A hot water ball tank electric heating intelligent control system and method

Abstract

The present application relates to electric heating control technical field, disclose a kind of hot water ball tank electric heating intelligent control system and method, its control method includes the following steps: collecting the heterogeneous element parameter and power grid state data of hot water ball tank heating system, generates heterogeneous element-power grid state feature matrix;Utilize graph convolutional neural network analysis element space topology data, calculate the dynamic thermal coupling coefficient between elements under different power grid disturbance mode;Each heating element is modeled as an independent reinforcement learning agent.The present application adopts MADDPG multi-agent collaborative learning architecture, dynamically calculates the thermal coupling coefficient between elements by graph convolutional neural network, accurately models the change law of spatial heat conduction relationship with power grid disturbance, searches the optimal topology reconfiguration scheme in real time by evolutionary algorithm, so that the element collaboration relationship can be dynamically adjusted according to the power grid disturbance characteristics, thereby solving the collaborative control failure problem of heterogeneous heating elements under power grid disturbance.

Description

Technical Field

[0001] This invention relates to the field of electric heating control, and more specifically, to an intelligent control system and method for electric heating of a hot water spherical tank. Background Technology

[0002] In the field of temperature control for industrial hot water spherical tanks, multiple electric heating elements of different power levels are typically configured for coordinated heating. However, in actual industrial environments, the power grid load fluctuates frequently, resulting in significant variations in grid quality, including disturbances such as frequency drift, harmonic interference, and voltage fluctuations. Traditional control methods, employing fixed component connection topologies and static control parameters, cannot adapt to the dynamic changes in power grid disturbances. Particularly when multiple heterogeneous heating elements need to work collaboratively, a single centralized controller struggles to effectively coordinate the complex interactions between the elements, easily leading to overload damage or control failure of some elements under disturbance conditions.

[0003] Therefore, there is an urgent need for an intelligent control method that can dynamically adjust the cooperative relationship of components according to the characteristics of power grid disturbances. Summary of the Invention

[0004] This invention provides an intelligent control system and method for electric heating of hot water spherical tanks, solving the technical problem in related technologies where the temperature control of spherical tanks cannot dynamically adjust the cooperative relationship of components according to the characteristics of power grid disturbances.

[0005] This invention provides an intelligent control method for electric heating of a hot water spherical tank. The method implements control based on a hot water spherical tank heating system composed of several heating elements and the connected power grid. The several heating elements in the hot water spherical tank heating system form a heating circuit. The control method includes the following steps:

[0006] S100: Collect heterogeneous component parameters and power grid status data of the hot water spherical tank heating system, and generate heterogeneous component-power grid status feature matrix;

[0007] S200 uses graph convolutional neural networks to analyze the spatial topology data of components and calculate the dynamic thermal coupling coefficient between components under different power grid disturbance modes.

[0008] S300 models each heating element as an independent reinforcement learning agent, and trains the Actor network of each agent and the shared Critic network based on the multi-agent deep deterministic policy gradient algorithm.

[0009] S400 uses an evolutionary algorithm to search for the optimal topology connection scheme based on real-time grid disturbance characteristics and component status, generating a set of candidate reconfiguration strategies.

[0010] The S500 inputs the topology reconfiguration strategy and component states into the Actor network of each agent, and outputs the cooperative power allocation scheme and PID parameter adjustment.

[0011] Furthermore, the parameters of the heterogeneous components include the rated power, aging coefficient, and spatial coordinate data of each heating element;

[0012] The power grid status data includes the power grid's frequency offset, harmonic content, and voltage fluctuation amplitude.

[0013] Furthermore, the processing steps of a graph convolutional neural network include:

[0014] Calculate the adjacency matrix based on the spatial location of the components;

[0015] Update node features using graph convolution operations;

[0016] The final output layer generates a thermal coupling coefficient matrix.

[0017] Furthermore, an independent Actor network is configured for each heating element to generate control actions for the heating element;

[0018] Configure a shared Critic network to evaluate the value of joint actions of all agents;

[0019] Network parameters are trained using a multi-agent deep deterministic policy gradient algorithm.

[0020] Furthermore, the training process of the multi-agent deep deterministic policy gradient algorithm includes:

[0021] Each agent's Actor network takes local observations and thermal coupling coefficients as inputs and outputs actions.

[0022] The shared Critic network takes the observations and actions of all agents as input and outputs action values.

[0023] The network parameters are updated by minimizing the TD error and storing the transition tuples in the experience replay buffer.

[0024] Furthermore, the Critic network employs a multilayer perceptron structure, with the hidden layer containing three fully connected layers. The first two layers each contain 512 neurons using the ReLU activation function, and the third layer contains 256 neurons. The output layer is a fully connected layer with an output dimension of 1, representing the value estimate of the joint state-action pairs of all agents.

[0025] Furthermore, the input of the multi-agent deep deterministic policy gradient algorithm includes vectors of temperature deviation, power state, and thermal coupling coefficients for each agent, and the output is a four-dimensional control vector containing the power allocation ratio and the incremental adjustment values ​​of the three PID parameters.

[0026] Furthermore, the current power grid disturbance characteristics and component operating status are input into the evolutionary algorithm, and the optimal cooperative topology among components is searched through population evolution operations, outputting a strategy set containing multiple candidate topology reconfiguration schemes;

[0027] Population evolution operations using genetic algorithms include the following three operations:

[0028] The selection operation uses a tournament selection method, randomly selecting 3 individuals from the population for comparison, and selecting the individual with the highest fitness.

