WO2020227983A1 - Architecture de réseau neuronal à apprentissage hybride - Google Patents

Architecture de réseau neuronal à apprentissage hybride Download PDF

Info

Publication number
WO2020227983A1
WO2020227983A1 PCT/CN2019/087083 CN2019087083W WO2020227983A1 WO 2020227983 A1 WO2020227983 A1 WO 2020227983A1 CN 2019087083 W CN2019087083 W CN 2019087083W WO 2020227983 A1 WO2020227983 A1 WO 2020227983A1
Authority
WO
WIPO (PCT)
Prior art keywords
neural network
input feature
feature parameters
output
components
Prior art date
Application number
PCT/CN2019/087083
Other languages
English (en)
Inventor
Zhan Li
Zhixing Ren
Yun Zhang
Jialong WANG
Original Assignee
Alibaba Group Holding Limited
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Alibaba Group Holding Limited filed Critical Alibaba Group Holding Limited
Priority to PCT/CN2019/087083 priority Critical patent/WO2020227983A1/fr
Priority to CN201980093428.6A priority patent/CN113518962A/zh
Publication of WO2020227983A1 publication Critical patent/WO2020227983A1/fr

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N5/00—Computing arrangements using knowledge-based models
    • G06N5/02—Knowledge representation; Symbolic representation
    • G06N5/022—Knowledge engineering; Knowledge acquisition
    • G06N5/025—Extracting rules from data
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/088—Non-supervised learning, e.g. competitive learning

