Energy-saving intelligent control system for multi-machine operation of air blower

Abstract

The application relates to the field of multi-machine operation control systems, and discloses a multi-machine operation energy-saving intelligent control system of a blower, which comprises the following modules: a data acquisition and fusion module for acquiring and fusing operation parameters and environmental parameters of the blower; a prediction modeling module for constructing a system load prediction model based on historical data; a multi-agent decision module for generating a control strategy according to a current state; a collaborative optimization module for coordinating non-cooperative strategies among multiple blowers; and a parameter updating module for updating the prediction model and the control strategy through control errors and training feedback. The modules work collaboratively to realize efficient energy-saving control and dynamic optimization of the system. The application adopts a multi-agent collaborative optimization technology, thereby guaranteeing the overall energy efficiency of the system, minimizing the load fluctuation of a single blower, prolonging the service life of the equipment, solving the problem of disordered scheduling of the traditional control method under large load fluctuation, and greatly reducing energy consumption.

Description

Technical Field

[0001] This invention relates to the field of multi-machine operation control systems, specifically to an energy-saving and intelligent control system for multi-machine operation of blowers. Background Technology

[0002] Energy conservation and system optimization have always been major technological demands in modern industry and daily life. Blowers, as one of the largest energy-consuming devices, are widely used in various industrial applications. Traditional single-unit control methods often cannot effectively coordinate the operation of multiple devices, resulting in low system efficiency, high energy consumption, and inability to adapt to changes in the external environment.

[0003] In existing technologies, optimization strategies based on single-machine control or static model prediction systems are typically used to manage equipment operation. Single-machine control methods mainly achieve energy-saving effects by adjusting the operating parameters of a single device, but they lack a global perspective and cannot take into account the collaborative optimization between multiple devices. Static model prediction systems operate based on pre-set models, which have poor adaptability. Although they can achieve good energy-saving effects under certain stable conditions, they often perform poorly in environments with large load fluctuations and cannot automatically correct model errors, resulting in control strategies that cannot flexibly respond to constantly changing system states.

[0004] However, existing technologies still have some significant shortcomings in practical applications. First, single-machine control strategies cannot coordinate the operation of multiple devices, leading to uneven load distribution within the system, resulting in energy waste and equipment wear. Second, static model prediction methods lack adaptability to long-term system changes; when the environment or equipment status changes, the system control accuracy drops significantly. Finally, the fault tolerance mechanisms in existing control methods are weak; once equipment malfunctions, manual intervention is often required, affecting the system's automation level and stability. Therefore, those skilled in the art propose an energy-saving intelligent control system for multi-machine operation of blowers to address these problems. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an energy-saving and intelligent control system for multi-unit operation of blowers, which solves problems such as poor coordination among multiple devices, poor system adaptability, and insufficient fault recovery capabilities in existing technologies.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an energy-saving intelligent control system for multi-unit operation of blowers, comprising:

[0007] The data acquisition and fusion module is used to collect the operating parameters and environmental parameters of the blower and perform fusion processing.

[0008] The predictive modeling module is connected to the data acquisition and fusion module and is used to build a system load prediction model based on historical input and output data.

[0009] A multi-agent decision-making module, connected to the data acquisition and fusion module and the predictive modeling module, is used to generate control strategies based on the current state vector;

[0010] The collaborative optimization module, connected to the multi-agent decision-making module, is used to coordinate non-cooperative strategies among multiple blowers;

[0011] The parameter update module, connected to the prediction modeling module and the multi-agent decision-making module, is used to update the prediction model and control strategy based on control error and training feedback.

[0012] Preferably, the data acquisition and fusion module includes:

[0013] The operating parameter acquisition unit is used to collect pressure, flow, speed, energy consumption and vibration spectrum data for each blower;

[0014] The environmental parameter acquisition unit is used to collect the external ambient temperature and real-time electricity price of the system.

[0015] The data fusion processing unit is used to perform principal component analysis on the vibration spectrum data, generate a dimension-reduced feature vector, and construct a system state vector. The state vector includes pressure value, flow rate value, rotational speed value, energy consumption value, vibration feature vector, ambient temperature value, and electricity price.

[0016] Preferably, the predictive modeling module includes:

[0017] Step response modeling unit, used to construct the response matrix of control input to system output;

[0018] The free response estimation unit is used to calculate the system output trend under conditions of no control input;

[0019] The multi-step prediction generation unit combines control input increments and free responses to output predicted system loads at several future time points.

