Flexible production line equipment cooperative energy-saving control system based on random ADMM
By employing the stochastic ADMM algorithm and federated learning technology, globally optimal energy-saving control of flexible production line equipment was achieved, solving the problems of information silos and data privacy in traditional systems, and improving the energy efficiency and system reliability of the production line.
Patent Information
- Application Number
- CN202511831063.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-06
- Publication Date
- 2026-03-17
AI Technical Summary
Traditional flexible production line equipment's energy-saving control system cannot effectively sense and adapt to dynamic changes in production load and environment, resulting in information silos, data privacy leaks, and the risk that mixed integer programming is prone to getting trapped in local optima and is difficult to achieve global high-efficiency energy saving under real-time requirements.
A flexible production line equipment collaborative energy-saving control system based on stochastic ADMM is adopted. Through multi-dimensional energy consumption sensing modules, data transmission modules, central control systems and dynamic correction modules, combined with cross-industry federated learning and non-convex optimization models, the system achieves global optimization of equipment start-up and shutdown strategies and privacy protection.
It achieves precise energy consumption matching of production line equipment, avoids energy waste, improves overall energy efficiency, reduces system maintenance costs and failure rate, and ensures data privacy protection and the stability of overall optimization.
Smart Images
Figure CN121680183A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of industrial automation and intelligent manufacturing, and particularly relates to a flexible production line equipment collaborative energy-saving control system based on a random alternating direction method of multipliers (ADMM). BACKGROUND
[0002] Under the background of Industry 4.0 and intelligent manufacturing, the flexible production mode for multi-variety and small batch has become mainstream. The production line equipment under this mode is frequently started and stopped, and the load fluctuates greatly, which puts forward very high requirements for energy efficiency management. The traditional production line energy-saving control mostly adopts independent control strategies based on fixed thresholds or simple timing rules, which has the following significant defects: first, it cannot perceive and adapt to the dynamic changes of production load and environment, often leading to energy waste during light load or idle running; second, each device or production line is independently optimized, forming an “information island”, which cannot share energy-saving experience and strategies, and the overall energy efficiency is limited; finally, the device collaborative scheduling is essentially a mixed integer non-convex optimization problem containing discrete start-stop variables and continuous power variables, and the traditional optimization algorithm has high complexity and is easy to fall into local optimum, so it is difficult to obtain a globally efficient energy-saving strategy under real-time requirements.
[0003] In recent years, the Internet of Things, big data and artificial intelligence technologies have brought new ideas for production line energy saving. For example, a centralized optimization model is used for scheduling, but it faces the risk of data privacy leakage and communication bandwidth pressure; the federated learning technology can realize collaborative learning on the premise of protecting data privacy, but when it is directly applied to dynamic and high real-time device control scenarios, there are challenges in model convergence speed and strategy output stability. The standard ADMM algorithm is suitable for distributed optimization, but when dealing with the non-convex problem model inherent in production line control, it also has the problems of uncertain convergence and easy to fall into local optimum.
[0004] Therefore, there is an urgent need for a flexible production line collaborative energy-saving control system that can balance data privacy protection, global optimization solving ability, high real-time response and strong working condition adaptability. SUMMARY
[0005] The purpose of the present application is to provide a flexible production line equipment collaborative energy-saving control system based on random ADMM, to solve the problems of multi-source data heterogeneity difficult to collaborate, mixed integer programming non-convex optimization easy to fall into local optimum, cross-line data sharing privacy and efficiency difficult to balance, poor energy-saving effect under random disturbance, and high system maintenance and redundant energy consumption in the traditional energy-saving production line and equipment start-stop control.
[0006] To achieve the above object, the application adopts the following technical scheme: a flexible production line equipment collaborative energy-saving control system based on random ADMM, comprising:
[0007] A multi-dimensional energy consumption perception module is configured to collect time sequence power data, equipment state transition data and environmental parameter data of the equipment in the production line.
[0008] A data transmission module is connected to the multi-dimensional energy consumption perception module and configured to transmit the time sequence power data, equipment state transition data and environmental parameter data to the central control system.
[0009] The central control system is connected to the data transmission module and internally integrated with a cross-industry federated learning module and a non-convex optimization model. The cross-industry federated learning module is configured to perform federated learning and collaborative optimization of multi-source energy consumption data based on a privacy protection mechanism. The non-convex optimization model is configured to use a random ADMM algorithm to perform non-convex optimization on the equipment start-stop strategy to generate a globally optimal equipment start-stop strategy.
[0010] A dynamic correction module is connected to the central control system and configured to receive the collaborative optimization data output by the cross-industry federated learning module and the globally optimal equipment start-stop strategy generated by the non-convex optimization model, perform fusion analysis and optimization calculation, and then feed back the optimized final control strategy to the production line actuator to regulate and control the start-stop state of the production line equipment.
