Hybrid CNN-SVM Regression Model for Accurate Energy Consumption Prediction in Smart Energy Systems by IRJET Journal

Hybrid CNN-SVM Regression Model for Accurate Energy Consumption Prediction in Smart Energy Systems

Kartikay1 , Dr.

Geena

sharma2 , Er.Mahender pal3

1,2,3Electrical Engineering

1,2,3Baddi university of Emerging Science and Technology , Baddi Himachal Pradesh ***

Abstract

The rapid increase in global energy demand, driven by urbanization and industrialization, highlights the need for accurate forecasting methods to ensure efficient energy management, cost reduction, and environmental sustainability. Traditional statistical and econometric models often fail to capture the complex, nonlinear dependencies in energy consumption patterns. In this study, advanced machine learning (ML) techniques are explored for energy consumption prediction, leveraging their ability to model nonlinearities, adapt to dynamic conditions, and utilize large datasets. A hybrid CNN-SVR approach is proposed, where Convolutional Neural Networks (CNN) extract meaningful spatial-temporal featuresfrom energy data,andSupportVectorRegression (SVR) performs accurate forecasting. The model is evaluated alongside conventional methods such as Linear Regression,SVM,RandomForest,andBoostedTreesusing performance metrics including MAE, MSE, RMSE, R², and MAPE. Results show that Boosted Trees outperformed others with the lowest error rates and high predictive accuracy, while the CNN-SVR model achieved the highest R² value, indicating strong generalization. Linear Regression also provided reliable results for linear dependenciesbutstruggledwithnonlinearpatterns.These findings emphasize the potential of hybrid and ensemble ML models in enhancing predictive accuracy, offering robust solutions for sustainable energy planning, smart gridmanagement,anddemand-sideoptimization.

Keywords: energy forecasting, machine learning, CNNSVR, boosted trees, smart grids

I INTRODUCTION

The global energy demand is rising exponentially due to rapid industrialization, urbanization, and population growth. Effective energy management and forecasting are essential to optimize energy use, reduce costs, and minimize environmental impact. Traditional energy

consumption forecasting methods rely on statistical and econometric models, which may lack precision when dealing with large, complex datasets. Machine Learning (ML), a subset of Artificial Intelligence (AI), offers advanced predictive capabilities that improve energy consumption forecasting accuracy. The integration of ML in energy consumption prediction enables real-time analysis,adaptivelearning,androbustpatternrecognition, ensuringefficientenergymanagementandsustainability.

Moreover,increasingconcernsregardingenergyefficiency andsustainabilityhavemadeaccurateforecastingacrucial aspect of modern power systems. Traditional forecasting models often struggle to adapt to the non-linear nature of energyconsumptionpatternsinfluencedbyvariousfactors such as weather conditions, economic activities, and behavioral aspects of consumers. ML-based methods, by contrast, offer dynamic, data-driven solutions that can continually improve in accuracy as more data is collected. Theuseofbigdataandcloudcomputingfurtherenhances the potential of ML in energy forecasting, making it a promising field of study for researchers and energy professionals. This paper explores the role of machine learninginenergyconsumptionprediction,highlightingits applications, advantages, and the latest advancements in the field.Global economic development is greatly accelerated by Energy Efficiency (EE) and Energy Conservation (EC) codes, standards, policies, and regulations, which increases demand for energy and addresses climate change. Because of its harmful consequences on the environment, energy waste is becoming a bigger problem. Therefore, at the level of residential buildings, legislators or decision-makers ought to be particularly aware of the needs of EE and EC [1]. Morethan2.5billionpopulationsofglobalurbanareasare increasing by 2050, so there is an urgent need for development and modernized existing city structures, so-called "smartcities.”So,wehavetomeetthegoalsofthe ParisAgreementonde-carbonizationby2050.Thebuilding sector's present energy performance is poor and needs to beupgraded.Thebuildingsectoraccountsforover40%of

International Research Journal of Engineering and Technology (IRJET) e-ISSN:2395-0056

Volume: 12 Issue: 09 | Sep 2025 www.irjet.net p-ISSN:2395-0072

all energy use in the European Union. Through necessary directives like the Energy Performance of Buildings Directive (EPBD), the "Clean Energy for all Europeans package," eco-design directives, energy labelling regulations, and boosting the manufacture of energyefficient household equipment, the EU started working on EEanditsconservation[2].

