Modular chiller load balancing scheduling method based on edge computing

Abstract

The application discloses a modular refrigeration machine load balancing scheduling method based on edge computing and relates to the technical field of load balancing scheduling. The method comprises the following steps: constructing a first state vector according to processing data of a refrigeration machine; inputting the first state vector into a load balancing intelligent agent model for processing; outputting an action vector at a current time; converting the action vector into a control instruction; performing control on the refrigeration machine; collecting post-execution data of the refrigeration machine; constructing a second state vector based on the post-execution data; calculating a reward function based on the second state vector; adopting a proximal optimization strategy to perform online update on parameters of the load balancing intelligent agent model based on the first state vector, the action vector and the reward function; and guaranteeing the real-time performance and reliability of scheduling based on the deployment of edge computing, supporting efficient cooperation of multiple refrigeration machines, prolonging the service life of the equipment, improving energy efficiency, and having remarkable energy-saving benefits and engineering practical value.

Description

Technical Field

[0001] This invention relates to the field of load balancing scheduling technology, and in particular to a modular chiller load balancing scheduling method based on edge computing. Background Technology

[0002] In recent years, with the increasing demands for energy efficiency, reliability, and precise temperature control in data centers, cold chain logistics, and industrial refrigeration, the deployment scale and architectural complexity of modular refrigeration equipment have increased significantly. Driven by both digital transformation and green and low-carbon development, efficient refrigeration has become a core element in ensuring the stable operation of critical facilities, reducing operating costs, and achieving energy conservation and consumption reduction. In particular, in large data centers, modern warehousing cold chains, and high-precision manufacturing environments, refrigeration units not only need to maintain strict environmental temperature and humidity standards, but also need to cope with dynamically changing load distribution, external climate fluctuations, and equipment heterogeneity. Their operating efficiency and scheduling strategies are directly related to overall energy efficiency, equipment lifespan, and operational safety, and have significant economic and social benefits.

[0003] However, traditional load balancing scheduling methods mostly rely on fixed rules or offline optimization strategies based on historical data. The core of these methods lies in preset thresholds and static matching, which are difficult to adapt to dynamically changing ambient temperatures, regional load differences, and equipment status fluctuations. This leads to problems such as uneven energy consumption, insufficient temperature control accuracy, and response lag during the scheduling process, which seriously affects the overall energy efficiency and operational reliability of the chiller units. When dealing with the above challenges, existing scheduling methods often lack a closed-loop framework that can continuously perceive, make real-time decisions, and self-optimize. Most methods do not incorporate the coupling relationship between ambient temperature fluctuations, regional temperature differences, and equipment power consumption into a unified state representation, and also lack a mechanism for online policy updates based on real-time interactive data, resulting in rigid scheduling strategies and long-term performance degradation. Summary of the Invention

[0004] The technical problem addressed by this invention is that traditional load balancing scheduling methods often rely on fixed rules or offline optimization strategies based on historical data. The core of these methods lies in preset thresholds and static matching, which are difficult to adapt to dynamically changing ambient temperatures, regional load differences, and equipment status fluctuations. This leads to uneven energy consumption, insufficient temperature control accuracy, and lag in response during the scheduling process, seriously affecting the overall energy efficiency and operational reliability of the chiller units. When addressing the above challenges, existing scheduling methods often lack a closed-loop framework that can continuously sense, make real-time decisions, and self-optimize. Most methods do not incorporate the coupling relationship between ambient temperature fluctuations, regional temperature differences, and equipment power consumption into a unified state representation, and also lack a mechanism for online strategy updates based on real-time interactive data, resulting in rigid scheduling strategies and long-term performance degradation.

[0005] To address the aforementioned technical problems, this invention provides the following technical solution: a modular chiller load balancing scheduling method based on edge computing, comprising the following steps:

[0006] Step S1: Construct the first state vector based on the processing data of the refrigeration machine;

[0007] Step S2: Input the first state vector into the load balancing agent model for processing, and output the action vector at the current moment;

[0008] Step S3: Convert the action vector into control commands to control the refrigerator and collect post-execution data of the refrigerator.

[0009] Step S4: Construct a second state vector based on the post-execution data, and calculate the reward function based on the second state vector;

[0010] Step S5: Based on the first state vector, action vector, and reward function, the parameters of the load balancing agent model are updated online using a proximal optimization strategy.

[0011] As a preferred embodiment of the modular chiller load balancing scheduling method based on edge computing described in this invention, step S1 involves constructing a first state vector based on the chiller's processing data.

[0012] Step S1 includes steps S101, S102 and S103;

[0013] Step S101: Obtain processing data, which includes real-time power consumption data of the chiller, set temperature data, operating status data, ambient temperature data, zone temperature feedback data, rated maximum power consumption data, and zone set temperature value.

[0014] Step S102: Calculate the load level data for each refrigeration unit and the temperature deviation data for each area based on the processing data;

[0015] The calculation process for load level data is as follows: the ratio of the real-time power consumption data of the i-th chiller to the rated maximum power consumption data is recorded as the load level data.

[0016] The absolute value of the difference between the regional temperature feedback data and the regional set temperature value for each region is recorded as the temperature deviation data.

[0017] The load level data, temperature deviation data, and ambient temperature data are normalized to obtain the load vector, temperature deviation vector, and ambient temperature vector.

[0018] Step S103: Based on the load vector, temperature deviation vector, and ambient temperature vector, a first state vector is constructed. The expression for the first state vector is:

[0019] ;

[0020] in, Represents the first state vector. Represents the load vector. Represents the temperature deviation vector. Represents the ambient temperature vector. This represents the action vector at the previous time step, which is initially set to 0.

[0021] As a preferred embodiment of the modular chiller load balancing scheduling method based on edge computing described in this invention, step S2 involves inputting the first state vector into the load balancing agent model for processing and outputting the action vector at the current moment.

[0022] Step S2 includes steps S201, S202 and S203;

[0023] Step S201, the load balancing agent model adopts an actor-commentator architecture;

[0024] The actor network in the actor-commentator architecture consists of a series of fully connected feature extraction layers, a mean generation network, and a variance generation network connected in sequence. The commentator network in the actor-commentator architecture is used to receive the first state vector and output the corresponding value function estimate.

[0025] Feature data is obtained by extracting features from the first state vector through multiple fully connected layers;

[0026] The feature data are subjected to a first processing and a second processing, respectively.

