Heating primary network and secondary network task scheduling and control method based on edge computing

Abstract

The application discloses a heat supply primary network and secondary network task scheduling and control method, comprising the following steps: setting intelligent edge terminals and local computing terminals between the heat supply primary network and the secondary network; dividing the heat supply primary network and secondary network cooperative control task into multiple control subtasks; setting the execution mode of the control subtasks as execution by the local computing terminals or migration to the intelligent edge terminals for execution, forming a scheduling scheme for the local and edge cooperative execution of the heat supply primary network and secondary network control task; taking the delay and energy consumption minimization of the local computing terminal execution of the control task and the intelligent edge terminal execution of the control task as the target, establishing a control task scheduling optimization model, and solving the optimal scheduling scheme of the control task; performing data dimension and magnitude processing and transformation on multi-source heterogeneous data information, and establishing a control prediction model after feature extraction and model training, model compression processing by machine learning.

Description

Technical Field

[0001] This invention belongs to the field of smart heating technology, specifically relating to a method for task scheduling and control of primary and secondary heating networks based on edge computing. Background Technology

[0002] The heating system consists of hot water or steam supplied by a heat source, which enters the community's heat exchange station through the primary network and then flows into the heat exchanger through the primary water supply pipeline. Simultaneously, water from the secondary network enters the heat exchanger in the reverse direction. Heat exchange is completed in the heat exchanger, and the water in the secondary network is heated. It then flows into each household's home through the secondary water supply pipeline. After heat exchange, the water from the primary network returns to the main network through the primary return pipeline, circulating back to the heat source station. The primary network refers to the pipelines from the heat source to the valves at the interfaces of each heat-consuming unit. The secondary network refers to the pipelines from the heat exchange station of each heat-consuming unit to the individual buildings within that unit.

[0003] The primary and secondary networks typically require control over heating temperature, circulating pumps, and makeup water pumps to meet the heating needs of the heating system. However, the control variables between the primary and secondary networks are numerous, the regulation is complex, and the delay is significant. If the data is processed by the cloud center, it will result in a large amount of data and slow data processing. How to scientifically and rationally schedule and allocate complex control tasks and achieve coordinated control between the primary and secondary networks are urgent problems to be solved.

[0004] Based on the aforementioned technical issues, a new method for task scheduling and control of primary and secondary heating networks based on edge computing needs to be designed. Summary of the Invention

[0005] The technical problem to be solved by this invention is to overcome the shortcomings of the prior art and provide a task scheduling and control method for primary and secondary heating networks based on edge computing. This method enables task scheduling and collaborative control of primary and secondary networks through the collaboration of local computing terminals and intelligent edge terminals. It selects the minimum weighted sum of energy consumption and latency to reduce data storage, communication and processing costs, avoids the burden of large amounts of data on traditional cloud computing centers, effectively utilizes intelligent edge terminals and local computing terminals for control task scheduling and allocation, and uses machine learning algorithms to establish a collaborative control prediction model for primary and secondary networks to achieve collaborative control between primary and secondary networks.

[0006] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0007] This invention provides a method for task scheduling and control of primary and secondary heating networks based on edge computing, comprising:

[0008] Intelligent edge terminals and local computing terminals are set up between the primary and secondary networks of the heating system to form an edge computing network;

[0009] The coordinated control task of the primary and secondary heating networks is divided into multiple control sub-tasks;

[0010] The execution mode of the control subtasks is set to be executed through the local computing terminal or migrated to the intelligent edge terminal for execution, forming a scheduling scheme for the coordinated execution of the primary and secondary network control tasks of the heating system by local and edge operations.

[0011] With the goal of minimizing the latency and energy consumption of control tasks executed by local computing terminals and intelligent edge terminals, a control task scheduling optimization model is established, and the optimal scheduling scheme for control tasks is obtained by solving the model.

[0012] Based on the optimal scheduling scheme of the control task, the multi-source heterogeneous data information of the primary and secondary network equipment of the heating system is processed and transformed in terms of data dimensions and magnitude through local computing terminals and intelligent edge terminals. Machine learning is used for feature extraction and model training. After the model is compressed using channel pruning and knowledge distillation methods, a control prediction model is established to carry out coordinated control of the primary and secondary networks of the heating system.

[0013] Furthermore, the number of intelligent edge terminals and local computing terminals is at least one; the local computing terminal is equipped with a task processing component, which has lower computing resources and capabilities than the intelligent edge terminal, and is used for small-scale data processing, feature extraction, and model training. The intelligent edge terminal is equipped with an edge infrastructure unit, an edge computing service unit, and an edge application unit; the edge infrastructure unit is configured with network resources, storage resources, computing resources, and an operating system for low-level computation, storage, and network communication; the edge computing service unit is configured with a data transmission component, a data storage component, a data preprocessing component, an algorithm model component, a feature engineering component, and configuration management, operation and maintenance management, and device access management components, which are used to provide business support for the collaborative control tasks of the primary and secondary networks of the heating system, and to establish a control prediction model after multi-source heterogeneous data processing, data fusion, and feature extraction; the edge application unit uses the components in the edge computing service unit to realize the collaborative control functions of the primary and secondary networks of the heating system, as well as the management of the control prediction model and data management.

