| 摘要: |
| 在碳排放双控背景下，高校教学建筑
运行阶段的节能减排问题愈发受到关注，构建
兼顾精度与效率的短期建筑碳排放预测模型成
为关键任务。当前研究多采用长短期记忆网
络（LSTM）处理该任务，但对建模数据量、
人员占用数据特征、计算成本等方面的系统性
评估不足，并且混合模型研究有待深入开展。
为此，以广州市某高校教学建筑组团为研究对
象，基于逐时气象数据、人员占用数据与碳排
放数据， 对基础LSTM 及五种混合模
型（LSTM-CNN、CNN-LSTM、LSTM-ATT、
CNN-ATT-LSTM、LSTM-ATT-CNN） 展开系
统性评估。结果表明：数据量≤1 年时，趋势
预测场景宜优先选用CNN-LSTM，精度优先
场景宜选用基础LSTM； 数据量＞1 年时，
CNN-ATT-LSTM 宜作为优先方案。该成果可
为模型选型与部署提供决策依据，助力高校推
进绿色低碳高质量发展。 |
| 关键词: 高校教学建筑 碳排放预测 模型选
型与部署 长短期记忆网络（LSTM） 卷积神
经网络（CNN） 注意力机制（ATT） |
| DOI：10.13791/j.cnki.hsfwest.20250925002 |
| 分类号: |
| 基金项目:广东省基础与应用基础研究基金项目（2024A1515012129）；国家自然科学基金项目（52108011）；
亚热带建筑与城市科学全国重点实验室自主研究课题（2024ZB06）；中央高校基本科研业务费专项资金项目（2024ZYGXZR048） |
|
| Research on carbon emission prediction of university teaching buildings in the operationstage based on LSTM hybrid model |
|
LIU Xiao,LIANG Zhaokai
|
| Abstract: |
| In the context of the carbon peaking and carbon neutrality initiative, enhancing the
operational carbon performance of university teaching buildings has emerged as an urgent
engineering imperative. These buildings demonstrate significant temporal variability, as their
carbon emissions are influenced by academic calendars, class timetables, examination periods,
holidays, weather conditions, and fluctuating occupancy levels. Short-term forecasting can facilitate
operational scheduling and campus carbon management. Long short-term memory networks are
extensively employed for building time-series prediction; however, their application in university
teaching buildings has not been adequately investigated. Existing research seldom evaluates the
combined impacts of training data volume, occupancy feature resolution, and computational cost,
and there is limited evidence regarding hybrid architectures. This study aims to bridge these gaps by
systematically comparing LSTM and several LSTM-based hybrid models for short-term operational
carbon emission prediction in university teaching buildings.The case study pertains to a cluster of
teaching buildings at a university in Guangzhou, China, which is situated in the hot-summer and
warm-winter climate zone. The building complex is primarily composed of ordinary classrooms
and offices. These buildings predominantly rely on purchased electricity as the main operational
energy source, without any gas consumption or on-site renewable energy generation. Hourly
electricity data for the years 2023 and 2024 were acquired from the campus logistics service center
and transformed into hourly carbon emissions by utilizing the electricity emission factor. The study
established a dataset that integrates hourly carbon emissions, occupancy data, and meteorological
variables.A significant contribution resides in the development of high-resolution occupancy inputs
specifically designed for teaching buildings. Classroom occupancy was reconstructed from course
and examination schedules and adjusted according to attendance assumptions. Self-study
occupancy and staff presence were obtained from field surveys carried out during teaching weeks,
pre-examination periods, examination weeks, holidays, summer and winter vacations, and the
Spring Festival closure period. Meteorological inputs were extracted from the ERA5 reanalysis
dataset. Pearson correlation analysis retained dry-bulb temperature, relative humidity, and shortwave
radiation. Variance inflation factor testing verified that the selected variables did not display
serious multicollinearity. After missing-value handling, dataset partitioning, normalization, and
sliding-window processing, the data were employed for multi-input single-step forecasting with a
24-hour look-back window and a one-hour forecasting horizon.Six models were assessed within a
unified training strategy. The baseline model was an independent LSTM. Subsequently, five hybrid
variants were constructed to explore diverse approaches for enhancing sequence learning. A
convolutional neural network module was positioned either prior to or subsequent to the LSTM toinvestigate whether local pattern extraction is more advantageous at the input phase or the feature phase. An attention module was incorporated after the LSTM
to accentuate informative time steps. Moreover, two composite architectures, namely CNN-ATT- LSTM and LSTM-ATT-CNN, were designed to examine
potential synergies among convolution, attention, and recurrent learning. All models were trained with optimized hyperparameters under identical experimental
conditions. Performance was evaluated using the coefficient of determination, root mean square error, mean absolute error, mean absolute percentage error, and
computational time. Each model was independently executed twenty times for each data condition, and mean values were reported.The findings demonstrate
that the suitability of the model is highly contingent upon data availability and management objectives. When the training data encompass a period of one year
or less, the CNN-LSTM model exhibits the optimal trend fitting, while the baseline LSTM attains the highest numerical precision. This implies that a limited
amount of historical data suffices for the baseline LSTM to generate stable point estimates, whereas the front-end convolution is more effective in capturing
short-term fluctuations and cyclical patterns. When the data span extends beyond one year, the CNN-ATT-LSTM model presents the best overall performance.
In comparison with the baseline LSTM, it enhances the coefficient of determination by approximately 0.02, reduces the root-mean-square error (RMSE) by
around 15%, and decreases the mean absolute error (MAE) by roughly 12%. Moreover, it incurs a lower computational cost, with an average runtime
approximately 0.4 times that of the baseline LSTM under the one-year data condition and 0.6 times under the two-year data condition. Generally, a greater
volume of historical data enhances trend consistency, as evidenced by higher coefficients of determination under the two-year condition. Nevertheless, larger
datasets do not ensure lower absolute errors for every architectural model, suggesting that data expansion and model structure interact in a non-trivial way.
From the perspective of engineering deployment, a differentiated selection strategy is advisable. When historical data are scarce and the task accentuates point
prediction accuracy, the baseline LSTM remains a viable option. When the task places priority on trend tracking or rapid rolling forecasting, CNN-LSTM is
more appropriate under limited-data circumstances. For buildings with relatively comprehensive long-term data, CNN-ATT-LSTM presents the most wellbalanced
solution in terms of accuracy, robustness, and deployment cost. These findings offer guidance for model selection in campus building energy
management systems.This research is confined to a relatively homogeneous cluster of teaching buildings within a single climate region. The generalizability of
the conclusions to other types of campus buildings necessitates further verification. Future research could explore transfer learning under small-sample
conditions, more refined spatiotemporal features, and integration with building energy management systems, and thus enhancing the control of operational
carbon in higher education facilities. |
| Key words: university teaching buildings carbon emission prediction model selection and deployment Long Short-Term Memory (LSTM) Convolutional
Neural Network (CNN) Attention Mechanism (ATT) |