[0027] The first processing specifically includes: performing forward propagation calculation on the feature data through a mean generation network to obtain a first calculation result; and using the tanh activation function in the mean output layer to process the first calculation result to obtain the mean of the power adjustment ratio of each refrigerator and the mean of the target temperature fine-tuning value.

[0028] The average of the power adjustment ratio and the average of the target temperature fine-tuning value constitute the first set of control parameters;

[0029] The second processing specifically includes performing a second forward propagation calculation on the feature data through a variance generation network to obtain a second calculation result, and using a softplus activation function in the variance output layer to process the second calculation result to obtain the variance of the power adjustment ratio of each refrigerator and the variance of the target temperature fine-tuning value.

[0030] The variance of the power adjustment ratio and the variance of the target temperature fine-tuning value constitute the second set of control parameters.

[0031] As a preferred embodiment of the modular chiller load balancing scheduling method based on edge computing described in this invention, in step S202, the first set of control parameters and the second set of control parameters are sampled according to the normal distribution to obtain the power adjustment ratio normal dataset and the target temperature fine-tuning value normal dataset.

[0032] The mathematical expression is,

[0033] ;

[0034] ;

[0035] in, Obeying the mean of the power adjustment ratio Variance of power adjustment ratio The normal distribution represents the power adjustment ratio of the i-th chiller. Obey the mean of the target temperature fine-tuning value Variance of target temperature fine-tuning value , which follows a normal distribution, and represents the target temperature fine-tuning value of the i-th chiller;

[0036] By sampling the normal distribution, the power adjustment ratio of each chiller can be obtained. and target temperature fine-tuning value ;

[0037] The action vector is composed of the power adjustment ratio and the target temperature fine-tuning value, and its expression is:

[0038] ;

[0039] in, This represents the action vector at the current moment. This represents the power adjustment ratio of the i-th chiller. This represents the target temperature fine-tuning value for the i-th chiller. Indicates the total number of refrigeration units;

[0040] Store the generated action vector At that time, the parameters of the actor network in the load balancing agent model are used as the parameters of the old strategy;

[0041] Step S203: Set constraints on the action vector, expressed as follows:

[0042] like ,but ;

[0043] like ,but ;

[0044] in, This indicates the maximum adjustment range of the preset power adjustment ratio. This indicates the maximum adjustment range of the preset target temperature fine-tuning value.

[0045] As a preferred embodiment of the modular chiller load balancing scheduling method based on edge computing described in this invention, the action vector is... It is converted into control commands to control the refrigeration unit and to collect post-execution data from the refrigeration unit;

[0046] Step S3 includes steps S301 and S302;

[0047] Step S301, convert the constraint-processed action vector According to the instruction mapping rules corresponding to the communication protocol of the chiller controller, the instructions are converted into a control instruction set and sent to each chiller controller through the communication network. The power adjustment ratio is then used accordingly. Adjust the operating power of the refrigeration unit and fine-tune the value according to the target temperature. Adjust the target temperature of the refrigeration unit;

[0048] Step S302: Collect post-execution data of each chiller after the action is performed. The post-execution data includes updated real-time power consumption data, ambient temperature data, regional temperature feedback data of each area, and chiller operating status data.

[0049] As a preferred embodiment of the modular chiller load balancing scheduling method based on edge computing described in this invention, step S4 involves constructing a second state vector based on the data after execution and calculating a reward function based on the second state vector.

[0050] Step S4 includes steps S401 and S402;

[0051] Step S401: Calculate the standard deviation of the load vector in the second state vector to obtain the load balancing index.

[0052] The average temperature deviation vector in the second state vector is calculated to obtain the average temperature deviation index.

[0053] The real-time total power consumption data is calculated based on the real-time power consumption data of each chiller, and the maximum total power consumption data is calculated based on the rated maximum power consumption data of each chiller. The ratio of the real-time total power consumption data to the maximum total power consumption data is then calculated as the real-time total energy consumption ratio.

[0054] As a preferred embodiment of the modular chiller load balancing scheduling method based on edge computing described in this invention, step S402 involves calculating a reward function through weighted summation based on load balancing indicators, average temperature deviation indicators, and real-time total energy consumption ratio. The expression for the reward function is as follows:

[0055] ;

[0056] in, Represents the reward function, This represents a load balancing metric. Indicates the average temperature deviation index. This represents the real-time total power consumption data. This represents the maximum total power consumption data. , and This indicates the preset weighting coefficient.

[0057] As a preferred embodiment of the modular chiller load balancing scheduling method based on edge computing described in this invention, step S5 involves updating the parameters of the load balancing agent model online using a near-end optimization strategy based on the first state vector, action vector, and reward function.

[0058] Step S5 includes steps S501, S502, S503, S504, S505 and S506.

[0059] Step S501: The advantage value is obtained by using the generalized advantage estimation method based on the first state vector, action vector and reward function.

[0060] Step S502: Record the strategy defined by the parameters of the old strategy stored in step S202 as the old strategy;

[0061] The parameters to be updated in the load balancing agent model include the parameters of the actor network and the commentator network, where the policy defined by the parameters of the actor network is denoted as the new policy.

[0062] Calculate the ratio of the joint probability density of the action vector under the new policy to the joint probability density under the old policy, denoted as the new-old policy probability ratio.

[0063] As a preferred embodiment of the modular chiller load balancing scheduling method based on edge computing described in this invention, step S503 involves calculating the strategy substitution loss based on the probability ratio of the new and old strategies, expressed as follows:

[0064] ;

[0065] in, Indicates the strategy substitutes for loss. This represents the ratio of the probability of the new strategy to the probability of the old strategy. This indicates the preset truncation parameter. Indicates the advantage value.

[0066] As a preferred embodiment of the modular chiller load balancing scheduling method based on edge computing described in this invention, step S504 involves calculating the value loss based on the first state vector St. The expression is,

[0067] ;

[0068] in, Indicates a loss of value. Represents the first state vector The value function estimate, This represents the estimated expected return calculated based on the reward function;

[0069] Step S505: The strategy substitution loss, value loss, and entropy regularization term are weighted and summed to obtain the total loss L, expressed as follows:

[0070] ;

[0071] in, Indicates the total loss. , This represents the preset weighting coefficients. This represents the entropy calculated from the policy distribution of the load balancing agent model;

[0072] Step S506: Minimize the total loss using gradient descent. Update the parameters of the actor network and the commentator network in the load balancing agent model.