[0029] The crossover operation uses a single-point crossover method, performing crossover on the chromosome matrix row by row;

[0030] The mutation operation uses bit-flip mutation, flipping each bit in the chromosome with a probability of 0.05. After mutation, matrix symmetry must be ensured. The population size is set to 50, the number of generations is 100, and the convergence criterion is that the improvement in optimal fitness is less than [a certain value] for 20 consecutive generations. .

[0031] Furthermore, the Actor network of each agent receives the selected topology reconfiguration strategy and the current element state information, calculates the power distribution ratio that the heating element should bear, and generates the corresponding PID controller parameter adjustment, including the proportional coefficient increment, integral coefficient increment, and derivative coefficient increment.

[0032] This invention also proposes an intelligent control system for electric heating of a hot water spherical tank, comprising: a data acquisition module, a thermal coupling analysis module, an intelligent control module, a topology optimization module, and a control execution module;

[0033] The data acquisition module inputs the processed feature matrix into the thermal coupling analysis module; the thermal coupling analysis module inputs the thermal coupling coefficient into the intelligent control module and the topology optimization module; the topology optimization module inputs the optimal topology scheme into the intelligent control module; the intelligent control module outputs control commands to the control execution module; and the control execution module realizes the coordinated control of the heating element.

[0034] The beneficial effects of this invention are as follows:

[0035] This invention employs the MADDPG multi-agent collaborative learning architecture, enabling each heating element to act as an independent agent and make autonomous decisions based on local observations and global feedback, thus overcoming the computational bottleneck of traditional centralized controllers when dealing with complex interactions among multiple elements.

[0036] By dynamically calculating the thermal coupling coefficient between components using graph convolutional neural networks, the variation law of spatial heat conduction relationship with power grid disturbances was accurately modeled, overcoming the limitation of fixed coupling models that cannot adapt to dynamic environments.

[0037] By using an evolutionary algorithm to search for the optimal topology reconfiguration scheme in real time, the cooperative relationship between components can be dynamically adjusted according to the characteristics of power grid disturbances. This overcomes the fragility of static topology structures and solves the problem of cooperative control failure of heterogeneous heating elements under power grid disturbances. Attached Figure Description

[0038] Figure 1 This is a flowchart of an intelligent control method for electric heating of a hot water spherical tank proposed in this invention;

[0039] Figure 2 This is a structural block diagram of an intelligent control system for electric heating of a hot water spherical tank proposed in this invention.

[0040] In the diagram: 101, Data Acquisition Module; 102, Thermal Coupling Analysis Module; 103, Intelligent Control Module; 104, Topology Optimization Module; 105, Control Execution Module. Detailed Implementation

[0041] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, features described in some examples may be combined in other examples.

[0042] like Figure 1 As shown, a smart control method for electric heating of a hot water spherical tank includes the following steps:

[0043] S100: Collect heterogeneous component parameters and power grid status data of the hot water spherical tank heating system, and generate heterogeneous component-power grid status feature matrix;

[0044] The system acquires the rated power, aging coefficient, and spatial coordinates of each heating element, while simultaneously collecting real-time information on the power grid's frequency offset, harmonic content, and voltage fluctuation amplitude. The collected element parameters and power grid status data are then normalized to generate a heterogeneous element-power grid status feature matrix.

[0045] The data preprocessing includes: normalizing the power data to the range of [0,1] using the maximum-minimum method; normalizing the spatial coordinates using the Z-score method; normalizing the frequency offset and voltage fluctuation amplitude using the maximum-minimum method; and ensuring that the aging coefficient and harmonic content are dimensionless proportional values ​​and do not require preprocessing.

[0046] Furthermore, the aging coefficient of the heating element The calculation employs a comprehensive evaluation method based on runtime and performance degradation:

[0047] ;

[0048] in, Let be the aging coefficient of the i-th heating element, with a value ranging from [0,1]. A smaller value indicates a more severe degree of aging. For components Cumulative runtime, For component design life, For the current power loss, Rated power, and These are the time degradation coefficient and the performance degradation coefficient, respectively, obtained by fitting historical statistical data, with typical values ​​of [value missing]. , .

[0049] It should be noted that the generation of the heterogeneous component-grid state characteristic matrix includes:

[0050] The first The feature vector of each heating element is represented as follows: ,in Rated power, The aging factor is... Spatial location coordinates;

[0051] The power grid state vector is represented as ,in This is the frequency offset value. Total harmonic distortion (THD) This refers to the voltage fluctuation amplitude.

[0052] A complete feature matrix is ​​formed by data splicing. .

[0053] S200 uses graph convolutional neural networks to analyze the spatial topology data of components and calculate the dynamic thermal coupling coefficient between components under different power grid disturbance modes.

[0054] Each heating element is treated as a graph node, and the spatial distance between elements is used as the edge weight. The input is fed into a graph convolutional neural network, and combined with the current power grid disturbance characteristics, the dynamic thermal coupling coefficient matrix between each pair of elements is output.