[0020] Preferably, the multi-agent decision-making module includes:

[0021] The state processing unit is used to input the state vector of each blower into its respective policy network;

[0022] The strategy generation unit is used to output the control actions of the blower, the control actions including start / stop signals and speed changes;

[0023] The reward construction unit is used to construct the local reward value for a single blower based on energy consumption value, speed deviation value, and number of start-stop cycles.

[0024] Preferably, the collaborative optimization module includes:

[0025] The revenue function construction unit is used to define the non-cooperative revenue function of multiple blowers, which is constructed based on their respective reward functions and the system prediction error.

[0026] An optimized coordination unit is used for joint strategy optimization using the alternating direction multiplier method.

[0027] The instruction coordination output unit is used to output the final consistent control strategy among multiple blowers.

[0028] Preferably, the parameter update module includes:

[0029] The model update unit is used to correct the response matrix in real time using the prediction error;

[0030] The strategy training unit is used to train the control strategy network based on simulation data;

[0031] The policy synchronization unit is used to synchronize and update the policy parameters obtained from offline training to the control network in the edge device.

[0032] Preferably, the dimensionality-reduced feature vector of the data fusion processing unit is the product of the vibration spectrum data and the principal component matrix, calculated using the following formula:

[0033] v pca =W·v;

[0034] Where: v is the original vector of the vibration spectrum; W is the principal component dimensionality reduction weight matrix; v pca This is the output dimension-reduced vibration eigenvector.

[0035] Preferably, the output prediction value of the multi-step prediction generation unit satisfies the following expression:

[0036] y pred =G·Δu+y free ;

[0037] Where: G represents the step response matrix obtained by constructing the unit through the response matrix; Δu represents the input increment vector for the future prediction period; y free y is the free response vector; pred Output for future predictions.

[0038] Preferably, the reward value of the reward construction unit is calculated according to the following expression:

[0039]

[0040] Where: r ie is the reward value for blower i; i The energy consumption value of blower i; e base As the baseline energy consumption; ω i This refers to the current speed of blower i. ω is the average rotational speed; max δ represents the maximum permissible speed; δ is the start / stop change indicator of blower i between the current step and the previous time step, which is 1 if the start / stop state changes, otherwise it is 0; α, β, and γ are preset weighting factors, all of which are positive numbers.

[0041] Preferably, the response matrix of the model update unit is updated using the following expression:

[0042]

[0043] Among them: G new G represents the updated response matrix; old Represents the original response matrix; η is the learning rate parameter, which is dimensionless and can be dynamically adjusted according to the actual application; y real y is the output vector actually observed by the system. pred Δu is the prediction vector output by the prediction modeling module; Δu is the increment vector of the control input.

[0044] This invention provides an energy-saving intelligent control system for multi-unit operation of blowers. It has the following beneficial effects:

[0045] 1. This invention employs multi-agent collaborative optimization technology, which minimizes load fluctuations of individual blowers while ensuring overall system energy efficiency, thereby extending the service life of the equipment. Compared to existing single-optimization control schemes, this invention effectively avoids the negative impact of individual machine "selfish" behavior on the overall system and solves the problem of disordered scheduling in traditional control methods when load fluctuations are large, thus significantly reducing energy consumption.

[0046] 2. This invention introduces an incremental correction online update mechanism, which monitors and adjusts the prediction model in real time to ensure the stability and adaptability of the system in long-term operation. This is different from the static model update method in the prior art, avoids the accumulation of model deviation, and successfully solves the problem of reduced control effect caused by model distortion in traditional methods.

[0047] 3. The present invention combines offline training and online updating of the strategy, enabling the control strategy to be continuously optimized according to the actual operation of the system and kept synchronized with each intelligent agent in the system. Unlike the existing technology that can only rely on a single data source for updates, the present invention can make full use of historical data and real-time feedback, avoid the problem of strategy update lag, and improve decision-making efficiency.

