
==== Front
Heliyon
Heliyon
Heliyon
2405-8440
Elsevier

S2405-8440(24)13010-1
10.1016/j.heliyon.2024.e36979
e36979
Research Article
Based on the improved fuzzy analytic hierarchy and the TSE-MLR model energy consumption prediction of university: A case study
Chen Xiao 521020@cdnu.edu.cn
a⁎
Peng Xiaobo b
Li Yanzi a
He Baiju a
a College of Physics and Engineering Technology, Chengdu Normal University, China
b Department of Logistics and Infrastructure Construction, Chengdu Normal University, China
⁎ Corresponding author. 521020@cdnu.edu.cn
28 8 2024
15 9 2024
28 8 2024
10 17 e369797 5 2024
6 8 2024
26 8 2024
© 2024 The Authors
2024
https://creativecommons.org/licenses/by-nc/4.0/ This is an open access article under the CC BY-NC license (http://creativecommons.org/licenses/by-nc/4.0/).
The accurate prediction of building energy consumption on university campuses is a significant research area. Current studies often focus on predicting the energy consumption of specific building areas or individual equipment, and typically consider only one factor, limiting the accuracy and applicability of the predictions. This study introduces the Time Segmented Energy-Multiple Linear Regression (TSE-MLR) prediction model, which integrates the improved fuzzy analytic hierarchy and the multiple linear regression algorithm. The model is compared with traditional (MLR, BP) and advanced (RNN) models, and their various indexes are discussed and analyzed. By collecting meteorological and energy consumption data from the study site over the past 12 years, the key factors affecting energy consumption on the university campus were identified using the improved fuzzy analytic hierarchy. Subsequently, the TSE-MLR model was trained using energy consumption data from 2010 to 2016 and validated using data from 2017 to 2019. The prediction results of the TSE-MLR model were compared with those obtained through Multiple linear regression, BP neural networks, and RNN. The results demonstrated that the TSE-MLR model significantly reduced the prediction error by 13.8 % and exhibited higher accuracy compared to the other models. Therefore, the TSE-MLR model introduced in this study offers a new and effective approach to predicting university energy consumption and supporting energy management using existing data from university building operations across different periods.

Keywords

University energy consumption prediction
TSE-MLR model
Improved fuzzy analytic hierarchy
Linear regression
BP neural network
RNN model
==== Body
pmc1 Introduction

Since the 1970s, the global economy has positively correlated with carbon emissions. Despite the increased attention of various countries in recent years, the growth rate has remained on the rise albeit at a slower pace. Within the global carbon emission reduction industries, the construction sector accounts for 36 % of the world's consumption and 39 % of the world's carbon emissions. A research report on China's building energy consumption reveals that the total building energy consumption and the total building area in China demonstrated a steady upward trend between 2014 and 2017. Furthermore, the annual power consumption of large public buildings in China is 1.5–2 times that of similar buildings in developed countries, such as Europe and the United States. As a significant energy consumer for public buildings, effective energy management and consumption reduction have become a key area of interest for researchers and businesses alike.

1.1 Factors affecting building energy consumption

Xue Cui [1] made use of a residential energy consumption survey data set and three tree-based machine learning algorithms to predict building energy consumption. The results show that the prediction model based on LightGBM has the best prediction effect on apartment buildings, while the prediction model based on CatBoost has the best prediction effect on single-family houses. Dadi Zhang [2] pointed out that building systems and occupants' behaviors contribute to the energy needed to create a comfortable environment. Research shows that teachers' proactive regulation of light may lead to reduced energy consumption and reduce students' complaints about indoor environmental quality. Lingling Zhang [3] adopted the significance analysis method to analyze the influence degree of 14 influencing factors on building energy consumption and refrigeration energy consumption. Mohamed M. Ouf [4] measured the historical energy consumption of 30 school buildings in Manitoba, Canada. The new school consumed less natural gas, but more electricity than the old school, indicating that the age of the school affected the energy consumption of the building. Karen A. Kellogg [5] found that factors such as energy prices, heating and cooling days, and population density explained A lot of changes in per capita electricity and natural gas consumption (71.2 % and 73.1 %, respectively), while strict building regulations explained 6.5 % and 1.6 % of additional changes, respectively. Lei Lei [6] adopted the EWKM coupled random forest method to select building energy consumption impact factors, which reduced the data dimension and improved the overall computational efficiency of the prediction model. Jingzheng Ren [7] used the fuzzy analytic hierarchy Process (AHP) as a decision-making tool to determine the most important influencing factors of national energy security, then determined the priority of strategic measures for each influencing factor, and summarized the performance of each strategic measure into a general index. Nidia Bucarelli and Nora El-Gohary [8] placed sensors in different locations to collect data on indoor physical parameters of office buildings, tested three clustering methods and defined four feature combinations. The results show that the sensor configuration can affect the prediction performance by 35–76 %. After considering the correlation between solar power generation, working hours, output, and past electricity consumption data, Hamdi Giray Resat [9] et al. proposed a hybrid prediction model for short-term and medium-term energy prediction process of an energy management system designed and developed by ARIMA and artificial neural network. The proposed hybrid model improved by about 39.9 % compared to the predicted data obtained using ARIMA alone.

1.2 Building energy consumption calculation

The forms and types of building energy supply in different countries are also very different, which will lead to differences in the calculation methods of building energy consumption. For example, about 80 % of the heating supply in the Russian construction field is provided by the central heating system, while about 58 % of the total heating supply in Norway is provided by household electric-driven heating equipment, and the heating mode in Germany, Britain, and France is dominated by household fossil fuel boilers. The European Union is currently updating the EPBD (Building Energy Efficiency) Directive, whose main goal is to increase refurbishment rates and reduce energy consumption and greenhouse gas emissions. For example, in Poland, the primary energy index of newly designed multifamily residential buildings cannot be higher than 65 kWh/m.

At present, the mainstream energy consumption calculation methods in China include the DOE 2 energy consumption calculation method, DeST calculation, and IBE load calculation method, but these methods are mostly used in academic and research fields and are rarely used in architectural design. Robert Cichowicz [10] used two different computational and consumption methods to analyze 15 family residential buildings located in Poland, the consumption method is about 10 % lower than the energy index calculated by using the calculation method. Ana Rodriguez [11] used the Morris sensitivity analysis method for low power load estimation, to determine which inputs the output estimates are most sensitive in the static computational model, solve the problem that the power load is difficult to monitor in detail and the monitoring equipment is expensive, and by providing uncertainty estimates during the audit process, Auditors can choose the best energy assessment strategy for low-power load estimation. Lei Liu [12] explored the intrinsic interaction mechanism between energy-related factors from a macro perspective, and the comprehensive Dynamic Energy Assessment (IDEA) model is established, it provides a more accurate interaction system and detailed life cycle energy consumption and internal interconnection for assessing national energy consumption. Sabri Ciftci [13] et al. designed a 9.9kWp, roof-mounted, and grid-connected PV system using PVsyst software. Based on the generation results, the system can generate 13.13 MWH per year, of which 6.43 MWH are used by consumers and 6.70 MWH can be sold to the network. The results show that the efficiency of the photovoltaic cells is 17.09 % and the total system loss is about 26 %.

1.3 Building energy consumption predict

Energy consumption prediction is the core of low-carbon campus construction. At present, artificial intelligence models such as neural networks and support vector machines have a high potential ability to achieve accurate nonlinear mapping between input and output in real environments. Yiran Yang [14] used a support vector machine, gated cycle unit, extreme learning machine, long short-term memory, and random frog jump algorithm as optimizers to predict the energy consumption of buildings. For cooling load prediction, LSTM-SFLA has an R of 0.9761, whereas for heat load prediction, SSR-SFLA has the best performance. R is 0.9583. The results show that using an SFLA optimizer can improve the prediction performance. Sadegh Afzal [15] et al. expanded four different models, including three artificial neural network frameworks and a regression model, and the results showed that the ELM-BBO model had the best performance in predicting cooling and heating loads. Abul Abrar Masrur Ahmed [16] proposed two models based on VMD, BiGRU, and LSTSE-MLR, which showed excellent prediction accuracy, especially in the short and medium term. However, the prediction accuracy could not be guaranteed if the prediction time was longer than one year. Shidrokh Goudarzi [17] proposed an improved hybrid model based on ARIMA, support vector regression (SVRS), and particle swarm optimization (PSO), which is used to predict accurate energy consumption from the provided data. This method is easier to achieve in different building Spaces, but the convergence speed still needs to be further improved. Ibtissam Amalou [18] et al. have conducted a comparative analysis of several deep learning models, especially recurrent neural network (RNN) architectures such as basic RNN, Long short-term memory (LSTM), and gated loop unit (GRU). These architectures were trained and tested on the energy dataset of the Smart Grid Smart City (SGSC) project (2010–2014) and evaluated using various metrics, and the results showed that the GRUs outperformed the basic RNN and LSTM. Shobhit Chaturvedi [19] et al. compared the performance of four-time series models, and Fb Prophet performed well in terms of monthly total volume and peak demand. Therefore, FB Prophet was selected to formulate future energy forecasts from 2019 to 2024. This suggests that India's annual total energy demand and peak energy demand are likely to grow at an annual rate of 3.9 % and 4.5 %, respectively.