[0073] The beneficial effects of this invention are as follows: By introducing an edge computing-based intelligent load balancing scheduling method, it effectively overcomes the limitations of traditional methods that rely on fixed rules and are difficult to dynamically respond to changes in the environment and load. This invention constructs a multi-dimensional state vector that integrates load, temperature deviation, and environmental conditions, and uses an agent with an actor-commentator architecture for real-time decision-making. This enables coordinated and precise adjustment of the power and temperature of each chiller. Through near-end policy optimization and online learning mechanisms, the strategy can be continuously updated during operation to adapt to dynamic operating conditions, thereby significantly improving the load balancing level and temperature control accuracy, and reducing overall energy consumption. In addition, the edge computing-based deployment ensures the real-time performance and reliability of scheduling, supports efficient collaboration among multiple chillers, and ultimately extends equipment life and improves energy efficiency while ensuring stable ambient temperature and humidity, demonstrating significant energy-saving benefits and practical engineering value. Attached Figure Description

[0074] Figure 1 This is a flowchart illustrating the steps of a modular chiller load balancing scheduling method based on edge computing, provided as an embodiment of the present invention. Detailed Implementation

[0075] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0076] Example, refer to Figure 1 As an embodiment of the present invention, a modular chiller load balancing scheduling method based on edge computing is provided, comprising the following steps:

[0077] Step S1: Construct the first state vector based on the processing data of the refrigeration machine;

[0078] Step S2: Input the first state vector into the load balancing agent model for processing, and output the action vector at the current moment;

[0079] Step S3: Convert the action vector into control commands to control the refrigerator and collect post-execution data of the refrigerator.

[0080] Step S4: Construct a second state vector based on the post-execution data, and calculate the reward function based on the second state vector;

[0081] Step S5: Based on the first state vector, action vector, and reward function, the parameters of the load balancing agent model are updated online using a proximal optimization strategy.

[0082] In one embodiment, targeting typical scenarios such as data centers, cold chain warehouses, or industrial plants where multiple chillers operate collaboratively, the core design of this embodiment lies in constructing a closed-loop, adaptive, and low-latency intelligent decision-making framework to optimize the collaborative operation of multiple chillers in real time. First, by collecting multi-dimensional data such as real-time power consumption data, set temperature data, operating status data, ambient temperature data, regional temperature feedback data, rated maximum power consumption data, and regional set temperature values ​​from the chillers, a first state vector is constructed, forming a comprehensive and structured perception of the current operating conditions, providing accurate input for intelligent decision-making. Next, the first state vector is input to a load-balancing intelligent agent model deployed on edge computing nodes for processing, outputting an action vector to achieve low-latency, localized real-time decision-making, effectively avoiding the limitations of traditional cloud-based centralized scheduling. The system addresses the response lag issue caused by signal delay. Next, action vectors are converted into control commands and sent to each chiller for execution, while post-execution data is collected to complete a closed-loop control from decision-making to execution. Based on this, a second state vector is constructed using the post-execution data, and a designed reward function is used to comprehensively evaluate load balancing, temperature control accuracy, and energy consumption, forming a quantitative evaluation mechanism for decision-making effectiveness. Finally, using a near-end strategy optimization method, the agent model's parameters are updated online based on the first state vector, action vector, and reward function, enabling it to continuously adapt to dynamically changing environments and loads, gradually improving scheduling performance. Through this coherent closed-loop process, low-latency, adaptive, and continuously optimized intelligent scheduling at the edge is achieved, significantly improving the energy efficiency balance and control reliability of the chiller cluster in dynamic environments.

[0083] Step S1: Construct the first state vector based on the processing data of the refrigeration machine;

[0084] Step S1 includes steps S101, S102 and S103;

[0085] Step S101: Obtain processing data, which includes real-time power consumption data of the chiller, set temperature data, operating status data, ambient temperature data, zone temperature feedback data, rated maximum power consumption data, and zone set temperature value.

[0086] Step S102: Calculate the load level data for each refrigeration unit and the temperature deviation data for each area based on the processing data;

[0087] The calculation process for load level data is as follows: the ratio of the real-time power consumption data of the i-th chiller to the rated maximum power consumption data is recorded as the load level data.

[0088] The absolute value of the difference between the regional temperature feedback data and the regional set temperature value for each region is recorded as the temperature deviation data.

[0089] The load level data, temperature deviation data, and ambient temperature data are normalized to obtain the load vector, temperature deviation vector, and ambient temperature vector.

[0090] Step S103: Based on the load vector, temperature deviation vector, and ambient temperature vector, a first state vector is constructed. The expression for the first state vector is:

[0091] ;

[0092] in, Represents the first state vector. Represents the load vector. Represents the temperature deviation vector. Represents the ambient temperature vector. This represents the action vector at the previous time step, which is initially set to 0.

[0093] In one embodiment, the processing data acquired in step S101 specifically includes: real-time power consumption data of each chiller (in kilowatts); set temperature data (in degrees Celsius); operating status data (0 indicates shutdown, 1 indicates operation); ambient temperature data (in degrees Celsius); and a region defined here as a physical environment space partition with independent temperature control requirements, served by multiple modular chillers. For example, in a data center scenario, this corresponds to a server rack aisle or server room; in a cold chain warehouse, it corresponds to a warehouse with a specific temperature zone; and in an industrial plant, it corresponds to a production or storage area. The acquired data also includes: real-time temperature feedback data of each controlled region (in degrees Celsius); rated maximum power consumption data of each chiller (in kilowatts); and the regional set temperature value of each controlled region (in degrees Celsius). The real-time power consumption data, set temperature data, operating status data, ambient temperature data, regional temperature feedback data, rated maximum power consumption data, and regional set temperature value are collected in real time by temperature sensors, power metering devices, and controllers deployed on-site and aggregated to the edge computing node.