[0048] 4. In the process of outputting control commands, the present invention adopts a fault-tolerant mechanism. When abnormal feedback occurs in the system, the erroneous command can be stopped in time to ensure the stable operation of the system. Compared with the traditional method, the present invention enhances the robustness of the system. Especially under high load or equipment failure, it can avoid global problems caused by single point of failure, thereby improving the overall reliability of the system. Attached Figure Description

[0049] Figure 1 This is a schematic diagram of the system architecture of the present invention;

[0050] Figure 2 This is a schematic diagram of the data acquisition and fusion module architecture of the present invention;

[0051] Figure 3 This is a schematic diagram of the predictive modeling module architecture of the present invention;

[0052] Figure 4 This is a schematic diagram of the multi-agent decision-making module architecture of the present invention;

[0053] Figure 5 This is a schematic diagram of the collaborative optimization module architecture of the present invention;

[0054] Figure 6 This is a schematic diagram of the parameter update module architecture of the present invention. Detailed Implementation

[0055] The technical solutions in the embodiments 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, and 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.

[0056] Please see the appendix Figure 1 -Appendix Figure 6 This invention provides an energy-saving intelligent control system for multi-unit operation of blowers, comprising:

[0057] The data acquisition and fusion module is used to collect the operating parameters and environmental parameters of the blower and perform fusion processing.

[0058] Specifically, in the energy-saving intelligent control system for multi-blower operation involved in this embodiment, the data acquisition and fusion module, as the system's information input end, is responsible for the full acquisition and preprocessing of the real-time status and environmental information of the blowers. This module is not only tightly coupled with the downstream predictive modeling module, but also directly affects the model accuracy and control strategy effectiveness of the overall system. Therefore, during system implementation, it is necessary to ensure that this module has high-frequency, high-precision data acquisition capabilities and can perform real-time dimensionality reduction and fusion of high-dimensional data streams to form a complete state description.

[0059] Typically, the data acquisition and fusion module includes an operating parameter acquisition unit, an environmental parameter acquisition unit, and a data fusion processing unit. These units collaborate to complete the real-time monitoring and state vector construction of the blower group. In one possible implementation of this invention, the module first acquires the key operating parameters of each blower in real time through a sensor network, then integrates data from different sources and performs dimensionality reduction analysis to provide a unified input for subsequent model prediction.

[0060] Alternatively, the system can employ various hardware components such as industrial-grade pressure sensors, flow meters, Hall effect speed sensors, and energy meters to achieve comprehensive monitoring of the blower's operating status. For example, flow rate can be acquired using turbine flow meters or electromagnetic flow meters to ensure measurement accuracy; vibration spectrum data is obtained through a high-sensitivity accelerometer array, with a vibration frequency range of 20Hz to 20kHz, to meet the needs of mechanical health monitoring.

[0061] In this embodiment, the operating parameter acquisition unit is mainly used to acquire the following data:

[0062] The pressure value 'p' of the blower is expressed in kilopascals (kPa).

[0063] The flow rate q is expressed in cubic meters per hour (m³ / h). 3 / h);

[0064] The rotational speed value n is expressed in revolutions per minute (rpm).

[0065] Real-time energy consumption e, in kilowatts (kW);

[0066] The vibration spectrum vector v is a multidimensional array that represents the acceleration value of the blower at different frequency points.

[0067] In some embodiments, to ensure real-time performance, the acquisition period for pressure and flow data is set to 1 second, and the vibration spectrum data can be sampled in batches with a window of 10 seconds to achieve both real-time and accurate state perception.

[0068] The environmental parameter acquisition unit is used to collect key influencing factors in the working environment of the blower, specifically including:

[0069] External temperature T, in degrees Celsius (°C);

[0070] Real-time electricity price E, in yuan per kilowatt-hour (yuan / kWh).

[0071] Electricity price data is obtained through an API interface with the power grid, and the update frequency is usually set to refresh every 15 minutes.

[0072] The data fusion processing unit combines the aforementioned operating parameters with environmental parameters and performs dimensionality reduction on the high-dimensional vibration spectrum vector v. Specifically, Principal Component Analysis (PCA) is used to reduce the dimensionality of the vibration spectrum data. The calculation steps of PCA include: first, normalizing the mean of the original vibration spectrum matrix; then, extracting the most important eigenvectors through singular value decomposition (SVD); and finally, reconstructing the dimensionality-reduced state vector.

[0073] In one possible implementation, assuming the original vibration spectrum vector v has a dimension of 1024, the system reduces its dimension to a 20-dimensional feature vector using the PCA weight matrix W (20×1024) obtained during training. Its mathematical expression is as follows:

[0074] v pca =W·v;

[0075] Where: v is the original vector of the vibration spectrum; W is the principal component dimensionality reduction weight matrix with dimensions m×n, where m=20 and n=1024 in this example; v pca This is the output dimension-reduced vibration eigenvector.