[0014] Furthermore, the collaborative control task of the primary and secondary heating networks is divided into multiple control sub-tasks, including the secondary network water supply temperature control task, the circulating pump control task, and the makeup water pump control task. Each control task also includes data acquisition, data preprocessing, feature extraction, and model training sub-tasks.

[0015] The secondary network water supply temperature control task is to establish a secondary network water supply temperature prediction and control model based on collected meteorological data, secondary side water supply temperature, secondary side return water temperature, primary side water supply temperature, primary side return water temperature, primary side temperature control valve data, and set secondary side water supply temperature data.

[0016] The circulating pump control task is to establish a predictive control model for the circulating pump based on the collected primary side water supply pressure, primary side return water pressure, secondary side water supply pressure, secondary side return water pressure, secondary side supply and return water pressure difference, circulating pump frequency converter data, and set secondary side supply and return water pressure difference data.

[0017] The water supply pump control task involves establishing a predictive control model for the water supply pump based on the collected primary side water supply pressure, primary side return water pressure, secondary side water supply pressure, secondary side return water pressure, water supply flow rate, water supply pump frequency converter data, and set secondary side return water pressure data.

[0018] Furthermore, the execution mode of the control subtasks is set to be either executed through a local computing terminal or migrated to an intelligent edge terminal for execution, forming a scheduling scheme for the coordinated execution of primary and secondary network control tasks of the heating system via local and edge methods, including:

[0019] The execution mode of each control subtask is set to be either executed through a local computing terminal or migrated to a smart edge terminal for execution; the information of each control subtask includes the computational amount of task data, the data task size migrated to the smart edge terminal, and the delay time;

[0020] Based on the execution methods of different control sub-tasks, a set of scheduling schemes is formed to coordinate the execution of primary and secondary network control tasks of the heating system in a local and edge manner.

[0021] Furthermore, the step of establishing a control task scheduling optimization model with the goal of minimizing the latency and energy consumption of control tasks executed by local computing terminals and intelligent edge terminals, and solving for the optimal scheduling scheme of control tasks, includes:

[0022] The latency and energy consumption for executing control tasks via the local computing terminal are expressed as follows:

[0023]

[0024]

[0025] Calculate the delay locally; x m，0 For the identifier of the task computed locally; l m This refers to the total amount of task data calculated locally. Data computing capabilities provided for local computing terminals; η is the local computing energy consumption; m is the energy consumption coefficient; m is the number of local computing terminals.

[0026] The latency of executing control tasks through a smart edge terminal is expressed as:

[0027]

[0028] The total latency is calculated at the edge. Delay for data upload; The time required for intelligent edge terminals to perform computations; x m,n This serves as an identifier for the subtask's computation on the intelligent edge terminal; c m The amount of data uploaded to the intelligent edge terminal; R m,n For the transmission rate of tasks migrated to intelligent edge terminals; f r The data computing power provided for intelligent edge terminals; N is the number of intelligent edge terminals;

[0029] The energy consumption for performing control tasks through intelligent edge terminals is expressed as:

[0030]

[0031] Calculate the total energy consumption at the edge; Energy consumption for uploading task data; Energy consumption for performing computations on intelligent edge terminals; p m p represents the communication power consumption when data is uploaded to the intelligent edge terminal. r The power consumption for migrating tasks to intelligent edge terminals;

[0032] The latency for the local computing terminal to execute control tasks and the latency for the intelligent edge terminal to execute control tasks are expressed as follows:

[0033]

[0034] The energy consumption of local computing terminals performing control tasks and intelligent edge terminals performing control tasks is expressed as follows:

[0035]

[0036] The objective is to minimize the latency and energy consumption of both the local computing terminal and the intelligent edge terminal performing control tasks, expressed as:

[0037]

[0038] Cost m,nTo minimize the weighted total cost of delay and energy consumption; w1 and w2 are non-negative weighting factors, and w1 + w2 = 1; These represent the minimum and maximum tolerable delays, respectively. The minimum and maximum energy consumption that can be tolerated;

[0039] The constraints include: the communication power consumption when uploading data to the intelligent edge terminal is less than the maximum power consumption when the device is uploading; the computing power of the intelligent edge terminal does not exceed its own maximum allowable computing power; and the latency of both the local computing terminal and the intelligent edge terminal meets the maximum latency that the device can tolerate.

[0040] With the goal of minimizing the latency and energy consumption of control tasks executed by local computing terminals and intelligent edge terminals, and with constraints set, a control task scheduling optimization model is established.

[0041] An improved multiverse optimization algorithm is used to solve the control task scheduling optimization model to obtain the optimal scheduling scheme for the control task.

[0042] Furthermore, the improved multiverse optimization algorithm includes: introducing a spiral to update the universe position and find the optimal individual; and introducing an adaptive compression factor to change the position of the current optimal universe.

[0043] The implementation process of the improved multiverse optimization algorithm includes:

[0044] Initialize the number of universes and the maximum number of iterations;

[0045] The universe is arranged according to its expansion rate, and a white hole is selected using a roulette wheel mechanism. Let NI(X) be the j-th parameter of the i-th universe. i ) represents the normalized expansion rate of the i-th universe; r1 is a random number between [0,1]. Let j be the parameter of the k-th universe chosen by the roulette wheel mechanism;

[0046] Update the probability of wormhole existence W EP and travel distance rate T DR , W EP min W EP max W respectively EP The minimum and maximum values; l is the current iteration number; L is the maximum iteration number; p is the development precision during the iteration process. The higher the value, the faster the local search.