[0055] It should be noted that the processing steps of a graph convolutional neural network include:

[0056] Calculate the adjacency matrix based on the spatial location of components:

[0057] ;

[0058] in, Calculate the adjacency matrix for the spatial location of the components. and Components and The position vector, For spatial scale parameters, The square of the Euclidean distance;

[0059] Through graph convolution operations Update node features, where Let l be the node feature matrix of the l-th layer. To add self-connected adjacency matrices, This is the corresponding degree matrix. For the first The weight matrix of the layer, The ReLU activation function is defined as follows: ;

[0060] The final output layer generates a thermal coupling coefficient matrix. .

[0061] Furthermore, spatial scale parameters The range of values ​​is constrained to Within the interval, The maximum spatial distance between any two components in the system is defined by the spatial scale parameter. The range of values ​​for this parameter ensures that adjacent components can effectively sense each other's thermal conduction effects, while avoiding the loss of local spatial features due to excessive smoothing.

[0062] The input layer of the aforementioned graph convolutional neural network receives the node feature matrix. ,in For the number of heating elements, For each element, a feature dimension is defined, including the element's rated power, aging factor, spatial location coordinates, and current grid disturbance characteristics; the output layer is a fully connected layer that convolves the hidden representation of the last graph layer. The mapping is to a thermal coupling coefficient matrix, which is:

[0063] ;

[0064] in, This is the thermal coupling coefficient matrix. The sigmoid function ensures that the coupling coefficients are in the range [0,1] for the output layer weight matrix. for transpose, This is the hidden representation of the last layer of graph convolution.

[0065] The aforementioned graph convolutional neural network employs supervised learning training, uses the Adam optimizer as the optimization strategy, and sets the learning rate to 0.001. The loss function includes a time consistency term, defined as:

[0066] ;

[0067] ;

[0068] in, For loss function, This is the predicted thermal coupling coefficient matrix. The actual coupling coefficient is calculated based on historical operating data and a heat conduction physics model. The squared L2 norm of the weight matrix. It is the Frobenius norm. The L2 regularization coefficient is... For time consistency loss, The time window length, , For time consistency weight, .

[0069] Furthermore, the introduction of time consistency loss ensures the temporal smoothness of thermal coupling coefficient prediction, avoiding drastic fluctuations in coupling relationship caused by instantaneous changes in grid disturbances. By minimizing the difference between predicted values ​​at adjacent times, the stability of system control and the reliability of practical engineering applications are improved.

[0070] Furthermore, the actual coupling coefficient The calculations combine physical heat conduction models and data-driven methods:

[0071] ;

[0072] ;

[0073] in, The physical coupling coefficient is based on Fourier's law of heat conduction. The thermal conductivity coefficient, For components and The effective heat transfer area between them For spatial distance, Thermal resistance; To determine the data-driven coupling coefficient based on historical temperature correlation analysis, calculation elements are used. and The Pearson correlation coefficient of the temperature time series was obtained; To integrate weights, It is dynamically adjusted based on data quality and the accuracy of the physical model.

[0074] Furthermore, the definitions and calculation methods for each parameter in the physical coupling coefficient are as follows: thermal conductivity coefficient The choice of material depends on the wall material of the spherical tank; for steel spherical tanks, the choice is... Stainless steel spherical tank Effective heat transfer area ,in, For the effective radius of the smaller element, For components and The angle between the line and the horizontal plane, For the effective radius of the smaller element, The radii of elements i and j are respectively; thermal resistance It consists of three parts: convective thermal resistance, conductive thermal resistance, and radiative thermal resistance. Indicates thermal resistance. The convective heat transfer coefficient is... , For wall thickness, For the thermal conductivity of the wall material, The radiative heat transfer coefficient is... , For emission rate, , The Stefan-Boltzmann constant is... , This represents the average temperature.

[0075] Furthermore, the L2 regularization coefficient The range of values ​​is constrained by Within the interval, specific values ​​are determined through grid search or Bayesian optimization methods. The range of L2 regularization coefficients balances the model's fitting ability and generalization performance, preventing overfitting.

[0076] Furthermore, the selection of parameter optimization methods is based on computational resources and accuracy requirements: when the number of training samples is less than 10,000 or the computation time requirement is strict, a grid search method is used to uniformly sample 5-8 candidate values ​​on a logarithmic scale for verification; when the training samples are sufficient and the accuracy requirement is high, a Bayesian optimization method with a Gaussian process as the surrogate model is used, and the acquisition function (using the EI criterion) is used to balance exploration and utilization, with the number of iterations set to 20-50.

[0077] S300 models each heating element as an independent reinforcement learning agent, and trains the Actor network of each agent and the shared Critic network based on the MADDPG (Multi-Agent Deep Deterministic Policy Gradient) algorithm.

[0078] For each heating element Configure an independent Actor network This is used to generate control actions for the heating elements; a shared Critic network is configured. It is used to evaluate the value of the joint actions of all agents; the network parameters are trained using a multi-agent deep deterministic policy gradient algorithm.

[0079] It should be noted that the MADDPG training process includes: each agent Actor networks with local observation and thermal coupling coefficient Input, output action , For the action to be performed, For the observation of the i-th agent, Let i be the policy function of the i-th agent. The thermal coupling coefficient is used; the Critic network is shared for observations from all agents. and actions For input, These represent the 1st, ..., nth actions, respectively. These represent the 1st, ..., nth observations executed, and the output action value is... , Represents the set of observation vectors of all agents.