II RELATED WORK

Mishra et al. (2023) introduced DECODE, a data-driven energy consumption prediction model leveraging historicaldataandenvironmentalfactorsinbuildings.The study employed Long Short-Term Memory (LSTM) networks to forecast energy usage, achieving high accuracy in both residential and commercial settings. The model's ability to handle temporal dependencies and integrate various data sources marked a significant advancement in predictive analytics for energy management.

Arsene (2023) developed deep convolutional neural networks for short-term multi-energy demand prediction in integrated energy systems. The study addressed the need for accurate forecasting in systems combining electrical, heat, and gas networks. The proposed models demonstrated superior performance in predicting energy demands across multiple energy vectors, contributing to moreefficientoperationofintegratedenergysystems.

Silva and Meneses (2023) conducted a comparative study between Long Short-Term Memory (LSTM) and Bidirectional LSTM (BLSTM) deep neural networks for power consumption prediction. The research evaluated the models across various datasets, finding that BLSTM outperformed LSTM in terms of prediction accuracy. This study highlighted the importance of model selection in achievingoptimalforecasting

Li et al. (2024) proposed a regression prediction algorithm for energy consumption in cloud computing environments. They optimized a Convolutional Neural Network-Bidirectional Gated Recurrent Unit (CNN-BiGRU) model using the horned lizard optimization algorithm. Through Spearman correlation analysis, they identified that power consumption has the highest positive correlationwithenergyefficiency,whileCPUusagehasthe highest negative correlation. Their optimized model outperformed traditional random forest models, demonstrating lower mean square error (MSE) and mean absolute error (MAE), indicating improved prediction accuracyandreliability.(arxiv.org)

Deifalla (2024) examined the application of machine learning algorithms in smart buildings to enhance energy efficiency. The study focused on analyzing data related to energy usage, occupancy patterns, and environmental conditions to develop predictive models. The research concluded that tailored ML models could significantly improve energy management strategies in educational buildings, leading to optimized operations and reduced energyconsumption.(researchgate.net)

Zhang et al. (2024) developeda hybridmachinelearning model combining Convolutional Neural Networks (CNN) and Gated Recurrent Units (GRU) for short-term energy consumption forecasting. Tested on data from industrial facilities, the model demonstrated superior accuracy compared to traditional methods. The study emphasized the effectiveness of hybrid models in capturing both spatialandtemporalfeaturesofenergyconsumptiondata.

Wang et al. (2024) research investigated the use of reinforcementlearningforoptimizingenergyconsumption in smart grids. The study developed a reinforcement learning-based strategy that dynamically adjusts energy distribution to minimize consumption while maintaining system stability. The proposed approach showed promise inenhancingtheefficiencyofsmartgridoperations.

Chen et al. (2024) exploredtheuseofbigdataandneural networks for city electric power consumption forecasting under the smart grid background. The research established a neural network model considering various nonlinear factors affecting city power consumption. The model achieved high prediction accuracy, providing valuable insights for power supply regulation and smart cityservices.

Silva and Meneses (2024) conducted a comparative study between Long Short-Term Memory (LSTM) and Bidirectional LSTM (BLSTM) deep neural networks for power consumption prediction. Evaluating the models across various datasets, they found that BLSTM outperformed LSTM in terms of prediction accuracy, highlighting the importance of model selection in achievingoptimalforecastingperformance.

IIIPROPOSEDAPPROACH

International Research Journal of Engineering and Technology (IRJET) e-ISSN:2395-0056

Volume: 12 Issue: 09 | Sep 2025 www.irjet.net p-ISSN:2395-0072

Fig1: Proposed Approach Flow chart

Step 1: Data Preprocessing

Givenanenergyconsumptiondataset ,wedefineitas: *( ) + where:

 represents the input features such as historical energy consumption, temperature, humidity, and occupancy.

 isthetargetenergyconsumptionvalueattime Normalization: To ensure consistent feature scaling, we apply Min-Max normalization:

where and are the minimum and maximum valuesoffeature

Step 2: Feature Extraction Using CNN CNNisappliedtolearnmeaningfulpatternsfromtheinput time-series energy data. The feature extraction process involves convolutional layers, activation functions, and pooling layers

2.1 Convolutional Layer

Each input time-series data is reshaped into a 2D matrix representation and passed through convolutional filters:

where:

( )( )

 aretheconvolutionalfilterweights,

 isthebiasterm,

 isthefiltersize,

 istheresultingfeaturemapatposition( )

2.2 Activation Function

Weapply ReLU (Rectified Linear Unit) tointroducenonlinearity: ( ) ( )

2.3 Pooling Layer

A max pooling operation reduces feature dimensionality whilepreservingimportantpatterns: ( ( )( ))

After multiple convolutional and pooling layers, the extracted features are flattened into a feature vector forregressionprediction.