[0076] It should be noted that the weight matrix of W was obtained by training with a large amount of historical data in the offline stage. The training process includes steps such as solving the covariance matrix and eigenvalue decomposition, which ensures that the eigenvectors after dimensionality reduction can maximize the retention of the main information in the vibration spectrum.

[0077] Specifically, the final state vector x includes the following components:

[0078] x = [p, q, n, e, T, E, v] pca ];

[0079] Where: p is the current pressure value; q is the current flow rate value; n is the current rotational speed value; e is the current energy consumption value; T is the external temperature value; E is the current real-time electricity price; v pca This is the eigenvector of the vibration spectrum after dimensionality reduction.

[0080] This structure allows downstream modules to receive a fixed-length state vector, eliminating the need for repeated decoding of the high-dimensional original signal and thus significantly reducing computational burden.

[0081] In a preferred implementation, the data fusion processing unit incorporates a multi-threaded processing mechanism to ensure that, in high-concurrency scenarios, all data sources can be collected and merged into the state vector in a timely manner. As a scalable solution, the edge computing unit can also be integrated on the blower side to perform real-time dimensionality reduction of the vibration spectrum and push the state vector to the central control system via an industrial bus.

[0082] In some embodiments, the system uses the TCP / IP network protocol to communicate with a remote server to ensure that environmental parameters such as real-time electricity price data can be synchronized regularly. Blower status data is reported in real time via the local area network to avoid data delays caused by external network fluctuations.

[0083] In general, the data acquisition and fusion module of this invention can be applied not only to blower clusters within a single factory, but also to remote monitoring scenarios of blowers across regions, enabling broader data integration and processing.

[0084] The predictive modeling module is connected to the data acquisition and fusion module and is used to build a system load prediction model based on historical input and output data.

[0085] Specifically, in the energy-saving intelligent control system for multi-unit blower operation involved in this invention, the predictive modeling module, as a downstream module of the data acquisition and fusion module, undertakes the task of dynamically predicting the system load and is the foundation for the implementation of the entire control strategy. This module directly receives the state vector input from the data acquisition and fusion module and outputs predicted load data to guide control decisions.

[0086] Generally, the predictive modeling module includes a response matrix construction unit, a free response estimation unit, and a multi-step prediction calculation unit. These units together constitute the system's dynamic prediction mechanism. Specifically, this module mainly extracts dynamic features of the system from historical input and output data, constructs a predictive model, and generates multi-step prediction results through this model for use by the multi-agent decision-making module.

[0087] In this embodiment, the response matrix construction unit is used to establish a dynamic prediction relationship based on historical system data. Alternatively, this unit first receives the historical control input sequence {Δu(k)} of the blower and the corresponding system response output {y(k)}. By processing the input and output sequences, the step response matrix G of the system is solved using the least squares estimation method or the recursive least squares algorithm.

[0088] In one possible implementation, assume the system's input increment vector at time k is Δu(k), and the corresponding output vector is y(k). The system's step response matrix G is defined as:

[0089]

[0090] Wherein: g mn This represents the step response coefficient between the m-th output and the n-th input; m is the dimension of the output variable; and n is the dimension of the input variable.

[0091] These step response parameters are obtained through system identification training and reflect the dynamic influence of input on output.

[0092] The free response estimation unit is used to predict the natural evolution trajectory of the system's future state under the condition that there are no new input increments. Generally, this unit estimates the free response vector y using the current state vector x(k) and historical output data. free In some embodiments, the free response can be calculated using a recursive model, for example, by modeling the output dynamics using a first-order or higher-order difference model.

[0093] In this embodiment, the estimation of the free response follows the following expression:

[0094] y free (k+1)=A·y(k)+B·x(k);

[0095] Where: A is the state transition matrix; B is the state input matrix; y(k) is the output vector at time k; and x(k) is the state vector at time k.

[0096] This structure not only takes into account the system's inertia, but also incorporates the potential impact of current environmental changes on the system's output.

[0097] The multi-step prediction computation unit is responsible for combining the response matrix and the free response to generate a complete future prediction output. Specifically, this unit uses the control input increment vector Δu within the prediction period to calculate the prediction output using the following formula:

[0098] y pred =G·Δu+y free ;

[0099] Where: G represents the step response matrix obtained by constructing the unit through the response matrix; Δu represents the input increment vector for the future prediction period, with a dimension of n×1; y free The free response vector has a dimension of m×1; y pred The output is a future prediction, with a dimension of m×1.