[0080] , The set of action vectors for all agents represents the value assessment of the current state-action combination; transition tuples are stored using an experience replay buffer. , For the reward, For the next observation, the network parameters are updated by minimizing the TD error.

[0081] The aforementioned Actor network employs a multilayer perceptron structure, with the input layer receiving dimension being... The input vector, where This is a local observation dimension (including the current temperature deviation of components, power output, operating status, etc.). The thermal coupling coefficient vector dimension is used; input data preprocessing includes: temperature deviation is normalized to the range of [-1,1] using maximum-minimum normalization, power output is normalized to the range of [0,1] using maximum-minimum normalization, and operating status is converted into numerical data using one-hot encoding; the hidden layer contains two fully connected layers, each containing 256 neurons, and the activation function is ReLU; the output layer is a fully connected layer with an output dimension of 4, corresponding to the power allocation ratio and the adjustment of the three parameters of the PID controller. ,in For power adjustment amount, This is the adjustment amount of the proportional coefficient. This is the adjustment amount for the integral coefficient. The differential coefficient adjustment is where the power distribution ratio is mapped to the range [0,1] via the sigmoid function, and the PID parameter adjustment is mapped to the range [-1,1] via the tanh function.

[0082] The aforementioned Actor network uses a policy gradient training mode, with the Adam optimizer used to optimize the policy and a learning rate of 0.0001.

[0083] The loss function is defined as:

[0084] ;

[0085] in, Indicates the first A loss function, For experience replay buffer, The output of the Critic network is used to optimize the policy parameters by maximizing the expected cumulative reward. These represent the actions of agents 1 through i-1, respectively. These represent the actions of the (i+1)th to the nth agents, respectively.

[0086] The aforementioned Critic network employs a multilayer perceptron structure, with the input layer receiving dimension being... The input vector, where For the number of agents, For each agent's observation dimension, The action dimension is 4; the hidden layer contains three fully connected layers, the first two each contain 512 neurons using the ReLU activation function, and the third layer contains 256 neurons; the output layer is a fully connected layer with an output dimension of 1, representing the value estimate of the joint state-action pairs of all agents. .

[0087] The aforementioned Critic network employs a temporal difference learning training mode, uses the Adam optimizer as the optimization strategy, and sets the learning rate to 0.001; the loss function is defined as follows: , where the target value , As a discount factor, For instant rewards, This is the current state. For the current action, For the next state, For the target Critic network, The next action generated by the target Actor network. Let i represent the objective policy function of the i-th agent. For the next observation by the i-th agent, the target network parameters are updated using a soft update method. ,in, For the current network parameters, For the target network parameters, This is the soft update coefficient.

[0088] Furthermore, discount factor The value of the constraint is in Within a given range, the value reflects the system's emphasis on future rewards; higher values ​​are suitable for long-term control tasks. (Soft update coefficient) The value of the constraint is in Within the range, the update rate of the target network parameters is controlled, and smaller values ​​ensure the stability of the training process.

[0089] The improvement of the aforementioned MADDPG algorithm in this invention lies in using the thermal coupling coefficient as an additional input for each agent, enabling each agent to consider the thermal conduction interaction with other components when making decisions. The algorithm's input includes the local observation vector (including temperature deviation and power state) and the thermal coupling coefficient vector of each agent, and the output is a four-dimensional control vector, including the power allocation ratio and the incremental adjustment values ​​of the three PID parameters. The reward function includes two parts: instantaneous reward and cumulative time reward.

[0090] ;

[0091] ;

[0092] ;

[0093] ;

[0094] ;

[0095] in, For the total reward function, For instant reward, The time reward represents the temporal smoothness of temperature tracking. For the length of the time reward window, , Weighted by time reward , For normalized temperature tracking error, This represents the normalized power change.

[0096] in This is the temperature error weighting coefficient. The power change weighting coefficient, To influence the weighting coefficients of the coupling effect, Let be the normalized temperature error of agent i at time t. Let be the normalized power change of agent i at time t. Let i be the thermal coupling coefficient between elements i and j. Let be the normalized temperature error of agent j at time t. For time indexing, and For agent i, respectively and The normalized temperature error at time t, where The change in power For maximum power, where For temperature tracking error, This is the maximum permissible error;

[0097] Furthermore, the cumulative time reward is used to calculate long-term control performance using an exponentially weighted moving average method:

[0098] ;

[0099] in, Let be the exponentially weighted moving average reward at time t. The exponentially weighted moving average reward at time t-1. For smoothing coefficients, The total reward function is modified to , The baseline expected reward, For long-term performance weights, The cumulative time reward mechanism incentivizes agents to pursue long-term stable control effects rather than short-term optimization.

[0100] Among them, Let be the total reward of agent i at time t. Basic rewards, These are long-term performance weighting coefficients;

[0101] Furthermore, the weighting coefficients of the reward function , , The value constraints are as follows , , And it satisfies the normalization condition. ,in The reward function weight coefficient is set as the primary weight, emphasizing the importance of temperature tracking accuracy. The constraint on the value of the reward function weight coefficient ensures that the system prioritizes the achievement of the control objective while taking into account energy efficiency and collaborative performance.

[0102] S400 uses an evolutionary algorithm to search for the optimal topology connection scheme based on real-time grid disturbance characteristics and component status, generating a set of candidate reconfiguration strategies.