[0011] Further, the multi-dimensional energy consumption perception module includes an intelligent electric energy metering device configured to collect system instantaneous power data and construct a power time sequence matrix P, which has the following expression:
[0012]
[0013] In the formula, P represents a power time sequence matrix containing instantaneous power data of channels, each row of the power time sequence matrix corresponds to a power time sequence of a channel, and the dimension of the power time sequence matrix is ; P represents the instantaneous power value of the th channel at time ; ;
[0014] The multi-dimensional energy consumption perception module reads the device switch state quantity in the programmable logic controller (PLC) register and generates a device running state transition matrix, which has the following expression:
[0015]
[0016] In the formula, P represents the transition from state to state The probability or weight; Indicates the system in time In state ; Indicates the system at the next moment. Transition to state ; Indicates that the current state is known. Next, the next state is The probability of;
[0017] The multi-dimensional energy consumption sensing module includes a temperature sensor and a humidity sensor. The temperature sensor and humidity sensor collect real-time data on changes in ambient temperature and humidity to correct the parameters of the device's thermodynamic model. The expression for this is as follows:
[0018]
[0019]
[0020] In the formula, Indicates time The temperature value at that time; This indicates the calibration reference temperature of the temperature sensor; Indicates the temperature sensitivity coefficient; This represents the raw signal output by the temperature sensor. Indicates time The relative humidity value at that time; Indicates the humidity sensitivity coefficient; This represents the raw signal output by the humidity sensor. Indicates the humidity calibration offset;
[0021] The thermodynamic model of the equipment is expressed as follows:
[0022]
[0023] In the formula, This indicates the system's power loss or energy loss; Indicated with temperature The relevant loss coefficient; This indicates that power loss is proportional to the square of the power.
[0024] Furthermore, the cross-industry federated learning module includes a federated learning submodule, which performs the following process:
[0025] (1) Server initializes global model And distribute it to all clients;
[0026] (2) Each client Using local data Train the model and obtain the model update parameters. ;
[0027] (3) Client Upload only encrypted model update parameters Data is shared to the server, not locally. ;
[0028] (4) The server aggregates updates from all clients and generates a new global model. ;
[0029] (5) The server will update the global model Distribute to all clients and continue iterating until convergence.
[0030] Furthermore, the cross-industry federated learning module includes a knowledge accumulation sub-module, which adopts a teacher-student model framework, wherein the teacher model generates soft labels for the input data, the expression of which is as follows:
[0031]
[0032] In the formula, Indicates the first The probability of soft labels for a class; The teacher model represents the category. The original logit output; Indicates temperature parameter; Represents the natural exponential function;
[0033] The student model needs to fit both the true labels and the teacher model's soft labels. The total loss of the student model is a weighted sum of the hard and soft losses, expressed as follows:
[0034]
[0035] In the formula, This represents hard loss, specifically cross-entropy loss; This represents the soft loss, specifically the distillation loss of the student model on the soft labels of the teacher model. This represents the total loss function of the student model; These are the weighting coefficients.
[0036] Furthermore, the cross-industry federated learning module includes a collaborative optimization sub-module, which forces the local model to align with the global model through regularization terms, thereby integrating the local models uploaded by the client into a global model; its optimization objective is expressed as follows:
[0037]
[0038] In the formula, Indicates the first Parameters of a client-side local model; Indicates the first Loss function for each client-side local model; These are hyperparameters used to balance the weights of the loss function term and the regularization term; Used to measure the Client-side local model parameters With global model parameters The degree of difference between them;
[0039] Meanwhile, the basic optimization objective of the local model is to minimize the loss of local data:
[0040]
[0041] In the formula, Indicates the local model parameters Perform minimization optimization; Indicates to the client Local dataset Summing all samples in the dataset; Represents the loss function; Indicates the first Input features of each sample; Indicates the first The true label of each sample; Indicates the client Local model parameters; Indicates the client Private dataset.
[0042] Furthermore, the objective function of the non-convex optimization model is:
[0043]
[0044] In the formula, A binary variable representing a pattern Is it activated? For pattern The fixed activation cost, and Binding; For pattern Main resource input; Sigmoid function To simulate nonlinear activation costs; For pattern Resources The amount of input; For resources In mode The unit cost coefficient in; For pattern With pattern The interaction penalty coefficient; The product term for the main resource inputs under different modes;
[0045] The non-convex optimization model has a non-convex constraint on pattern activation logic: the difference between the activation identifier of each pattern and the corresponding factor and the subtraction critical value must satisfy the following non-negative condition:
[0046]
[0047] In the formula, This is the critical value;
[0048] The non-convex optimization model is optimized using the stochastic ADMM algorithm. The optimized model is as follows:
[0049]
[0050] In the formula, For equipment At any moment The actual power consumption; For equipment Fixed operating cost coefficient; It is a binary variable representing the device. At any moment Is it running? Temperature-dependent power loss; For equipment At any moment Temperature; This is the penalty factor for power loss; For the first The parameter vector of a local model; This is the parameter vector of the global model; These are the weighting coefficients for the federal regularization terms; The power matrix has dimensions. ; for The low-rank decomposition introduces randomness; The weighting coefficients for the ADMM penalty term;
[0051] The non-convex optimization model has constraints on production timing: the cumulative production volume within a certain time range must meet the requirements corresponding to the time interval and equipment operation schedule, and at the same time, it involves the probability of transitions between different states;
[0052]
[0053]
[0054] In the formula, It is a binary variable representing the device. At any moment Is there a production plan? For equipment The minimum continuous running time; These are the elements of the state transition probability matrix; Indicates from state Transferred to The probability of.