[0047] Calculate the current universe expansion rate. If the current universe expansion rate is better than the current universe expansion rate, then update the current universe expansion rate; otherwise, keep the current universe expansion rate.

[0048] Perform individual position updates within the universe, searching for the optimal individual when r² < W. EP Update cosmic position:

[0049]

[0050] When r2≥W EP Update cosmic position x j Let ub be the j-th parameter of the currently formed optimal universe; j lb j r1, r2, r3, and r4 are the upper and lower bounds of variable j, respectively; r2, r3, and r4 are random numbers between [0,1]; λ is the adaptive compression factor. b is a constant defined for the logarithmic spiral shape; m is a constant between [0,1].

[0051] Termination condition determination: If the termination condition is met, output the corresponding result, which is the optimal scheduling scheme for the control task; otherwise, increment the iteration count by 1 and return to select a new white hole.

[0052] Furthermore, the optimal scheduling scheme based on the control task processes and transforms the multi-source heterogeneous data information of the primary and secondary network equipment of the heating system in terms of data dimensions and magnitude through local computing terminals and intelligent edge terminals, including:

[0053] Based on the optimal scheduling scheme of the control task, the execution mode of the task is set. The multi-source data of the collected primary network and secondary network devices are decomposed according to the time sequence characteristics through the local computing terminal and intelligent edge terminal. Let X be the standardized transformation input, and X exists in the form of multi-dimensional data, matrix form and vector form.

[0054] Let X = (X1, X2, ..., X...) p When presented in matrix form, it is represented as:

[0055]

[0056] Set up generalized power transform BC-Zscore normalization transform schemes for different data formats:

[0057] If X exists in the form of a vector, return the transformed result vector BC_Z = [X - mean(X)] / std(X);

[0058] If X exists in the form of a matrix, the mean and standard deviation of the column vectors of X are used to standardize the data of the corresponding columns one by one, and the transformed result matrix BC_Z is returned.

[0059] If X exists in the form of a multidimensional array, its mean and standard deviation are calculated along multiple dimensions of X, and then X is standardized to return the transformed high-dimensional array BC_Z.

[0060] X = (X1, X2, ..., X...) p After BC-Z score normalization transformation:

[0061]

[0062] This is the average value; The standard deviation is X; after BC-Z score standardization, X = (X1, X2, ..., X...). p In each column of )

[0063] The data is iteratively processed according to the selected multi-source data processing scheme. The above process is executed repeatedly, and the data transformation and processing results are aggregated.

[0064] The multi-source data standardization transformation ends after all the data sources input to the task have undergone unified transformation in terms of dimensions and magnitudes. The task output is then saved for use in multi-source data fusion, edge data computation, or data storage.

[0065] Furthermore, the machine learning algorithm used to establish the control prediction model is a convolutional neural network, and channel pruning and knowledge distillation methods are used to compress the convolutional neural network model.

[0066] Furthermore, the method of compressing the convolutional neural network model using channel pruning includes:

[0067] Initialize the convolutional neural network;

[0068] The channels are sparsified under the constraint of L1 norm regularization, and the neural network is trained. The importance of a channel is determined by using a scaling parameter in the Batch Normalization layer. The smaller the scaling parameter, the less impact the pruning of that channel has on the convolutional neural network. The loss function after adding L1 regularization is expressed as: L represents the total loss predicted by the convolutional neural network model after sparsification; W represents the training weights, i.e., the first summation term corresponds to the loss l′ predicted by the model after normal training of the convolutional neural network; γ represents the scaling factor on each channel; g(γ) represents the penalty function for the scaling factor γ; λ′ represents the hyperparameter balancing the two summation terms.

[0069] Channel pruning: After training the model using L1 norm sparsification, prune the channels with smaller scaling parameters in the convolutional blocks according to the γ value of the BN layer.

[0070] After pruning, the network is fine-tuned based on the actual performance of the convolutional neural network model until a convolutional neural network model that meets the actual needs is achieved.

[0071] Furthermore, the model compression process of the convolutional neural network model using the knowledge distillation method includes:

[0072] Train the teacher neural network model and use it as the teacher network model for knowledge distillation; use the pruned convolutional neural network as the student network model.

[0073] The loss function for knowledge distillation is defined as: L = αL soft +βL hard L soft The loss function for when the teacher model guides the student model; L hard α is the original loss function of the student model; α and β are hyperparameters balancing the two loss functions.

[0074] Training the student network: The data features extracted through feature extraction are input into the student network model and the teacher network model. The loss function of knowledge distillation is then used as the loss function for the entire training process. By adjusting α and β, the optimal training effect is obtained.