[0094] In step S102, the load level data of the i-th chiller is first calculated. Specifically, the real-time power consumption data of the chiller is divided by its rated maximum power consumption data to obtain a load level value ranging from 0 to 1. The temperature deviation data is calculated by taking the absolute value of the difference between the regional temperature feedback data and the regional set temperature value for each area to obtain a non-negative temperature deviation value. Subsequently, the load level data of all chillers, the temperature deviation data of all areas, and the ambient temperature data are normalized. The maximum and minimum values ​​of each data point used for normalization are preset fixed parameters, which are determined by, for the load level data, The theoretical range [0,1] is directly used as the normalization parameter. For temperature deviation data, the upper limit of normalization is determined based on the maximum temperature deviation value that occurred in the system's historical operation, taking into account a safety margin. For example, it is set to [0,10] (unit: degrees Celsius). For ambient temperature data, the normalization range is determined based on the historical extreme meteorological values ​​of the installation location or the ambient temperature range expected to be used by the system. For example, it is set to [-10,40] (unit: degrees Celsius). The load level data, temperature deviation data, and ambient temperature data are mapped to the interval between 0 and 1 using a linear mapping formula, respectively, to obtain the load vector L, temperature deviation vector ΔT, and ambient temperature vector. ;

[0095] In step S103, the first state vector The construction method is to combine the load vector L, temperature deviation vector ΔT, and ambient temperature vector. and the action vector of the previous moment (Dimension 2m×1) Concatenate the vectors in this order to form a complete state representation vector with a total dimension of (m+n+1+2m)×1, where m represents the total number of modular refrigeration units in the system, and n represents the total number of independently controlled physical environment areas. Both m and n are positive integers, and their specific values ​​are determined according to the actual deployment scenario. For example, in a cold chain warehouse with 5 refrigeration units and 3 independent temperature zones to be monitored, m is set to 5 and n is set to 3. The previous action vector... The vector is initially set to all zeros by default.

[0096] From the perspective of the design of the steps themselves, this part constructs a unified, comparable state representation that includes time-series information through systematic data collection, calculation and normalization processing, providing a reliable and structured input for subsequent intelligent decision-making. From the core design perspective, this step, as the perception and preprocessing link of the entire scheduling process, directly supports the edge-side load balancing agent to make real-time, low-latency decisions. It is an important foundation for realizing a scheduling closed loop of dynamic response and continuous optimization, effectively improving the system's ability to integrate multi-source heterogeneous data and the accuracy of state expression.

[0097] Step S2: Input the first state vector into the load balancing agent model for processing, and output the action vector at the current moment;

[0098] Step S2 includes steps S201, S202 and S203;

[0099] Step S201: The load balancing agent model adopts an actor-commentator architecture;

[0100] The actor network in the actor-commentator architecture consists of a series of fully connected feature extraction layers, a mean generation network, and a variance generation network connected in sequence. The commentator network in the actor-commentator architecture is used to receive the first state vector and output the corresponding value function estimate.

[0101] Feature data is obtained by extracting features from the first state vector through multiple fully connected layers;

[0102] The feature data are subjected to a first processing and a second processing, respectively.

[0103] The first processing specifically includes: performing forward propagation calculation on the feature data through a mean generation network to obtain a first calculation result; and using the tanh activation function in the mean output layer to process the first calculation result to obtain the mean of the power adjustment ratio of each refrigerator and the mean of the target temperature fine-tuning value.

[0104] The average of the power adjustment ratio and the average of the target temperature fine-tuning value constitute the first set of control parameters;

[0105] The second processing specifically includes performing a second forward propagation calculation on the feature data through a variance generation network to obtain a second calculation result, and using a softplus activation function in the variance output layer to process the second calculation result to obtain the variance of the power adjustment ratio of each refrigerator and the variance of the target temperature fine-tuning value.

[0106] The variance of the power adjustment ratio and the variance of the target temperature fine-tuning value constitute the second set of control parameters.

[0107] Step S202: Sample the first set of control parameters and the second set of control parameters according to the normal distribution to obtain the normal dataset of power adjustment ratio and the normal dataset of target temperature fine-tuning value;

[0108] The mathematical expression is,

[0109] ;

[0110] ;

[0111] in, Obeying the mean of the power adjustment ratio Variance of power adjustment ratio The normal distribution represents the power adjustment ratio of the i-th chiller. Obey the mean of the target temperature fine-tuning value Variance of target temperature fine-tuning value , which follows a normal distribution, and represents the target temperature fine-tuning value of the i-th chiller;

[0112] By sampling the normal distribution, the power adjustment ratio of each chiller can be obtained. and target temperature fine-tuning value ;

[0113] The action vector is composed of the power adjustment ratio and the target temperature fine-tuning value, and its expression is:

[0114] ;

[0115] in, This represents the action vector at the current moment. This represents the power adjustment ratio of the i-th chiller. This represents the target temperature fine-tuning value for the i-th chiller. Indicates the total number of refrigeration units;

[0116] Store generated action vectors At that time, the parameters of the actor network in the load balancing agent model are used as the parameters of the old policy.

[0117] Step S203: Set constraints on the action vector, expressed as follows:

[0118] like ,but ;

[0119] like ,but ;

[0120] in, This indicates the maximum adjustment range of the preset power adjustment ratio. This indicates the maximum adjustment range of the preset target temperature fine-tuning value.

[0121] In one embodiment, in step S201, the load balancing agent model is specifically deployed on the processor of the edge computing node. Its actor-critic architecture includes two independent feedforward neural networks, an actor network and a critic network. The actor network first extracts features from the input first state vector through shared multi-layer fully connected layers to obtain abstracted feature data. Subsequently, this feature data is fed in parallel to two independent fully connected sub-networks for processing. Specifically, in the mean generation network, after the feature data undergoes forward propagation, its output is processed through a hyperbolic tangent (tanh) activation function to generate a 2m×1 mean vector. The first m elements of this vector sequentially represent the respective refrigerators. The mean of the power adjustment ratio, and the last m elements represent the mean of the target temperature fine-tuning values ​​of each refrigerator, together forming the first set of control parameters (mean vector); In the variance generation network, after the feature data is calculated through forward propagation, its output is processed by the softplus activation function (which makes the output value always positive) to generate a variance vector with the same dimension of 2m×1. The first m elements of this vector represent the variance of the power adjustment ratio of each refrigerator, and the last m elements represent the variance of the target temperature fine-tuning values ​​of each refrigerator, together forming the second set of control parameters (variance vector); At the same time, the commentator network independently receives the first state vector and processes it through its own multi-layer fully connected layers, finally outputting a scalar value representing the current state value estimate;

[0122] In step S202, based on the first set of control parameters (mean) and the second set of control parameters (variance) generated in step S201, sampling is performed from the probability distribution defined by the first set of control parameters and the second set of control parameters to generate specific control actions. Specifically, the power adjustment ratio of the i-th refrigerator is randomly sampled from a normal distribution constructed with its corresponding mean and variance parameters; its target temperature fine-tuning value is randomly sampled from a normal distribution constructed with another set of corresponding mean and variance parameters. The power adjustment ratios and target temperature fine-tuning values ​​of all m refrigerators are arranged in an alternating order to form the action vector at the current moment. After the action vector is generated, the values ​​of all adjustable parameters in the actor network at this moment are immediately recorded and stored. This set of parameter values ​​is defined as the parameters of the "old strategy".

