According to statistics, modern urban people’s life and work are closely related to buildings, and more than 80% of their time is spent in urban buildings. Building intelligence is undoubtedly a key research topic with far-reaching influence.
In recent years, with the rise of edge computing technology, the application expansion of edge intelligence-related scenarios has become a path for technology companies to compete to show technological innovation and commercial value. Various edge AI solutions have emerged, such as HUAWEI Cloud intelligent edge platform IEF and HiLens, a one-stop multi-mode AI development platform for end-cloud collaboration. According to statistics, modern urban people’s life and work are closely related to buildings, and more than 80% of their time is spent in urban buildings. Building intelligence is undoubtedly a key research topic with far-reaching influence. This article will revolve around one of the most intelligent building is one of the important subjects of central air-conditioning energy efficiency forecast to expand with the management, at present, the subject is facing the biggest bottleneck is: most of the existing prediction and management of energy efficiency method is limited to single task, the clouds unable to support the central air-conditioning energy efficiency model on the edge of implied a lot of complex scenes.
It is well known that hVAC systems (including heating, ventilation and air conditioning) dominate the electricity consumption of commercial buildings. Existing research on HVAC systems has shown that it is important to accurately quantify the energy efficiency ratio of chiller units (the higher the value, the more energy saving), and the recently proposed data-driven energy efficiency ratio prediction can be applied to the cloud. However, due to different types of air conditioners or sensors in different parks, projects at different edges differ greatly in features, models and other aspects. In the case of small samples, it is difficult to use a general model to adapt to all projects.
In recent years, huawei cloud edge cloud innovation lab and from the Hong Kong polytechnic university, IBM research, huazhong university of science and technology, tongji university, shenzhen university and other famous team close cooperation between colleges and continue to carry out technical research, based on edge field of intelligent building scene, hope to gradually solve the reality implied a lot of complex scenes edge intelligence issues. Interested readers are welcome to take a look at the history of multitasking learning, scheduling, and multitasking applications published between 2018 and 2020:
General algorithm: multi-task migration and edge scheduling
Multi-task migration relationship discovery based on metadata
Zheng, Z., Wang Y., Dai Q., Zheng H., Wang, D. “Metadata-driven task relation discovery for multi-task learning.” In Proceedings of IJCAI (CCF-A), 2019.
In this paper, there is a practical application case of multi-task. Different edge intelligent projects use different devices to make the edge side model different, which can be applied to multi-task setting. The highlight of this paper is the introduction of metadata, which is the description information of data sets and is used for daily system operation in complex systems and contains expert information. Based on metadata extraction task attributes, this paper designed a multi-task universal AI algorithm combining metadata task attributes and sample task attributes hierarchally (Figure 1). The expert review of relevant papers also believes that this technology has shown practical value in the application practice, which is of great significance to the real implementation of machine learning projects and will become an interesting technology for today’s large organizations.
In FIG. 1, colors represent different clusters and numbers represent different device models. The method based on sample attributes can easily lead to negative migration (confusing different device models in the same cluster, left figure), while the method based on metadata can avoid negative migration (right figure).
Edge task assignment system and implementation of multi-task transfer learning
Zheng, Z., Chen, Q., Hu, C., Wang, D., & Liu, F. “On-edge Multi-task Transfer Learning: Model and Practice with Data-driven Task Allocation.” In Proceedings of IEEE TPDS (CCF-A), 2019.
Chen, Q., Zheng, Z., Hu, C., Wang, D., & Liu, F. “Data-driven task allocation for multi-task transfer learning on the edge. ” In Proceedings of IEEE ICDCS (CCF-B), 2019.
Multi-task transfer learning is a typical approach to solve the problem of insufficient samples on the edge. However, the current task assignment scheduling work on edge usually assumes that different tasks are equally important, resulting in inefficient resource allocation at the task level. In order to improve system performance and service quality, we found that the importance of different tasks to decision-making is an important indicator that needs to be measured. We show that materials-based task assignment is a variant of the NP-complete knapsack problem, and that the solution of this complex problem needs to be recalculated frequently in variable edge scenarios. Therefore, we propose an AI-driven algorithm to solve the edge calculation problem, and test the algorithm in the actual variable edge scenarios (FIG. 2). Compared with the SOTA algorithm, the algorithm can reduce processing time by more than 3 times and energy consumption by nearly 50%.