Fanyue Qian [20] using a small amount of actual data, predicted the target cold and hot loads through simulation software and used the transfer learning algorithm TrAdaboost as auxiliary training data to predict the cold season loads of the next year. Although this method requires less data, it has a short time to predict future energy consumption, which is limited to the next year. Hangxin Li [21] proposed a dual-time-scale coordinated optimal control strategy to optimize energy scheduling among building energy systems. Considering the uncertainty of prediction, the study conducted tests on the photovoltaic and battery platforms built in zero-carbon buildings in Hong Kong. The results showed that the starting time of the optimal scheduling optimization horizon was the starting point of the off-peak period. This strategy can reduce costs by up to 14.7 % compared to existing strategies; Muhammad Faiq [22] established a new Long short-term memory (LSTSE-MLR) energy prediction method by using the predicted weather data and energy dependence relationship and proved that this prediction method is superior to support vector regression and Gaussian process regression, but LSTSE-MLR requires a large amount of historical data to accurately predict. Milan Protiac [23] modeled energy consumption using the data of Serbia from 1997 to 2014. The results showed that both MLR and ANN could simulate energy consumption well, but the overall performance of ANN was relatively good in terms of error size. Sadegh Afzal [24] used the multi-layer perceptron neural network to predict the values of cooling and heating loads. They adopted a hybrid method and combined the multi-layer perceptron with eight meta-heuristic algorithms to effectively adjust and optimize the hyperparameters of the multi-layer perceptron model. The research shows that the MLP-PSOGWO model achieved the highest total R-value under cooling load and heat load, which were 0.966 and 0.998, respectively. Hamed Khajavi and Amir Rastgoo [25] combined support vector regression (SVR) with six meta-heuristic algorithms to optimize the hyperparameters of the SVR algorithm. The results show that the hybrid SVR-Battle. The results show that the hybrid SVR-Battle Royale Optimization model has the best performance among all the hybrid models studied, with R2 values equal to 0.999386 and 0.998898, respectively, which are the highest values among all models.

For the research of building energy consumption prediction, many scholars use machine learning models to predict, which has the advantage of improved accuracy, but it cannot guarantee the accuracy of long-term prediction. Some scholars use all the factors affecting building energy consumption for prediction, which is more accurate and comprehensive, but the difficulty of data processing is also greatly increased, resulting in low efficiency. Some scholars also use advanced neural network models to predict energy consumption, which indeed greatly improves the prediction accuracy, but to train an accurate model, a large number of historical data are needed to support it.

This paper sets out to address the challenge of forecasting university building energy consumption with precision and efficiency. Employing an enhanced fuzzy analytic hierarchy process, we identify the pivotal factors exerting substantial influence on energy usage. Grounded in the empirical electrical consumption data from Chengdu Normal University spanning 2010 to 2017, we introduce the Time-Segmented Energy-Multiple Linear Regression (TSE-MLR) model—a pioneering approach to campus-wide energy consumption prediction. To validate the efficacy of our model, we conduct a comparative analysis against conventional linear regression and Back Propagation (BP) neural network models, utilizing a suite of performance metrics. The results unequivocally demonstrate the TSE-MLR model's superiority in predictive accuracy.

The core innovation of the TSE-MLR model lies in its temporal segmentation strategy, dividing the year into vacation and non-vacation periods, reflective of the seasonal fluctuations inherent to academic calendars. This novel time series matrix construction enables the model to adeptly capture long-term dependencies and cyclical patterns within the data, leading to heightened forecasting accuracy.

Through the implementation of the TSE-MLR model, we achieve a decade-long projection of energy consumption trends at Chengdu Normal University. This foresight empowers administrators to implement strategic macro-level energy management, paving the way for sustainable reductions in building energy consumption.

In essence, this research not only advances the methodology of energy consumption prediction but also contributes tangible benefits to university sustainability initiatives, showcasing the model's potential for broader application in academic and institutional settings.

2 Methods

2.1 Research method

The objective of this paper's research is to determine the factors affecting energy consumption in university buildings, conduct a thorough analysis and quantification of these factors, and develop a time-division prediction model for energy consumption based on the theory of multiple linear regression. This model will take into account the unique energy usage patterns during long university holidays, using relevant data and parameters. By applying this model, we aim to accurately predict the energy consumption of Chengdu Normal University at different times. The predicted values can serve as important references for colleges and universities to create effective energy management strategies, reduce building energy consumption, and enhance energy utilization efficiency. The technology roadmap for achieving this goal is illustrated in Fig. 1.Fig. 1 Research roadmap for university energy consumption prediction based on improved fuzzy analysis hierarchy and TSE-MLR model.

Fig. 1

2.2 Improved fuzzy analytic hierarchy

Numerous factors impact building energy consumption, including but not limited to weather patterns, seasonal variations, time of day, and electricity pricing. These interconnected factors demonstrate substantial randomness and can potentially exert a nonlinear influence on building electricity consumption. Mohammad Reza Zare Banadkouki [26] The entropy weight method (EWM) and fuzzy ideal solution priority technology (fTOPSIS) are used to assign the standard weight, and rank the given strategy, To improve the energy consumption of the ceramic tile industry in Iran, the results of the study showed, "The cost of implementation strategy" has the highest weight in the evaluation criteria. Cengiz Kahraman [27] Proposed a fuzzy multi-criterion decision-making approach, for the selection of renewable energy alternatives, the innovation point of the article is that the fuzzy AD application, selection of the best renewable energy alternatives and a comparison with the fuzzy AHP, and when applying the proposed method, identified the most appropriate alternative to renewable energy for Turkey. Xianhui Mao [28] used an improved fuzzy comprehensive evaluation model to analyze the stability of the surrounding rock, the results show that the proposed model can accurately and effectively predict the stability of the surrounding rock, consistent with the field findings. Improved fuzzy analytic hierarchy can effectively solve these problems, and the weight of each influencing factor is determined by establishing the fuzzy consistency judgment matrix, to help the decision-makers choose the optimal scheme.

The influencing factors of energy consumption in university buildings are not only scattered but also diverse. There are also a lot of uncertain policy factors, so it is difficult to include all the factors in the model. Improved fuzzy analytic hierarchy can not only reduce subjective factors but also make quantitative and qualitative analyses of uncertain factors. Therefore, based on the research scope and data of this paper, this thesis uses the method in literature to build an index system of influencing factors and uses the improved fuzzy analytic hierarchy to calculate the weights of different influencing factors. The specific calculation process is as follows:

First, the three-scale method is used to establish the complementary fuzzy judgment matrix, and the formula of the complementary fuzzy judgment matrix is shown in Eq. (1):(1) F=(fij)n×n

The evaluation method of the three-scale method is shown in Table 1:Table 1 Three-scale evaluation method.

Table 1Scale	Implication	
0	The former is worse than the latter.	
0.5	The former and the latter are equal.	
1	The former is superior to the latter.	

Second, the weights and ri = ∑j=1nfij of each row are calculated, and then the fuzzy judgment matrix is transformed into a fuzzy consistent judgment matrix R=(rij)n×n by conversion formula, whose formula is shown in Eq. (2).

The conversion formula is:(2) rij=ri−rj2n+0.5

Then the weight of each row is summed by the line summation method to obtain the sorting vector W(0)=(w1,w2,……,wn)T, whose Eq. (3) as follows:(3) w‾i=∏j=1maijm

The complementary judgment matrix R=(rij)n×n is changed into reciprocal matrix E=(eij)n×n by the conversion formula eij = rijrji.