[0103] The current power grid disturbance characteristics and component operating status are input into the evolutionary algorithm. The optimal cooperative topology among components is searched through population evolution operations, and a strategy set containing multiple candidate topology reconfiguration schemes is output.

[0104] Furthermore, the population evolution operation employs the three basic operations of the genetic algorithm:

[0105] The selection operation uses a tournament selection method, randomly selecting from the population. By comparing each individual, the individual with the highest fitness is selected. ;

[0106] The crossover operation uses a single-point crossover method, performing crossover row by row on the chromosome matrix, with a crossover probability of... ;

[0107] The mutation operation uses bit-flipping mutation, which modifies each bit in the chromosome with a probability. Perform a flip, and ensure matrix symmetry after mutation; set the population size to [value missing]. The number of generations is The convergence criterion is that the improvement in optimal fitness over 20 consecutive generations is less than [a certain value]. .

[0108] The aforementioned evolutionary algorithm uses a binary matrix encoding method to represent topological connections, where chromosomes are encoded as follows: Symmetric binary matrix , Indicator element and There are collaborative connections between them. Indicates no connection;

[0109] The algorithm input includes the current power grid disturbance feature vector. Set of operating state vectors for each component And historical performance evaluation data, output as Candidate topology matrices ;

[0110] The fitness function includes a dynamic evaluation over time and is defined as follows:

[0111] ;

[0112] ;

[0113] ;

[0114] ;

[0115] ;

[0116] ;

[0117] in, for, Energy efficiency indicators within a time window. To evaluate window length, ; Overload risk within the time window; It is a static robustness index; For dynamic stability indicators, among which For temperature variance, For stability weights, .

[0118] in For element i at time... Power output, For components In topology Next moment efficiency, The maximum allowable power of component i. The total power of the system, Let be the element in the i-th row and j-th column of the topological matrix. Represents the smallest eigenvalue. for For the topological matrix The calculated graph Laplacian matrix, Set the target temperature;

[0119] Furthermore, the weighting coefficients for the time dimension satisfy... ,in Specifically designed for dynamic performance evaluation, the remaining weights are adjusted accordingly. , , This ensures a balance between dynamic adaptability and static performance.

[0120] Furthermore, the component operating state vector The composition is defined as follows:

[0121] ;

[0122] in, This is the component's operating state vector. For components Current temperature, This represents the current actual power output. For current efficiency, The aging factor is... This is the current current value;

[0123] The constraint range of each parameter is: ,in To the upper limit of safe temperature, , Indicates the efficiency range. This is the current constraint range.

[0124] Furthermore, the component efficiency coefficient The calculation is based on the influence of topology on the operating state of components:

[0125] ;

[0126] ;

[0127] in, The component efficiency coefficient, For components The basic efficiency is obtained from the component's technical specifications. For cooperative gain coefficient, The thermal coupling coefficient is... To set the temperature, For ambient temperature, The rated power of component i;

[0128] Historical performance evaluation data is obtained through a sliding time window method, collecting past data. The system's temperature tracking accuracy, energy efficiency ratio, and stability indicators within each control cycle, among which... The cycle is determined based on the system response time.

[0129] Furthermore, the values ​​of the fitness function weight coefficients are subject to the following constraints: , , And it satisfies the normalization condition. ,in As the dominant weight, system energy efficiency should be given priority. Control the severity of penalties for overload risk. To ensure system robustness, the weight coefficients of the fitness function are constrained to balance the importance of various performance indicators in multi-objective optimization.

[0130] S500 inputs the topology reconfiguration strategy and component states into the Actor network of each agent, and outputs the cooperative power allocation scheme and PID parameter adjustment amount;

[0131] Each agent's Actor network receives the selected topology reconfiguration strategy and the current element state information, calculates the power distribution ratio that the heating element should bear, and generates the corresponding PID controller parameter adjustment, including the proportional coefficient increment, integral coefficient increment, and derivative coefficient increment.

[0132] The network output decoding process includes: multiplying the power allocation ratio in the range [0,1] by the total system power to obtain the specific power command. ;

[0133] in The set power for the i-th element, For power adjustment amount, Total system power

[0134] The PID parameter adjustment range in the range of [-1,1] is proportionally mapped to the actual parameter range;

[0135] ;

[0136] ;

[0137] ;

[0138] in This is the updated scaling factor. This is the original proportionality coefficient. This is the adjustment amount for the proportional coefficient. is the scaling factor, where The updated integral coefficients, The original integral coefficients, This is the adjustment amount for the integral coefficient. is the scaling factor for the integral coefficients, where These are the updated differential coefficients. These are the original differential coefficients. This is the adjustment amount for the differential coefficient. The scaling factor for the differential coefficients. , , These are the scaling factors for each parameter.

[0139] Furthermore, the constraints on power allocation include total power balance constraints. , among which, among which The set power for the i-th element, For the total power requirement of the system, Total number of components; upper and lower bound constraints on component power. ,in To minimize power consumption and avoid frequent start-stop cycles, , For safe power limits, , The rated power of the i-th component; PID parameter stability constraints. , , To prevent excessive parameter adjustments from causing system instability;

[0140] in The initial value of the scaling factor is , where Here are the initial values ​​of the integration coefficients, where These are the initial values ​​for the differential coefficients;

[0141] Temperature safety constraints Safe operation is ensured by predicting temperature using a heat conduction model. Let i be the predicted temperature of the i-th element. To the upper limit of safe temperature, To set the temperature

[0142] Furthermore, the constraint range of the PID parameter scaling factor is: , , ,in , , These are the initial values ​​for each PID parameter. The PID parameter scaling factor constraint range ensures that the parameter adjustment range is moderate and avoids system instability caused by drastic changes in controller parameters.