[0123] In step S203, to ensure the feasibility of the control command, amplitude constraints are applied to each element in the action vector. Specifically, the maximum allowable amplitude of the power adjustment ratio is set to [value missing]. The maximum allowable range for the target temperature fine-tuning value is , and These are constants pre-set based on the specific safe operating parameters of the refrigeration equipment and the accuracy indicators of the temperature control system, where, Set to a percentage of the chiller's rated power, such as 10%, to limit the range of power changes in a single adjustment and prevent damage to the equipment; Set to an absolute value, such as 0.5°C, to limit the adjustment range of the temperature setpoint in a single operation, thus ensuring response speed while preventing system oscillation. and The specific value needs to be determined during initialization based on the allowable operating range of the chiller model and the temperature stability requirements of the controlled environment. Specifically, each element in the action vector is judged and restricted. If its absolute value exceeds the corresponding maximum amplitude, its value is set to the maximum amplitude and the original sign is retained.

[0124] From the design of the steps themselves, this step generates control parameters through a deterministic neural network, generates specific actions based on parameterized distribution sampling, and then applies physical constraints. This process achieves the generation of precise control instructions that both conform to the optimization direction and ensure the safe operation of the equipment while considering environmental uncertainties. The mean generation uses the tanh function to ensure a reasonable output range, the variance generation uses the softplus function to ensure effective quantification of exploration, and the final hard constraints directly ensure execution safety.

[0125] From a design perspective, step S2 is the core real-time decision-making stage of the entire scheduling method. By deploying the complete intelligent agent model on edge nodes and utilizing their efficient forward computing capabilities, the entire process from state perception to control command generation is completed directly on-site at the device, achieving millisecond-level low-latency decision-making. This localized decision-making mode completely avoids the communication latency and uncertainty brought about by traditional cloud-based centralized scheduling. It is the key mechanism for this solution to achieve high real-time performance and high reliability adaptive scheduling, providing the core driver for the accurate execution of control commands and continuous optimization of strategies in subsequent steps.

[0126] Step S3, convert the action vector It is converted into control commands to control the refrigeration unit and to collect post-execution data from the refrigeration unit;

[0127] Step S3 includes steps S301 and S302;

[0128] Step S301, convert the constraint-processed action vector According to the instruction mapping rules corresponding to the communication protocol of the chiller controller, the instructions are converted into a control instruction set and sent to each chiller controller through the communication network. The control instruction is then adjusted according to the power ratio. Adjust the operating power of the refrigeration unit and fine-tune the value according to the target temperature. Adjust the target temperature of the refrigeration unit;

[0129] Step S302: Collect post-execution data of each chiller after the action is executed. The post-execution data includes updated real-time power consumption data, ambient temperature data, regional temperature feedback data of each area, and chiller operating status data.

[0130] In one embodiment, in step S301, firstly, the action vector after the constraint processing in step S203 is... This is converted into control commands that can directly drive the refrigeration equipment. Specifically, based on the pre-configured communication protocol between the edge computing node and each refrigeration controller, such as the Modbus TCP / IP protocol, each power adjustment ratio in the action vector is converted into control commands. and target temperature fine-tuning value According to the data format specified in the protocol and the predefined mapping relationship that associates specific control parameters, such as power setpoints and temperature setpoints, with specific communication register addresses, the data is converted into corresponding power setpoint adjustment instructions and temperature setpoint adjustment instructions, which together constitute a control instruction set. Subsequently, this control instruction set is sent in real-time to the corresponding chiller controller via the local industrial Ethernet. Upon receiving the instructions, each controller executes its internally embedded control program, which parses the instructions and drives power and temperature regulation mechanisms, such as compressors, fans, and valves. Based on the power adjustment instructions, the settings are adjusted according to the chiller's rated maximum power and temperature setpoint. The product of these factors is used to adjust the setpoint of its operating power; based on the temperature fine-tuning command, the value is increased from the original target temperature setpoint. To update its target temperature setpoint;

[0131] In step S302, after the control command is issued and a preset short delay has elapsed, for example, a sampling interval matching the control cycle, the edge computing node synchronously collects data from all refrigerators after executing the new command through the sensor network and controller interface. This data is the post-execution data. The preset short delay is set to be equal to the system's control cycle T to ensure that feedback data is collected after a complete decision-execution-perception closed loop. For example, T is set to 30 seconds. The control cycle T is set based on the thermodynamic inertia of the refrigerator and the controller response time, aiming to balance the real-time performance of the control with the stability of the system, avoiding data fluctuations due to excessively frequent data collection or affecting the response speed due to excessively long intervals. This data includes the updated real-time power consumption data (unit: kilowatts) of each refrigerator, the current ambient temperature data (unit: degrees Celsius), the latest regional temperature feedback data (unit: degrees Celsius) of each controlled area, and the current operating status data of each refrigerator, such as the status indicators of running, standby, and fault.

[0132] From the design of the steps themselves, this step completes the reliable conversion and safe execution from abstract decision-making instructions to specific equipment control actions. By explicitly relying on existing industrial communication protocols and mapping relationships, it ensures that control instructions can be accurately identified and executed by heterogeneous chiller controllers, realizing precise direct control of power and temperature settings. The synchronous collection of data after execution provides real and timely feedback information for closed-loop evaluation.

[0133] From a core design perspective, step S3 is a crucial bridge in the decision-making-execution-perception closed loop. By completing the localized conversion and distribution of instructions at the edge, the control loop delay is greatly shortened, ensuring the real-time performance of scheduling instructions. At the same time, precise execution and synchronous data collection enable continuous updates to the system state, providing accurate and reliable data input for the subsequent quantitative evaluation of effects in step S4 and the online optimization of strategies in step S5. It is a necessary link to maintain the stable operation and continuous improvement of the entire adaptive scheduling system.

[0134] Step S4: Construct a second state vector based on the post-execution data, and calculate the reward function based on the second state vector;

[0135] Step S4 includes steps S401 and S402;

[0136] Step S401: Calculate the standard deviation of the load vector in the second state vector to obtain the load balancing index.

[0137] The average temperature deviation vector in the second state vector is calculated to obtain the average temperature deviation index.

[0138] The real-time total power consumption data is calculated based on the real-time power consumption data of each chiller, and the maximum total power consumption data is calculated based on the rated maximum power consumption data of each chiller. The ratio of the real-time total power consumption data to the maximum total power consumption data is then calculated as the real-time total energy consumption ratio.