[0075] The beneficial effects of this invention are:

[0076] This invention forms an edge computing network by setting up intelligent edge terminals and local computing terminals between the primary and secondary networks of a heating system; it divides the collaborative control tasks of the primary and secondary networks into multiple control sub-tasks; it sets the execution mode of the control sub-tasks to be either executed through the local computing terminal or migrated to the intelligent edge terminal, forming a scheduling scheme for the collaborative execution of the control tasks of the primary and secondary networks of the heating system by local and edge computing; it establishes a control task scheduling optimization model with the goal of minimizing the latency and energy consumption of the control tasks executed by the local computing terminal and the intelligent edge terminal, and solves the optimal scheduling scheme for the control tasks; based on the optimal scheduling scheme for the control tasks, it processes and transforms the multi-source heterogeneous data information of the primary and secondary network equipment of the heating system through the local computing terminal and the intelligent edge terminal, uses machine learning for feature extraction and model training, and uses channel pruning and knowledge distillation methods to compress the model and establish a control... This system develops a predictive model for coordinated control of the primary and secondary heating networks. It enables control of these networks through collaboration between local computing terminals and intelligent edge terminals, minimizing the weighted sum of energy consumption and latency to reduce data storage, communication, and processing costs. This avoids relying on traditional cloud computing centers to handle large amounts of data, effectively utilizing intelligent edge terminals and local computing terminals for control task scheduling and allocation. Machine learning algorithms are employed to establish a predictive model for coordinated control between the primary and secondary networks, achieving collaborative control between them. The intelligent edge terminals are deployed at the interface between the primary and secondary sides, undertaking the collaborative work and enabling rapid, localized control. The system effectively unifies the dimensions and magnitudes of various heating data acquisition systems through the generalized power transform Zscore multi-source data transformation processing method. However, given the relatively limited computing resources of the intelligent edge terminals, the number of model parameters is restricted, necessitating channel pruning and knowledge distillation methods for model compression.

[0077] Other features and advantages will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention are realized and obtained through the structures particularly pointed out in the description and the drawings.

[0078] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0079] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0080] Figure 1 This is a flowchart of a task scheduling and control method for primary and secondary heating networks based on edge computing, according to the present invention.

[0081] Figure 2 This is a schematic diagram of the task scheduling structure of the primary and secondary heating networks of the present invention. Detailed Implementation

[0082] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0083] Example 1

[0084] Figure 1 This is a flowchart of a task scheduling and control method for primary and secondary heating networks based on edge computing, which is involved in this invention.

[0085] Figure 2 This is a schematic diagram of the task scheduling structure of the primary and secondary heating networks involved in this invention.

[0086] like Figure 1-2 As shown, this embodiment 1 provides a method for task scheduling and control of primary and secondary heating networks based on edge computing, which includes:

[0087] Intelligent edge terminals and local computing terminals are set up between the primary and secondary networks of the heating system to form an edge computing network;

[0088] The coordinated control task of the primary and secondary heating networks is divided into multiple control sub-tasks;

[0089] The execution mode of the control subtasks is set to be executed through the local computing terminal or migrated to the intelligent edge terminal for execution, forming a scheduling scheme for the coordinated execution of the primary and secondary network control tasks of the heating system by local and edge operations.

[0090] With the goal of minimizing the latency and energy consumption of control tasks executed by local computing terminals and intelligent edge terminals, a control task scheduling optimization model is established, and the optimal scheduling scheme for control tasks is obtained by solving the model.

[0091] Based on the optimal scheduling scheme of the control task, the multi-source heterogeneous data information of the primary and secondary network equipment of the heating system is processed and transformed in terms of data dimensions and magnitude through local computing terminals and intelligent edge terminals. Machine learning is used for feature extraction and model training. After the model is compressed using channel pruning and knowledge distillation methods, a control prediction model is established to carry out coordinated control of the primary and secondary networks of the heating system.

[0092] In this embodiment, the number of intelligent edge terminals and local computing terminals is at least one. The local computing terminal is equipped with a task processing component, which has lower computing resources and capabilities than the intelligent edge terminal, and is used for small-scale data processing, feature extraction, and model training. The intelligent edge terminal is equipped with an edge infrastructure unit, an edge computing service unit, and an edge application unit. The edge infrastructure unit is configured with network resources, storage resources, computing resources, and an operating system for low-level computation, storage, and network communication. The edge computing service unit is configured with a data transmission component, a data storage component, a data preprocessing component, an algorithm model component, a feature engineering component, and configuration management, operation and maintenance management, and device access management components. It is used to provide business support for the collaborative control tasks of the primary and secondary networks of the heating system, and to establish a control prediction model after processing multi-source heterogeneous data, data fusion, and feature extraction. The edge application unit uses the components in the edge computing service unit to realize the collaborative control functions of the primary and secondary networks of the heating system, as well as the management of the control prediction model and data.

[0093] In this embodiment, the collaborative control task of the primary and secondary heating networks is divided into multiple control sub-tasks, including the secondary network water supply temperature control task, the circulating pump control task, and the makeup water pump control task. Each control task also includes data acquisition, data preprocessing, feature extraction, and model training sub-tasks.

[0094] The secondary network water supply temperature control task is to establish a secondary network water supply temperature prediction and control model based on collected meteorological data, secondary side water supply temperature, secondary side return water temperature, primary side water supply temperature, primary side return water temperature, primary side temperature control valve data, and set secondary side water supply temperature data.

[0095] The circulating pump control task is to establish a predictive control model for the circulating pump based on the collected primary side water supply pressure, primary side return water pressure, secondary side water supply pressure, secondary side return water pressure, secondary side supply and return water pressure difference, circulating pump frequency converter data, and set secondary side supply and return water pressure difference data.

[0096] The water supply pump control task involves establishing a predictive control model for the water supply pump based on the collected primary side water supply pressure, primary side return water pressure, secondary side water supply pressure, secondary side return water pressure, water supply flow rate, water supply pump frequency converter data, and set secondary side return water pressure data.