[0055] Furthermore, the solution steps of the stochastic ADMM algorithm are as follows:
[0056] (1) Split the variables into local problems and global problems; where the expression for the local problem is:
[0057]
[0058] The expression for the global problem is:
[0059]
[0060] In the formula, For equipment At any moment The actual power; For equipment Fixed operating cost coefficient; It is a binary variable representing the device. At any moment Is it running? For global variables in the first... In the next iteration, The consistency constraint target value; For the dual variable of ADMM; For ADMM penalty parameters; This is the low-rank global representation of the power matrix; This is the original power matrix; is the basis matrix for low-rank decomposition; The regularization coefficient is used.
[0061] (2) Stochastic gradient update: from the power matrix The random sampling submatrix P Ω Ω is a random index set, which updates the global variable. The expression is as follows:
[0062]
[0063] In the formula, For the first The low-rank global table of the power matrix obtained after the next iteration update; Indicates the first The power matrix updated in the next iteration; Indicates the first ADMM dual variable matrix at the next iteration; From the power matrix A submatrix randomly sampled from the middle; To perform rank-r truncated singular value decomposition on the matrix; The pre-defined low-rank dimension; This is the weight adjustment factor during random updates;
[0064] (3) Inject federated knowledge, the expression of which is as follows:
[0065]
[0066] In the formula, For the first Model parameters for each local device; Global model parameters aggregated for the central server; For local devices The gradient of the loss function; The coefficients of the federal regularization term; This is the learning rate.
[0067] Furthermore, the dynamic correction module generates the final control strategy through a hybrid weight dynamic fusion mechanism, and the fusion formula is as follows:
[0068]
[0069] In the formula, This is the fused dynamic weight vector; To collaboratively optimize weights for cross-industry federated learning modules; These are the optimization weights for a non-convex optimization model; The strategy fusion coefficient;
[0070] And it is optimized using the stochastic ADMM-Proximal algorithm, which
[0071]
[0072] In the formula, The first sequence Decision variables for the next iteration ; The first sequence Decision variables for the next iteration ; The first sequence The Lagrange multiplier vector for the next iteration; For the proximal term coefficients, control Weights of quadratic regularization terms in the proximal operator during variable updates; For the proximal term coefficients, targeting Variable updates control the weights of the weighted quadratic terms in their proximal operators; The coefficient matrix represents the collaborative constraints of equipment in a flexible production line. For constant terms in the constraints; To augment the Lagrange penalty coefficient; The intensity of the reference trajectory tracking; These are Lagrange multiplier vectors; These are the coefficients of the proximal terms; These are the coefficients of the proximal terms; This is a weighted diagonal matrix; This represents the number of iterations.
[0073] Then, the control inputs are synthesized to form the final control output:
[0074]
[0075] In the formula, This indicates the decision variables after iteration. The optimal decision variables obtained upon convergence; This indicates the decision variables after iteration. The optimal decision variables obtained upon convergence; For the final control output; For discrete control weights; This is the feedback gain matrix; This represents the change in weight.
[0076] Furthermore, the dynamic correction module feeds back the optimized final control strategy to the production line actuators, effectively reducing the energy consumption of the industrial production line by precisely controlling the start-up and shutdown states of the equipment. This process comprehensively considers both operating energy consumption and start-up and shutdown energy consumption, resulting in a system energy consumption model that includes:
[0077] Energy consumption model for industrial production line operation:
[0078]
[0079] In the formula, This is a double summation symbol; This represents the total number of devices in the system. The total number of time intervals to be considered; For the first The operating power of a device in the t-th time interval; For the first The device in the Operating status indicator variables for each time interval; For the first Energy loss coefficient of each device; For the first The device in the Energy loss over a time interval;
[0080] Energy consumption model for starting and stopping industrial production lines:
[0081]
[0082] In the formula, For the first Energy consumption coefficient for starting and stopping individual equipment; For the first The device in the Operating status indicator variables for each time interval; For the first The device in the Operating status indicator variables for each time interval.
[0083] Compared with the prior art, the present invention has the following beneficial effects:
[0084] 1. Compared with traditional energy-saving systems, this invention constructs a unified optimization model through multi-source data perception, and dynamically adjusts start-up and shutdown strategies in real time by integrating equipment power, operating status and environmental parameters. This accurately matches production needs with energy consumption balance, avoiding excessive start-up and shutdown or energy waste caused by traditional systems relying on fixed thresholds or a single data source, resulting in a significant improvement in energy efficiency.
[0085] 2. Compared with traditional energy-saving systems, this invention is based on the stochastic ADMM algorithm combined with low-rank decomposition and random sampling, which effectively solves the non-convex optimization problem of mixed integer programming, ensures the stability of the global optimal solution, overcomes the defects of traditional methods that are prone to getting trapped in local optima or have high computational complexity, and achieves the accuracy of energy consumption control strategies in complex production line scenarios.