[0139] Step S402: Based on the load balancing index, average temperature deviation index, and real-time total energy consumption ratio, the reward function is calculated by weighted summation. The expression for the reward function is:

[0140] ;

[0141] in, Represents the reward function, This represents a load balancing metric. Indicates the average temperature deviation index. This represents the real-time total power consumption data. This represents the maximum total power consumption data. , and This indicates the preset weighting coefficient.

[0142] In one embodiment, in step S401, based on the post-execution data collected in step S302, a second state vector is constructed using the same method as in step S103. Specifically, using updated real-time power consumption data, regional temperature feedback data, and ambient temperature data, the load level data of each chiller and the temperature deviation data of each region are recalculated. After the same normalization processing, new load vectors, temperature deviation vectors, and ambient temperature vectors are obtained, which are then compared with the action vector at the current moment. The data is concatenated to form a second state vector. Then, the standard deviation of all elements of the load vector in the second state vector is calculated. The standard deviation value is the load balance index. The smaller the value, the more uniform the load distribution. The arithmetic mean of all elements of the temperature deviation vector is calculated. The mean value is the average temperature deviation index. The smaller the value, the higher the overall temperature control accuracy. At the same time, the real-time total power consumption data is obtained by summing the real-time power consumption data of each chiller. The maximum total power consumption data is obtained by summing the rated maximum power consumption data of each chiller. The ratio of the real-time total power consumption data to the maximum total power consumption data is the real-time total energy consumption ratio, which directly reflects the relative level of the current total power consumption.

[0143] In step S402, based on the load balancing index, average temperature deviation index, and real-time total energy consumption ratio calculated in step S401, a reward function is calculated by weighted summation. The specific form of the reward function is the reciprocal of the load balancing index multiplied by the weighting coefficient. Add the reciprocal of the average temperature deviation index multiplied by the weighting coefficient. In addition, the real-time total energy consumption ratio multiplied by the weighting factor Among them, the weighting coefficient , and This is a constant preset based on the priority given to load balancing, temperature control accuracy, and energy saving in specific application scenarios, and is set to meet the following requirements. The value of this non-negative real number needs to be determined based on the optimization objective during system initialization. For example, in scenarios where temperature control accuracy is the primary objective (such as data centers), it might be set to [value missing]. = In scenarios where energy conservation is the primary objective, the following settings are adopted: = In scenarios where load balancing is emphasized to extend equipment lifespan, the following settings are provided: = The larger the value of the reward function, the better the overall scheduling effect.

[0144] From the perspective of the design of the steps themselves, this step transforms multi-dimensional and heterogeneous post-execution data into three calculable and quantifiable core performance indicators: load balancing, temperature deviation, and total energy consumption, and finally integrates them into a scalar reward value. This establishes an objective and unified real-time evaluation system for control effect. The use of the reciprocal form means that the smaller the indicator value, the better the effect, and the greater the reward value it contributes.

[0145] From a core design perspective, step S4 is the evaluator for the entire adaptive scheduling closed loop. It quickly quantifies the effect at the edge and generates the reward function value. The first state vector generated in step S2 Action vectors Together, these constitute a complete set of empirical data, which serves as the direct basis for driving the online optimization of the strategy in step S5. This enables the agent to automatically and continuously adjust its decision-making strategy based on the actual effects of historical decisions, ultimately allowing the overall system performance to continuously approach the comprehensive goal of load balancing, precise temperature control, and energy saving.

[0146] Step S5: Based on the first state vector, action vector, and reward function, the parameters of the load balancing agent model are updated online using a near-end optimization strategy;

[0147] Step S5 includes steps S501, S502, S503, S504, S505 and S506.

[0148] Step S501: The advantage value is calculated using the generalized advantage estimation method based on the first state vector, action vector, and reward function.

[0149] Step S502: Record the strategy defined by the parameters of the old strategy stored in step S202 as the old strategy;

[0150] In the load balancing agent model, the parameters to be updated include the parameters of the actor network and the commentator network, where the policy defined by the parameters of the actor network is denoted as the new policy.

[0151] The ratio of the joint probability density of the action vector under the new policy to that under the old policy is calculated and denoted as the probability ratio of the new and old policies.

[0152] Step S503: Calculate the policy substitution loss based on the probability ratio of the old and new policies, expressed as follows:

[0153] ;

[0154] in, Indicates the strategy substitutes for loss. This represents the ratio of the probability of the new strategy to the probability of the old strategy. This indicates the preset truncation parameter. Indicates the advantage value.

[0155] Step S504: Calculate the value loss based on the first state vector St. The expression is,

[0156] ;

[0157] in, Indicates a loss of value. Represents the first state vector The value function estimate, This represents the estimated expected return calculated based on the reward function;

[0158] Step S505: The policy substitution loss, value loss, and entropy regularization term are weighted and summed to obtain the total loss L, expressed as follows:

[0159] ;

[0160] in, Indicates the total loss. , This represents the preset weighting coefficients. This represents the entropy calculated by the policy distribution of the load balancing agent model.

[0161] Step S506: Minimize the total loss using gradient descent. Update the parameters of the actor network and the commentator network in the load balancing agent model.

[0162] In one embodiment, in step S501, based on the first state vector Action vectors and reward function Calculate the dominance value The advantage value is used to measure the quality of a particular action relative to the average level. It is calculated based on the value estimate of the reward sequence and the commentator network output. It is estimated by the temporal difference method to reduce the estimation variance, thereby more stably evaluating the value of the action.

[0163] In step S502, the generated action vector stored in step S202 is... The policy defined by the parameters of the current actor network is denoted as the old policy. The parameters to be optimized in the load-balanced agent model include all adjustable parameters of both the actor network and the commentator network. The policy defined by the actor network parameters before optimization is denoted as the new policy, i.e., the policy to be updated. The action vector is then calculated. The ratio of the joint probability density under the new strategy to the joint probability density under the old strategy is called the new-old strategy probability ratio. This probability ratio reflects the difference in preference between the new and old strategies for the same decision.

[0164] In step S503, based on the ratio of the probability of the new strategy to the old strategy and the advantage value... Calculate the alternative loss strategy The core of its calculation is to first restrict the probability ratio to an interval centered at 1 with a radius of ε, i.e., a truncation operation. Then, the probability ratios before and after truncation are multiplied by the advantage value respectively, and the smaller of the two values ​​is taken as the loss. Here, ε is a preset small normal number, such as 0.2, as the truncation parameter. This value is the standard hyperparameter of the proximal policy optimization (PPO) algorithm in the field of reinforcement learning. It is preset according to the algorithm stability theory and is used to limit the update range of a single policy iteration within a reasonable range.