Step 3: Regression Prediction Using SVM

The extracted features are used as input for Support Vector Regression (SVR), a variant of SVM that finds an optimalfunction ( )tominimizepredictionerror.

3.1 SVR Objective Function

Given training samples ( ), SVR attempts to find a function: ( ) thatminimizesthefollowingoptimizationproblem:

subjectto: ( )

where:

 istheweightvector,

 isthepenaltyparameter,

 istheslackvariableallowingsomeerror,

 isthemarginoftolerance.

3.2 Kernel Trick for Nonlinear Regression

Sinceenergydataishighlynonlinear,weapplythe Radial Basis Function (RBF) kernel: ( ) ( )

where controls the influence of each training sample. This transforms the feature space, allowing SVR to performnonlinearregression.

International Research Journal of Engineering and Technology (IRJET) e-ISSN:2395-0056

Volume: 12 Issue: 09 | Sep 2025 www.irjet.net p-ISSN:2395-0072

Step 4: Model Evaluation

Toassessthemodel'sperformance,weuse:

1. Mean Absolute Error (MAE)

2. Root Mean Square Error (RMSE)

3. R-Squared Score ( )

where ̂ is the predicted value and is the mean actual energyconsumption.

IV RESULTS AND ANALYSIS

Table 1 Model Performance Comparison Table Model

The results presented in Table 1 and fig2 provide a comprehensive comparison of different machine learning models for energy consumption prediction using various error metrics. These include Mean Absolute Error (MAE), Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Coefficient of Determination (R²), and Mean AbsolutePercentageError(MAPE).Eachmetrichighlights a different aspect of predictive accuracy and reliability. From the table, the Boosted Trees model demonstrates the best overall performance, achieving the lowest MAE (0.0324), MSE (0.00337), RMSE (0.0581), and MAPE (2.95%),whilemaintainingaveryhighR²valueof0.9982. This indicates that the model can capture both linear and nonlinear relationships effectively, making it highly suitable for complex energy consumption patterns. In comparison, Linear Regression alsoperformsremarkably

well with the lowest MAE of 0.0368 and a high R² of 0.9978, showing that linear dependencies between features and energy demand are significant. However, its slightly higher RMSE and MAPE compared to Boosted Trees suggest limitations in handling nonlinearities. CNNSVR (Convolutional Neural Network with Support Vector Regression) also achieves strong performance, with an R² of 0.9987, the highest among all models, but its relatively higher MAE and MAPE (5.55%) indicate reduced consistency in percentage-based accuracy. On the other hand, Random Forest and SVM models demonstrate weakerperformancecomparedtothe others.Both exhibit similar RMSE values (~0.116), with higher MAE and MAPE,suggestingdifficultiesincapturingsubtlevariations in energy usage. While they still achieve high R² values (>0.99),theirpredictive reliabilityiscomparativelylower. Overall, the results emphasize that advanced ensemble models such as Boosted Trees, along with hybrid approacheslikeCNN-SVR,significantlyenhanceprediction accuracy. These findings highlight the importance of leveraging deep learning and ensemble strategies to improve the precision of energy consumption forecasting, which is vital for efficient energy management, demandsideplanning,andsustainableresourceallocation.