[0086] 3. Compared with traditional energy-saving systems, this invention designs a federated learning module, which achieves cross-production line privacy-preserving collaborative optimization through gradient deviation correction and knowledge accumulation. It shares optimization experience while protecting the privacy of production data, breaks through the information limitations of traditional systems, promotes the coordinated improvement of energy efficiency throughout the plant, and avoids resource waste caused by repeated trial and error.
[0087] 4. Compared with traditional energy-saving systems, this invention introduces ADMM-Proximal optimization and hybrid weight fusion mechanism to adapt to grid fluctuations and equipment state disturbances. Combined with modular integration to simplify line design, it reduces failure rate and redundant energy consumption, and balances energy efficiency improvement and system reliability. The overall maintenance cost and energy consumption fluctuation range are significantly lower than traditional solutions. Attached Figure Description
[0088] Figure 1 This is a schematic diagram of the structure of the flexible production line equipment collaborative energy-saving control system provided in an embodiment of the present invention.
[0089] Figure 2 This is a flowchart of the collaborative energy-saving control system for flexible production line equipment provided in an embodiment of the present invention.
[0090] Figure 3 This is an optimization flowchart of the random ADMM algorithm in the system according to an embodiment of the present invention.
[0091] Figure 4 This is a schematic diagram comparing the energy consumption prediction error of the system of this invention with that of the traditional PID control strategy.
[0092] Figure 5 This is a schematic diagram comparing the response speed of the system of this invention with that of the traditional PID control strategy.
[0093] Figure 6 This is a schematic diagram comparing the cumulative energy consumption of the system of this invention with that of the traditional PID control strategy.
[0094] Figure 7 This is a schematic diagram comparing the effects of the traditional start-stop strategy and the optimized strategy of this invention. Detailed Implementation
[0095] The present invention will be further described below with reference to the accompanying drawings and embodiments.
[0096] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.
[0097] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.
[0098] like Figures 1-2 As shown, this embodiment provides a flexible production line equipment collaborative energy-saving control system based on stochastic ADMM, including:
[0099] The multi-dimensional energy consumption sensing module is used to collect time-series power data, equipment state transition data and environmental parameter data of equipment in the production line;
[0100] The data transmission module connects to the multi-dimensional energy consumption sensing module and is used to transmit time-series power data, equipment state transition data and environmental parameter data to the central control system.
[0101] The central control system connects to the data transmission module and integrates a cross-industry federated learning module and a non-convex optimization model. The cross-industry federated learning module is used for federated learning and collaborative optimization of multi-source energy consumption data based on a privacy protection mechanism. The non-convex optimization model is used to solve the non-convex optimization of the equipment start-up and shutdown strategy using the stochastic ADMM algorithm to generate the globally optimal equipment start-up and shutdown strategy.
[0102] The dynamic correction module, connected to the central control system, receives collaborative optimization data output from the cross-industry federated learning module and the globally optimal equipment start-up and shutdown strategy generated by the non-convex optimization model. It then performs fusion analysis and optimization calculations, and feeds back the optimized final control strategy to the production line actuators to regulate the start-up and shutdown status of the production line equipment.
[0103] The multi-dimensional energy consumption sensing module includes an intelligent power metering device, a temperature sensor, and a humidity sensor.
[0104] The intelligent power metering device collects instantaneous power data from the system and constructs a power time-series matrix P to better monitor power changes. Its expression is as follows:
[0105]
[0106] In the formula, Indicates a containing A power time-series matrix of instantaneous power data for each channel, where each row of the power time-series matrix corresponds to the power time series of one channel, and its dimension is... ; Indicates the first Each channel in time The instantaneous power value, .
[0107] The multi-dimensional energy consumption sensing module reads the device switching status values from the programmable logic controller (PLC) registers and generates a device operation state transition matrix, the expression of which is as follows:
[0108]
[0109] In the formula, Indicates from state Transition to state The probability or weight; Indicates the system in time In state ; Indicates the system at the next moment. Transition to state ; Indicates that the current state is known. Next, the next state is The probability of.
[0110] The temperature and humidity sensors collect real-time data on changes in ambient temperature and humidity to correct the parameters of the equipment's thermodynamic model. Their expressions are as follows:
[0111]
[0112]
[0113] In the formula, Indicates time The temperature value at that time; This indicates the calibration reference temperature of the temperature sensor; Indicates the temperature sensitivity coefficient; This represents the raw signal output by the temperature sensor. Indicates time The relative humidity value at that time; Indicates the humidity sensitivity coefficient; This represents the raw signal output by the humidity sensor. Indicates the humidity calibration offset;
[0114] To improve the equipment's adaptability to different temperatures, a thermodynamic model of the equipment is constructed, with the following expression:
[0115]
[0116] In the formula, This indicates the system's power loss or energy loss; Indicated with temperature The relevant loss coefficient; This indicates that power loss is proportional to the square of the power.
[0117] The cross-industry federated learning module includes a federated learning submodule, a knowledge accumulation submodule, and a collaborative optimization submodule.