Finally, the ranking vector W(0) is used as the initial value V0 of the eigenvalue method to further find the ranking vector W(k) with higher accuracy, that is, V0 = V0(V01,V02,…,V0n)T as the initial value, using the overlapping formula Vk+1 = Evk, to find the eigenvector Vk+1, and find the infinite norm of Vk+1 ,‖ Vk+1‖ ∞.

If |‖ Vk+1‖ ∞−‖ Vk‖ ∞ |<ε, then ‖ Vk+1‖ ∞ is the largest eigenvalue λmax, and after the normalization of Vk+1, the obtained vector W(k) = Vi+1 is the scheme ordering vector, and the iteration ends. The formula for Vk+1 is shown in Eq. (4):(4) Vk+1=([Vk+1，1∑i=1nVk+1，i,Vk+1，2∑i=1nVk+1，i,……，Vk+1，n∑i=1nVk+1，i])T

Otherwise, Vk = Vk+1‖Vk+1‖∞ = ([Vk+1，1‖Vk+1‖∞,Vk+1，2‖Vk+1‖∞,……，Vk+1，n‖Vk+1‖∞])T was used as the new initial value and stacked again.

2.3 Building energy consumption calculation

There are many calculation methods for building energy consumption. Yun-Yi Zhang [29] developed an ontology-based method to automatically calculate KPI to support building energy assessment. Each KPI can be defined by input, formula, and output, and the formula consists of parameters and operators. Se-Hyun Lee [30] proposed a lighting control method based on RILL, which realizes the reduction of lighting energy consumption in buildings when the illumination increases from standard 500 lux, the energy consumption increases by 39 % and 51 % respectively.

The most important energy consumption in campus buildings is power consumption. Based on the basic formula of total energy consumption of buildings proposed by previous studies, the equation is simple and easy to implement without complicated calculations. Based on the characteristics of the data of Chengdu Normal University, this thesis calculates the building energy consumption according to the characteristics of the research university. The formula is as described in Eq. (5):(5) Ed=E1+E2+E3+E4+E5

where.Ed——Total energy consumption of buildings;

E1——Air conditioning energy consumption = air conditioning refrigeration energy consumption of air conditioning + air conditioning heating energy consumption = air conditioning energy efficiency ratio cooling condition total energy + heating efficiency total energy under heating condition;

E2——Lighting energy consumption = total power of lighting fixtures lighting use time total number of lighting fixtures lighting equipment efficiency;

E3——Energy consumption of office equipment = total power of office equipment use time of office equipment number of office equipment;

E4——Elevator energy consumption = elevator operating power elevator total operating time elevator number of elevators;

E5——Other energy consumption.

2.4 Interpretation of prediction logic

To elevate the precision in forecasting building energy consumption, this study innovatively segments the academic calendar into holiday and non-holiday periods for both winter and summer, specifically analyzing the consumption patterns at Chengdu Normal University. This approach acknowledges the dynamic nature of energy usage influenced by occupancy and activity levels.

Literature reveals a plethora of methodologies employed by scholars for energy consumption prediction, encompassing engineering approximations, statistical analyses, AI-driven techniques, and high-performance computing. Commonly, predictions are made across seasons, temperature thresholds, or broad temporal scales. Notably, Abrar Shahriar Pramanik [31] devised a Short-Term Load Forecasting (STLF) technique that leverages analogous historical data to account for typical loads, special events like holidays, and consumer behavior, achieving MAE and RMSE reductions exceeding 23 % and 24 %, respectively. Dandan Liu [32] introduced the EWT-Autoformer model for power demand forecasting, demonstrating its superiority in handling non-stationary time series over competing models. Radek Svoboda [33] explored the predictive prowess of machine and deep learning algorithms in managing high-frequency data, seasonal fluctuations, external variables, and non-stationarity in multi-step energy consumption forecasts, observing notable accuracy improvements.

Our research adopts a rigorous modeling strategy to anticipate energy demands within educational institutions. Initially, raw data undergo preprocessing to ensure quality and relevance. The timeline is segmented to align with academic rhythms, facilitating a nuanced understanding of consumption dynamics. Employing an enhanced fuzzy analytic hierarchy process, we construct a judgment matrix tailored to diverse criteria, pinpointing the most influential determinants of energy consumption. We evaluate the reliability of various predictive models, contrast their performances, and extrapolate future energy requirements for Chengdu Normal University, aiming to inform sustainable planning and resource allocation.

2.5 Time segmented energy-multiple linear regression model

Given the multifaceted influences on the energy consumption of college buildings, traditional single-variable linear regression models prove insufficient in delivering precise predictions. Multiple linear regression, conversely, stands as a robust and accurate tool, enabling the formulation of relationships between a dependent variable and a multitude of independent predictors. Despite its efficacy, Haekyung's [34] work on multiple regression analysis, which included a relative importance assessment, was confined to a segment of a university campus, neglecting a comprehensive campus-wide perspective. Yang H [35] addressed this gap by introducing the Joinpoint-Multiple Linear Regression (JP-MLR) model. Comparative evaluations against Jointpoint Regression (JPR), standard Multiple Linear Regression (MLR), Backpropagation (BP) neural networks, and Random Forest (RF) algorithms showcased the JP-MLR's superiority over machine learning-oriented BP models and the traditional JPR and MLR approaches. Its performance was found to be on par with the versatile RF model.

Building upon these insights, this study endeavors to bridge the limitations of prior studies. It undertakes a holistic view of university campuses, segmenting the academic year into vacation and non-vacation periods for an in-depth analysis of building energy consumption. Consequently, a time-segmented multiple linear regression prediction model is developed, tailored specifically to the unique energy consumption patterns of university settings. The mathematical representation of this predictive model is delineated in Eq. (6).

Let the random variable y change with n independent variables. x1, x2,… …, xn, and have the following linear relationship:(6) y=β0+β1x1+……+βnxn+ε

Where y is the dependent variable x1,x2,… …, xn is the independent variable, β0,β1,β2,… …, βn is n unknown parameters; ε is an unobservable random variable with zero mean and variance σ2 >0, called the error term, and generally assumes ε ～N(0, σ2).

For n(n ≥ p) independent observation, n group data (samples) are obtained, and the formula is shown in Eq. (7):(7) yi=β0+β1x1i+……+βnxni+εi

The matrix can be expressed as follows:Y=[y1y2…yn],B=[1x111x12……xk1……xk2………1x1n………………xkn],β=[β1β2…βk],u=[u1u2…uk]

The model can be shown in Eq. (8):(8) y=Xβ+ε

In the analysis of the TSE-MLR model, the data from 2010 to 2019 were selected as the research object, and the vacation and class period was divided into two stages to produce two linear regression models, whose formulas are shown in Eqs. (9)-(11).{Ed=β0+β1X1+β2X2+……+βnXn，T≤T0(9)Ed=β0′+β1′X1+β2′X2+……+βn′Xn，T＞T0(10)

(11) Namely,Ed=f(X1,X2,……,Xn)

where.T0——Cut-off point between holidays and classes;

Ed—— building energy consumption;

β0,β0′,β1,β1′,… …, βn,βn′—— unknown parameters;

X1,X2,… …, Xn—— influencing factors.

2.6 Model performance assessment

For assessing the predictive accuracy of various algorithms in building energy consumption forecasts, common metrics include the coefficient of determination (R2), root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE). These indicators provide a comprehensive view of model performance. For instance, Fan Yang [36] utilized MSE, RMSE, and MAE to gauge the efficacy of the LS-TSE-MLR-MKF model, the LSTSE-MLR model, and Gaussian process regression. Building upon precedents set by prior research, employing a suite of statistical measures is essential for a balanced evaluation. A model might excel in one metric but fall short in others. To this end, we employ conventional performance indicators — R2, MSE, MAE, and MAPE — whose calculations are outlined in Eq. (12) through (15).(12) R2=(∑t=1n(yot−yo‾)(ymt−ym‾))2(∑t=1n(yot−yo‾)(ymt−ym‾))2

(13) MSE=1n∑t=1n(yot−ymt)2

(14) MAE=1n∑t=1n|(yot−ymt)|

(15) RMSE=1n∑t=1n(yot−ymt)2

where.N—— the number of data;

Ym——prediction results;

Yo——ground truth;

ym‾——mean values of the predicted results;

yo‾ ——the average value of the true results.