Definitions

  • an environment control system such as a heating, ventilation, and air conditioning (HVAC) system
  • HVAC heating, ventilation, and air conditioning
  • the HVAC system may include many duplicative and/or similar components, such as coolers, fans, secondary pumps, air conditioners, refrigeration units, water pumps, such as cooling water pumps (CWPs) , secondary chilling water pumps (SCWPs) , and the like.
  • CRAC computer room air conditioning units
  • CFD computational fluid dynamics
  • the deep learning neural network does not depend on any of the physical models and does not distinguish various input features.
  • Classic neural networks obtain knowledge/relations only from historical data and does not have any domain knowledge. It is more difficult for general deep learning model to apply to a system having a large number of duplicate and similar devices such as a computer room of a data center.
  • these HVAC components have complex nonlinear correlations, the inputs from the sensors are treated equally from the perspective of a neural network structure, and information behind data from the inputs may be biased by the duplicative and/or similar inputs, which may result overfitting and eventual inaccuracy, causing inefficiency.
  • a popular solution is to manually aggregate the inputs based on a human expert’s domain knowledge, and to set the aggregated input as the input to the neural network.
  • this solution is room-specific and introduces extra manual work.
  • this solution relies on the experience and analysis of an HVAC expert, it is difficult to fully understand the most reasonable correlations among various HVAC components to achieve an energy efficient computer room condition in different operating conditions, such as outdoor temperature, outdoor humidity, computing load, and the like.
  • FIG. 1 illustrates an example block diagram of an environment control system used with a hybrid-learning neural network (HLNN) which may be utilized to predict power usage effectiveness (PUE) of a computer room.
  • HLNN hybrid-learning neural network
  • FIG. 2 illustrates an example detailed block diagram of the environment control system of FIG. 1 with associated levels.
  • FIG. 3 illustrates an example block diagram of the HLNN architecture.
  • FIG. 4 illustrates an example flowchart describing a process of predicting the PUE by the HLNN.
  • the two-tower deep learning architecture may learn embeddings from data and an ontology structure automatically, and may include simultaneous training of two sub-networks, an unsupervised Auto-encoder Net (AE-Net) that captures the domain knowledge embedded in the ontology and a supervised Prediction Net (P-Net) that predicts the PUE from inputs.
  • AE-Net Auto-encoder Net
  • P-Net supervised Prediction Net
  • the simultaneous learning of the AE-Net (unsupervised) and the P-Net (supervised) may be referred to as hybrid-learning, and the two-tower deep learning architecture may also be referred to as a hybrid-learning neural network (HLNN) architecture.
  • HLNN hybrid-learning neural network
  • machine learning methods may be used to learn from historical data to obtain complex relationships among various HVAC components and the energy efficiency of the computer room in different operating conditions.
  • the first and second neural networks may share a shared structure comprising one input layer and two concept layers.
  • Each of the AE-Net and the P-Net may have its own hidden layers and an output layer.
  • the AE-Net, the P-Net, and the shared structure may form a hybrid-learning neural network (HLNN) architecture.
  • the AE-Net may be an unsupervised-learning network, which may be trained to make its output copy input with a lowest possible error, while P-Net may be a deep feedforward neural network to predict the PUE.
  • Domain knowledge of the components associated with the HVAC system and the computing equipment of the computer room may be embedded into the HLNN architecture.
  • the design of the two-tower deep learning architecture of the input layer and concept layers may be guided by the domain ontology containing multiple levels of nodes where a top level may contain a root concept and a bottom level may contain multiple instance.
  • the instances in the bottom level of the ontology may be represented by the nodes in the input layer, and the concepts in the middle levels may also have corresponding nodes in concept layers of the shared structure.
  • the relations and/or connections between levels may be copied in the input and the concept layers.
  • FIG. 1 illustrates an example block diagram of an environment control system 100 used with a hybrid-learning neural network (HLNN) which may be utilized to predict power usage effectiveness (PUE) of a computer room 102.
  • HLNN hybrid-learning neural network
  • the environment control system 100 may include a plurality of components such as an equipment and data collection module 104 communicatively coupled to an HVAC group 106 and an outside equipment and data group 108.
  • the equipment and data collection module 104 may be configured to maintain profiles of components managed by the HVAC group 106 and the outside equipment and data group 108, receive input data from various sensors associated with those components, and transmit data to those components to, in part, control the environment of, and calculate a predicted PUE of, the computer room 102.
  • Some of the environment control system components may be located in the computer room 102, and other components may be located outside of a building in which the computer room 102 is located.
  • the environment control system 100 may monitor energy consumption of components associated with the computer room 102, the equipment and data collection module 104, the HVAC group 106, and the outside equipment and data group 108.
  • the environment control system 100 may be communicatively coupled to a computer 110.
  • the computer 110 may comprise one or more processors 112 and memory 114 communicatively coupled to the one or more processors 112, which may store computer-readable instructions to be executed by the computer 110 to perform functions of the HLNN described below.
  • the computer 110 may be located within the computer room 102 or may be remotely located from the computer room 102.
  • the computer room 102 may house computing equipment 116 including servers, power supplies, displays, routers, network and communication modules, and the like (not shown) .
  • the computing equipment 116 may be coupled to the environment control system 100 and may provide information regarding energy usage by the computing equipment 116 based on historical, current, and expected energy usage and computing loads for calculating the predicted PUE of the computer room 102.