[0097] In this embodiment, the execution mode of the control subtask is set to be either executed through a local computing terminal or migrated to a smart edge terminal for execution, forming a scheduling scheme for the coordinated execution of the primary and secondary network control tasks of the heating system via local and edge methods, including:

[0098] The execution mode of each control subtask is set to be either executed through a local computing terminal or migrated to a smart edge terminal for execution; the information of each control subtask includes the computational amount of task data, the data task size migrated to the smart edge terminal, and the delay time;

[0099] Based on the execution methods of different control sub-tasks, a set of scheduling schemes is formed to coordinate the execution of primary and secondary network control tasks of the heating system in a local and edge manner.

[0100] In this embodiment, the step of establishing a control task scheduling optimization model with the goal of minimizing the latency and energy consumption of control tasks executed by local computing terminals and intelligent edge terminals, and solving for the optimal scheduling scheme of control tasks, includes:

[0101] The latency and energy consumption for executing control tasks via the local computing terminal are expressed as follows:

[0102]

[0103]

[0104] Calculate the delay locally; x m，0 For the identifier of the task computed locally; l m This refers to the total amount of task data calculated locally. Data computing capabilities provided for local computing terminals; η is the local computing energy consumption; m is the energy consumption coefficient; m is the number of local computing terminals.

[0105] The latency of executing control tasks through a smart edge terminal is expressed as:

[0106]

[0107] The total latency is calculated at the edge. Delay for data upload; The time required for intelligent edge terminals to perform computations; x m,n This serves as an identifier for the subtask's computation on the intelligent edge terminal; c m The amount of data uploaded to the intelligent edge terminal; R m,n For the transmission rate of tasks migrated to intelligent edge terminals; f r The data computing power provided for intelligent edge terminals; N is the number of intelligent edge terminals;

[0108] The energy consumption for performing control tasks through intelligent edge terminals is expressed as:

[0109]

[0110] Calculate the total energy consumption at the edge; Energy consumption for uploading task data; Energy consumption for performing computations on intelligent edge terminals; p m p represents the communication power consumption when data is uploaded to the intelligent edge terminal. r The power consumption for migrating tasks to intelligent edge terminals;

[0111] The latency for the local computing terminal to execute control tasks and the latency for the intelligent edge terminal to execute control tasks are expressed as follows:

[0112]

[0113] The energy consumption of local computing terminals performing control tasks and intelligent edge terminals performing control tasks is expressed as follows:

[0114]

[0115] The objective is to minimize the latency and energy consumption of both the local computing terminal and the intelligent edge terminal performing control tasks, expressed as:

[0116]

[0117] Cost m,n To minimize the weighted total cost of delay and energy consumption; w1 and w2 are non-negative weighting factors, and w1 + w2 = 1; These represent the minimum and maximum tolerable delays, respectively. The minimum and maximum energy consumption that can be tolerated;

[0118] The constraints include: the communication power consumption when uploading data to the intelligent edge terminal is less than the maximum power consumption when the device is uploading; the computing power of the intelligent edge terminal does not exceed its own maximum allowable computing power; and the latency of both the local computing terminal and the intelligent edge terminal meets the maximum latency that the device can tolerate.

[0119] With the goal of minimizing the latency and energy consumption of control tasks executed by local computing terminals and intelligent edge terminals, and with constraints set, a control task scheduling optimization model is established.

[0120] An improved multiverse optimization algorithm is used to solve the control task scheduling optimization model to obtain the optimal scheduling scheme for the control task.

[0121] In this embodiment, the improved multiverse optimization algorithm includes: introducing a spiral to update the universe position and find the optimal individual; and introducing an adaptive compression factor to change the position of the current optimal universe.

[0122] The implementation process of the improved multiverse optimization algorithm includes:

[0123] Initialize the number of universes and the maximum number of iterations;

[0124] The universe is arranged according to its expansion rate, and a white hole is selected using a roulette wheel mechanism. Let NI(X) be the j-th parameter of the i-th universe. i ) represents the normalized expansion rate of the i-th universe; r1 is a random number between [0,1]. Let j be the parameter of the k-th universe chosen by the roulette wheel mechanism;

[0125] Update the probability of wormhole existence W EP and travel distance rate T DR , W EP min W EP max W respectively EP The minimum and maximum values; l is the current iteration number; L is the maximum iteration number; p is the development precision during the iteration process. The higher the value, the faster the local search.

[0126] Calculate the current universe expansion rate. If the current universe expansion rate is better than the current universe expansion rate, then update the current universe expansion rate; otherwise, keep the current universe expansion rate.

[0127] Perform individual position updates within the universe, searching for the optimal individual when r² < W. EP Update cosmic position:

[0128] When r2≥W EP Update cosmic position x j Let ub be the j-th parameter of the currently formed optimal universe; j lb jr1, r2, r3, and r4 are the upper and lower bounds of variable j, respectively; r2, r3, and r4 are random numbers between [0,1]; λ is the adaptive compression factor. b is a constant defined for the logarithmic spiral shape; m is a constant between [0,1].

[0129] Termination condition determination: If the termination condition is met, output the corresponding result, which is the optimal scheduling scheme for the control task; otherwise, increment the iteration count by 1 and return to select a new white hole.