Fig2ModelsComparisonProposedandExisting

Conclusion

This study demonstrates the effectiveness of advanced machine learning approaches for energy consumption prediction, addressing the limitations of traditional statistical models. The comparative evaluation highlights that the Boosted Trees model achieved the most robust performance with the lowest MAE (0.0324), MSE (0.00337), RMSE (0.0581), and MAPE (2.95%), along with averyhighR²valueof0.9982,confirmingitscapabilityto capturebothlinearandnonlineardependencies.Similarly, the CNN-SVR hybrid model achievedthehighestR²score of 0.9987, demonstrating superior generalization, though

International Research Journal of Engineering and Technology (IRJET) e-ISSN:2395-0056

Volume: 12 Issue: 09 | Sep 2025 www.irjet.net p-ISSN:2395-0072

itsMAPE(5.55%)wasslightlyhighercomparedtoBoosted Trees. Linear Regression also performed competitively with an MAE of 0.0368 and R² of 0.9978, though less effective with nonlinear data. On the other hand, Random Forest and SVM lagged behind with higher error rates. Overall, the findings confirm that hybrid and ensemblebased ML models significantly improve forecasting precision, supporting efficient energy management and sustainabilitygoals.

References

1. S. A. Sinaki, R. Younes, W. A. Weckman and Gary R, “Anintelligentmodel topredictenergyperformances of residential buildings based on deep neural networks,” Energies,13(3),2020,571.

2. M. Asaada, F. Ahmada, M. S. Alama and M. Sarfraza, “Smart grid and Indian experience: A review, ” Resources Policy,2019,101499.

3. L. Wena, K. Zhoua, and S. Yanga, “Load demand forecasting of residential buildings using a deep learning model,” Electric Power Systems Research, 179,2020,106073.

4. Y. Lia, S. Kubickia, A. Guerrieroa and Y. Rezguib, “Review of building energy performance certification schemes towards future improvement. Renewable andSustainableEnergyReviews,113,2019,109244.

5. I. Leobner, P. Smolek, B. Heinzl, P. Raich, A. et.al., “Simulation-based strategies for smart demand response,”. Journal of Sustainable Development of Energy, Water and Environment Systems, 6(1), 2018, 33-46.

6. T. Teeraratkul, D. Neill and S. Lall, “Shape-based approach to household electric load curve clustering and prediction,” Transactions on Smart Grid, 9(5), 2017,5196-5206.

7. Z. Wang and R. S. Srinivasan, “A review of artificial intelligence based building energy use prediction: Contrasting the capabilities of single and ensemble prediction models,”Renewable and Sustainable Energy Reviews,75,2015,796-808

8. R.Mishra,T. Sharma,and S. Gupta,“DECODE:Adatadriven energy consumption prediction model using LSTM,” IEEE Internet of Things Journal, vol. 10, no. 1, pp.890–904,2023.

9. P. Arsene, “Deep convolutional neural networks for multi-energy demand prediction in integrated systems,” Energy Reports,vol.9,pp.235–250,2023.

10. A. Silva and R. Meneses, “Comparative study of LSTM and BLSTM deep networks for power consumption prediction,” Springer Smart Energy Systems Journal, vol.14,no.5,pp.127–140,2023

11. H. Deifalla, “Machine learning algorithms for smart building energy consumption prediction,” ResearchGate,2023.

12. Y. Zhang, Q. Li, and C. Wu, “Hybrid CNN-GRU model forshort-termenergyconsumptionforecasting,” IEEE Transactions on Smart Grid, vol. 15, no. 2, pp. 367–380,2023.

13. X.Wang,L.Sun,andY.Feng,“Reinforcementlearningbased optimization for smart grid energy consumption,” Journal of Renewable and Sustainable Energy,vol.16,no.6,pp.214–229,2023.

14. M. Li, H. Zhang, and Y. Zhao, “Optimized CNN-BiGRU for energy consumption forecasting in cloud computing,” IEEE Transactions on Cloud Computing, vol.18,no.4,pp.605–619,2024.

15. H. Deifalla, “Application of ML algorithms in smart buildings for energy efficiency,” IEEE Access, vol. 12, pp.10234–10248,2024.

16. J. Zhang, T. Lee, and R. Kumar, “CNN-GRU hybrid modelforindustrialenergyconsumptionforecasting,” Springer Smart Energy Journal,vol.19,no.3,pp.450–463,2024.

17. X.Wang,H.Liu,andP.Chen,“Reinforcementlearning for smart grid energy optimization,” IEEE Transactions on Smart Grid, vol. 16, no. 1, pp. 124–138,2024.

18. L.Chen,J.Li,andM.Zhou,“Bigdataanddeeplearning for power consumption prediction in smart cities,” IEEE Transactions on Sustainable Computing, vol. 19, no.5,pp.145–160,2024.