3 Case study

3.1 Overview of the study subject

Chengdu Normal University has been selected as the focal point for this research. Spanning over 1100 mu, this institution caters to a population exceeding 17,000 individuals, with its primary building energy demands encompassing cooling, heating, and lighting. The school boasts a per capita building area of approximately 43.14 m2, and its energy consumption from 2010 to 2017 ranged from a minimum of 1.82 × 10⁴ kg to a maximum of 3.50 × 10⁴ kg of standard coal. Three pivotal reasons underpin the choice of Chengdu Normal University as the study subject. The university provides easy access to detailed energy consumption data. Given its size, which mirrors the average dimensions of global universities, it offers a representative case study for elucidating the generic patterns of energy usage in tertiary educational institutions. Chengdu, a prominent hub in southwestern China, is distinguished by its distinctive urban design and architectural heritage. As a nexus for education, research, and cultural exchange, the city serves as a model for sustainable practices in energy conservation, emissions reduction, and efficient energy management. In addition, characterized by a subtropical monsoon humid climate, Chengdu experiences a comfortable average annual temperature of 16 °C and receives around 1000 mm of precipitation annually, predominantly during July and August. Winter and spring months are relatively dry, with rare occurrences of extreme snow or ice. This climatic profile ensures that the energy consumption patterns studied are widely applicable and indicative of conditions encountered in similar environments, thereby enhancing the study's practical utility.

By focusing on Chengdu Normal University, this research aims to contribute valuable insights into the energy consumption dynamics prevalent in educational establishments, fostering informed decision-making and sustainable energy policies.

3.2 Determination of influencing factors

3.2.1 Selection of influencing factors

In the existing buildings, the envelope structure, building form, equipment configuration, and operation mode have been determined, and the main factors affecting the energy consumption of the operation are mostly variable external factors, such as climate change and human behavior, so the owner of this issues needs to study the variable influencing factors. In this paper, the two indexes of meteorological factors are the monthly average maximum temperature, monthly average minimum temperature, and monthly average humidity. The second indexes of human factors are lighting time, air conditioning time, and fan time. The secondary indexes of policy factors are implementation guidelines, publicity efforts, and energy conservation awareness.

3.2.2 Build the evaluation matrix

Based on the research scope and data of this paper, it constructs an index system of influencing factors, as shown in Table 2 below. To ensure that the selected influencing factors are positively correlated with building energy consumption and to improve the accuracy of prediction, we adopt an improved fuzzy analytic hierarchy to screen the influencing factors.Step 1 the three-scale method is used to establish the complementary fuzzy judgment matrix, which is shown below.

A=[0.51100.51000.5]

B1=[0.50110.51000.5]

B2=[0.51100.51000.5]

B3=[0.51000.50110.5]

Table 2 System of influencing factors.

Table 2Target layer	Criterion layer	Index level	
Factors influencing building energy consumption	Human factor	Lighting duration	
Air conditioning duration	
Fan length	
Meteorological factor	Average monthly minimum temperature	
Average monthly maximum temperature	
Mean monthly humidity	
Policy factor	Implementation guidelines	
Publicity efforts	
Energy saving consciousness	

A represents the fuzzy judgment matrix of the target layer to the criterion layer, B1 represents the fuzzy judgment matrix of the length of lighting, air conditioning and fan time of the index layer by the criterion layer. Similarly, B2 and B3 also represent the fuzzy judgment matrix of the criterion layer to the index layer.Step 2 the weight of each row is calculated:

(16) ri=∑j=1nfij

The weighted sum of each line from the target layer to the decision layer is:

r1 = 2.5,r2 = 1.5,r3 = 0.5.

The weights of each line from the human factor decision level to the scheme are respectively:

r1 = 1.5,r2 = 2.5,r3 = 0.5.

The weighted sum of each line from the climate factor decision level to the scheme is:

r1 = 2.5,r2 = 1.5,r3 = 0.5.

The weighted sum of each line from the policy factor decision level to the scheme is:

r1 = 1.5,r2 = 0.5,r3 = 2.5.Step 3 transforms the fuzzy judgment matrix to the fuzzy consistency judgment matrix by using the conversion formula:

R=(rij)n×n

R0=[0.50.66670.83330.33330.50.66670.16670.33330.5]

R1=[0.50.33330.83330.66670.50.83330.16670.16670.5]

R2=[0.50.66670.83330.33330.50.66670.16670.33330.5]

R3=[0.50.66670.33330.33330.50.16670.66670.83330.5]

Similar to the above fuzzy judgment matrix, it represents the fuzzy consistency judgment matrix of the target layer to the criterion layer, and the fuzzy consistency judgment matrix of the criterion layer's human factors to the lighting time, air conditioning time, and fan time of the indicator layer. Similarly, it also represents the fuzzy consistency judgment matrix of the criterion layer to the indicator layer. The same is true of the reciprocal matrix described below, so the article will not repeat the meaning of the reciprocal matrix letters.

The fourth step is to obtain the sorting vector by using the first method or the square root method.

The ranking vector from the target layer to the decision layer is:W(0)=(0.4444,0.3333,0.2222)T

The ranking vector from the human factor decision layer to the scheme layer is:W(0)=(0.3704，0.4444，0.1852)T

The ranking vector from the climate factor decision layer to the scheme layer is:W(0)=(0.4444,0.3333,0.2222)T

The ranking vector from the policy factor decision layer to the scheme layer is:W(0)=(0.3333，0.2222，0.4444)T

In the fifth step, the complementary type judgment matrix is changed into a mutual type matrix by using the conversion formula eij = rijrji .E0=[12.00034.99880.499912.00030.20000.49991]

E1=[10.49994.99882.000314.99880.20000.20001]

E2=[12.00034.99880.499912.00030.20000.49991]

E3=[12.00030.49990.499910.20002.00034.99881]

In the sixth step, the ranking vector W(0) is taken as the initial iterative value V0 of the eigenvalue method, and the ranking vector W(k) with higher accuracy is further obtained. When the number of iterations reaches a calculation accuracy of less than 0.0001, the iteration ends.

The sorting vector obtained from the target layer to the decision layer iteration four times:W0(4)=(0.59537，0.27635，0.12827)T

The sorting vector is obtained four times from the decision layer to the scheme layer, where the human factor needs to iterate five times to achieve its calculation accuracy.W1(5)=(0.35217，0.55908，0.08875)T

W2(4)=(0.59537，0.27635，0.12827)T

W3(4)=(0.27635，0.12824，0.59538)T

Finally, the comprehensive weight of all the scheme indicators relative to the total target is calculated as follows:W=(0.20967，0.33286，0.05284，0.16453，0.07640，0.03545，0.03545，0.01645，0.07637)T

therefore, the results of the nine influencing factors are air conditioning time > lighting time > monthly average minimum temperature > monthly average maximum temperature > energy saving awareness > fan time > fan time > Implementation guidelines = monthly average humidity > publicity.

Due to the lack of energy-saving consciousness in policy factors, lack of detailed guidelines, and policy propaganda data can not be quantified, the correlation analysis of building energy consumption, and the correlation of humidity is not significant, and according to the change of climate, open fan period generally more in holiday, and most students home and fan generally in the teaching building and office building. Therefore, the first four influencing factors, namely air conditioning time, lighting time, monthly average minimum temperature and monthly average maximum temperature, to predict the energy consumption of university buildings.

3.2.3 Future weather data acquisition

By analyzing the trend of meteorological data from 2010 to 2023, the law of meteorological change in recent years, local climate characteristics, and global warming, the future weather conditions are predicted in this paper. A statistical analysis of meteorological data from 2010 to 2023 (using the highest and lowest monthly average temperatures in January and August from 2010 to 2023 as an example, as shown in Fig. 2 shows that there is little variation in temperature and humidity in the same month of each year, but much variation in different months. So we can use the weather data of previous years to get the monthly average maximum temperature and monthly average minimum temperature trend lines to predict future weather conditions.Fig. 2 Forecast of future temperature trend.

Fig. 2

3.3 Energy consumption data collection

Monthly energy consumption data spanning from 2010 to 2021 was acquired and uniformly converted into standard coal equivalents for analysis, as illustrated in Fig. 3. A discernible pattern emerges, revealing a downward trend in energy consumption during holidays, juxtaposed with an upward trajectory during non-holiday periods. This dichotomy can be attributed to the fluctuation in campus population: fewer occupants during vacations contrast sharply with the dense student body present during academic sessions, driving the observed energy consumption dynamics.Fig. 3 Building energy consumption in 2010–2020.