[0118] The federated learning submodule performs the following process:
[0119] (1) Server initializes global model And distribute it to all clients;
[0120] (2) Each client Using local data Train the model and obtain the model update parameters. ;
[0121] (3) Client Upload only encrypted model update parameters Data is shared to the server, not locally. ;
[0122] (4) The server aggregates updates from all clients and generates a new global model. ;
[0123] (5) The server will update the global model Distribute to all clients and continue iterating until convergence.
[0124] The cross-industry federated learning module includes a knowledge accumulation sub-module, which adopts a teacher-student model framework. The teacher model generates soft labels for the input data, with the following expression:
[0125]
[0126] In the formula, Indicates the first The probability of soft labels for a class; The teacher model represents the category. The original logit output; Indicates temperature parameter; Represents the natural exponential function;
[0127] To accurately capture knowledge, the student model needs to fit both the true labels and the teacher model's soft labels. The total loss of the student model is a weighted sum of the hard and soft losses, expressed as follows:
[0128]
[0129] In the formula, This represents hard loss, specifically cross-entropy loss; This represents the soft loss, specifically the distillation loss of the student model on the soft labels of the teacher model. This represents the total loss function of the student model; These are the weighting coefficients.
[0130] The cross-industry federated learning module includes a collaborative optimization sub-module, which uses regularization terms to force the local model to align with the global model, thereby integrating the local models uploaded by the client into the global model; its optimization objective is expressed as follows:
[0131]
[0132] In the formula, Indicates the first Parameters of a client-side local model; Indicates the first Loss function for each client-side local model; These are hyperparameters used to balance the weights of the loss function term and the regularization term; Used to measure the Client-side local model parameters With global model parameters The degree of difference between them.
[0133] Meanwhile, the basic optimization objective of the local model is to minimize the loss of local data:
[0134]
[0135] In the formula, Indicates the local model parameters Perform minimization optimization; Indicates to the client Local dataset Summing all samples in the dataset; Represents the loss function; Indicates the first Input features of each sample; Indicates the first The true label of each sample; Indicates the client Local model parameters; Indicates the client Private dataset.
[0136] To address the nonconvexity of the equipment start-up and shutdown strategy, this system constructs a nonconvex optimization model. The objective function of the nonconvex optimization model is:
[0137]
[0138] In the formula, A binary variable representing a pattern Is it activated? For pattern The fixed activation cost, and Binding; For pattern Main resource input; Sigmoid function To simulate nonlinear activation costs; For pattern Resources The amount of input; For resources In mode The unit cost coefficient in; For pattern With pattern The interaction penalty coefficient; The product of the main resource inputs for different modes.
[0139] The non-convex optimization model has a non-convex constraint on pattern activation logic: the difference between the activation identifier of each pattern and the corresponding factor and the subtraction critical value must satisfy the following non-negative condition:
[0140]
[0141] In the formula, This is the critical value.
[0142] Non-convex optimization models suffer from three main drawbacks in industrial energy-saving control: the solution space contains multiple local extrema, making the algorithm prone to getting trapped in suboptimal solutions; computational complexity is high, and real-time performance deteriorates significantly when the number of devices is large; and algorithm stability is poor, with the choice of penalty coefficient directly affecting convergence, and sudden changes in device state leading to numerical instability. To address these issues, the stochastic ADMM algorithm is used to optimize the non-convex optimization model, and the optimized model is as follows:
[0143]
[0144] In the formula, For equipment At any moment The actual power consumption; For equipment Fixed operating cost coefficient; It is a binary variable representing the device. At any moment Is it running? Temperature-dependent power loss; For equipment At any moment Temperature; This is the penalty factor for power loss; For the first The parameter vector of a local model; This is the parameter vector of the global model; These are the weighting coefficients for the federal regularization terms; The power matrix has dimensions. ; for The low-rank decomposition introduces randomness; The weighting coefficients for the ADMM penalty term.
[0145] The non-convex optimization model also has constraints on production timing: the cumulative production volume within a certain time range must meet the requirements corresponding to the time interval and equipment operation schedule, and at the same time, it involves the probability of transitions between different states;
[0146]
[0147]
[0148] In the formula, It is a binary variable representing the device. At any moment Is there a production plan? For equipment The minimum continuous running time; These are the elements of the state transition probability matrix; Indicates from state Transferred to The probability of.
[0149] Non-convex hybrid models contain 0-1 variables and continuous power, making them difficult to solve in real time using traditional algorithms. The ADMM algorithm decouples them into device subproblems and global low-rank projection, performs SVD on randomly sampled data, and completes a single iteration in milliseconds; the dual variable dynamically corrects the solution trajectory to ensure that the optimal start-stop strategy is output within the cycle.
[0150] The optimization process of the stochastic ADMM algorithm is as follows: Figure 3 As shown.