  • FIG. 2 illustrates an example detailed block diagram of the environment control system 100 of FIG. 1 with associated levels (levels 1-4 shown) .
  • the HVAC group 106 may comprise an HVAC control module 202 communicatively coupled to the equipment and data collection module 104, an air conditioning group 204, and a refrigeration group 206.
  • the HVAC control module 202 may be configured to receive operating information from various sensors and controllers of the air conditioning group 204 and from the refrigeration group 206.
  • the HVAC control module 202 may forward the operating information to the equipment and data collection module 104 for calculation by the HLNN.
  • the HVAC control module 202 may also be configured to transmit control information received from the equipment and data collection module 104 to the air conditioning group 204 and the refrigeration group 206 for adjusting various parameters of the air conditioning group 204 and the refrigeration group 206 to optimize a desired parameter for predicting the PUE.
  • the HVAC group 106 may further comprise a secondary pump group (not shown) and may similarly communicate associated operating information to and from the HVAC control module 202.
  • the air conditioning group 204 may comprise N air conditioners (two, AC-1 208 and AC-N 210, shown) .
  • each of N air conditioners may comprise several controls and sensors, such as a corresponding switch, a corresponding fan speed controller/sensor, a corresponding air conditioner output air temperature sensor, and a corresponding air conditioner return air temperature sensor.
  • Each of N air conditioners may be configured to receive AC operating information from the corresponding controls and sensors and forward the AC operating information to the air conditioning system 204, which, in turn, forwards the AC operating information to the HVAC control module 202.
  • Each of N air conditioners may also be configured to transmit AC control information received from the air conditioning system 204 to the corresponding controls to optimize a desired parameter for predicting the PUE.
  • the refrigeration group 206 may comprise a plurality of refrigeration systems including a plurality of coolers (cooler-1 212 shown) and a plurality of cooling towers (tower-1 214 shown) .
  • each of the plurality of coolers may comprise associated switch, cooling mode controller, outflow cooling water temperature controller/sensor
  • each of the plurality of cooling towers may comprise associated cooling tower fan speed controller/sensor, outflow cooling water temperature controller/sensor, and return cooling water temperature controller/sensor.
  • Each of the plurality of refrigeration systems may be configured to receive refrigeration operating information from the corresponding controls, switches, and sensors (not shown) and forward the refrigeration operating information to the HVAC control module 202 via the refrigeration group 206.
  • Each of the plurality of refrigeration systems may also be configured to transmit refrigeration control information received from the refrigeration group 206 to the corresponding controls, switches, and sensors to optimize the desired parameter for predicting the PUE.
  • the outside equipment and data group 108 may comprise an outside equipment monitoring module 216 communicatively coupled to the equipment and data collection module 104, an outside humidity module 218, an outside wet bulb temperature module 220, and other modules (not shown) .
  • the outside humidity module 218 may be communicatively coupled to M humidity sensors (two humidity sensors, humidity sensor-1 222 and humidity sensor-M 224, shown) .
  • the outside wet bulb temperature module 220 may be communicatively coupled to M wet bulb temperature sensors (two wet bulb temperature sensors, wet bulb temperature sensor-1 226 and wet bulb temperature sensor-M 228, shown) .
  • the outside equipment monitoring module 216 may receive humidity and wet bulb temperature information from the corresponding sensors and forward the information to the equipment and data collection module 104 for optimizing the desired parameter for predicting the PUE.
  • Each block illustrated in FIG. 2 may be associated with one of a plurality of levels of the domain ontology.
  • the domain ontology having four levels is illustrated herein as an example, however, the number of the levels of the domain ontology may not be limited to four and may be more or less than four levels.
  • Level 1 may include the equipment and data collection module 104, which may be referred to as D1.
  • Level 2 may include q modules including the HVAC control module 202 and the outside equipment monitoring module 216, which may be referred to as C_1, C_2, ...C_q, respectively.
  • Level 3 may include p modules including the air conditioning group 204, the refrigeration group 206, the outside humidity module 218, and the outside wet bulb temperature module 220, which may be referred to as B_1, B_2, ...B_p, respectively.
  • Level 4 may include k modules including the AC-1 208, the AC-N 210, the cooler-1 212, the tower-1 214, the humidity sensor-1 222, the humidity sensor-M 224, the wet bulb temp sensor-1 226, and the wet bulb temp sensor-M 228, which may be referred to as A_1, A_2, ...A_k, respectively.
  • FIG. 3 illustrates an example block diagram of the HLNN architecture 300.
  • the HLNN structure 300 may comprise a domain ontology 302, a shared structure 304, a first neural network, such as an AE-Net 306, and a second neural network, such as a P-Net 308.
  • There may be a plurality of levels in the ontology and four levels corresponding to the blocks illustrated in FIG. 2 are shown in the domain ontology 302 as examples.
  • the top level, Level 1 may contain the root concept, D_1 310, and the bottom level, Level 4, may contain a plurality of instances, of which four instances, A_1 312, A_2 314, A_n 316, and A_k 318, are shown.
  • These four instances in Level 4 of the domain ontology 302 may be represented as nodes, A_1 320, A_2 322, A_n 324, and A_k 326, respectively, in an input layer 328 of the shared structure 304.
  • the second level, Level 2, and the third level, Level 3, of the domain ontology 302 may represent a plurality of instances, of which two concepts, C_1 330 and C_q 332, in Level 2 and three concepts, B_1 334, B_2 336, and B_p 338, in Level 3 are shown. These instances in Level 2 and Level 3 of the domain ontology 302 may also have corresponding nodes, C_1 340, C_q 342, B_1 344, B_2 346, and B_p 348, respectively, in concept layers 350 of the shared structure 304. Additionally, relations/connections between levels may also be copied in the input layer 328 and the concept layers 350.