Figure 2 dynamic task allocation scheduling based on edge scenarios
Edge applications: Building intelligence
Cooling machine load control based on multi-task
Zheng, Z., Chen, Q., Fan, C., Guan, N., Vishwanath, A., Wang, D., & Liu, F. “Data Driven Chiller Sequencing for Reducing HVAC Electricity Consumption in Commercial Buildings.” In Proceedings of ACM e-Energy, 2018. Best Paper Award.
Zheng, Z., Chen, Q., Fan, C., Guan, N., Vishwanath, A., Wang, D., & Liu, F. “An Edge Based Data-Driven Chiller Sequencing Framework for HVAC Electricity Consumption Reduction in Commercial Buildings.” IEEE Transactions on Sustainable Computing, 2019.
Multitasking can be applied to building energy efficiency. Chillers are big energy consumers in buildings. It is one of the most important research problems of building intelligence to predict and manage the cooling function efficiency, predict the energy efficiency ratio and optimize the cooling load decision. It is observed in this study that in the energy efficiency prediction of chiller decision-making, different equipment models and working conditions of different edge projects will lead to different models of final demand. In this case, the adoption of a single model in the cloud is likely to lead to decreased accuracy and decision-making errors. In this work, a multi-task chiller load decision-making framework based on edge-cloud collaboration is developed (FIG. 3), which can save more than 30% energy compared with the current industrial method under the condition of utilizing the existing edge-side nodes and not deploying additional hardware.
FIG. 3 Cooling load decision-making framework of edge cloud collaboration
Air conditioning comfort prediction based on multi-task
Zheng, Z., Dai Y., Wang D., “DUET: Towards a Portable Thermal Comfort Model.” In Proceedings of ACM BuildSys (Core rank A), 2019.
Yang, L., Zheng, Z., Sun, J., Wang, D., & Li, X. A domain-assisted data driven model for thermal comfort prediction in buildings. In Proceedings of ACM e-Energy. 2018.
Air conditioning comfort prediction is one of the important research topics in the long history of building intelligence. Current comfort estimation methods often require additional sensors or human intervention such as user feedback, making scale itself a challenge. Machine learn-based comfort prediction for air conditioners has been shown to reduce additional manual intervention. However, in different edge scenarios, factors such as the type of building cooling and the type of sensors installed will cause serious errors in a single universal model on the cloud. In this study, a multi-task method is proposed to predict air conditioning comfort, and the accuracy is 39% and 31% higher than that of mechanism model and single-task model, respectively.
Edge adaptive task definition
Based on the above projects, readers can learn about edge intelligence algorithms, systems and applications based on multi-task. It is important to note that before multitasking can be used, questions about how tasks are defined and divided need to be answered, such as determining the number of machine learning models required for different projects within an application and the scope of application of each model. Currently, this approach is usually only manual intervention by data scientists and domain experts, with a low degree of automation and difficulty in scale replication. Therefore, edge automatic definition machine learning task is an unsolved but important problem.
In order to define machine learning prediction tasks adaptively in various edge scenarios, Huawei Cloud Edge Cloud Innovation Lab recently published a research paper entitled MELODY: Adaptive Task Definition of COP Prediction with Metadata for HVAC Control and Electricity Saving. In this study, a multi-task prediction framework (MELODY) containing task definitions is proposed, in which task definitions can adaptively define and learn complex energy-ratio prediction tasks.
MELODY is the first method to define energy efficiency ratio prediction task adaptively according to various edge scenarios. This work provides an attractive mechanism for researchers and application developers seeking automatic and efficient edge machine learning methods, especially for complex systems with diverse metadata but insufficient data samples. MELODY’s key idea is to use metadata to dynamically divide multiple tasks. This paper proposes the mathematical definition of metadata and two sources and methods of extracting metadata.
The team evaluated the performance of the solution in a field application: a four-month experiment was conducted on nine chillers in eight buildings in two large industrial parks. The MELODY solution is superior to the latest energy-efficiency ratio prediction method and can save 252 MWh of electricity per month for both parks, which is more than 35% less energy than the current cooling system in buildings.
MELODY Paper accepted by ACM E-Energy 2020:
Jie Pu, Zimu Zheng, Daqi Xie. MELODY: Adaptive Task Definition of COP Prediction with Metadata for HVAC Control and Electricity Saving. ACM e-Energy 2020. Australia.
ACM E-Energy belongs to ACM EIG-Energy Interest Group, the flagship conference at the intersection of computing and Energy.