[0151] The solution steps of the random ADMM algorithm are as follows:
[0152] (1) Split the variables into local problems and global problems; where the expression for the local problem is:
[0153]
[0154] The expression for the global problem is:
[0155]
[0156] In the formula, For equipment At any moment The actual power; For equipment Fixed operating cost coefficient; It is a binary variable representing the device. At any moment Is it running? For global variables in the first... In the next iteration, The consistency constraint target value; For the dual variable of ADMM; For ADMM penalty parameters; This is the low-rank global representation of the power matrix; This is the original power matrix; is the basis matrix for low-rank decomposition; This is the regularization coefficient, which balances low-rank error with ADMM consistency.
[0157] (2) Stochastic gradient update: from the power matrix The random sampling submatrix P Ω (Ω is a random index set), thereby updating the global variable, the expression of which is as follows:
[0158]
[0159] In the formula, For the first The low-rank global table of the power matrix obtained after the next iteration update; Indicates the first The power matrix updated in the next iteration; Indicates the first ADMM dual variable matrix at the next iteration; From the power matrix A submatrix randomly sampled from the middle; To perform rank-r truncated singular value decomposition on the matrix; The pre-defined low-rank dimension; This is the weight adjustment factor during random updates.
[0160] (3) Inject federated knowledge, the expression of which is as follows:
[0161]
[0162] In the formula, For the first Model parameters for each local device; Global model parameters aggregated for the central server; For local devices The gradient of the loss function; The coefficients of the federal regularization term; This is the learning rate.
[0163] The dynamic correction module generates the final control strategy through a hybrid weight dynamic fusion mechanism, and the fusion formula is as follows:
[0164]
[0165] In the formula, This is the fused dynamic weight vector; To collaboratively optimize weights for cross-industry federated learning modules; These are the optimization weights for a non-convex optimization model; This represents the strategy fusion coefficient.
[0166] And it is optimized using the stochastic ADMM-Proximal algorithm, which
[0167]
[0168] In the formula, The first sequence Decision variables for the next iteration ; The first sequence Decision variables for the next iteration ; The first sequence The Lagrange multiplier vector for the next iteration; For the proximal term coefficients, control Weights of quadratic regularization terms in the proximal operator during variable updates; For the proximal term coefficients, targeting Variable updates control the weights of the weighted quadratic terms in their proximal operators; The coefficient matrix represents the collaborative constraints of equipment in a flexible production line. The constant term in the constraint is the "target value" of the constraint equation; To augment the Lagrange penalty coefficient; The intensity of the reference trajectory tracking; These are Lagrange multiplier vectors; These are the coefficients of the proximal terms; These are the coefficients of the proximal terms; This is a weighted diagonal matrix; This represents the number of iterations.
[0169] Then, the control inputs are synthesized to form the final control output:
[0170]
[0171] In the formula, This indicates the decision variables after iteration. The optimal decision variables obtained upon convergence; This indicates the decision variables after iteration. The optimal decision variables obtained upon convergence; For the final control output; For discrete control weights; This is the feedback gain matrix; This represents the change in weight.
[0172] The dynamic correction module feeds back the optimized final control strategy to the production line actuators, effectively reducing the energy consumption of the industrial production line by precisely controlling the start-up and shutdown states of the equipment. This process comprehensively considers operating energy consumption and start-up and shutdown energy consumption, resulting in a system energy consumption model that includes:
[0173] Energy consumption model for industrial production line operation:
[0174]
[0175] In the formula, This is a double summation symbol; This represents the total number of devices in the system. The total number of time intervals to be considered; For the first The operating power of a device in the t-th time interval; For the first The device in the Operating status indicator variables for each time interval; For the first Energy loss coefficient of each device; For the first The device in the Energy loss over a time interval.
[0176] Energy consumption model for starting and stopping industrial production lines:
[0177]
[0178] In the formula, For the first Energy consumption coefficient for starting and stopping individual equipment; For the first The device in the Operating status indicator variables for each time interval; For the first The device in the Operating status indicator variables for each time interval.
[0179] Figure 4 , 5 Figures 6 and 7 are comparisons of this system and the traditional PID control strategy in terms of energy consumption prediction error, response speed, and cumulative energy consumption. Figure 7 The invention visually demonstrates the difference between the traditional strategy and the optimized strategy of this invention in terms of device start-up and shutdown sequence. The start-up and shutdown of this strategy is smoother and more coordinated, effectively avoiding peak loads caused by multiple devices starting at high power at the same time.
[0180] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.
Claims
1. A flexible production line device collaborative energy-saving control system based on random ADMM, characterized in that, The application relates to a production line energy consumption control system, which comprises the following parts: a multi-dimensional energy consumption perception module for collecting time sequence power data, equipment state transition data and environmental parameter data of equipment in a production line; a data transmission module connected with the multi-dimensional energy consumption perception module and used for transmitting the time sequence power data, the equipment state transition data and the environmental parameter data to a central control system; a central control system connected with the data transmission module and internally integrated with a cross-industry federal learning module and a non-convex optimization model; the cross-industry federal learning module is used for federal learning and collaborative optimization of multi-source energy consumption data based on a privacy protection mechanism; the non-convex optimization model is used for non-convex optimization solving of a device start-stop strategy by adopting a random ADMM algorithm to generate a globally optimal device start-stop strategy; a dynamic correction module connected with the central control system, used for receiving collaborative optimization data output by the cross-industry federal learning module and the globally optimal device start-stop strategy generated by the non-convex optimization model, performing fusion analysis and optimization calculation, and then feeding back an optimized final control strategy to a production line execution mechanism to control the start-stop state of the production line equipment.