[0143] Furthermore, the initial parameters of the PID controller are determined using the Ziegler-Nichols method based on system identification: firstly, the characteristic parameters of the system, including the delay time, are obtained through step response experiments. and time constant Calculate the critical gain Then set the initial parameters. , , ; The critical gain for system stability. For the system response characteristic time, The system response delay time, where These are the initial values ​​for the integral coefficients. This is the initial value of the proportionality coefficient. The initial value of the differential coefficients; the maximum value of the temperature tracking error. and maximum power Determined by the 95th percentile of historical operating data, i.e. , To ensure the stability of the normalization process;

[0144] in Let be the temperature tracking error at time t, where Let t be the power value at time t.

[0145] In this embodiment, to improve the response speed and stability of the control system, the TD3 algorithm is used to refine and update the network parameters of the MADDPG algorithm in S300. A centralized dual-Critic network evaluates the global control effect, while a distributed Actor network executes local control decisions. The TD3 algorithm, through its dual-Critic network and delayed policy update mechanism, reduces the overestimation problem of the value function and improves the stability of multi-agent collaborative learning.

[0146] Furthermore, the combination of MADDPG and TD3 algorithms adopts a hierarchical training strategy. First, the MADDPG algorithm is used for preliminary training to obtain basic network parameters. Then, the dual-Critic structure and delayed update mechanism of the TD3 algorithm are used to fine-tune the network parameters. The hierarchical training strategy combining MADDPG and TD3 algorithms fully leverages the advantages of MADDPG in multi-agent coordination and the improvements of TD3 in stability.

[0147] The input to the aforementioned TD3 algorithm in this invention includes the set of state observations of all agents at the current moment. Action set Reward Value And the observation of the next moment. The output is the updated Actor network parameters. and dual Critic network parameters The algorithm uses a dual Critic network to calculate the target value by taking the minimum value. ,in The target Actor network outputs the action with added noise, and the delayed update strategy is implemented every... The Actor network and the target network are updated only once per step.

[0148] in Let be the state observation of the i-th agent at time t, where Let be the action of the i-th agent at time t, where Let be the reward obtained by the i-th agent at time t, where Let be the state observation of the i-th agent at time t+1, where Let be the Actor network parameters for the i-th agent, where and These are the parameters of the two Critic networks, where... For the target value, For instant rewards, As a discount factor, For the target Critic network, For the next state, The action output by the target Actor network after adding noise;

[0149] Furthermore, delay the update parameters. The value of the constraint is in Within the integer set, it is usually set to 2 or 3. The delay update parameter controls the degree of delay in the update frequency of the Actor network relative to the Critic network. A smaller value can reduce the over-update of the Actor network and improve training stability.

[0150] In this embodiment, to achieve smooth topology switching and avoid abrupt changes in the control process, a soft switching matrix is ​​used to gradually adjust the cooperation weights between components, thereby achieving a smooth topology transition. Soft switching matrix ,in For time-varying switching coefficients, and These are the topology matrices before and after the switch, adjusted gradually. Achieve a smooth transition in topology from 0 to 1.

[0151] The aforementioned soft handover matrix method includes the following steps:

[0152] Calculate the switching trigger condition ,in To switch the trigger condition, where and These are the power grid disturbance feature vectors for the current and previous time moments, respectively. This is the trigger threshold;

[0153] Determine the switching duration ,in For switching duration, where For time coefficient, It is the Frobenius norm;

[0154] Generate time-varying switching coefficients Ensure that Within the interval, the transition is smooth from 0 to 1, where For time-varying switching coefficients, For the current moment, For switching duration;

[0155] Calculate the real-time soft handover matrix And normalize each row of elements so that the sum of the collaboration weights of all elements is 1;

[0156] Applying the soft handover matrix to power allocation yields the adjusted power command. .

[0157] in The power command adjusted for the i-th element. Original power command, Let be the element in the i-th row and j-th column of the soft switching matrix. Total number of components

[0158] Furthermore, trigger threshold The value of the constraint is in Within the interval, The historical standard deviation of the power grid disturbance feature vector serves as a trigger threshold constraint to avoid frequent topology switching and overreaction to minor disturbances; time coefficient The value of the constraint is in Within the second / unit norm range, ensure that the switching time is neither too short to cause system impact, nor too long to affect system response speed.

[0159] Furthermore, the historical standard deviation of the power grid disturbance eigenvector. Calculation using the sliding window method: ,in For historical standard deviation, The length of the sliding window. Let be the characteristic vector of power grid disturbance at time t. The mean within the window, where The length of the sliding window. This is the mean within the window; when historical data is insufficient, an empirical value is used. These correspond to the typical standard deviations of frequency offset, harmonic distortion rate, and voltage fluctuation, respectively; Figure: Number of layers in the convolutional neural network. The number of components is determined adaptively. Hidden Dimensions To ensure that the network has sufficient expressive power to handle complex spatial relationships, For the number of network layers, The number of components, of which To hide the dimension.