[0130] It should be noted that the introduction of the spiral update method enhances the algorithm's global search capability, avoids the algorithm from getting trapped in local optima, and effectively improves the algorithm's convergence accuracy; the introduction of the adaptive compression factor enables the algorithm to dynamically adjust the position of the optimal universe during the iteration process, meeting the optimization requirements of the algorithm at different stages.

[0131] In this embodiment, the optimal scheduling scheme based on the control task processes and transforms the multi-source heterogeneous data information of the primary and secondary network equipment of the heating system in terms of data dimensions and magnitude through local computing terminals and intelligent edge terminals, including:

[0132] Based on the optimal scheduling scheme of the control task, the execution mode of the task is set. The multi-source data of the collected primary network and secondary network devices are decomposed according to the time sequence characteristics through the local computing terminal and intelligent edge terminal. Let X be the standardized transformation input, and X exists in the form of multi-dimensional data, matrix form and vector form.

[0133] Let X = (X1, X2, ..., X...) p When presented in matrix form, it is represented as:

[0134]

[0135] Set up generalized power transform BC-Zscore normalization transform schemes for different data formats:

[0136] If X exists in the form of a vector, return the transformed result vector BC_Z = [X - mean(X)] / std(X);

[0137] If X exists in the form of a matrix, the mean and standard deviation of the column vectors of X are used to standardize the data of the corresponding columns one by one, and the transformed result matrix BC_Z is returned.

[0138] If X exists in the form of a multidimensional array, its mean and standard deviation are calculated along multiple dimensions of X, and then X is standardized to return the transformed high-dimensional array BC_Z.

[0139] X = (X1, X2, ..., X...) pAfter BC-Z score normalization transformation:

[0140]

[0141] This is the average value; The standard deviation is X; after BC-Z score standardization, X = (X1, X2, ..., X...). p In each column of )

[0142] The data is iteratively processed according to the selected multi-source data processing scheme. The above process is executed repeatedly, and the data transformation and processing results are aggregated.

[0143] The multi-source data standardization transformation ends after all the data sources input to the task have undergone unified transformation in terms of dimensions and magnitudes. The task output is then saved for use in multi-source data fusion, edge data computation, or data storage.

[0144] In this embodiment, the machine learning algorithm used to establish the control prediction model is a convolutional neural network, and channel pruning and knowledge distillation methods are used to compress the convolutional neural network model.

[0145] In this embodiment, the model compression process of the convolutional neural network model using channel pruning includes:

[0146] Initialize the convolutional neural network;

[0147] The channels are sparsified under the constraint of L1 norm regularization, and the neural network is trained. The importance of a channel is determined by using a scaling parameter in the Batch Normalization layer. The smaller the scaling parameter, the less impact the pruning of that channel has on the convolutional neural network. The loss function after adding L1 regularization is expressed as: L represents the total loss predicted by the convolutional neural network model after sparsification; W represents the training weights, i.e., the first summation term corresponds to the loss l′ predicted by the model after normal training of the convolutional neural network; γ represents the scaling factor on each channel; g(γ) represents the penalty function for the scaling factor γ; λ′ represents the hyperparameter balancing the two summation terms.

[0148] Channel pruning: After training the model using L1 norm sparsification, prune the channels with smaller scaling parameters in the convolutional blocks according to the γ value of the BN layer.

[0149] After pruning, the network is fine-tuned based on the actual performance of the convolutional neural network model until a convolutional neural network model that meets the actual needs is achieved.

[0150] In this embodiment, the model compression process of the convolutional neural network model using the knowledge distillation method includes:

[0151] Train the teacher neural network model and use it as the teacher network model for knowledge distillation; use the pruned convolutional neural network as the student network model.

[0152] The loss function for knowledge distillation is defined as: L = αL soft +βL hard L soft The loss function for when the teacher model guides the student model; L hard α is the original loss function of the student model; α and β are hyperparameters balancing the two loss functions.

[0153] Training the student network: The data features extracted through feature extraction are input into the student network model and the teacher network model. The loss function of knowledge distillation is then used as the loss function for the entire training process. By adjusting α and β, the optimal training effect is obtained.

[0154] It's important to note that pruning removes unimportant redundant parameters from the network model, thus compressing the model size. Simultaneously, pruning significantly increases parameter sparsity, and this high parameter sparsity brings two advantages to deep neural networks. First, pruned sparse parameters require less disk storage space because they can be stored in Compressed Sparse Row (CSR) or Compressed Sparse Column (CSC) formats. Second, the computational cost is greatly reduced due to parameter pruning. Therefore, parameter pruning can significantly reduce the storage size and computational complexity of deep neural network models. Knowledge distillation, on the other hand, involves using softmax transformation and learning the output class distribution of the teacher model to transfer knowledge from the complex teacher network model to a simpler student model. This allows the student network model to approximate or even surpass the teacher network's performance, achieving comparable predictive results with less computational complexity.

[0155] In the several embodiments provided in this application, it should be understood that the disclosed systems and methods can also be implemented in other ways. The system embodiments described above are merely illustrative; for example, the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code, which contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0156] In addition, the functional modules in the various embodiments of the present invention can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0157] If the functionality is implemented as a software module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0158] Based on the above-described preferred embodiments of the present invention, and through the foregoing description, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.