[0100] Alternatively, in the implementation, the prediction time step can be set according to the actual operating conditions, such as 5 steps, 10 steps or more, to meet the needs of multi-step prediction. This module allows for dynamic adjustment of the prediction step size to adapt to the real-time requirements of the system operation.

[0101] In some embodiments, an adaptive update mechanism is integrated within the prediction modeling module to improve prediction accuracy. When a change in the dynamic characteristics of the system is detected, the retraining process of the response matrix can be automatically triggered, thereby maintaining the timeliness and accuracy of the prediction model. As a preferred approach, a sliding window strategy can be adopted to update the historical sequence when new data arrives, ensuring the timeliness of the model training data.

[0102] Generally, the predictive modeling module and the multi-agent decision-making module exchange data through standardized interfaces, such as REST API or internal memory sharing mechanisms, to minimize the overall system latency.

[0103] In one implementation, to ensure stable operation under high concurrency scenarios, the predictive modeling module adopts a parallel computing architecture and uses matrix parallel multiplication to accelerate the predictive calculation process in order to support the real-time control requirements of large-scale blower systems.

[0104] A multi-agent decision-making module, connected to the data acquisition and fusion module and the predictive modeling module, is used to generate control strategies based on the current state vector;

[0105] Specifically, in the multi-blower operation energy-saving intelligent control system of the present invention, the multi-agent decision-making module is located downstream of the predictive modeling module. It receives the predictive output and state vector and generates the control actions for each blower. Through a distributed agent mechanism, this module enables each blower to have independent decision-making capabilities while coordinating with the global strategy, thereby achieving flexible and precise operation control.

[0106] Typically, the multi-agent decision-making module includes a state encoding unit, a policy output unit, and a reward function generation unit. These units work closely together to make dynamic control decisions regarding the start-up, shutdown, and speed of the blowers. The output of this module represents the local control actions of each blower, which are then further integrated into a globally optimal policy by the collaborative optimization module.

[0107] In this embodiment, the state coding unit is used to receive the state vector x provided by the predictive modeling module. i The state vector contains the blower's pressure, flow rate, speed, energy consumption, ambient temperature, electricity price, and vibration spectrum features after dimensionality reduction via principal component analysis. In one possible implementation, the state encoding unit will... i The input is fed into the policy network to extract state features and use them for subsequent control action calculations.

[0108] Alternatively, the policy network can employ a deep neural network structure, including an input layer, several hidden layers, and an output layer. The input layer receives the state vector, the hidden layers are used for non-linear feature extraction, and the output layer generates continuous or discrete values ​​of the control action. In this embodiment, the output of the policy output unit includes:

[0109] Start / stop control signal s i Its value range is {0,1}, indicating whether blower i is enabled;

[0110] Speed ​​adjustment increment Δn i The unit is revolutions per minute (rpm).

[0111] In a preferred implementation, the control actions of the blower can be viewed as a two-dimensional action vector:

[0112] a i =[s i ,Δn i ];

[0113] Where: s i Indicates the start / stop status of blower i; Δn i This represents the target speed increment of blower i.

[0114] The reward function generation unit is used to construct an individual reward function for each blower to guide the policy network in learning energy-efficient control strategies. Typically, this reward function combines multiple indicators such as energy consumption reduction, speed consistency, and smooth start-stop behavior to ensure that system control is both energy-efficient and smooth.

[0115] In this embodiment, the reward function r i The calculation formula is as follows:

[0116]

[0117] Where: r i e is the reward value for blower i; i Energy consumption value of blower i (unit: kW); e base Baseline energy consumption (unit: kW); ω i The current speed of blower i (unit: rpm); ω is the average rotational speed (unit: rpm); max δ represents the maximum permissible speed (unit: rpm); δ is the start / stop change indicator of blower i between the current step and the previous time step, which is 1 if the start / stop state changes, otherwise it is 0; α, β, and γ are preset weighting factors, all of which are positive numbers.

[0118] As one implementation method, the first term α·(e) in the reward function base -ei The first term represents the energy-saving driving force, the second term is used to punish the behavior of speed deviating from the global average, and the third term is used to suppress equipment wear caused by frequent start-stop.