2. The flexible production line device collaborative energy-saving control system based on random ADMM according to claim 1, characterized in that, The multi-dimensional energy consumption perception module comprises an intelligent electric energy metering device, the intelligent electric energy metering device collects system instantaneous power data and constructs a power time sequence matrix P, and the expression of the power time sequence matrix P is as follows: In the formula, represents a power time sequence matrix containing channel instantaneous power data, each row of the power time sequence matrix corresponds to the power time sequence of one channel, and the dimension is ; represents the instantaneous power value of the channel at time ; ; The multi-dimensional energy consumption perception module reads equipment switch state variables in a programmable logic controller (PLC) register and generates an equipment running state transition matrix, and the expression of the equipment running state transition matrix is as follows: wherein denotes the probability or weight of transitioning from state to state ; denotes the system being in state at time ; denotes the system transitioning to state at the next time instant ; denotes the probability of the next state being given the current state is known; The multi-dimensional energy consumption perception module comprises a temperature sensor and a humidity sensor, the temperature sensor and the humidity sensor collect environmental temperature and humidity change data in real time and are used for correcting equipment thermodynamic model parameters, and the expression of the equipment thermodynamic model parameters is as follows: wherein represents the temperature value at time ; represents the temperature sensor calibration reference temperature; represents the temperature sensitivity coefficient; represents the raw signal of the temperature sensor output; represents the relative humidity value at time ; represents the humidity sensitivity coefficient; represents the raw signal of the humidity sensor output; represents the humidity calibration offset; The expression of the equipment thermodynamic model is as follows: wherein represents the power loss or energy loss of the system; represents the loss coefficient related to temperature represents the loss coefficient related to temperature represents that the power loss is proportional to the square of the power. 3.The flexible production line device collaborative energy saving control system based on random ADMM of claim 1, wherein, The cross-industry federal learning module comprises a federal learning submodule, and the federal learning submodule performs the following processes: (1) The server initializes the global model and distributes it to all clients; (2) each client with local data training the model to obtain model update parameters ; (3) Client Upload only encrypted model update parameters to server, not sharing local data ; (4) The server aggregates all the updates from the clients to generate a new global model ; (5) The server will update the global model Distribute to all clients and continue iterating until convergence.
4. The flexible production line device collaborative energy-saving control system based on random ADMM of claim 1, wherein, The cross-industry federal learning module comprises a knowledge sedimentation submodule, and the knowledge sedimentation submodule adopts a teacher-student model framework, wherein a teacher model generates soft labels for input data, and the expression of the soft labels is as follows: wherein represents the soft label probability of the class ; represents the original logit output of the teacher model for the class ; represents the temperature parameter; represents the natural exponential function; A student model needs to simultaneously fit real labels and teacher model soft labels, and the total loss of the student model is a weighted sum of a hard loss and a soft loss, and the expression of the total loss is as follows: In the formula, represents a hard loss, which is a cross-entropy loss; represents a soft loss, which is a distillation loss of the student model on the soft label of the teacher model; represents a total loss function of the student model; is a weight coefficient.
5. The flexible production line device collaborative energy-saving control system based on random ADMM according to claim 1, characterized in that, The cross-industry federal learning module comprises a collaborative optimization submodule, and the collaborative optimization submodule aligns a local model with a global model through a regularization term to integrate the local model uploaded by a client into the global model; and the optimization objective of the collaborative optimization submodule is expressed as follows: In the formula, represents the parameters of the local model of the i-th client; represents the parameters of the local model of the i-th client; represents the loss function of the local model of the i-th client; represents the loss function of the local model of the i-th client; is a hyperparameter used to balance the weights of the loss function term and the regularization term; is used to measure the degree of difference between the parameters of the i-th client local model and the parameters of the global model and the parameters of the global model and the parameters of the global model Meanwhile, the basic optimization objective of the local model is to minimize the loss of local data: In the formula, Indicates the local model parameters Perform minimization optimization; Indicates to the client Local dataset Summing all samples in the dataset; Represents the loss function; Indicates the first Input features of each sample; Indicates the first The true label of each sample; Indicates client Local model parameters; Indicates client Private dataset.