Fig. 3

Notably, the exceptional circumstances posed by the pandemic from 2020 to 2021 have skewed energy consumption patterns, deviating markedly from historical norms. To ensure the accuracy and relevance of our predictions for university energy consumption, we have elected to concentrate on data from 2010 to 2019, a timeframe less influenced by such extraordinary events.

The following figure shows the values of the influencing factors. The duration of the weather factors is obtained according to the data of the meteorological Bureau. The average monthly lowest temperature and highest temperature are shown in Fig. 4, Fig. 5. Human factors of lighting, air conditioning, and fan length calculation rules are the campus is divided into 10 teaching buildings, 2 office buildings, 8 dormitories, 1 canteen, and a library, and the building, office each roughly divided into 10 rooms, dormitory each floor has 30 rooms, canteen a floor has a lobby and 20 rooms, library each floor has 30 rooms, assuming every day building, office building, library, the light time from 7:00, dormitory, canteen the light time from 7:00–23:30. The lighting time of each day is the number of rooms and the number of floors on the principle of light on the time of air conditioner and fan is the same as the calculation rule of lighting time. The specific data are shown in Fig. 6.Fig. 4 Monthly average minimum temperature from 2010 to 2019.

Fig. 4

Fig. 5 Monthly average maximum temperature from 2010 to 2019.

Fig. 5

Fig. 6 Estimation of lighting time and air conditioning time.

Fig. 6

3.4 The TSE-MLR model predicts the building energy consumption

3.4.1 Experiment protocol

(1) Experimental data

Collect the original electricity consumption data of Chengdu Normal University from 2010 to 2021, and uniformly convert the electricity consumption data into standard coal. The monthly temperature was collected from the Wenjiang temperature website in Chengdu.(2) Specific experimental scheme

To prove the superiority and accuracy of the proposed multiple linear regression, we conducted a detailed experiment. This experimental scheme consists of a comparison model and two parallel prediction models, namely TSE-MLR, MLR, and BP. R2, MSE, MAE, and RMSE were used as evaluation indicators for the contrast model and parallel model predictions.

As can be seen from Fig. 3, college energy consumption from January to February and June to August was in a sharp decline, April to June month energy consumption was in a high and relatively flat state, but September to December energy consumption has been on the rise, because during the winter, air conditioning energy consumption is large, and the energy consumption is greater than the energy consumption of lighting, therefore, to improve the accuracy of building energy consumption prediction in colleges and universities, at the same time combined with the characteristics of winter and summer vacation, will study time holidays and vacation. According to the holiday characteristics of colleges and universities and the trend of energy consumption in Fig. 3, this paper, March to June and September to December are determined as non-holidays, and January to February and July to August are determined as holidays.

Due to the different energy consumption factors, different units and different magnitudes, to eliminate the differences between data, it is necessary to normalize the data. The data normalization methods mainly include min-max standardization, the z-core method, etc. In this paper, min-max standardization is used to process the data, and Eq. (17) as follows. To judge the accuracy of the model, the first 70 % of the 120 data is taken as the training set, and the last 30 % is taken as the prediction set.(17) x=(x0−xmin)/(xmax−xmin)

Where x is the normalized data and x0 is the raw data before normalization.(3) Model establishment

Using the original data from 2010 to 2017, the average monthly average maximum temperature x1, monthly average minimum temperature x2, lighting time x3, and air conditioning time x4 were set as four independent variables, building energy consumption y was set as the dependent variable, and multiple linear regression analysis was conducted. The results of the analysis output are output in table form, and the multiple linear regression model is finally obtained from the data output from the core table.

3.4.2 Experiment results and analysis

(1) Experimental results

The MLR analysis was performed according to the selected dependent variables, and to detect the significance of each independent variable, a correlation analysis was performed, along with a significance test, and a sum of squares and mean square check. Substituted into the university building energy consumption data, the multiple linear regression analysis, and then the multiple linear regression equation test, the results are reasonable. As shown in Table 3, Table 4, all models were found to pass the test of significance: p < 0.05, indicating that the model construction was meaningful. The final expression of the TSE-MLR Eqs.(18)-(20) as follows:{Ed=0.165+0.731X1−0.578X2+0.344X3−0.011X4，Non−Holiday(18)Ed=−0.156+0.154X1+0.105X2+0.256X3+0.590X4，Holiday(19)

Table 3 Results of the TSE-MLR analysis.

Table 3Section	Parameter	Non-Standardized Coefficients	F	
B	Standard Error	
Non-Holiday	b0	0.165	0.102	F (4,51) = 9.432 p = 0.000	
X1	0.731	0.208	
X2	−0.578	0.183	
X3	0.344	0.080	
X4	−0.011	0.040	
Holiday	b0	−0.156	0.066	F(4,23) = 68.329 p = 0.000	
X1	0.154	0.294	
X2	0.105	0.293	
X3	0.256	0.037	
X4	0.590	0.073	

Table 4 Analysis of variance of college building energy consumption.

Table 4Section	Quadratic Sum	Mean Square	F	p	
Non-Holiday	Regression	0.308	0.077	9.432	0.000	
Residual	0.417	0.008			
Aggregate	0.725				
Holiday	Regression	1.462	0.366			
Residual	0.123	0.005	68.329	0.000	
Aggregate	1.585				

Namely,(20) Ed=f(X1,X2,X3,X4)

(2) Experimental analysis

It can be seen from Table 3 that the monthly average maximum temperature has a significant positive impact on building energy consumption during non-holiday periods. It can be seen that the higher the temperature, the more building energy consumption. The second is the lighting time, the slope is larger than the holiday lighting time because the number of students on non-holidays is larger, and the demand for lighting is larger than the demand for holidays. The monthly mean minimum temperature and air conditioning time have a significant negative impact on building energy consumption, but the negative impact of the monthly mean minimum temperature is greater than the air conditioning time, indicating that the lower the temperature, the less building energy consumption.

During holidays, air conditioning time has a significant positive impact on building energy consumption, and the slope is larger than that of non-holiday air conditioning time. This is because there are fewer students in school during holidays, and the demand for lighting time decreases sharply. On the contrary, the demand for air conditioning increases, because only libraries and office buildings are still in operation during holidays.

According to the ANOVA in Table 4, F = 9.41 for non-holidays and F = 68.329 for holidays, and the P-value of significance for the two periods is approximately zero. From Table 5, the degree of fit R2 is 0.425 and 0.922, respectively, so the predictive value of the model is reliable.Table 5 Summary of university building energy consumption prediction model.

Table 5Section	R	R2	Adjust R2	RMSE	
Non-Holiday	0.652	0.425	0.380	0.086	
Holiday	0.960	0.922	0.909	0.066	

4 Comparison and discussion of the prediction model evaluation indicators

4.1 The prediction of the different models

The raw data from 2017 to 2019 was used as the training set to train the multivariate linear regression MLR model and the machine learning-based BP neural network model, respectively, to compare the predictive performance of the three models.

(1) MLR model. MLR analysis was performed according to the selected dependent variables and the results are shown in Table 6 below. According to Table 6, the MLR model is:(21) Ed=0.208+0245X1−0.309X2+0.285X3−0.046X4

(2) BP neural network model.

Table 6 Results of the MLR analysis.

Table 6Parameter	Non-Standardized Coefficients	F	
B	Standard Error	
b0	0.208	0.034	F (4,79) = 27.338 p = 0.000	
X1	0.245	0.186	
X2	−0.309	0.180	
X3	0.285	0.035	
X4	−0.046	0.031	

The first 70 % of the real data from 2010 to 2019 is the training set, and the last 30 % is the test set. The BP neural network model adopts 4 variables as input layer variables, and the output layer variable is the monthly energy consumption of building energy consumption. The data are normalized. Parameters were adjusted by trial and error. After debugging, it is found that when the number of hidden layer nodes is 11, the minimum mean square error is 0.066.(3) The TSE-MLR model expression is:

{Ed=0.165+0.731X1−0.578X2+0.344X3−0.0.011X4，Non−Holiday(22)Ed=−0.156+0.154X1+0.105X2+0.256X3+0.590X4，Holiday(23)

(24) Namely,Ed=f(X1,X2,X3,X4)

(4) RNN model.