[0119] In some embodiments, the strategy output unit also includes an action constraint mechanism to ensure that the output speed adjustment increment Δn i It does not exceed the preset safety threshold range. For example, setting Δn i ∈[-50,+50]rpm, to avoid exceeding the mechanical safety range of the blower.

[0120] Specifically, in the multi-agent architecture, the agents of each blower are independent of each other. The system adopts a centralized training and distributed execution mechanism, enabling agents to share environmental information during the training phase, while making autonomous decisions based only on local states during the execution phase. In one possible implementation, the policy network is trained using the Deep Deterministic Policy Gradient (DDPG) algorithm, which can handle continuous action spaces.

[0121] In general, the multi-agent decision-making module and the collaborative optimization module exchange control action data through shared memory or a fast data bus to ensure that the control loop is completed within milliseconds.

[0122] As a preferred option, the module can also integrate a policy replay cache to store historical state-action-reward sequences for subsequent offline retraining, thereby enhancing the robustness of the policy.

[0123] In some embodiments, the module implements a motion smoothing mechanism, which introduces a low-pass filter between two adjacent decisions to smooth the speed adjustment signal Δn. i It is smoother in time series and reduces mechanical shock.

[0124] The collaborative optimization module, connected to the multi-agent decision-making module, is used to coordinate non-cooperative strategies among multiple blowers;

[0125] Specifically, in the multi-blower operation energy-saving intelligent control system of the present invention, the collaborative optimization module, as the downstream of the multi-agent decision-making module, is responsible for integrating the local control actions of each blower and performing joint optimization processing at the system level to solve for the optimal overall control strategy. This module forms a high coupling relationship with the multi-agent decision-making module, achieving coordination and consistency between local and global strategies through real-time data exchange.

[0126] Generally, the collaborative optimization module includes a reward function construction unit, an equilibrium solution unit, and a control command output unit. These units work together logically to coordinate the scheduling of multiple blowers. Specifically, the core of this module lies in combining the rewards of local agents with the global system objective, forming a strategy optimization problem within a multi-agent game framework.

[0127] In this embodiment, the revenue function construction unit is used to construct a non-cooperative revenue function for each blower based on the control actions output by the multi-agent decision-making module. Alternatively, the revenue function not only considers individual rewards but also incorporates system load balancing constraints into the revenue model to encourage local actions to align with the global objective.

[0128] In one possible implementation, assume that the individual reward value of blower i is r. i The load deviation of the entire system is defined as ΔL. Then the revenue function U... i It can be represented as:

[0129] U i =r i -λ·|ΔL|;

[0130] Among them: U i r represents the final revenue value of blower i; i ΔL is the individual reward value of blower i (unit: dimensionless, derived from the reward function output of the multi-agent module); ΔL is the deviation between the current total load of the system and the target load (unit: kW); λ is the load constraint weight coefficient, in the form of a dimensionless parameter.

[0131] This structure couples individual rewards with global load, effectively suppressing system performance degradation caused by the "selfish" behavior of individual blowers.

[0132] The equilibrium solution unit is used to find the optimal solution for the control action of each blower within the framework of the aforementioned reward function. Generally, this unit employs distributed optimization algorithms such as the Alternating Direction Multiplier Method (ADMM) to achieve cooperative convergence of solutions among multiple agents through an iterative process.

[0133] In this embodiment, the basic update steps of ADMM are as follows:

[0134] For each agent i, while keeping the actions of other blowers fixed, update its own action a. i To minimize the local loss function;

[0135] Perform a Lagrange multiplier update to correct the system consistency constraints;

[0136] Repeat the above steps until the system meets the convergence condition.

[0137] As one implementation method, the update formula of ADMM can be expressed as:

[0138]

[0139] in: z is the control action vector of blower i in the (k+1)th iteration; (k) u is the system's average action vector; (k) ρ is the scaling Lagrange multiplier for blower i; ρ is the penalty parameter used to balance local losses and global constraints.

[0140] Specifically, control action vector a i Including start / stop signals s i and speed increment Δn i ,Right now:

[0141] a i =[s i ,Δn i ];

[0142] In a preferred implementation, the equalization unit dynamically adjusts the value of ρ in each iteration to accelerate algorithm convergence, which can significantly improve response speed, especially when the load fluctuates greatly.