  • the concept B _1 334 is shown to be connected to a set of instances, A_1 312, A_2 314, and A_n 316, and in the concept layer 350, the corresponding node, B_1 344 is also shown to be connected to the corresponding nodes A_1 320, A_2 322, and A_n 324 in the input layer 328.
  • the P-Net 308 may be a deep feedforward neural network and may comprise hidden layers 352 and a one-node output layer 354 to output PUE parameters 356, plus the input layer 328 and the concept layers 350 of the shared structure 304.
  • An example feed-forward operation of the P-Net 308 is described below.
  • a neutron and a node may be interchangeable used.
  • the connections may be guided by the domain knowledge, which may not fully connect the nodes in the concept layers 350.
  • the concept relation weight between two concept nodes i.e., node j and node i
  • the weighted input of node i may then be expressed as:
  • Each layer of the concept layers 350 may be mapped from a corresponding level of the concepts in the domain ontology 302.
  • L PN (a, d p )
  • o P denote the calculated output of the neural network
  • d P denote the desired output
  • the AE-Net 306 may be an unsupervised learning model comprising hidden layers 358 and an output layer 360, plus the input layer 328 and the concept layers 350 of the shared structure 304.
  • the AE-Net 306 may be designed to minimize the difference between the input from the input layer 328 of the shared structure 304 and the output from the output layer 360.
  • an input vector a from the input layer 328, a representation vector c from the top concept layer of the concept layers 350, and an output vector r (R_1 362 and R_k 364 shown) from the output layer 360 a mapping that transforms a into c may be called an encoder, and a mapping that transforms c back to r may be called a decoder.
  • the encoder may be composed of the input layer 328 and the concept layers 350, while the decoder may be composed of the hidden layers 358 and the output layer 360.
  • the training process in the AE-Net 306 may help the encoder preserve the domain knowledge in the domain ontology 302.
  • the output vector r ⁇ r 1 , r 2 , ..., r k ⁇
  • the encoder function, f ⁇ , and the decoder function, g ⁇ may be expressed as:
  • the encoder function f ⁇ and decoder function g ⁇ may then be expressed, respectively, as:
  • s f and s g are the encoder and decoder activation functions.
  • r is not an exact reconstruction of a but the parameters of a distribution p (A
  • R r) that generates a with high probability.
  • the AE-Net 306 may be trained to find a parameter set to minimize reconstruction error in the equation below:
  • L AE is the loss function or the reconstruction error.
  • Both s f and s g may be sigmoid functions.
  • the HLNN 300 may be trained in a manner similar to standard neural networks. The only difference may be that the loss function L Model may be composed of two components: the loss of the AE-Net 306, L AE , and the prediction loss L PN of the P-Net 308:
  • L PN may be biased or weighted by another constant ⁇
  • the derivatives of the loss may be expressed as
  • the equation (15) may be transformed to:
  • Equation (15) may be transformed to:
  • the equation (15) may be transformed to:
  • Th equation (16) , (17) , and (18) show that the derivatives of the loss function L Model are back propagated for learning both the AE-Net 306 and the P-Net 308.
  • the solution for the PUE may be optimized by minimizing the loss calculated by the loss function L Model as expressed by the equation (11) , which may be accomplished by setting the derivative, such as the equations (16) , (17) , and (18) , of the loss function L Model to zero and solving for the variables. Because the solution may not always converge to zero, or may take longer than an acceptable time or a number of iterations, the value for the derivative may be set to a sufficiently small and acceptable threshold value.
  • FIG. 4 illustrates an example flowchart 400 describing a process of predicting the power utilization effectiveness (PUE) by the HLNN 300.
  • the HLNN 300 may create an ontology having a plurality of levels, such as the domain ontology 302, of the components associated with the environment control system 100 associated with the computer room 102 as illustrated in FIGs. 1-3.
  • the HLNN 300 may receive information of the components associated with the environment control system 100 automatically including corresponding associated historical data, locations and physical connections, and hierarchy among the components as illustrated in FIGs. 1-3.
  • the computing equipment 116 may include servers, power supplies, displays, routers, network and communication modules (telephone, internet, wireless devices, etc. ) , and the like.
  • the relationships among components of the environment control system 100 and the computing equipment 116 may be based on loading of the computing equipment 116, such as a workload, or computing load, of the servers and an electrical load of the servers as a function of the workload of the servers.
  • the HLNN 300 may receive input feature parameters of the components associated with the environment control system 100. More specifically, the input layer 328 of the shared structure 304 may receive k instances, A_1 312, A_2 314, A_n 316, and A_k 318, from the domain ontology 302, where k is an integer. Each of k instances may have a corresponding input feature parameter in the input layer 328 (A_1 320, A_2 322, A_n 324, and A_k 326 as illustrated in FIG. 3) may belong to one or more corresponding upper concepts of a plurality of upper concepts as illustrated hierarchically in the concept layers 350.
  • both the first neural network, such as the AE-Net 306, and the second neural network, such as the P-Net 308, may be trained simultaneously.
  • a mapping that transforms a into c may be called an encoder, and a mapping that transforms c back to r may be called a decoder.
  • the encoder may be composed of the input layer 328 and the concept layers 350, while the decoder may be composed of the hidden layers 358 and the output layer 360.
  • the training process in the AE-Net 306 may help the encoder preserve the domain knowledge in the domain ontology 302.
  • the HLNN 300 may minimize the loss based on the loss function L Model by utilizing the trained AE-Net 306 and the trained P-Net 308 and predict a power usage effectiveness (PUE) of the computer room 102 at block 410.
  • the derivative of the loss function L Model such as the equations (16) , (17) , and (18) , may be to zero for solving for the variables. Because the solution may not always converge to zero, or may take longer than an acceptable time, the value for the derivative may be set to a sufficiently small and acceptable threshold value.