1. The paper acceptance rate is 23.2%, and the acceptance rate over the years is about 20%;
2. Ubicomp with CCF-A; Ccf-b had the same ECAI and TKDD H5-index.
3. Among the 55 members of the Evaluation Program Committee, 8 ACM/IEEE fellows (about 15%) include Andrew A. Chien, Klara Nahrstedt, Prashant Shenoy, etc.
4. Members of the evaluation committee come from IBM Research Institute, University of Illinois at Urbana-Champaign, University of Cambridge, University of Washington, Purdue University, University of Massachusetts Amherst, Simon Fraser University, Nanyang Technological University, Tsinghua University, Hong Kong Polytechnic University and other internationally renowned universities and enterprises;
5. STOC, ISCA and PLDI of CCF-A; The IWQos, SIG Metric, COLT, HPDC, ICS, LCTES, SPAA of CCF-B belong to the 13 conferences of ACM Federated Computing Research Conference (FCRC) series. ACM FCRC Summit series is sponsored by Google, Microsoft, IBM, Huawei, ARM, Xilinx and other international well-known enterprises.
Energy efficiency ratio prediction
Chiller-based HVAC systems are commonly used in commercial buildings and consume 40 to 70 percent of a building’s total electricity, depending on the consumption of the HVAC system. The electricity paid by commercial buildings, much of which is attributed to hVAC systems, is typically in the top three of an organization’s operating expenses. This trend puts enormous pressure on facility managers to improve the energy efficiency of buildings by reducing the power consumption associated with hVAC systems.
The main consumption of HVAC comes from chillers (see Figure 4). The effectiveness of the typical chiller load control depends largely on the performance of the chiller in operation, that is, the energy efficiency ratio under different cooling load conditions. Energy efficiency ratio is an index to measure the efficiency of cooling function, which refers to the output cooling capacity of unit input power consumption. Energy efficiency is more than 1 normally, value is bigger, mean efficiency is higher. In practice, facility managers typically measure initial information on the energy efficiency ratio during the initial testing and commissioning of the chiller during its deployment to the building, and use this initial information to perform chiller load control. Cooling load is usually regarded as the only parameter in initial information testing. However, these initial information cannot capture the effects of actual parameters and have been proved to be inaccurate by recent studies.
FIG. 4 Schematic diagram of cooler
This study takes energy efficiency ratio prediction as a case study. The energy efficiency ratio is highly dependent on a variety of factors, such as operating conditions, cooling requirements, equipment aging, weather, etc. In order to capture these factors in chiller, existing work has proposed a data-driven approach. The energy efficiency ratio prediction problem can be regarded as learning a “formula” called model in the training stage, which can output energy efficiency ratio with given characteristics in the reasoning stage.
Adaptive task definition
Existing approaches generally assume that the configuration of prediction tasks, such as the number of prediction models and the scope of application of prediction models within the same application, is defined and fixed by data scientists or domain experts. Here is a comparison of three widely accepted Settings: single-tasking, multi-tasking, and expert-assisted multi-tasking.
Single task setting
One of the most typical and widely accepted approaches to forecasting task configuration is based on fixed, single-task setting: this means combining all data sets as a whole and training a single prediction model. Researchers can use any machine learning algorithm (such as SVM, neural network, Boosting, etc.) to learn this model and apply the trained single model to the reasoning stage in any scenario.
Single task assumption For different data sets within different projects in the same application, a single model should be sufficient to describe the relationship between selected features and energy efficiency ratios. However, this assumption may not always be true.
For example, there are two parks that use two types of chiller: Park H uses Trane CVHG1100 chiller and Park J uses Carrier W3C100 chiller, so the thermodynamic model between characteristics and energy efficiency ratio should be adjusted according to the type of chiller. Edge users often expect to see different models applied to two edge projects: even if the two chiller input the same characteristic values such as water temperature, the energy efficiency ratio of the output should be different. But if two data sets are conjoined and the same energy-efficiency model is trained, it is often difficult to ensure this without human intervention.
Previous studies by the authors of this paper also showed that, in addition to the different models of different edge projects with different cold models, examples that may lead to different models also include: Different projects use different working conditions and parameter configurations, different types of sensors used in different projects, and different characteristics used in different seasons, etc. (the reasons for this will not be described in detail, but interested readers can refer to the historical work of the research team mentioned at the beginning of the article). The application range of models trained in different edge scenarios may be radically different. Therefore, using single-task Settings for different scenarios is not always the best choice, which may lead to major errors in practice, especially in some edge intelligence projects where the size of the training sample is not enough to automatically distinguish the scenarios among a large number of features.