[0160] like Figure 2 As shown, based on the steps of the above-mentioned intelligent control method for electric heating of a hot water spherical tank, an intelligent control system for electric heating of a hot water spherical tank is proposed, including the following modules:

[0161] The data acquisition module 101 is used to acquire heterogeneous component parameters and power grid status data of the hot water spherical tank heating system. It includes: a component parameter acquisition unit, used to acquire the rated power, aging coefficient and spatial location coordinate data of each heating element; a power grid status acquisition unit, used to acquire the frequency offset value, harmonic content and voltage fluctuation amplitude information of the power grid in real time; and a data preprocessing unit, used to normalize the acquired data and generate a heterogeneous component-power grid status feature matrix.

[0162] The thermal coupling analysis module 102 includes: a graph convolutional neural network unit for analyzing component spatial topology data; a thermal coupling calculation unit for calculating the dynamic thermal coupling coefficient between components under different power grid disturbance modes; and a parameter optimization unit for optimizing network parameters and thermal coupling coefficients.

[0163] The intelligent control module 103 includes: an Actor network unit, which configures an independent Actor network for each heating element to generate control actions; a Critic network unit, which configures a shared Critic network to evaluate the joint action value of all agents; and a MADDPG training unit for training the parameters of each network.

[0164] The topology optimization module 104 includes: an evolutionary algorithm unit for searching for the optimal topology connection scheme; a policy generation unit for generating a set of candidate reconstruction policies; and a fitness evaluation unit for evaluating the performance of the topology scheme.

[0165] The control execution module 105 includes: a power distribution unit for calculating the power distribution ratio of each heating element; a PID parameter adjustment unit for generating the adjustment amount of the PID controller parameters; and a constraint processing unit for ensuring that the control output meets the system constraints.

[0166] It should be noted that the data flow relationship between the modules is as follows:

[0167] The data acquisition module 101 inputs the processed feature matrix into the thermal coupling analysis module 102;

[0168] The thermal coupling analysis module 102 inputs the thermal coupling coefficient into the intelligent control module 103 and the topology optimization module 104;

[0169] The topology optimization module 104 inputs the optimal topology scheme into the intelligent control module 103;

[0170] The intelligent control module 103 outputs control commands to the control execution module 105;

[0171] The control execution module 105 realizes the coordinated control of the heating element.

[0172] Based on the above-mentioned intelligent control method and system for electric heating of hot water spherical tanks, the following example is given, in which electric heating is controlled according to the following steps:

[0173] In this example, a hot water tank heating system in a chemical plant is used as a specific application example. The plant needs to provide a stable supply of hot water to the production line, requiring the water temperature to be maintained at 80±1℃. The system uses three electric heating elements with different power for coordinated heating control. In actual operation, due to large fluctuations in the plant's power load and unstable power grid quality, precise temperature control is a challenge.

[0174] Step 1: Collect heterogeneous component parameters and power grid status data of the hot water spherical tank heating system, and generate a heterogeneous component-power grid status feature matrix;

[0175] The system comprises three heating elements with different power ratings: 10kW, 15kW, and 20kW. Operating parameters of each element are collected in real time via sensors, including:

[0176] Component 1: Power 10kW, aging factor 0.9, location coordinates (0,0,0);

[0177] Component 2: Power 15kW, aging coefficient 0.85, location coordinates (0.5,0,0);

[0178] Component 3: Power 20kW, aging factor 0.95, location coordinates (0,0.5,0);

[0179] Simultaneously collect power grid status data:

[0180] Frequency offset: 0.2Hz;

[0181] Harmonic content: 3%;

[0182] Voltage fluctuation: 5%;

[0183] The feature matrix is ​​generated after the above data is normalized.

[0184] Step 2: Analyze the spatial topology data of the components using a graph convolutional neural network to calculate the thermal coupling coefficient.

[0185] Construct a diagram structure based on the spatial relationship of the three elements:

[0186] Distance between components 1 and 2: 0.5m;

[0187] Distance between components 1 and 3: 0.5m;

[0188] Distance between components 2 and 3: 0.7m;

[0189] The thermal coupling coefficient matrix is ​​obtained by calculating using a graph convolutional network.

[0190] Components 1-2: 0.8;

[0191] Components 1-3: 0.8;

[0192] Components 2-3: 0.6;

[0193] Step 3: Configure the reinforcement learning agent network

[0194] Configure Actor networks for each of the three heating elements, while sharing a single Critic network:

[0195] Actor1: Controls a 10kW component;

[0196] Actor2: Controls a 15kW component;

[0197] Actor3: Controls a 20kW component;

[0198] Shared Critic: Evaluating the overall control effectiveness;

[0199] Step 4: Search for the optimal topology connection scheme

[0200] Based on the current state, the optimal topology connection scheme is obtained through evolutionary algorithm search:

[0201] Components 1-2: Direct connection;

[0202] Components 1-3: Direct connection;

[0203] Components 2-3: Indirect connection;

[0204] Step 5: Generate control commands

[0205] Each Actor network outputs control commands according to the topology scheme:

[0206] Component 1: Power distribution 30%, PID parameter fine-tuning +5%;

[0207] Component 2: Power distribution 40%, PID parameter fine-tuning -3%;

[0208] Component 3: Power distribution 30%, PID parameter fine-tuning +2%;

[0209] The above steps achieve intelligent collaborative control of the hot water spherical tank heating system. After adopting this control method, the system can still maintain a temperature control accuracy of ±0.5℃ under power grid disturbance conditions, which is an improvement over the traditional method. At the same time, the service life of the heating element is extended and energy consumption is reduced.