Claims

1. An edge computing-based heat supply primary network and secondary network task scheduling and control method, characterized in that, It comprises: An intelligent edge terminal and a local computing terminal are arranged between a primary network and a secondary network of a heating system to form an edge computing network; A control task of the primary network and the secondary network of the heating system is divided into multiple control subtasks; An execution mode of the control subtasks is set to be executed by the local computing terminal or migrated to the intelligent edge terminal for execution, forming a scheduling scheme for executing the control task of the primary network and the secondary network of the heating system in a local and edge collaborative manner; A control task scheduling optimization model is established to minimize the delay and energy consumption of the local computing terminal and the intelligent edge terminal in executing the control task, and an optimal scheduling scheme of the control task is obtained by solving the model, including: The delay and energy consumption of the local computing terminal in executing the control task are respectively represented as: ； ； a local computing delay; an identification of a task computed locally; a total amount of task data for local computing; a data computing capability provided by the local computing terminal; a local computing energy consumption; an energy consumption coefficient; andm is the number of local computing terminals. The delay of the intelligent edge terminal in executing the control task is represented as: ； total edge computing latency; data upload latency; time for intelligent edge terminal to perform computation; identification of subtasks computed at intelligent edge terminal; amount of data uploaded to intelligent edge terminal; transfer rate of tasks migrated onto intelligent edge terminal; data computation capability provided by intelligent edge terminal; number of intelligent edge terminals; The energy consumption of the intelligent edge terminal in executing the control task is represented as: ； total energy consumption for edge computing; energy consumption for task data upload; energy consumption for performing computation by intelligent edge terminal; energy consumption for communication when data is uploaded to intelligent edge terminal; energy consumption for task migration to intelligent edge terminal; The delay of the local computing terminal and the intelligent edge terminal in executing the control task is represented as: ； The energy consumption of the local computing terminal and the intelligent edge terminal in executing the control task is represented as: ； The delay and energy consumption of the local computing terminal and the intelligent edge terminal in executing the control task are minimized, represented as: ； to minimize a weighted total of delay and energy consumption; , are non-negative weighting factors, and ; , are a minimum delay and a maximum delay that can be tolerated, respectively; , are a minimum energy consumption and a maximum energy consumption that can be tolerated, respectively. The constraint conditions include: the communication energy consumption power when uploading data to the intelligent edge terminal is less than the maximum energy consumption power when uploading; the computing capacity of the intelligent edge terminal does not exceed the maximum computing capacity allowed by itself; the delay of the local computing terminal and the intelligent edge terminal in computing satisfies the maximum delay that the device can tolerate; The control task scheduling optimization model is established to minimize the delay and energy consumption of the local computing terminal and the intelligent edge terminal in executing the control task, and the constraint conditions are set; An improved multi-universe optimization algorithm is used to solve the control task scheduling optimization model to obtain the optimal scheduling scheme of the control task; Based on the optimal scheduling scheme of the control task, the multi-source heterogeneous data information of the primary network and the secondary network devices of the heating system is processed and transformed in data dimension and order by the local computing terminal and the intelligent edge terminal, feature extraction and model training are performed by machine learning, and a control prediction model is established after model compression processing by channel reduction and knowledge distillation method, and the heating system is controlled in a collaborative manner.

2. The method of claim 1, wherein, The number of the intelligent edge terminal and the local computing terminal is at least one; the local computing terminal is provided with a task processing component, which is lower than the computing capacity of the intelligent edge terminal, and is used for performing a small amount of data processing, feature extraction and model training; the intelligent edge terminal is provided with an edge infrastructure unit, an edge computing service unit and an edge application unit; the edge infrastructure unit is configured with network resources, storage resources, computing resources and an operating system, and is used for performing bottom-layer operation, storage and network communication; the edge computing service unit is configured with a data transmission component, a data storage component, a data preprocessing component, an algorithm model component, a feature engineering component, and a configuration management, operation and maintenance management and device access management component, and is used for providing business support for the heating system primary network and secondary network collaborative control task, performing multi-source heterogeneous data processing, data fusion and feature extraction, and establishing a control prediction model; the edge application unit realizes the heating system primary network and secondary network collaborative control function, and control prediction model management and data management by using the components in the edge computing service unit.

3. The method of claim 1, wherein the primary and secondary network task scheduling and control method is characterized by, The heating system primary network and secondary network collaborative control task is divided into multiple control subtasks, including a secondary network water supply temperature control task, a circulating pump control task and a water supply pump control task, each of which further includes data acquisition, data preprocessing, feature extraction and model training subtasks; The secondary network water supply temperature control task is to establish a secondary network water supply temperature prediction control model according to collected meteorological data, secondary side water supply temperature, secondary side return water temperature, primary side water supply temperature, primary side return water temperature, primary side temperature control valve data and set secondary side water supply temperature data; The circulating pump control task is to establish a circulating pump prediction control model according to collected primary side water supply pressure, primary side return water pressure, secondary side water supply pressure, secondary side return water pressure, secondary side water supply and return water pressure difference and circulating pump frequency converter data and set secondary side water supply and return water pressure difference data; The water supply pump control task is to establish a water supply pump prediction control model according to collected primary side water supply pressure, primary side return water pressure, secondary side water supply pressure, secondary side return water pressure, water supply flow, water supply pump frequency converter data and set secondary side return water pressure data.