We also selected the first 70 % of data from 2010 to 2019 for training and the last 30 % for testing to make time-series predictions of the RNN model. The RNN model receives four variables as inputs to predict monthly building energy consumption. After the data is normalized, the RNN hyperparameters, including the number of hidden layer elements, are adjusted by trial and error. Finally, an optimal configuration is determined, the number of input hidden layer nodes is 50, and the mean square error is 0.011, which verifies the effectiveness of RNN in complex time series prediction tasks.

Due to the epidemic due to the impact of 2020, universities often teach students through the Internet during the epidemic period, and students cannot attend classes on campus, making the energy consumption study during this period not representative. In this paper, the real data from 2010 to 2017 are selected to predict the building energy consumption in the next 3 years. The prediction results of the four algorithms are shown in Table 7.Table 7 Predictive values of the four algorithms from 2017 to 2019.

Table 7Time/month	Y1(TSE-MLR)/kg	Y2(MLR)/kg	Y3(BP)/kg	Y4(RNN)/kg	
2017.01	168504	186573	191099	264945	
2017.02	142819	157988	156058	242001	
2017.03	274459	267836	122147	64842	
2017.04	269919	263408	179765	257789	
2017.05	239457	233684	204361	306586	
2017.06	242632	236772	224262	238728	
2017.07	111181	122750	126980	123121	
2017.08	108976	120312	129935	83641	
2017.09	269230	262745	180544	172870	
2017.10	295212	288104	213644	80856	
2017.11	307384	299982	250458	185317	
2017.12	315606	308002	230086	428797	
2018.01	168504	186573	182342	126511	
2018.02	142819	157988	166258	263465	
2018.03	274460	267837	95368	235417	
2018.04	269920	263408	196273	342410	
2018.05	239457	233684	198129	314033	
2018.06	242632	236772	228598	165381	
2018.07	111181	122749	126964	60458	
2018.08	108976	120313	132516	265976	
2018.09	269230	262745	180544	172870	
2018.10	295212	288105	236602	219686	
2018.11	307384	299983	253242	262526	
2018.12	315604	308001	237385	177276	
2019.01	168504	186573	182342	126511	
2019.02	142819	157988	137040	78184	
2019.03	274459	267836	112353	145746	
2019.04	269919	263408	206287	258795	
2019.05	239456	233684	199390	219598	
2019.06	242632	236772	224262	238728	
2019.07	111181	122750	131214	123979	
2019.08	108976	120312	131870	146582	
2019.09	269229	262745	184405	162055	
2019.10	295212	288104	225451	150289	
2019.11	307384	299982	251912	192626	
2019.12	315605	308002	231489	297456	

4.2 Error analysis and comparison

Fig. 7, Fig. 8, Fig. 9, Fig. 10 shows the fitting of the traditional multiple linear regression model, time-segment multiple linear regression model, BP neural network, RNN model, and the actual value, and it is not difficult to see from Fig. 11 that the TSE-MLR model has a higher fitting degree.Fig. 7 The fitting of TSE-MLR to the true value.

Fig. 7

Fig. 8 The fitting of MLR to the true value.

Fig. 8

Fig. 9 The fitting of BP to the true value.

Fig. 9

Fig. 10 The fitting of RNN to the true value.

Fig. 10

Fig. 11 The fitting of four models to the true value.

Fig. 11

According to Table 8, All indexes of the TSE-MLR model performed well. For example, the prediction accuracy of R2 was 13.8 %, 6.79 %, and 29.8 % higher than that of the MLR model, BP neural network model, and RNN model, respectively. According to the error maps of the four models in Fig. 12, we can see that the TSE-MLR model has the smallest error, and the MLR model has the largest error, followed by the BP model. These values strongly suggest that the fitting effect of the TSE-MLR model is slightly better than the MLR model and the BP neural network models, indicating the feasibility of the TSE-MLR model. In addition, we can also find that through reasonable calculation, the regression analysis method can also achieve the prediction accuracy of the machine learning model, and the BP neural network model needs to master complex programming knowledge, while TSE-MLR model is easy to operate, so TSE-MLR model is more practical application value. This model has profound implications, emphasizing the importance of improving the standard model to improve performance, predicting the accuracy of the prediction model, and further validating its ability to predict building energy consumption.Table 8 Comparison of the prediction results for the three different models.

Table 8Model	MSE	RMSE	R2	MAE	
TSE-MLR	0.006	0.077	0.674	0.076	
MLR	0.011	0.094	0.581	0.092	
BP	0.009	0.086	0.627	0.084	
RNN	0.011	0.106	0.473	0.087	

Fig. 12 Error of the four models.

Fig. 12

It is worth noting that the research object of the TSE-MLR model is located in the subtropical monsoon climate zone, which has the climate characteristics of early spring, hot summer, cool autumn, and warm winter. Different campuses may have different climate characteristics, but the university's characteristics of building energy consumption must cycle by year. Although the research object of the TSE-MLR model is mainly for buildings with fixed long holidays every year, other buildings with similar characteristics to universities can also be predicted using this model, which shows the universality and specificity of the model.

In this paper, the R2 of the proposed TSE-MLR model is compared with that of the prediction model constructed by other scholars. Table 9 shows that when the TSE-MLR model is compared with more advanced models, it does not have absolute advantages in terms of prediction accuracy. However, it can be seen from Table 8 that the prediction accuracy of this model is higher than that of traditional prediction models. In addition, most other scholars predict the energy consumption of cooling load and heating load respectively, while this study aims to predict the overall energy consumption of colleges and universities. At the same time, TSE-MLR model construction difficulty is low, the user does not need to have high computer power, the model construction time is short, the running time is long, the consumption of resources is less, can be relatively easy to predict, can provide a simple way for most people to predict energy consumption.Table 9 Comparison of prediction accuracy of TSE-MLR with some models mentioned in the paper.

Table 9	Model	
Index	TSE-MLR	LSTM-SFLA	SVR-SFLA	MLP-PSOGWO	SVR-Battle Royale Optimization	
R2	0.674	0.953	0.918	0.996	0.999	

It is worth noting that the research object of the TSE-MLR model is in the subtropical monsoon climate zone with early spring, hot summer, cool autumn, and warm winter. Different campuses may have different climate characteristics, but the characteristics of building energy consumption must be cycled by year. Although the research object of the TSE-MLR model is mainly for buildings with fixed long holidays every year, such as universities, other buildings with similar characteristics to universities can also be predicted using this model, which shows the universality and specificity of the model.

4.3 Discussion on future trends

4.3.1 Policies and the current situation

In recent years, China has adopted a series of policy plans to promote energy conservation, emission reduction, and energy structure transformation, including setting a target of reducing national energy consumption per unit of GDP by 13.5 % from 2020 by 2025, increasing the proportion of non-fossil energy consumption and power generation to about 20 % and 39 %, and increasing the proportion of non-fossil energy in energy consumption year by year. To achieve the long-term goal of energy saving, emission reduction, and green and low-carbon development.

With the improvement of the national governance system, it is expected that more detailed and efficient policies will be introduced to promote energy conservation and emission reduction and accelerate progress towards the Sustainable Development Goals. However, according to the data of the Economic Operation Report of the Power Industry in 2021–2023, the national social electricity consumption continues to grow: 8.31 trillion KWH in 2021; To 8.64 trillion KWH by 2022; By 2023, it will reach 9.22 trillion KWH. Although the national government has actively introduced guidance, electricity consumption in 2021, 2022, and 2023 is still growing, and this trend of continuous growth in electricity consumption is consistent with our forecast growth trend of university building energy consumption.

4.3.2 Factors influencing energy consumption

This study systematically analyzed the influence of temperature fluctuation and building energy consumption behavior on total energy consumption, especially the differential performance in two different situations: non-holiday and holiday. Through quantitative analysis, four influencing factors are revealed: the monthly average maximum temperature, minimum temperature, and the change of lighting and air conditioning use time affect the level of energy consumption.

In non-holidays, the monthly average maximum temperature showed an upward trend. When the average maximum temperature increased by 1°, the energy consumption increased by 112 % and 16 % in holidays. In non-holidays, when the monthly average minimum temperature is reduced by 1°, the energy consumption will decrease by 89 %, and the energy consumption will increase by 11 % during holidays. At any time, the lighting duration is a more important factor, each hour increase, non-holiday energy consumption increased by 53 %, and holiday energy consumption increased by 27 %; During holidays, energy consumption increased by 62 % for every hour of air conditioning time, and decreased by 1.7 % during non-holidays.