[0165] In step S504, based on the first state vector Calculate value loss Value loss is the value estimate of the current state by the commentator network. The square of the difference between the target value estimate and the target value estimate (Vtarget), where Vtarget is the target value estimate. Based on the reward function value obtained at the current moment The calculation of the next state by the commentator network is used to provide a more accurate learning target for the commentator network.

[0166] In step S505, the policy is replaced with the loss. Value loss And an entropy regularization term We perform a weighted summation to obtain the total loss L used for parameter updates, where the entropy regularization term is calculated from the current actor network policy, i.e., the probability distribution, and is used to measure the randomness and exploration capability of the policy. , The preset weighting coefficients are the weights of the value loss and the entropy regularization term, respectively. Set between 0.5 and 2.0. Set between 0.01 and 0.1, for example, a common setting is [ , ]=[1.0,0.01], and The specific numerical basis is preset based on the trade-off between the accuracy of strategy updates, the importance of value estimation, and the need for exploration. In the period after the start of operation, for example, the first N control cycles, where N is 1000, the adjustment needs are determined by calculating the statistical characteristics of the probability ratio of the new and old strategies or evaluating the performance improvement of the reward function Rt. The adjustment follows clear rules. Specifically, if the calculated ratio value falls within the preset reasonable range [1-ε, 1+ε] in the most recent h control cycles, for example, h=100, and the proportion of values ​​outside the range [0.8, 1.2] exceeds a threshold O, for example, O=0.3, then it is determined that the update magnitude is continuously too large, and the α value is multiplied by a decay coefficient less than 1, for example, 0.8, to reduce it.

[0167] If within M consecutive control cycles, for example, M=200, the reward function The moving average is calculated using the most recent W values ​​as the calculation window. For example, if the absolute change of W=100 is less than a preset minimum threshold δ, such as δ=0.001, then performance improvement is considered stagnant. The value is increased by multiplying it by an enhancement factor greater than 1, such as 1.5;

[0168] After each adjustment, the coefficients α and β are updated within their defined range, i.e., α∈[0.5,2.0] and β∈[0.01,0.1].

[0169] In step S506, the gradient descent method is used to calculate the gradient of the total loss L with respect to the parameters of the actor network and the commentator network in the load balancing agent model, and the parameters are slightly adjusted in the opposite direction of the gradient to minimize the total loss L, thereby completing the online update of the model parameters.

[0170] From the perspective of the design of the steps themselves, this step constructs a composite loss function that integrates policy performance (advantage value), policy update stability (truncation probability ratio), value estimation accuracy (value loss), and policy exploration (entropy regularization term), and optimizes it using gradient descent. This achieves a complete, stable, and efficient online policy iteration mechanism. The weighted synthesis and truncation operations of each loss term jointly ensure that the update process moves in the direction of improving performance while avoiding policy collapse due to excessive single-step updates.

[0171] From a core design perspective, step S5 is the learning and evolution engine of the entire adaptive scheduling method. It runs on edge computing nodes and uses locally generated real-time state-action-reward data to directly perform incremental optimization on the agent model without relying on the cloud or historical datasets. This on-site online update mechanism enables the scheduling strategy to continuously adapt to the dynamically changing operating environment and load demand. It is the core of achieving long-term autonomous optimization of the system and continuously improving energy efficiency balance and temperature control accuracy. It is also a key feature that distinguishes it from traditional static or offline optimization scheduling schemes.

[0172] This invention introduces an edge computing-based intelligent load balancing scheduling method, effectively overcoming the limitations of traditional methods that rely on fixed rules and struggle to dynamically respond to changes in environment and load. This invention constructs a multi-dimensional state vector integrating load, temperature deviation, and environmental conditions, and employs an agent-commentator architecture for real-time decision-making. This enables coordinated and precise adjustment of the power and temperature of each chiller. Through near-end policy optimization and online learning mechanisms, the strategy is continuously updated during operation, adapting to dynamic operating conditions, thereby significantly improving load balancing and temperature control accuracy, and reducing overall energy consumption. Furthermore, the edge computing-based deployment ensures the real-time performance and reliability of scheduling, supports efficient collaboration among multiple chillers, and ultimately extends equipment life and improves energy efficiency while maintaining stable environmental temperature and humidity, demonstrating significant energy-saving benefits and practical engineering value.

[0173] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0174] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the protection scope of the present invention.

Claims

1. A modular chiller load balancing scheduling method based on edge computing, characterized in that, Includes the following steps, Step S1: Construct the first state vector based on the processing data of the refrigeration machine; Step S2: Input the first state vector into the load balancing agent model for processing, and output the action vector at the current moment; Step S3: Convert the motion vector into control commands, execute control on the refrigerator, and collect post-execution data of the refrigerator; Step S4: Construct a second state vector based on the post-execution data, and calculate the reward function based on the second state vector; Step S5: Based on the first state vector, action vector, and reward function, the parameters of the load balancing agent model are updated online using a proximal optimization strategy.

2. The modular chiller load balancing scheduling method based on edge computing as described in claim 1, characterized in that, Step S1: Construct the first state vector based on the processing data of the refrigeration machine; Step S1 includes steps S101, S102 and S103; Step S101: Obtain processing data, which includes real-time power consumption data of the chiller, set temperature data, operating status data, ambient temperature data, zone temperature feedback data, rated maximum power consumption data, and zone set temperature value. Step S102: Calculate the load level data for each refrigeration unit and the temperature deviation data for each area based on the processing data; The calculation process for load level data is as follows: the ratio of the real-time power consumption data of the i-th chiller to the rated maximum power consumption data is recorded as the load level data. The absolute value of the difference between the regional temperature feedback data and the regional set temperature value for each region is recorded as the temperature deviation data. The load level data, temperature deviation data, and ambient temperature data are normalized to obtain the load vector, temperature deviation vector, and ambient temperature vector. Step S103: Based on the load vector, temperature deviation vector, and ambient temperature vector, a first state vector is constructed. The expression for the first state vector is: ; in, Represents the first state vector. Represents the load vector. Represents the temperature deviation vector. Represents the ambient temperature vector. This represents the action vector at the previous time step, which is initially set to 0.