[0143] The control command output unit distributes the optimized global optimal control strategy to each blower actuator after the equalization solution unit has completed its solution. Typically, this unit uses an industrial bus or high-efficiency network interface to send data, ensuring millisecond-level response latency.

[0144] In some embodiments, the control command output unit integrates a fault-tolerant mechanism. When the system detects an abnormality in feedback from a blower, it can output a stop signal according to the current strategy to avoid affecting the execution of the overall control strategy due to a single point of failure.

[0145] As a preferred approach, this module can also introduce other constraints, such as noise suppression or temperature control, into the payoff function to enhance the system's scalability and robustness. For example, the payoff function can be expanded to:

[0146] U i =r i -λ1|ΔL|-λ2|T i -T ref |;

[0147] Wherein: T i T represents the motor temperature of blower i (unit: °C). ref λ1 and λ2 are the target reference temperature (unit: °C); λ1 and λ2 are weighting parameters used to balance multiple constraints.

[0148] Generally, the collaborative optimization module and the parameter update module achieve dynamic adaptation through a feedback mechanism. Specifically, optimization results and execution status are fed back to the parameter update module in real time for further correction of the model and strategy.

[0149] In one implementation, the module supports a parallel optimization mechanism, which is particularly suitable for application scenarios with a large-scale blower cluster (such as more than 50 units), using distributed computing to accelerate the global optimization process.

[0150] The parameter update module, connected to the prediction modeling module and the multi-agent decision-making module, is used to update the prediction model and control strategy based on control error and training feedback.

[0151] Specifically, in the multi-blower operation energy-saving intelligent control system of the present invention, the parameter update module is downstream of the collaborative optimization module. Its main function is to monitor the deviation between the system prediction and the actual execution in real time, dynamically correct the model parameters, and periodically update the control strategy. This module forms a bidirectional coupling relationship with the predictive modeling module and the multi-agent decision-making module, ensuring that the system maintains continuous optimization of prediction accuracy and control effect during long-term operation.

[0152] Generally, the parameter update module includes an online model correction unit, an offline policy training unit, and a policy synchronization unit. These units work together to build an update mechanism that is both online adaptive and supports offline enhancement.

[0153] In this embodiment, the online model correction unit dynamically adjusts the response matrix G in the predictive modeling module based on the error between the system's predicted output and the actual executed output. Alternatively, when the system detects a significant error, this unit triggers an online update process to suppress model drift.

[0154] In one possible implementation, the online update of the response matrix employs an incremental correction strategy, as expressed mathematically below:

[0155]

[0156] Among them: G new G represents the updated response matrix; old Represents the original response matrix; η is the learning rate parameter, which is dimensionless and can be dynamically adjusted according to the actual application; y real The actual output vector observed by the system (unit: kW or m) 3 / h, based on the control objective); y pred The prediction vector is the output of the predictive modeling module; Δu is the incremental vector of the control input (the unit depends on the control dimension and can be rpm, start / stop signal, etc.).

[0157] Generally, the learning rate η is set to a positive number less than 1 to avoid instability caused by the model updating too quickly.

[0158] The policy offline training unit is used to periodically retrain the multi-agent policy network using simulation datasets. Specifically, during long-term operation, this unit collects system state-action-reward sequences and generates a large-scale training sample library. Alternatively, an experience replay mechanism can be used to prioritize the inclusion of historically best samples into the training set to accelerate the convergence speed of the policy network.

[0159] In some embodiments, offline training of the policy network employs a combination of supervised learning and reinforcement learning. For example, the Deep Deterministic Policy Gradient (DDPG) algorithm can be used to maximize the reward function r. i The strategy is optimized. The training objective function is:

[0160]

[0161] Where: J(θ) represents the expected reward of the policy network parameter θ; x i Let π be the state vector of blower i; θ (x i ) represents the action generated by the policy network; r i The reward function has been defined in the aforementioned module.

[0162] The role of the policy synchronization unit is to distribute the new policy parameters obtained from offline training to the edge computing unit in real time, ensuring that the policy remains synchronized across all agents in the system. As an implementation, this unit employs a dual-channel communication mechanism: on the one hand, it transmits the new policy parameters via the internal network; on the other hand, it uses a checksum mechanism to confirm the integrity of the parameter synchronization.

[0163] Specifically, in large-scale blower clusters, this unit can adopt a batch push strategy to gradually update strategy parameters by region in order to reduce instantaneous network load.

