HyperAI超神经

In recent years, the rapid increase in carbon emissions caused by the accelerated urbanization process has seriously threatened the global environment. Many countries have given clear time points for "carbon peak and carbon neutrality", and a "green revolution" covering the world and all industries has begun. Among all industries, buildings are well-deserved energy consumers, among which heating, ventilation, and air-conditioning (HVAC) systems are the "hardest hit area". Relevant data show that heating, ventilation and air conditioning (HVAC) accounts for 38% of global building energy consumption.

In response to the high energy consumption of buildings, the industry often optimizes equipment operating efficiency and intelligent control to regulate energy consumption in real time. Among them, in terms of chiller operation control, cold load prediction is an important way to optimize chiller sequencing control. It can ignore temporary changes in cold load and avoid unnecessary switching of HVAC equipment and chillers, thereby reducing consumption caused by start-stop.

In addition, different regions have different climatic conditions, resulting in different HVAC consumption. For example, Hong Kong is in tropical climate conditions, and HVAC energy consumption accounts for a higher proportion. In the "Global Artificial Intelligence Challenge for Building Electrical and Mechanical Facilities" organized by the Electrical and Mechanical Services Department of the Hong Kong Special Administrative Region,Researchers from Lingnan University and City University of Hong Kong proposed a new dynamically engineered multi-modal feature learning (DEMMFL) model for accurately predicting building cooling loads in the long term, thereby achieving energy conservation goals..

Get the paper:

https://www.sciencedirect.com/science/article/pii/S0306261923015477

Reply "Cooling Load" to get the full paper PDF

Dataset: Cooling load data for two office buildings

In this study, the researchers focused on the building energy consumption of two office buildings in Hong Kong (South_Tower and North_Tower, ST&NT) and established a detailed dataset covering the time range from April 1, 2020 to September 30, 2021. The dataset collects data at a sampling interval of every 15 minutes, ensuring a detailed capture of the dynamics of building energy consumption.

Two office buildings used to collect data

Model Architecture: DEMMFL Model

As shown in the figure below, in order to predict the building cooling load, the collected data features are input into the DEMMFL model and the Deep Learning model, and the common cooling load data of ST and NT are output.

Using the input and output of the DEMMFL model

In this study, the research team developed a prediction model called DEMMFL (Dynamic Engineered Multimodal Feature Learning) specifically for predicting building cooling loads. The model structure is built using convolution of past input data without autoregressive output terms.

In order to achieve long-term prediction accuracy, the DEMMFL model uses a regularized statistical learning method to achieve the best variance and bias trade-off for prediction. Lasso, ridge, and the recently developed Lasso-ridge regression were used in the study to optimize the learning hyperparameters through cross-validation (CV). With this approach, the research team was able to effectively scale and process all features and data, including cooling load, to improve the accuracy and efficiency of the model.

In addition, this study also explored a variety of deep learning models, including XGBoost and LightBoost implemented by AutoGluon, as well as LSTM and GRU. All of these models were trained using the mean squared error as the loss function, and the training method was ADAM.

Best overall performance: Lasso-ridge regression

The researchers compared the statistical learning performance of the DEMMFL model on the training set and the test set. The results showed that Lasso-ridge outperformed the other three techniques in all modes except non-operating periods, and its overall performance on the test set was 4.2% higher than the second-ranked method.

This result shows that when using the Lasso-ridge method, the DEMMFL model performs well in the long-term prediction of building cooling loads, not only with significant improvement in accuracy but also more efficient in variable selection, which provides an effective tool for building energy consumption management.

Predicting Sensitivity: Cooling Load and OAT

When analyzing the weekday operation period patterns of NT and ST, the researchers found that the OAT (outdoor air temperature) feature plays a dominant role in cooling load prediction.

After building the models, the research team evaluated the sensitivity of cooling load to OAT in each model. While keeping the other variables constant, they increased OAT by one degree Celsius and calculated other OAT-related variables accordingly.

Detailed prediction results of the four models on the test dataset

The experimental results show that there are significant differences in the sensitivity of cooling loads to OAT in different buildings.The cooling load of the South Tower is more sensitive to the changes in OAT due to its larger size. In addition, the difference in sensitivity between the two towers increases in all modes except the operation period mode, which may be related to the different characteristics of the buildings and the operation modes.

DEMMFL model: high accuracy, low error

The researchers used deep learning models such as LSTM, GRU, and AutoGluon to optimize the same training dataset and compared them with the DEMMFL model on the same test set in September 2021. Due to the use of knowledge-driven engineered features, the DEMMFL model showed a clear advantage and achieved the best RMSE, while the AutoGluon model achieved the second best.

Comparing the actual cooling load data with the cooling load data predicted by the four models, the results are shown in the figure below.

LSTM and GRU showed significantly worse prediction results in the first three days of the month, and on September 22 (the Mid-Autumn Festival holiday), the deep learning model had a large prediction error, while the DEMMFL model predicted very accurately.

In summary, the DEMMFL model has higher prediction accuracy and smaller error in predicting building cooling load.

Starting from energy management: AI and the future of urban construction

Using the DEMMFL model, accurate energy consumption prediction and optimization can be achieved for commercial buildings, residential areas and public facilities. The expansion of this technology will help us better understand and manage the overall energy consumption of urban buildings, thereby promoting more efficient and sustainable urban development.

Regarding the global environment, carbon reduction and energy consumption reduction require the joint efforts of every household, every enterprise, and every industry. In recent years, the rapidly growing demand for energy consumption optimization has also spawned a number of mature solutions, which are continuously iterating during implementation. IBM, Google DeepMind, Schneider Electric, as well as domestic companies such as SenseTime and Midea Building Technology all use artificial intelligence to assist in energy management.

For example, Google DeepMind applied machine learning algorithms to a 700-megawatt wind farm in the central United States. Based on a neural network, it was trained using widely available weather forecasts and historical turbine data to predict wind output 36 hours before actual power generation, thereby providing staff with more accurate energy supply plans based on how much electricity to deliver each hour one day in advance.

Based on the powerful architecture of its large-scale device SenseCore and SenseNova large-scale model system, SenseTime continues to output high-quality AI algorithms and computing power to enable multi-domain intelligent upgrades of power systems.

Midea Building Technology combines AI technology with the experience of experts in the HVAC field to develop Smart Control, an algorithm engine for optimizing the operation of HVAC systems. It can accurately match heating and cooling needs. Combined with the iBUILDING cloud platform, it can achieve energy saving and carbon reduction of 15%-30% and comfort improvement of more than 30%.

In the future, AI technology will be closely integrated with urban planning and management to create a highly integrated, intelligent and interconnected ecosystem, in which AI will play the role of not only a technology user, but also a leader in urban development.

As AI technology continues to advance, we look forward to a more efficient, sustainable, and inclusive urban future.

References:

https://www.sciencedirect.com/science/article/pii/S0306261923015477?via%3Dihub
https://www.marketsandmarkets.com/industry-news/AI-Powered-Energy-Sector-in-2023-Products-Companies-and-Innovations