  • the trained neural networks may be generated automatically, and the training of the trained neural networks may be performed by using a gradient descent algorithm to implement learning of the input feature parameters for corresponding concepts.
  • An architecture of the trained neural networks may reflect deep learning of the plurality of components and associated concepts based on the relationships among the plurality of components.
  • the trained neural networks may comprise hierarchical concept layers, such as the concept layers 350, coupled between the input layer, such as the input layer 328, and an output layer, such as the output layer 354 or 360.
  • the concept layers 350 may be added between the input layer 328 and the hidden layers 352 and 358 as illustrated in FIG. 3.
  • the concept layer 350 may be embedded with domain knowledge from the domain ontology 302.
  • the concept layer 350 may construct a concept structure based on relationships among the plurality of components.
  • the concept structure may be created manually or automatically with smart components capable of communicating with each other.
  • the training portion of the HLNN 300 and the prediction the PUE utilizing the HLNN 300 may be performed separately and/or by different parties.
  • a general deep learning network may not be capable of reasonably distinguishing all duplicative and/or similar input features, and may identify the importance of each feature based entirely on historical data.
  • a structure such as the computer room 102 with a large number of duplicative and similar devices, if these duplicate and/or similar input features parameters were not categorized, aggregated or abstracted, the complexity of the network and space for learning and searching would greatly increase, requiring higher quality and quantity of data. Although, it may be easy to obtain unreasonable overfitting, it would decrease prediction accuracy.
  • Computer-readable instructions include routines, applications, application modules, program modules, programs, components, data structures, algorithms, and the like.
  • Computer-readable instructions can be implemented on various system configurations, including single-processor or multiprocessor systems, minicomputers, mainframe computers, personal computers, hand-held computing devices, microprocessor-based, programmable consumer electronics, combinations thereof, and the like.
  • the computer-readable storage media may include volatile memory (such as random-access memory (RAM) ) and/or non-volatile memory (such as read-only memory (ROM) , flash memory, etc. ) .
  • volatile memory such as random-access memory (RAM)
  • non-volatile memory such as read-only memory (ROM) , flash memory, etc.
  • the computer-readable storage media may also include additional removable storage and/or non-removable storage including, but not limited to, flash memory, magnetic storage, optical storage, and/or tape storage that may provide non-volatile storage of computer-readable instructions, data structures, program modules, and the like.
  • a non-transient computer-readable storage medium is an example of computer-readable media.
  • Computer-readable media includes at least two types of computer-readable media, namely computer-readable storage media and communications media.
  • Computer-readable storage media includes volatile and non-volatile, removable and non-removable media implemented in any process or technology for storage of information such as computer-readable instructions, data structures, program modules, or other data.
  • Computer-readable storage media includes, but is not limited to, phase change memory (PRAM) , static random-access memory (SRAM) , dynamic random-access memory (DRAM) , other types of random-access memory (RAM) , read-only memory (ROM) , electrically erasable programmable read-only memory (EEPROM) , flash memory or other memory technology, compact disk read-only memory (CD-ROM) , digital versatile disks (DVD) or other optical storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transmission medium that can be used to store information for access by a computing device.
  • communication media may embody computer-readable instructions, data structures, program modules, or other data in a modulated data signal, such as a carrier wave, or other transmission mechanism. As defined herein, computer-readable storage media do not include communication media.
  • the computer-readable instructions stored on one or more non-transitory computer-readable storage media that, when executed by one or more processors, may perform operations described above with reference to FIGs. 1-4.
  • computer-readable instructions include routines, programs, objects, components, data structures, and the like that perform particular functions or implement particular abstract data types.
  • the order in which the operations are described is not intended to be construed as a limitation, and any number of the described operations can be combined in any order and/or in parallel to implement the processes.
  • a method comprising: receiving input feature parameters of a plurality of components associated with at least one computer room; raining a first neural network and a second neural network based on the input feature parameters; and predicting the power usage effectiveness (PUE) of the at least one computer room based on an output of the first neural network and an output of the second neural network.
  • PUE power usage effectiveness
  • the method as paragraph A recites, wherein the first neural network is an unsupervised neural network and the second network is a supervised prediction neural network.
  • training the first neural network and the second neural network based on the input feature parameters includes simultaneously training the first neural network and the second neural network based on the input feature parameters.
  • receiving the input feature parameters of the plurality of components associated with the at least one computer room includes: creating an ontology having a plurality of levels associated with the plurality of components; and receiving information of the plurality of components based on the on the ontology including corresponding associated concept, historical data, locations, physical connections, and hierarchy among the plurality of components.
  • training the first neural network and the second neural network based on the input feature parameters includes using a gradient descent algorithm to implement learning of the input feature parameters for corresponding concepts.