4.3.3 Energy supply source

The energy supply structure of various countries is diverse, among which the oil and gas energy system is more stable, the coal system is unique, and the renewable energy structure is more environmentally friendly. Chengdu Normal University is located in the Sichuan Basin and has abundant water resources. Dujiangyan Hydropower Station is one of the important sources of electricity supply in Chengdu, with hydropower generation accounting for 78%–80 %. (Data source: China Electric Power Statistical Yearbook). In addition, Chengdu Normal University's electricity supply also includes thermal power generation and natural gas. There is a significant difference in the conversion efficiency of various types of energy, which directly affects the energy consumption of buildings. Electric energy and natural gas have a high efficiency in terminal use, while coal and oil have a large loss in the conversion and transmission process, and their efficiency is relatively low. In terms of energy supply, central heating and cooling systems are more energy efficient than decentralized individual systems. Given Chengdu's mild climate, Chengdu Normal University uses a distributed heating and cooling system, resulting in low overall energy efficiency. Therefore, in the multi-energy utilization, careful differentiation of energy types and their conversion efficiency has a profound impact on building energy consumption, which hydropower, thermal power generation, and natural gas are all energy supplies of Chengdu Normal University. Among them, hydropower is the main source of electricity, and hydropower can significantly reduce energy consumption because of its efficient and clean characteristics.

With the increasing global energy demand, the problem of building energy consumption has become increasingly prominent. To effectively respond to this challenge, we use the TSE-MLR model to more accurately understand the future trend of building energy consumption, and thus take targeted measures to reduce energy consumption. The prediction results of the TSE-MLR model are shown in Fig. 13. In a longitudinal comparison from year to year, building energy consumption shows an upward trend, increasing by 114 % in 2029 compared with 2020. This data reveals the seriousness of the building energy consumption problem and highlights the urgency of reducing the building energy consumption. The large energy consumption of university buildings is a common problem in the world. Doing a good job on the energy consumption of university buildings and improving the energy efficiency level will effectively alleviate the shortage of power energy supply.Fig. 13 Trend of energy consumption from 2020 to 2029.

Fig. 13

5 Conclusion

Drawing from the meticulous analysis of historical energy consumption data across university campuses, it's evident that a myriad of factors, encompassing climatic conditions, lighting requisites, and air conditioning utilization, significantly sway energy expenditure. These influences manifest differently across various temporal phases, underscoring the need for nuanced temporal segmentation in predictive models.

A salient observation from the energy consumption records of Chengdu Normal University (2010–2019) is the stark contrast in energy demands between academic sessions and vacation periods. Capitalizing on these insights, we innovatively segregated the time continuum into holiday and non-holiday epochs, constructing the Time Segmented Energy-Multiple Linear Regression (TSE-MLR) model. To ascertain the model's predictive acumen, we juxtaposed the TSE-MLR model against the Multiple Linear Regression (MLR) model and the Backpropagation (BP) Neural Network model. Our findings unequivocally underscored the TSE-MLR model's supremacy, boasting a Mean Square Error (MSE) of 0.006, a Root Mean Square Error (RMSE) of 0.077, a Coefficient of Determination (R2) of 0.674, and a Mean Absolute Error (MAE) of 0.076. Compared to the MLR, BP Neural Network, and Recurrent Neural Network (RNN) models, the TSE-MLR model outshone competitors with an R2 accuracy enhancement of 13.8 %, 6.79 %, and 29.8 % respectively, affirming its superior predictive precision and reliability.

These outcomes substantiate the TSE-MLR model's promising applicability in forecasting energy consumption for collegiate buildings. By meticulously incorporating the impacts of climate, lighting, and air conditioning, and leveraging temporal segmentation, the model's accuracy is notably amplified. This capability renders invaluable support for bolstering energy management and the advancement of low-carbon campuses within academic institutions.

However, the model's efficacy is subject to limitations. Our temporal segmentation, restricted to monthly intervals, inadvertently omitted weekends and statutory holidays from the designated holiday periods, potentially skewing forecast accuracy. Moreover, the absence of prospective meteorological data due to technical constraints further constrains the model's predictive capacity.

Notwithstanding these caveats, the TSE-MLR model's approach holds significant relevance for energy consumption analysis in universities across varying climates. Despite regional disparities in climate, building typology, and energy consumption patterns, the foundational principles and framework of the TSE-MLR model remain adaptable. Tailoring and refining the model by local specifics could further augment its predictive accuracy and versatility. In conclusion, the TSE-MLR model presents a robust, adaptable solution for energy consumption prediction in university settings, capable of informing strategic energy management and fostering sustainable, eco-conscious campus environments.

Data availability statement

Data will be made available on request.

CRediT authorship contribution statement

Xiao Chen: Writing – review & editing, Supervision, Methodology, Conceptualization. Xiaobo Peng: Writing – review & editing, Methodology, Data curation. Yanzi Li: Writing – original draft, Validation, Software, Investigation, Formal analysis. Baiju He: Writing – original draft, Validation, Software, Investigation, Formal analysis.

Declaration of competing interest

The authors declare the following financial interests/personal relationships which may be considered as potential competing interests: Xiao Chen reports financial support was provided by Chengdu Normal University. If there are other authors, they declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Appendix To improve the reliability of the results, we analyzed the model training process in more detail and processed the hyperparameters, including the learning rate, batch size, and the number of iterations. Among the four models compared in this paper, only the BP neural network model and the RNN model involve the setting of hyperparameters. The detailed setting process is as follows.① Learning rate

Considering that the BP model may require a large learning rate at the early stage of training to rapidly reduce the loss, we set the initial learning rate to 0.01. Due to the sensitivity of RNNS, we chose a relatively small initial learning rate of 0.001.② Lot size

Larger batch sizes can speed up convergence, but large batch sizes can also lead to longer training times because more samples need to be processed. To ensure the fairness and consistency of the comparison between the BP neural network model and the RNN model in the training process, for both the BP neural network model and the RNN model, we chose the energy consumption data from 2010 to 2016 as the training set, and the energy consumption data from 2017 to 2019 as the prediction set, with a total of data in monthly units.③ Number of iterations

In BP neural networks and RNN models, the number of iterations determines the number of times the entire training data set is traversed and learned. We chose RMSE as the observation index and selected the number of iterations corresponding to the minimum RMSE value, as shown in Fig. A1. As can be seen from Fig. A1, for the BP neural network model, the number of iterations with the smallest error is 700–1000, and for the RNN model, the number of iterations is 1100. Therefore, we determine the number of iterations of the BP neural network model to be 700 times, and the number of iterations of the RNN model to be 1100 times.Fig. A1 RMSE values corresponding to different iterations.

Fig. A1

Acknowledgments

This research is funded by 10.13039/501100004498 Chengdu Normal University , Grant numbers: 2023SJYLKC03, XJYLKC2021, and 2021JG15 .
==== Refs
References