[0210] The embodiments of the present invention have been described above, but the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention, all of which are within the protection scope of the present invention.

Claims

1. A smart control method for electric heating of a hot water spherical tank, comprising a hot water spherical tank heating system composed of several heating elements and a connected power grid, wherein the several heating elements in the hot water spherical tank heating system form a heating circuit, characterized in that... Its control method includes the following steps: S100: Collect heterogeneous component parameters and power grid status data of the hot water spherical tank heating system, and generate heterogeneous component-power grid status feature matrix; S200 uses graph convolutional neural networks to analyze the spatial topology data of components and calculate the dynamic thermal coupling coefficient between components under different power grid disturbance modes. S300 models each heating element as an independent reinforcement learning agent, and trains the Actor network of each agent and the shared Critic network based on the multi-agent deep deterministic policy gradient algorithm. S400 uses an evolutionary algorithm to search for the optimal topology connection scheme based on real-time grid disturbance characteristics and component status, generating a set of candidate reconfiguration strategies. The S500 inputs the topology reconfiguration strategy and component states into the Actor network of each agent, and outputs the cooperative power allocation scheme and PID parameter adjustment.

2. The intelligent control method for electric heating of a hot water spherical tank according to claim 1, characterized in that, The parameters of the heterogeneous components include the rated power, aging coefficient and spatial coordinate data of each heating element; The power grid status data includes the power grid's frequency offset, harmonic content, and voltage fluctuation amplitude.

3. The intelligent control method for electric heating of a hot water spherical tank according to claim 2, characterized in that, The processing steps of a graph convolutional neural network include: Calculate the adjacency matrix based on the spatial location of the components; Update node features using graph convolution operations; The final output layer generates a thermal coupling coefficient matrix.

4. The intelligent control method for electric heating of a hot water spherical tank according to claim 3, characterized in that, Each heating element is configured with an independent Actor network to generate control actions for the heating element; Configure a shared Critic network to evaluate the value of joint actions of all agents; Network parameters are trained using a multi-agent deep deterministic policy gradient algorithm.

5. The intelligent control method for electric heating of a hot water spherical tank according to claim 4, characterized in that, The training process of the multi-agent deep deterministic policy gradient algorithm includes: Each agent's Actor network takes local observations and thermal coupling coefficients as inputs and outputs actions. The shared Critic network takes the observations and actions of all agents as input and outputs action values. The network parameters are updated by minimizing the TD error and storing the transition tuples in the experience replay buffer.

6. The intelligent control method for electric heating of a hot water spherical tank according to claim 5, characterized in that, The Critic network employs a multilayer perceptron structure. The hidden layer contains three fully connected layers. The first two layers each contain 512 neurons and use the ReLU activation function. The third layer contains 256 neurons. The output layer is a fully connected layer with an output dimension of 1, representing the value estimate of the joint state-action pairs of all agents.

7. The intelligent control method for electric heating of a hot water spherical tank according to claim 6, characterized in that, The input to the multi-agent deep deterministic policy gradient algorithm includes vectors for each agent containing temperature deviation, power state, and thermal coupling coefficients. The output is a four-dimensional control vector containing the power allocation ratio and the incremental adjustment values ​​of the three PID parameters.

8. The intelligent control method for electric heating of a hot water spherical tank according to claim 7, characterized in that, The current power grid disturbance characteristics and component operating status are input into the evolutionary algorithm. The optimal cooperative topology among the components is searched through population evolution operations, and a strategy set containing multiple candidate topology reconfiguration schemes is output. Population evolution operations using genetic algorithms include the following three operations: The selection operation uses a tournament selection method, randomly selecting 3 individuals from the population for comparison, and selecting the individual with the highest fitness. The crossover operation uses a single-point crossover method, performing crossover on the chromosome matrix row by row; The mutation operation uses bit-flip mutation, flipping each bit in the chromosome with a probability of 0.

05. After mutation, matrix symmetry must be ensured. The population size is set to 50, the number of generations is 100, and the convergence criterion is that the improvement in optimal fitness is less than [a certain value] for 20 consecutive generations. .

9. The intelligent control method for electric heating of a hot water spherical tank according to claim 8, characterized in that, Each agent's Actor network receives the selected topology reconfiguration strategy and the current element state information, calculates the power distribution ratio that the heating element should bear, and generates the corresponding PID controller parameter adjustment, including the proportional coefficient increment, integral coefficient increment, and derivative coefficient increment.

10. A smart control system for electric heating of a hot water spherical tank, used to execute the steps in the smart control method for electric heating of a hot water spherical tank as described in any one of claims 1-9, characterized in that, include: The system includes a data acquisition module, a thermal coupling analysis module, an intelligent control module, a topology optimization module, and a control execution module. The data acquisition module inputs the processed feature matrix into the thermal coupling analysis module; the thermal coupling analysis module inputs the thermal coupling coefficient into the intelligent control module and the topology optimization module. The topology optimization module inputs the optimal topology scheme into the intelligent control module; the intelligent control module outputs control commands to the control execution module; and the control execution module realizes the coordinated control of the heating element.

Images