  • predicting the PUE of the at least one computer room based on the output of the first neural network and the output of the second neural network includes minimizing a total loss calculated based on the biased loss associated with the first neural network and an unbiased loss associated with the second neural network.
  • a system comprising: one or more processors; and memory communicatively coupled to the one or more processors, the memory storing computer-readable instructions executable by one or more processors, that when executed by the one or more processors, cause the one or more processors to perform operations comprising: receiving input feature parameters of a plurality of components associated with at least one computer room; training a first neural network and a second neural network based on the input feature parameters; and predicting the power usage effectiveness (PUE) of the at least one computer room based on an output of the first neural network and an output of the second neural network.
  • PUE power usage effectiveness
  • the system as paragraph M recites, wherein the first neural network is an unsupervised neural network and the second network is a supervised prediction neural network.
  • training the first neural network and the second neural network based on the input feature parameters includes simultaneously training the first neural network and the second neural network based on the input feature parameters.
  • receiving the input feature parameters of the plurality of components associated with the at least one computer room includes: creating an ontology having a plurality of levels associated with the plurality of components; and receiving information of the plurality of components based on the on the ontology including corresponding associated concept, historical data, locations, physical connections, and hierarchy among the plurality of components.
  • training the first neural network and the second neural network based on the input feature parameters includes using a gradient descent algorithm to implement learning of the input feature parameters for corresponding concepts.
  • predicting the PUE of the at least one computer room based on the output of the first neural network and the output of the second neural network includes minimizing a total loss calculated based on the biased loss associated with the first neural network and an unbiased loss associated with the second neural network.
  • a non-transitory computer-readable storage medium storing computer-readable instructions executable by one or more processors, that when executed by the one or more processors, cause the one or more processors to perform operations comprising: receiving input feature parameters of a plurality of components associated with at least one computer room; training a first neural network and a second neural network based on the input feature parameters; and predicting the power usage effectiveness (PUE) of the at least one computer room based on an output of the first neural network and an output of the second neural network.
  • PUE power usage effectiveness
  • the non-transitory computer-readable storage medium as paragraph Y recites, wherein the first neural network is an unsupervised neural network and the second network is a supervised prediction neural network.
  • non-transitory computer-readable storage medium as paragraph Y recites, wherein training the first neural network and the second neural network based on the input feature parameters includes simultaneously training the first neural network and the second neural network based on the input feature parameters.
  • receiving the input feature parameters of the plurality of components associated with the at least one computer room includes: creating an ontology having a plurality of levels associated with the plurality of components; and receiving information of the plurality of components based on the on the ontology including corresponding associated concept, historical data, locations, physical connections, and hierarchy among the plurality of components.
  • non-transitory computer-readable storage medium as paragraph AB recites, wherein the relationships among the plurality of components are based, at least in part, on loading of computing equipment in the computer room.
  • AD The non-transitory computer-readable storage medium as paragraph AC recites, wherein the loading of the computing equipment includes a workload of the computing equipment and an electrical load used by the computing equipment.
  • the non-transitory computer-readable storage medium as paragraph AD recites, wherein the computing equipment includes a server and a power supply for the server.
  • non-transitory computer-readable storage medium as paragraph AB recites, wherein training the first neural network and the second neural network based on the input feature parameters includes using a gradient descent algorithm to implement learning of the input feature parameters for corresponding concepts.
  • non-transitory computer-readable storage medium as paragraph Y recites, wherein predicting the PUE of the at least one computer room based on the output of the first neural network and the output of the second neural network includes biasing a loss associated with the first neural network with a constant value.
  • non-transitory computer-readable storage medium as paragraph AG recites, wherein predicting the PUE of the at least one computer room based on the output of the first neural network and the output of the second neural network t includes minimizing a total loss calculated based on the biased loss associated with the first neural network and an unbiased loss associated with the second neural network.
  • AI The non-transitory computer-readable storage medium as paragraph AH recites, wherein minimizing the loss calculated based on the loss function by utilizing the first neural network and the second neural network includes solving for a derivative of the loss function equaling zero.
  • non-transitory computer-readable storage medium as paragraph AH recites, wherein minimizing the loss calculated based on the loss function by utilizing the trained first neural network and the trained second neural network includes solving for a derivative of the loss function being less than or equal to a threshold value.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • Software Systems (AREA)
  • Computational Linguistics (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • Mathematical Physics (AREA)
  • Computing Systems (AREA)
  • Molecular Biology (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Biomedical Technology (AREA)
  • General Health & Medical Sciences (AREA)
  • Biophysics (AREA)
  • Health & Medical Sciences (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