3. The modular chiller load balancing scheduling method based on edge computing as described in claim 2, characterized in that, Step S2: Input the first state vector into the load balancing agent model for processing, and output the action vector at the current moment; Step S2 includes steps S201, S202 and S203; Step S201, the load balancing agent model adopts an actor-commentator architecture; The actor network in the actor-commentator architecture consists of a series of fully connected feature extraction layers, a mean generation network, and a variance generation network connected in sequence. The commentator network in the actor-commentator architecture is used to receive the first state vector and output the corresponding value function estimate. Feature data is obtained by extracting features from the first state vector through multiple fully connected layers; The feature data are subjected to a first processing and a second processing, respectively. The first processing specifically includes: performing forward propagation calculation on the feature data through a mean generation network to obtain a first calculation result; and using the tanh activation function in the mean output layer to process the first calculation result to obtain the mean of the power adjustment ratio of each refrigerator and the mean of the target temperature fine-tuning value. The average of the power adjustment ratio and the average of the target temperature fine-tuning value constitute the first set of control parameters; The second processing specifically includes performing a second forward propagation calculation on the feature data through a variance generation network to obtain a second calculation result, and using a softplus activation function in the variance output layer to process the second calculation result to obtain the variance of the power adjustment ratio of each refrigerator and the variance of the target temperature fine-tuning value. The variance of the power adjustment ratio and the variance of the target temperature fine-tuning value constitute the second set of control parameters.

4. The modular chiller load balancing scheduling method based on edge computing as described in claim 3, characterized in that, Step S202: Sample the first set of control parameters and the second set of control parameters according to the normal distribution to obtain the normal dataset of power adjustment ratio and the normal dataset of target temperature fine-tuning value; The mathematical expression is, ; ; in, Obeying the mean of the power adjustment ratio Variance of power adjustment ratio The normal distribution represents the power adjustment ratio of the i-th chiller. Obey the mean of the target temperature fine-tuning value Variance of target temperature fine-tuning value , which follows a normal distribution, and represents the target temperature fine-tuning value of the i-th chiller; By sampling the normal distribution, the power adjustment ratio of each chiller can be obtained. and target temperature fine-tuning value ; The action vector is composed of the power adjustment ratio and the target temperature fine-tuning value, and its expression is: ; in, This represents the action vector at the current moment. This represents the power adjustment ratio of the i-th chiller. This represents the target temperature fine-tuning value for the i-th chiller. Indicates the total number of refrigeration units; Store the generated action vector At that time, the parameters of the actor network in the load balancing agent model are used as the parameters of the old strategy; Step S203: Set constraints on the action vector, expressed as follows: like ,but ; like ,but ; in, This indicates the maximum adjustment range of the preset power adjustment ratio. This indicates the maximum adjustment range of the preset target temperature fine-tuning value.

5. The modular chiller load balancing scheduling method based on edge computing as described in claim 4, characterized in that, Step S3, convert the action vector It is converted into control commands to control the refrigeration unit and to collect post-execution data from the refrigeration unit; Step S3 includes steps S301 and S302; Step S301, the constraint-processed action vector According to the instruction mapping rules corresponding to the communication protocol of the chiller controller, the instructions are converted into a control instruction set and sent to each chiller controller through the communication network. The power adjustment ratio is then used accordingly. Adjust the operating power of the refrigeration unit and fine-tune the value according to the target temperature. Adjust the target temperature of the refrigeration unit; Step S302: Collect post-execution data of each chiller after the action is performed. The post-execution data includes updated real-time power consumption data, ambient temperature data, regional temperature feedback data of each area, and chiller operating status data.

6. The modular chiller load balancing scheduling method based on edge computing as described in claim 5, characterized in that, Step S4: Construct a second state vector based on the post-execution data, and calculate the reward function based on the second state vector; Step S4 includes steps S401 and S402; Step S401: Calculate the standard deviation of the load vector in the second state vector to obtain the load balancing index. The average temperature deviation vector in the second state vector is calculated to obtain the average temperature deviation index. The real-time total power consumption data is calculated based on the real-time power consumption data of each chiller, and the maximum total power consumption data is calculated based on the rated maximum power consumption data of each chiller. The ratio of the real-time total power consumption data to the maximum total power consumption data is then calculated as the real-time total energy consumption ratio.

7. The modular chiller load balancing scheduling method based on edge computing as described in claim 6, characterized in that, Step S402: Based on the load balancing index, average temperature deviation index, and real-time total energy consumption ratio, the reward function is calculated by weighted summation. The expression for the reward function is: ; in, Represents the reward function, This represents a load balancing metric. Indicates the average temperature deviation index. This represents the real-time total power consumption data. This represents the maximum total power consumption data. , and This indicates the preset weighting coefficient.

8. The modular chiller load balancing scheduling method based on edge computing as described in claim 7, characterized in that, Step S5: Based on the first state vector, action vector, and reward function, the parameters of the load balancing agent model are updated online using a near-end optimization strategy; Step S5 includes steps S501, S502, S503, S504, S505 and S506. Step S501: The advantage value is obtained by using the generalized advantage estimation method based on the first state vector, action vector and reward function. Step S502: Record the strategy defined by the parameters of the old strategy stored in step S202 as the old strategy; The parameters to be updated in the load balancing agent model include the parameters of the actor network and the commentator network, where the policy defined by the parameters of the actor network is denoted as the new policy. Calculate the ratio of the joint probability density of the action vector under the new policy to the joint probability density under the old policy, denoted as the new-old policy probability ratio.

9. The modular chiller load balancing scheduling method based on edge computing as described in claim 8, characterized in that, Step S503: Calculate the policy substitution loss based on the probability ratio of the old and new policies, expressed as follows: ; in, Indicates the strategy substitutes for loss. This represents the ratio of the probability of the new strategy to the probability of the old strategy. This indicates the preset truncation parameter. Indicates the advantage value.

10. The modular chiller load balancing scheduling method based on edge computing as described in claim 9, characterized in that, Step S504: Calculate the value loss based on the first state vector St. The expression is, ; in, Indicates a loss of value. Represents the first state vector The value function estimate, This represents the estimated expected return calculated based on the reward function; Step S505: The strategy substitution loss, value loss, and entropy regularization term are weighted and summed to obtain the total loss L, expressed as follows: ; in, Indicates the total loss. , This represents the preset weighting coefficients. This represents the entropy calculated from the policy distribution of the load balancing agent model; Step S506: Minimize the total loss using gradient descent. Update the parameters of the actor network and the commentator network in the load balancing agent model.

Images