1 Cui Xue Lee Minhyun Koo Choongwan Hong Taehoon Energy consumption prediction and household feature analysis for different residential building types using machine learning and SHAP:Toward energy-efficient buildings Energy Build. 309 2024 113997 ISSN 0378-7788
2 Zhang Dadi Philomena M. Bluyssen,Energy consumption, self-reported teachers' actions and children's perceived indoor environmental quality of nine primary school buildings in The Netherlands Energy Build. 235 2021 110735 ISSN 0378-7788
3 Zhang Lingling Zhang Jiran Ren Panpan Ding Libin Wengang Hao An Chaofeng Xu Ao Analysis of energy consumption prediction for office buildings based on GA-BP and BP algorithm Case Stud. Therm. Eng. 50 2023 103445 10.1016/j.csite.2023.103445 ISSN 2214-157X
4 Mohamed M. Ouf Issa Mohamed H. Energy consumption analysis of school buildings in Manitoba, Canada International Journal of Sustainable Built Environment 6 Issue 2 2017 359 371 ISSN 2212-6090
5 Kellogg Karen A. Cumbre-Gibbs Nicol La The impact of state level residential building code stringency on energy consumption in the United States Energy Build. 278 2023 112607 ISSN 0378-7788
6 Lei Lei Shao Suola Liang Lixia An evolutionary deep learning model based on EWKM, random forest algorithm, SSA and BiLSTSE-MLR for building energy consumption prediction Energy,Volume,288 2024 129795 10.1016/j.energy.2023.129795 ISSN，0360-5442
7 Ren Jingzheng Sovacool Benjamin K. Enhancing China's energy security: determining influential factors and effective strategic measures Energy Convers. Manag. 88 2014 589 597 10.1016/j.enconman.2014.09.001 ISSN 0196-8904
8 Bucarelli Nidia El-Gohary Nora Sensor deployment configurations for building energy consumption prediction Energy Build. 308 2024 113888 10.1016/j.enbuild.2024.113888 ISSN0378-7788
9 Resat H.G. Design and development of hybrid forecasting model using artificial neural networks and ARIMA methods for sustainable energy management systems: a case study in tobacco industry Journal of the Faculty of Engıneerıng And Archıtecture of Gazı Unıversıty 35 3 2020 1129 1140
10 Cichowicz Robert Jerominko Tomasz Comparison of calculation and consumption methods for determining Energy Performance Certificates (EPC) in the case of multi-family residential buildings in Poland (Central-Eastern Europe) Energy 282 2023 128393 10.1016/j.energy.2023.128393 ISSN 0360-5442
11 Rodriguez Ana Smith Stefán Thor Potter Ben Sensitivity analysis for building energy audit calculation methods: handling the uncertainties in small power load estimation Energy 238 Part A 2022 121511 ISSN 0360-5442
12 Liu Lei Tam Vivian W.Y. Almeida Laura Le Khoa N. Dynamically assessing life cycle energy consumption of buildings at a national scale by 2020: An empirical study in China Energy and Buildings 296 2023 113354 10.1016/j.enbuild.2023.113354 ISSN 0378-7788
13 Çiftçi S. Solak M. Kuncan M. Powered by the sun: designing and analyzing technical and economic aspects of a school sustained by photovoltaics Journal of Mechatronics and Artificial Intelligence in Engineering 1 1 2020 21 32
14 Yang Yiran Li Gang Luo Tao Al-Bahrani Mohammed Al-Ammar Essam A. Sillanpaa Mika Ali Shafaqat Leng Xiujuan The innovative optimization techniques for forecasting the energy consumption of buildings using the shuffled frog leaping algorithm and different neural networks Energy 268 2023 126548 ISSN 0360-5442
15 Afzal Sadegh Shokri Afshar Ziapour Behrooz M. Shakibi Hamid Sobhani Behnam Building energy consumption prediction and optimization using different neural network-assisted models; comparison of different networks and optimization algorithms Eng. Appl. Artif. Intell. 127 Part B 2024 107356 10.1016/j.engappai.2023.107356 ISSN 0952-1976
16 Abul Abrar Masrur Ahmed Bailek Nadjem Abualigah Laith Bouchouicha Kada Kuriqi Alban Sharifi Alireza Sareh Pooya Khatib Abdullah Mohammad Ghazi Al Mishra Pradeep Colak Ilhami El-kenawy El-Sayed M. Global control of electrical supply: a variational mode decomposition-aided deep learning model for energy consumption prediction Energy Rep. 10 2023 2152 2165 10.1016/j.egyr.2023.08.076 ISSN 2352-4847
17 Goudarzi S Anisi M H Soleymani S A An IoT-based prediction technique for efficient energy consumption in buildings IEEE Transactions on Green Communications and Networking 5 4 2021 2076 2088
18 Amalou Ibtissam Mouhni Naoual Abdali Abdelmounaim Multivariate time series prediction by RNN architectures for energy consumption forecasting Energy Rep. 8 Supplement 9 2022 1084 1091 10.1016/j.egyr.2022.07.139 ISSN 2352-4847
19 Chaturvedi Shobhit Rajasekar Elangovan Natarajan Sukumar McCullen Nick A comparative assessment of SARIMA, LSTM RNN and Fb Prophet models to forecast total and peak monthly energy demand for India Energy Pol. 168 2022 113097 10.1016/j.enpol.2022.113097 ISSN 0301-4215
20 Qian Fanyue Gao Weijun Yang Yongwen Yu Dan Potential analysis of the transfer learning model in short and medium-term forecasting of building HVAC energy consumption Energy 193 2020 116724 10.1016/j.energy.2019.116724 ISSN 0360-5442
21 Li Hangxin Wang Shengwei Two-time-scale coordinated optimal control of building energy systems for demand response considering forecast uncertainties Energy 253 2022 124204 ISSN 0360-5442
22 Faiq Muhammad Tan Kim Geok Pao Liew Chia Hossain Ferdous Tso Chih-Ping Lim Li Li Wong Adam Yoon Khang Shah Zulhilmi Mohd Prediction of energy consumption in campus buildings using long short-term memory Alex. Eng. J. 67 2023 65 76 ISSN 1110-0168
23 Protić Milan Shamshirband Shahaboddin Petković Dalibor Abbasi Almas Kiah Miss Laiha Mat Akhtar Unar Jawed Živković Ljiljana Raos Miomir Forecasting of consumers heat load in district heating systems using the support vector machine with a discrete wavelet transform algorithm Energy 87 2015 343 351 ISSN 0360-5442
24 Afzal Sadegh Ziapour Behrooz M. Shokri Afshar Shakibi Hamid Sobhani Behnam Building energy consumption prediction using multilayer perceptron neural network-assisted models; comparison of different optimization algorithms Energy 282 2023 128446 ISSN 0360-5442
25 Khajavi Hamed Rastgoo Amir Improving the prediction of heating energy consumed at residential buildings using a combination of support vector regression and meta-heuristic algorithms Energy 272 2023 127069 ISSN 0360-5442
26 Mohammad Reza Zare Banadkouki Selection of strategies to improve energy efficiency in industry: a hybrid approach using entropy weight method and fuzzy TOPSIS Energy 279 2023 128070 ISSN 0360-5442
27 Kahraman Cengiz Kaya İhsan Cebi Selcuk A comparative analysis for multiattribute selection among renewable energy alternatives using fuzzy axiomatic design and fuzzy analytic hierarchy process Energy 34 10 2009 1603 1616 ISSN 0360-5442
28 Mao X. Hu A. Zhao R. Wang F. Wu M. Evaluation and application of surrounding rock stability based on an improved fuzzy comprehensive evaluation method Mathematics 11 14 2023 3095 10.3390/math11143095
29 Zhang Yun-Yi Hu Zhen-Zhong Lin Jia-Rui Zhang Jian-Ping Linking data model and formula to automate KPI calculation for building performance benchmarking Energy Rep. 7 2021 1326 1337 10.1016/j.egyr.2021.02.044 ISSN 2352-4847
30 Lee Se-Hyun Oh Seung-Taek Lim Jae-Hyun Lighting control method based on RIIL to reduce building energy consumption Energy Rep. 11 2024 2090 2098 10.1016/j.egyr.2024.01.075 ISSN 2352-4847
31 Shahriar Pramanik Abrar Sepasi Saeed Nguyen Tung-Lam Roose Leon An ensemble-based approach for short-term load forecasting for buildings with high proportion of renewable energy sources Energy Build. 308 2024 113996 10.1016/j.enbuild.2024.113996 ISSN 0378-7788
32 Liu Dandan Wang Hanlin Time series analysis model for forecasting unsteady electric load in buildings Energy and Built Environment 2023 10.1016/j.enbenv.2023.07.003 ISSN 2666-1233
33 Svoboda Radek Kotik Vojtech Platos Jan Data-driven multi-step energy consumption forecasting with complex seasonality patterns and exogenous variables: model accuracy assessment in change pointneighborhoods AppliedSoftComputing,Volume150 2024 111099 10.1016/j.asoc.2023.111099 ISSN1568-4946
34 Im Haekyung Srinivasan Ravi S. Maxwell Daniel Steiner Ruth L. Karmakar Sayar The impact of climate change on a university campus' energy use: use of machine learning and building characteristics Buildings 12 2 2022 108 10.3390/buildings12020108
35 Yang H. Ran M. Zhuang C. Prediction of building electricity consumption based on Joinpoint−Multiple linear regression Energies 15 22 2022 8543 10.3390/en15228543
36 Yang F. Mao Q. Auto-evaluation model for the prediction of building energy consumption that combines modified kalman filtering and long short-term memory Sustainability 15 22 2023 15749 10.3390/su152215749