L'invention concerne des systèmes et des procédés pour prédire l'efficacité énergétique d'une salle informatique dans un centre de données, et plus particulièrement pour prédire un indicateur d'efficacité énergétique (PUE) d'une salle informatique avec des paramètres optimisés à l'aide d'une architecture d'apprentissage en profondeur à deux tours. L'architecture d'apprentissage en profondeur à deux tours peut apprendre automatiquement des incorporations à partir de données et d'une structure d'ontologie, et peut inclure un entraînement de deux sous-réseaux, tels qu'un premier réseau neuronal qui capture les connaissances de domaine incorporées dans l'ontologie et un second réseau neuronal qui prédit le PUE à partir d'entrées. L'apprentissage du premier réseau neuronal, qui peut être non supervisé, et celui du second réseau neuronal, qui peut être supervisé, peuvent être simultanés et appelés apprentissage hybride, et l'architecture d'apprentissage en profondeur à deux tours peut également être appelée architecture de réseau neuronal à apprentissage hybride (HLNN).
PCT/CN2019/087083 2019-05-15 2019-05-15 Architecture de réseau neuronal à apprentissage hybride WO2020227983A1 (fr)

Priority Applications (2)

Application Number Priority Date Filing Date Title
PCT/CN2019/087083 WO2020227983A1 (fr) 2019-05-15 2019-05-15 Architecture de réseau neuronal à apprentissage hybride
CN201980093428.6A CN113518962A (zh) 2019-05-15 2019-05-15 混合学习神经网络架构

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2019/087083 WO2020227983A1 (fr) 2019-05-15 2019-05-15 Architecture de réseau neuronal à apprentissage hybride

Publications (1)

Publication Number Publication Date
WO2020227983A1 true WO2020227983A1 (fr) 2020-11-19

Family

ID=73290100

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/087083 WO2020227983A1 (fr) 2019-05-15 2019-05-15 Architecture de réseau neuronal à apprentissage hybride

Country Status (2)

Country Link
CN (1) CN113518962A (fr)
WO (1) WO2020227983A1 (fr)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113465139A (zh) * 2021-05-28 2021-10-01 山东英信计算机技术有限公司 一种制冷优化方法、***、存储介质及设备
CN115907202A (zh) * 2022-12-13 2023-04-04 中国通信建设集团设计院有限公司 一种双碳背景下的数据中心pue计算分析方法及***

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113721151B (zh) * 2021-11-03 2022-02-08 杭州宇谷科技有限公司 基于双塔深度学习网络的电池容量预估模型及方法

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20090201293A1 (en) * 2008-02-12 2009-08-13 Accenture Global Services Gmbh System for providing strategies for increasing efficiency of data centers
US20100076607A1 (en) * 2008-08-08 2010-03-25 Osman Ahmed Data center thermal performance optimization using distributed cooling systems
CN103645795A (zh) * 2013-12-13 2014-03-19 浪潮电子信息产业股份有限公司 一种基于人工神经网络的云计算数据中心节能方法