4. The method of claim 1, wherein, The execution mode of the control subtask is set to be executed by the local computing terminal or migrated to the intelligent edge terminal for execution, forming a scheduling scheme for the local and edge collaborative execution of the heating system primary network and secondary network control task, including: The execution mode of each control subtask is set to be executed by the local computing terminal or migrated to the intelligent edge terminal for execution; the information of each control subtask includes the calculation amount of task data, the data task size and the delay time of migration to the intelligent edge terminal; According to the execution mode of different control subtasks, a scheduling scheme set for the local and edge collaborative execution of the heating system primary network and secondary network control task is formed.

5. The method of claim 1, wherein, The improved multi-universe optimization algorithm comprises: introducing a spiral update universe position and searching for an optimal individual; and introducing an adaptive compression factor to change the position of a current optimal universe; The implementation process of the improved multi-universe optimization algorithm comprises: initializing the number of universes and the maximum number of iterations; arrange the universe according to the rate of cosmic expansion and select a white hole through a roulette mechanism, , is the jth parameter of the ith universe; is the normalized expansion rate of the ith universe; is a random number between [0, 1]; is the jth parameter of the kth universe selected by the roulette mechanism; Updating wormhole existence probability And travel distance rate , , ; , Respectively The minimum and maximum values; Is the current iteration number; Is the maximum iteration number; Is the development precision in the iteration process, the higher the value, the faster the local search; calculating a current universe expansion rate, updating the current universe expansion rate if the universe expansion rate is better than the current universe expansion rate, or keeping the current universe otherwise; Perform individual position update of the universe, find the optimal individual, when , update universe position: ; When , update the cosmic position ; is the jth parameter of the best universe formed so far; , are the upper and lower limits of the j variable, respectively; , , is a random number between [0, 1]; is an adaptive compression factor, ; is a constant defining the logarithmic spiral shape; is a constant between [0, 1]; termination condition determination: if the termination condition is met, the corresponding result, i.e., the optimal scheduling scheme of the control task, is output; otherwise, the number of iterations is increased by 1, and a white hole is reselected.

6. The heating primary and secondary network task scheduling and control method according to claim 1, characterized in that, Based on the optimal scheduling scheme of the control task, the multi-source heterogeneous data information of the heating system primary network and secondary network equipment is processed and transformed in terms of data dimension and order by a local computing terminal and an intelligent edge terminal, comprising: Based on the optimal scheduling scheme of the control task, the execution mode of the task is set, and the multi-source data collected from the primary network and secondary network equipment are decomposed according to the timing characteristics by the local computing terminal and the intelligent edge terminal, where X is a standardized transformation input, and X exists in the form of multi-dimensional data, a matrix and a vector; Let us set In matrix form, this is written as: ； A generalized power transformation BC-Zscore standardization transformation scheme of different data formats is set: If X is in vector form, return the transformed result vector ; If X exists in the form of a matrix, the mean and standard deviation of the column vector of X are used to perform data standardization processing on the corresponding column one by one, and a transformed result matrix BC_Z is returned; If X exists in the form of a multi-dimensional array, the mean and standard deviation of X are solved along multiple dimensions of X, and then X is subjected to data standardization processing, and a transformed high-dimensional array BC_Z is returned; After BC-Zscore normalization transformation: ； ; ; ; is the average value; is the standard deviation; after BC-Zscore standardization transformation in each column of , ; According to the selected multi-source data processing scheme, data iteration processing is performed, and the above process is repeatedly executed to gather data transformation processing results; Until the data sources input by the task are subjected to unified transformation in terms of dimension and order, the task output is saved for multi-source data fusion, edge data calculation or data storage, and the multi-source data standardization transformation is completed.

7. The method of claim 1, wherein, The machine learning algorithm used to establish the control prediction model is a convolutional neural network, and a channel pruning and knowledge distillation method is used to perform model compression processing on the convolutional neural network model.

8. The method of claim 7, wherein, The channel pruning method for model compression processing of the convolutional neural network model comprises: initializing the convolutional neural network; The channels are sparsified under the constraint of L1 norm regularization, and the neural network is trained. The importance of a channel is determined by using a scaling parameter in the Batch Normalization layer. The smaller the scaling parameter, the less impact the pruning of that channel has on the convolutional neural network. The loss function after adding L1 regularization is expressed as: ; The total loss predicted by the convolutional neural network model after sparsification; The training weights, specifically the first summation term, correspond to the loss predicted by the model after normal training of the convolutional neural network. ; Scaling factor on each channel; For scaling factor The penalty function; To balance the hyperparameters of the two summation terms; Channel pruning: train the model after sparsification using L1 norm, prune the channels with smaller scaling parameters in the convolution block according to the BN layer's values. After pruning is completed, the network is fine-tuned according to the actual effect of the convolutional neural network model until the convolutional neural network model that meets the actual needs is obtained.

9. The heating primary and secondary network task scheduling and control method according to claim 7, characterized in that, The knowledge distillation method for model compression processing of the convolutional neural network model comprises: training a teacher neural network model, and taking it as a teacher network model for knowledge distillation; and taking the convolutional neural network after pruning as a student network model; The loss function of knowledge distillation is set, denoted as: ; The loss function when the teacher model guides the student model; The original loss function of the student model; , The hyperparameter for balancing the two loss functions; Training the student network: input the data features extracted by the feature extraction into the student network model and the teacher network model, and take the knowledge distillation loss function as the loss function of the entire training, and obtain the optimal training effect by adjusting , .

Images