[0164] In general, the parameter update module also has an anomaly detection mechanism. For example, if the model update causes a sudden increase in the system's prediction error, the module can trigger a rollback mechanism to restore the response matrix to the previous stable version, preventing divergence in online updates.

[0165] In some embodiments, this module allows the integration of rules from an expert system. For example, if the blower's operating environment undergoes a sudden change (such as a sharp rise in temperature), a function that prioritizes safety policies can be triggered to avoid control anomalies caused by extreme conditions.

[0166] As an extension, the parameter update module can upload the collected data to a cloud server for remote experts to further analyze and provide feedback on improvements for future version iterations.

[0167] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A multi-unit operation energy-saving intelligent control system for blowers, characterized in that, include: The data acquisition and fusion module is used to collect the operating parameters and environmental parameters of the blower and perform fusion processing. The data acquisition and fusion module includes: The operating parameter acquisition unit is used to collect pressure, flow, speed, energy consumption and vibration spectrum data for each blower; The environmental parameter acquisition unit is used to collect the external ambient temperature and real-time electricity price of the system. The data fusion processing unit is used to perform principal component analysis on the vibration spectrum data, generate a dimension-reduced feature vector, and construct a system state vector, which includes pressure value, flow rate value, rotational speed value, energy consumption value, vibration feature vector, ambient temperature value, and electricity price. The predictive modeling module is connected to the data acquisition and fusion module and is used to build a system load prediction model based on historical input and output data. A multi-agent decision-making module, connected to the data acquisition and fusion module and the predictive modeling module, is used to generate control strategies based on the current state vector; The multi-agent decision-making module includes: The state processing unit is used to input the state vector of each blower into its respective policy network; The strategy generation unit is used to output the control actions of the blower, the control actions including start / stop signals and speed changes; The reward construction unit is used to construct the local reward value of a single blower based on the energy consumption value, speed deviation value, and number of start-stop cycles; The collaborative optimization module, connected to the multi-agent decision-making module, is used to coordinate non-cooperative strategies among multiple blowers; The collaborative optimization module includes: The revenue function construction unit is used to define the non-cooperative revenue function of multiple blowers, which is constructed based on their respective reward functions and the system prediction error. An optimized coordination unit is used for joint strategy optimization using the alternating direction multiplier method. The instruction coordination output unit is used to output the final consistent control strategy among multiple blowers; The parameter update module, connected to the prediction modeling module and the multi-agent decision-making module, is used to update the prediction model and control strategy based on control error and training feedback.

2. The multi-unit operation energy-saving intelligent control system for blowers according to claim 1, characterized in that, The predictive modeling module includes: Step response modeling unit, used to construct the response matrix of control input to system output; The free response estimation unit is used to calculate the system output trend under conditions of no control input; The multi-step prediction generation unit combines control input increments and free responses to output predicted system loads at several future time points.

3. The multi-unit operation energy-saving intelligent control system for blowers according to claim 1, characterized in that, The parameter update module includes: The model update unit is used to correct the response matrix in real time using the prediction error; The strategy training unit is used to train the control strategy network based on simulation data; The policy synchronization unit is used to synchronize and update the policy parameters obtained from offline training to the control network in the edge device.

4. The multi-unit operation energy-saving intelligent control system for blowers according to claim 1, characterized in that, The dimensionality-reduced feature vector of the data fusion processing unit is the product of the vibration spectrum data and the principal component matrix, calculated using the following formula: ； in: This is the original vector of the vibration spectrum; Principal component dimensionality reduction weight matrix; This is the output dimension-reduced vibration eigenvector.

5. The multi-unit operation energy-saving intelligent control system for a blower according to claim 2, characterized in that, The output prediction value of the multi-step prediction generation unit satisfies the following expression: ； in: This represents the step response matrix obtained by constructing the unit from the response matrix; This represents the input increment vector for the future prediction period; The free response vector; Output for future predictions.

6. The multi-unit operation energy-saving intelligent control system for blowers according to claim 1, characterized in that, The reward value of the reward construction unit is calculated according to the following expression: ； in: For blower The reward value; For blower Energy consumption value; Baseline energy consumption; For blower Current rotational speed; The average rotational speed; Maximum permissible speed; For blower The start / stop change flag between the current step and the previous time step is 1 if the start / stop state changes, and 0 otherwise. , , These are preset weighting factors, all of which are positive numbers.

Images