6. The flexible production line device collaborative energy-saving control system based on random ADMM according to claim 1, characterized in that, The objective function of the non-convex optimization model is as follows: wherein, is a binary variable indicating whether the mode is activated; is a fixed activation cost of the mode bound to ; is the amount of primary resource input of the mode ; a Sigmoid function is a simulated non-linear activation cost; is the amount of resource input of the mode ; is the unit cost coefficient of the resource in the mode ; is the interaction penalty coefficient of the mode with the mode ; is the product item of the primary resource input of different modes The non-convex optimization model has a non-convex constraint of mode activation logic: the activation identification quantity of each mode and the difference between the corresponding factors and the critical value need to meet the following non-negative condition: In the formula, is a critical value; The non-convex optimization model is optimized by adopting a random ADMM algorithm, and the optimization model is as follows: wherein is the device is the fixed operating cost coefficient of the device at time is the actual power consumption at time is the fixed operating cost coefficient of the device is a binary variable indicating whether the device is operating at time is the temperature of the device is the temperature-dependent power loss is the temperature of the device at time is the penalty coefficient for power loss is the parameter vector of the th local model is the parameter vector of the global model is the weight coefficient of the federal regularization term is the power matrix of dimension ; ; is the low-rank decomposition of , introducing randomness is the weight coefficient of the ADMM penalty term The non-convex optimization model has a production time sequence constraint: the accumulation of production quantity in a certain time range needs to meet the requirements corresponding to the time interval and the equipment running arrangement, and the probability conditions of the transition between different states are involved. wherein is a binary variable indicating the equipment at time whether there is a production plan; is the equipment minimum continuous run time; is a state transition probability matrix element; denotes the probability of transitioning from state to .
7. The flexible production line device collaborative energy-saving control system based on random ADMM according to claim 6, characterized in that, The solving steps of the random ADMM algorithm are as follows: (1) the variables are split into local problems and global problems; wherein the expression of the local problem is: The expression of the global problem is: In the formula, For equipment At any moment The actual power; For equipment Fixed operating cost coefficient; It is a binary variable representing the device. At any moment Is it running? For global variables in the first... In the next iteration, The consistency constraint target value; For the dual variable of ADMM; For ADMM penalty parameters; This is the low-rank global representation of the power matrix; This is the original power matrix; is the basis matrix for low-rank decomposition; The regularization coefficient is used. (2) Stochastic gradient update: Randomly sample a sub-matrix P from the power matrix P Ω , Ω is a random index set, and thus update the global variable, whose expression is as follows: ； wherein, is the power matrix after the first iteration update; is the power matrix after the second iteration update; is the power matrix after the third iteration update; is the power matrix after the fourth iteration update; is the ADMM dual variable matrix at the fourth iteration; is the ADMM dual variable matrix at the third iteration; is a sub-matrix randomly sampled from the power matrix is a sub-matrix randomly sampled from the power matrix is a rank-r truncated singular value decomposition of the matrix is a pre-defined low-rank dimension; is a weight adjustment factor at the random update. (3) inject federal knowledge, the expression is as follows: wherein, is the model parameter of the th local device; is the global model parameter aggregated by the central server; is the loss function gradient of the th local device; is the federated regularization term coefficient; is the learning rate.
8. The flexible production line device collaborative energy-saving control system based on random ADMM of claim 1, wherein, The dynamic correction module generates the final control strategy through a hybrid weight dynamic fusion mechanism, and the fusion formula is: In the formula, is the dynamic weight vector after fusion; is the collaborative optimization weight of the cross-industry federated learning module; is the optimization weight of the non-convex optimization model; is the strategy fusion coefficient; And a random ADMM-Proximal algorithm is used for optimization, and its wherein are the decision variables of the first are the decision variables of the first are the decision variables of the first iteration; are the decision variables of the first are the decision variables of the first iteration; ; are the Lagrange multiplier vectors of the first are the Lagrange multiplier vectors of the first iteration; is the proximal term coefficient, controlling the weight of the quadratic regularizer in the proximal operator for the variable update; is the proximal term coefficient, controlling the weight of the weighted quadratic term in the proximal operator for the variable update; is the coefficient matrix of the flexible production line device coordination constraint; is the constant term in the constraint; is the augmented Lagrange penalty coefficient; is the reference trajectory tracking strength; is the Lagrange multiplier vector; is the proximal term coefficient; is the proximal term coefficient; is the weight diagonal matrix; is the iteration number; Then the control quantity is synthesized to form the final control output: wherein, denotes the optimal decision variable converged after iteration denotes the optimal decision variable converged after iteration denotes the optimal decision variable converged after iteration denotes the optimal decision variable converged after iteration is the final control output; is the discrete control weight; is the feedback gain matrix; is the weight variation amount. 9.The flexible production line device collaborative energy saving control system based on random ADMM of claim 1, wherein, The dynamic correction module feeds back the optimized final control strategy to the production line actuator, accurately controls the start and stop state of the equipment, and effectively reduces the energy consumption of the industrial production line; this process comprehensively considers the running energy consumption and the start-stop energy consumption, and the system energy consumption model formed includes: Industrial production line running energy consumption model: In the formula, This is a double summation symbol; This represents the total number of devices in the system. The total number of time intervals to be considered; For the first The operating power of a device in the t-th time interval; For the first The device in the Operating status indicator variables for each time interval; For the first Energy loss coefficient of each device; For the first The device in the Energy loss over a time interval; Industrial production line start-stop energy consumption model: In the formula, For the first Energy consumption coefficient for starting and stopping individual devices; For the first The device in the Operating status indicator variables for each time interval; For the first The device in the Operating status indicator variables for each time interval.