Family Cites Families (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109670623A (zh) * 2017-10-16 2019-04-23 优酷网络技术（北京）有限公司 神经网络预测方法及装置
CN109002942A (zh) * 2018-09-28 2018-12-14 河南理工大学 一种基于随机神经网络的短期负荷预测方法

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20090201293A1 (en) * 2008-02-12 2009-08-13 Accenture Global Services Gmbh System for providing strategies for increasing efficiency of data centers
US20100076607A1 (en) * 2008-08-08 2010-03-25 Osman Ahmed Data center thermal performance optimization using distributed cooling systems
CN103645795A (zh) * 2013-12-13 2014-03-19 浪潮电子信息产业股份有限公司 一种基于人工神经网络的云计算数据中心节能方法

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
LIU, WEI ET AL.: "Research of Mutual Learning Neural Network Training Method", CHINESE JOURNAL OF COMPUTERS, vol. 40, no. 6, 30 June 2017 (2017-06-30), XP055752928, ISSN: 0254-4164 *

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113465139A (zh) * 2021-05-28 2021-10-01 山东英信计算机技术有限公司 一种制冷优化方法、***、存储介质及设备
CN113465139B (zh) * 2021-05-28 2022-11-08 山东英信计算机技术有限公司 一种制冷优化方法、***、存储介质及设备
CN115907202A (zh) * 2022-12-13 2023-04-04 中国通信建设集团设计院有限公司 一种双碳背景下的数据中心pue计算分析方法及***
CN115907202B (zh) * 2022-12-13 2023-10-24 中国通信建设集团设计院有限公司 一种双碳背景下的数据中心pue计算分析方法及***

Also Published As

Publication number Publication date
CN113518962A (zh) 2021-10-19

Similar Documents

Publication Publication Date Title
Li et al. Intelligent multi-zone residential HVAC control strategy based on deep reinforcement learning
Wei et al. Multi-objective optimization of the HVAC (heating, ventilation, and air conditioning) system performance
Dash et al. Efficient stock price prediction using a self evolving recurrent neuro-fuzzy inference system optimized through a modified differential harmony search technique
Fallah-Mehdipour et al. Real-time operation of reservoir system by genetic programming
Hsu Using improved grey forecasting models to forecast the output of opto-electronics industry
WO2020227983A1 (fr) Architecture de réseau neuronal à apprentissage hybride
EP3835895A1 (fr) Apprentissage par renforcement profond multi-agents pour la commande dynamique des équipements électriques dans les bâtiments
CN110866592B (zh) 模型训练方法、装置、能效预测方法、装置和存储介质
Wahid et al. An efficient approach for energy consumption optimization and management in residential building using artificial bee colony and fuzzy logic
CN110826784B (zh) 能源使用效率的预测方法、装置及存储介质和终端设备
US11835928B2 (en) Adaptive mixed integer nonlinear programming for process management
Lissa et al. Transfer learning applied to reinforcement learning-based hvac control
Behrang et al. Assessment of electricity demand in Iran's industrial sector using different intelligent optimization techniques
Cheng et al. Forecasting monthly runoff time series by single-layer feedforward artificial neural network and grey wolf optimizer
Raghavendra et al. Artificial humming bird with data science enabled stability prediction model for smart grids
Cao et al. PSO-Stacking improved ensemble model for campus building energy consumption forecasting based on priority feature selection
WO2019227273A1 (fr) Modèle de réseau neuronal basé sur un concept hiérarchique destiné à une prédiction d'efficacité d'utilisation de puissance de centre de données
Gruber et al. Efficient Bayesian network learning for system optimization in reliability engineering
Cho et al. Rule reduction for control of a building cooling system using explainable AI
CN112330044A (zh) 一种基于迭代聚集网格搜索算法的支持向量回归模型
CN116717839A (zh) 供热控制方法、控制装置和供热***
CN116954329A (zh) 制冷***的状态调节方法、装置、设备、介质及程序产品
CN112234599B (zh) 一种多元复杂城市电网超前动态自适应分区方法及其***
Guo et al. Fruit Fly Optimization Algorithm Based on Single‐Gene Mutation for High‐Dimensional Unconstrained Optimization Problems
Genkin et al. A transfer learning approach to minimize reinforcement learning risks in energy optimization for automated and smart buildings

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 19928522

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 19928522

Country of ref document: EP

Kind code of ref document: A1