R2, MAPE, and RMSE; other reviewed SVM studies also reported accurate consumption estimates.The comparisons included data subsets without lighting data, without date data, and with weather-only data.Source: §4.2 · p. 20
4.3 Hybrid based methods
- For building energy consumption, reviewed hybrid models outperformed single or ensemble alternatives in some comparisons and were reported as accurate against PSO-SVR and ARIMA.One comparison ranked hybrid above ensemble and ensemble above single ANN; another compared ARIMA-SVR and PSO with the hybrid method.Source: §4.3 · pp. 20–21
4.4 Ensemble based studies
- Among reviewed ensemble approaches, XGBoost and random forest outperformed SVM and linear regression in one comparison, while GBRTs improved average energy consumption in another.The review also describes ensemble methods as a possible solution to short-term forecasting problems.Source: §4.4 · p. 21
4.5 Results and discussions
- DL-based, hybrid, and ensemble models received the highest robustness scores for energy consumption forecasting, while ANN, SVM, single ML, and LR models received lower ratings.Figure 7 presents the review's analytical comparison across high, good, medium, and low robustness categories.Source: Figure 7 · p. 23, §4.5 · p. 24
5 BE load prediction
Building energy load is a power measure indexed in watts. The section introduces predictive methods for forecasting building-sector energy load.
5 BE load prediction
- BE load is indexed in watts and refers to the electricity required to operate an electrical device at a given moment.The section presents predictive forecasting methods and identifies studies using ML methods for building-sector energy load.Source: §5 · p. 24
5.1 ANN based studies
The review covers ANN, hybrid, and ensemble methods for building energy load prediction. Its robustness assessment rates hybrid and ensemble methods higher than LR, with DL and SVM rated good and ANN medium.
5.1 ANN based studies
- Reviewed ANN approaches forecast electricity, cooling, and medium-term cooling loads, with GPR followed by ANN providing higher estimation accuracy than LR in one comparison.The reviewed cooling-load ANN used data from five previous days to estimate next-day consumption.Source: §5.1 · p. 25
5.2 Hybrid based studies
- For building heating and cooling loads, reviewed hybrid models improved on single-method accuracy, and ANN-GA had the highest performance among techniques in one study.A GA-M5rule hybrid was also reported to increase cooling-load forecasting accuracy against ANN and M5rule.Source: §5.2 · pp. 25–26
5.3 Ensemble based studies
- Ensemble methods performed best on complex simulated building datasets, whereas single SVM performed best on simple datasets; other comparisons favored RF or ensemble bagging ANN.In the complex dataset comparison, GBRT was followed by XGBoost among the highest-performing methods.Source: §5.3 · p. 26
5.4 Results and discussion
- Hybrid and ensemble methods received higher robustness scores than LR for energy load forecasting; DL and SVM received good scores, while ANN received medium scores.Figure 8 presents the review's method-by-method robustness indicator, divided into high, good, medium, and low categories.Source: Figure 8 · p. 29, §5.4 · p. 29
6 Deep learning methods in building information sector
The review surveys deep-learning applications in building-energy information and compares their reported performance across recurrent, LSTM, and CNN studies. It concludes that DL methods generally score higher in robustness, while results vary by task and model.
6.1 Recurrent neural networks (RNN) based studies
- RNN studies reported better performance than LR, ANN, and SVM in several comparisons, while CNN outperformed RNN in one energy-load study.The comparisons used measures including MAE, MAPE, and RMSE; the CNN comparison also considered processing time.Source: §6.1 · p. 32
6.2 Long short term memory (LSTM) based studies
- Two-layer LSTM performed best among the tested one-, two-, and three-layer variants, and other LSTM studies reported gains in load, consumption, and gas forecasting.Reported improvements included about 40% over LR and 23.7% over ANN for building-energy load, up to 21% for air-conditioner consumption accuracy, and about 10 to 15% for gas prediction.Source: §6.2 · pp. 33–34
- LSTM provided the highest robustness among LSTM, ELMAN, and NARX methods, while XGBoost followed by LSTM outperformed ARIMA and ANN for electricity-demand prediction.A hybrid LSTM-GA method also improved energy-consumption prediction accuracy and sustainability over common DL methods.Source: §6.2 · p. 34
6.3 Convolutional neural network (CNN) based studies
- CNN studies reported successful heating-demand estimation and energy-load forecasting; one study reported accuracy 62% better than before.Source: §6.3 · p. 35
6.4 Results and discussion
- The review’s Table 9 evaluates DL methods for accuracy, reliability, and sustainability across the included studies.Its listed method ratings include ANN, RNN, SVM, CNN, LSTM, and hybrid approaches.Source: §6.4 · pp. 35–36, Table 9 · p. 35
- DL-based techniques generally had higher robustness scores; hybrid and ensemble models scored well, ANN and SVM were medium, and LR was lowest.Figure 12 summarizes the review’s robustness evaluation.Source: Figure 12 · p. 37, §6.4 · p. 37
7 Discussions
Forecasting is the dominant building-energy application discussed, with LSTM, ANN, and CNN most represented among the models. The review also relates model choice to data scale and prediction horizon.
7 Discussions
- Forecasting is the main trend in building-sector ML applications, alongside optimization and clustering.Source: §7 · p. 38, Figure 13 · p. 38
- LSTM has the highest allocation in building-energy studies, followed by ANN and CNN; LSTM and ANN address load, consumption, and demand, while CNN addresses load and demand.Source: §7 · p. 39, Figure 14 · p. 39
- DL and ensemble techniques may suit long-term records with large data volumes, while single or hybrid ML methods may work effectively with smaller datasets or short-term predictions.Source: Table 10 · p. 39, §7 · p. 40
- Building-energy prediction and management are framed as balancing occupant comfort with economic and environmental concerns.The discussion also connects smart energy systems with coordinated building subsystems and IoT-based information exchange.Source: §7 · pp. 40–41
8 Conclusion
The conclusion compares robustness by forecasting task and model family, finding different rankings for demand, consumption, and load. It also identifies computational cost and building-process uncertainty as barriers to wider soft-computing use.
8 Conclusion
- For energy-demand forecasting, hybrid and ensemble methods rank high in robustness, SVM good, ANN medium, and linear regression low.Source: §8 · p. 41
- For energy-consumption forecasting, DL, hybrid, and ensemble models have the highest robustness, while ANN, SVM, and single ML models range from good to medium and LR is lower.Source: §8 · p. 41
- For energy-load forecasting, hybrid and ensemble models score higher in robustness; DL and SVM score good, ANN medium, and LR lowest.Source: §8 · p. 41
- High computational cost and uncertainty in building processes and design parameters remain barriers to wider soft-computing use.The paper calls for efficient DL and ensemble methods to be evaluated in realistic optimization studies.Source: §8 · p. 42
9 Nomenclature
The nomenclature lists abbreviations for neural-network, machine-learning, optimization, and error-measure terms used in the paper.
9 Nomenclature
- Neural-network and learning abbreviations include ANN, FFNN, WNN, MLP, BPNN, LSTM, RNN, CNN, ML, DL, and ANFIS.Source: §9 · pp. 42–43
- Modeling and optimization abbreviations include GA, PCA, LDA, DA, RSM, MCDM, LS, MLR, SB, GBR, GPR, NARX, GBRT, ABC, CART, ANP, RF, ERT, DT, and PLS.Source: §9 · pp. 42–43
- The listed performance and forecasting terms include MAPE, MAE, RMSE, R2, and ARIMA.Source: §9 · pp. 42–43