Multitask setting
However, the configuration of prediction tasks, such as the number of models required and the scope of application of models, is still an open question. To delve deeper into this issue, the team went further to validate multi-tasking rather than single-tasking, which is to observe the performance of multiple models on multiple test sets. In a real building, five models (hereinafter referred to as M1-M5) were trained using the training data set from chiller 1 to chiller 5. The performance of the five models was then tested in different scenarios in another five test data sets (T1-T5). The experiment and results are shown in Figure 5-1 and 5-2 respectively.
Figure 5-1 Experimental schematic diagram of complex chiller training model under different chiller test sets
Figure 5-2 Comparison results of prediction accuracy and sample collection time of complex chiller training model under different chiller test sets
Observations show that,
1) accuracy
Despite training based on different data sets, chiller 1’s model worked well on chiller 2 and Chiller 3’s test sets, but resulted in serious errors on Chiller 4 and Chiller 5’s test sets. Similar observations can be seen for chiller 2 through chiller 5 models. This is because chillers 1 to 3 are from the same type of chillers, while Chillers 4 and 5 are a different type.
2) Sample collection time
If tasks are divided by chillers, at least 81 days of samples are required for each chillers task. However, if the task is divided into two tasks according to the model, the sample of each model only needs 30 days. This is because each model task contains data collected by multiple chillers.
According to the results of the precision and sampling time, instead of considering five cold machine so as to define five cold machine, in this data set is the two models (1-3 cooler and cooler 4-5) to define two types of tasks, in the example above can reduce the sampling time of about 63%, while increasing the accuracy of nearly 10%.
Expert-assisted multi-tasking
In fact, it is not only the cold type, but also the environment (e.g., weather conditions) and operating conditions (e.g., water supply temperature) that can change the energy efficiency ratio model over time. With the knowledge of domain experts, fixed tasks can be defined in the built environment and applied to different buildings.
For example, a recent work by the team, based on domain expertise in built environment research, presented fixed 50 tasks for multi-task chillers’ energy efficiency ratio prediction based on working conditions in three buildings; Another recent work by the team gave fixed four tasks based on season and type of cooling in 160 buildings for multi-task thermal comfort prediction.
However, the number of models required and their scope of application can vary according to different edge project scenarios, and the configuration of domain experts is difficult to scale dynamically with different edge projects. For example, in a small data set of a building, it is best to have three tasks, namely training three different models for energy-efficiency ratio prediction. But in another large data set with 1000 buildings, it would be nice to have 75 tasks. Manually defining machine learning tasks to be predicted in edge scenarios often leads to excessive cost or reduced accuracy, especially when tasks change dynamically from project to project and time. Therefore, it is necessary to define tasks adaptively for different scenarios.
MELODY
The purpose of this study is to solve the problem of adaptive task definition, that is, to automatically define different tasks in different scenarios, for example, to determine the number of models to be used in different scenarios and the application scope of models. The team encountered three major challenges and came up with a multi-task Prediction framework (MELODY) using an adaptive task definition approach.
Challenge 1: The goals of the current project are unknown, and often worse, the set of possible task candidates is also unknown.
MELODY solves the first challenge by proposing task mining. It adaptively defines tasks based on novel structures and algorithms such as task forests, as shown in Figure 6. This allows MELODY to scale to predict the energy efficiency of many buildings and environments.
Figure 6 Example of task forest: data representation model training sample, attribute representation model application scope; Nodes represent subtasks, including data, attributes, and models (if any); Each root node of the forest, that is, the vertex of each tree, represents a task combined by each subtask. For detailed implementation such as task forest initialization and maintenance and proof of algorithm complexity, readers who are interested can read the appendix of the paper.
Challenge 2: The attributes of the application scope of the signage ENERGY efficiency ratio model are unknown, and the sources of such attributes are also under study.
MELODY solves the second challenge by using metadata as a source for task attributes.
Metadata is defined by domain experts for day-to-day control of the building management system. For example, the name of a sensor and the type of building are metadata. In the MELODY framework, the team came up with a way to extract two types of metadata from two sources in the database.
Metadata contains potential domain information that enables adaptive extraction of domain-knowledge tasks and opens the door to automated and powerful task definitions, as shown in Figure 7.
Figure 7 Task definition based on metadata extraction (see paper for specific implementation)
Challenge 3: The number of task combinations increases exponentially with the number of attributes; Therefore, chiller samples are insufficient for all combinatorial training models.
MELODY overcomes the third challenge by using multi-tasking transfer learning. In multitasking optimization, learning tasks can use knowledge from other different tasks, thus reducing the need for data volumes.
Multi-task evaluation
In this study, the performance of the scheme was evaluated by applying it to actual data. Experiments were carried out on 9 chillers in 8 buildings in 2 large industrial parks over a period of 4 months. The campus is shown in Figure 8.
FIG. 8 8 buildings and their chillers in 2 large industrial parks
Table 1 Task definition output results
Table 1 shows the overall information of the tasks mined by the task definition algorithm. Two different task sets are found in Park J and Park H. Observations show different project models in different numbers and in different scenarios when models are used. With the five-minute interval data, 33 tasks can be mined in Park J. The application scope of these task models is mainly judged by the rated power and average humidity of chillers. With the one-hour interval data, there are only two tasks in Park H, and the scope of application needs to be determined by power rating and cooling capacity rating. You can see that the sample size in each task is small. Of the total 35 tasks, 13 tasks had less than 100 samples, and the remaining 22 tasks had less than 1000 samples.
Several typical methods applied to the prediction of cooling efficiency are compared:
(1) The industry’s current approach: Initial configuration file (IP) Estimates the future energy efficiency ratio using the initial configuration file measured at installation, which is currently being used in the industry.
(2) Common methods in academia: single-task learning (STL) learns a model by summarizing the data of all tasks from each data set;
(3) Recent research work: Independent multitasking learning (IMTL) for data sources, which learns each task independently of data sources. For example, nine tasks can be fixed for nine chillers without sharing any samples or knowledge between tasks;
(4) Recent research work: Multi-task learning with domain knowledge (MTL), which learns task clustering defined by domain knowledge. For example, fix 50 tasks with 10 load ratios and 5 chillers.
Table 2 Improvement of error rate of each method
Table 2 shows that MELODY’s task definition can be improved over STL (single-task approach). However, incorrect task definitions (i.e., IMTL and MTL) did not improve the single-task approach. This is mainly because IMTL and MTL generate smaller data sets after task partitioning compared with adaptive task in different data sets (such as MELODY), which results in lack of training samples in some tasks. As the number of tasks increases with the number of attributes and time, the effect becomes worse as the task migration relationship becomes more complex. In this case, sharing knowledge between tasks becomes more challenging and leads to an effect known as negative migration, which is the error of sharing knowledge from unrelated source domains to target domains. MELODY was able to solve the related problems and thus outperform the latest energy-efficiency ratio prediction method, reducing the energy-efficiency ratio prediction error rate by 18.18-61.70%, resulting in 252 MWh energy savings per month in the two parks, and saving 36.75% more energy than the current cooling water cooler operation in the building.
Author: Dr. Zheng, senior research engineer of Huawei Cloud Edge Cloud Innovation Lab, graduated from The Hong Kong Polytechnic University, his main research interests are edge intelligence and AIoT. He has published more than ten papers on top international conferences and journals (TPDS, IJCAI, ICDCS, CIKM, TOSN, TIST, etc.), won the best conference paper award for many times, and won the outstanding Contribution awards of Huawei for key technology breakthrough, high-value patents and new service incubation for many times.
Huawei Cloud Edge Cloud Innovation Lab: Our vision is to explore the key technologies of edge-edge cloud collaboration and build ubiquitous intelligent edge cloud with extreme experience. We will work together with industrial partners and academic institutions to study edge cloud innovative technologies, incubate edge cloud innovative applications and build a prosperous ecosystem of edge cloud. His research interests include large-scale intelligent edge cloud platform, edge cloud collaborative AI, edge cloud collaborative rendering and video acceleration. At present, huawei edge computing platform IEF has been launched, and contributed to the first kubernetes-based cloud native edge computing platform KubeEdge, won the Peak open source technology innovation Award, the best intelligent edge computing technology innovation platform and a number of awards; The industry’s first edge cloud collaborative incremental learning workflow will be launched soon, huawei Cloud HiLens service and IEF service; Academically, 7 papers related to edge cloud collaborative AI and cloud native edge computing have been published in the past two years and won many awards of best paper and excellent paper.
Click to follow, the first time to learn about Huawei cloud fresh technology ~