Intelligent water vapor sampling control method and system for thermal power plant

Abstract

The application relates to the technical field of water vapor monitoring of thermal power plants, and particularly discloses a kind of intelligent water vapor sampling control method and system for thermal power plants, the method comprises the following steps: collecting and processing multi-source information to obtain a standardized data set, constructing a reinforcement learning neural network model, optimizing PID parameters, outputting control signals for adjustment, and iteratively optimizing the reinforcement learning neural network model; the system comprises a real-time information acquisition module for multi-source information, a control module for constructing a reinforcement learning neural network model and outputting the control output of a PID controller, and an execution module for adjusting the opening degree of each regulating valve of a sampling device to complete water vapor sampling; the application fully utilizes the PSO algorithm-BP neural network to optimize the PID parameters, improves the precision and efficiency of the PID control, combines the reinforcement learning algorithm, realizes the dynamic optimization of the PID parameters and the accurate regulation and control of the sampling process, can guarantee the representativeness of the water vapor sampling while meeting the requirements of the online chemical instrument, and realizes the accurate monitoring of the water vapor quality.

Description

Technical Field

[0001] This invention relates to the field of water and steam monitoring technology in thermal power plants, and in particular to an intelligent water and steam sampling control method and system for thermal power plants. Background Technology

[0002] Water vapor sampling is a core component of water vapor detection in thermal power plants. The stability of the process and the accuracy of the sampling directly determine the accuracy and reliability of the analytical results, thus affecting the economic efficiency and safety of the unit operation. Currently, most water vapor sampling is done manually, which presents the following main problems: When manually operating the high-temperature sampling rack to discharge sewage, improper operation can easily lead to erosion and internal leakage of the sewage valve, which may affect the personal safety of the operators and the safe operation of the unit. Manual operation of high-temperature sampling racks results in low frequency of sewage discharge and incomplete discharge, leading to the deposition of corrosion products in pipelines, valves and other components. This not only affects the regulating characteristics of valves but also causes the sampling test data to be distorted and lose its representativeness. The flow rate and pressure of water samples fluctuate greatly due to the influence of unit load. Manual operation has a large lag, making it difficult to ensure that the water sample parameters meet the requirements of accurate measurement by online chemical instruments, and may even cause damage to related instruments.

[0003] With the normalization of deep peak shaving in thermal power units, it is necessary to establish an intelligent water and steam sampling method to ensure the safe and stable operation of the units. To this end, this invention proposes an intelligent water and steam sampling control method and system for thermal power plants to solve the above problems. Summary of the Invention

[0004] The purpose of this invention is to address the data inaccuracies and safety risks to personnel and equipment caused by manual operation in the existing steam and water detection process in thermal power plants. This invention proposes an intelligent steam and water sampling control method and system for thermal power plants, which can achieve accurate monitoring of steam and water quality and reduce the operational burden on operators while ensuring their personal safety.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, this invention proposes an intelligent water and steam sampling control method for thermal power plants, the method comprising: Collect information from multiple sources, preprocess the information from multiple sources, obtain a standardized dataset, and divide it into a training set and a validation set; Optimize the initial weights and thresholds of the BP neural network to obtain the optimal weights and thresholds, train the reinforcement learning agent based on the convergence of the reward function, and construct the reinforcement learning neural network model. The data in the preprocessed standardized dataset is input into the reinforcement learning neural network model to obtain the initial optimization parameters of the PID controller and the correction amount of the PID parameters. Based on the fused optimized PID parameters and the real-time data deviation, the control signal is output to adjust the opening degree of each valve. The operating status of the sampling device is determined based on the fault warning coefficient, and the fault type is located by combining historical data. Operating data and fault records are collected regularly to train the reinforcement learning neural network model and update the parameters of the reinforcement learning neural network model.

[0006] Furthermore, the multi-source information includes unit load, sampling inlet water vapor temperature, sampling inlet water vapor pressure, sampling inlet water vapor flow rate, cooling water inlet temperature of the cooling device, cooling water outlet temperature of the cooling device, cooling water flow rate of the cooling device, cooling sample water outlet temperature, pressure reducing device inlet pressure, pressure reducing device outlet pressure, constant temperature device outlet sample water temperature, and total actual flow rate.

[0007] Furthermore, the multi-source information is preprocessed in the following manner: use The criterion for removing outliers from multi-source information is as follows: ; In the formula, For the first The original data from multiple sources The mean of the original dataset containing multi-source information. The standard deviation of the original dataset of multi-source information is used. When the data of the multi-source information to be processed meets this formula, it is judged as an outlier and removed. The missing values ​​are filled using linear interpolation, as shown in the following formula: ; In the formula, For the first The imputation value for missing data points from multiple sources. , These are the valid data adjacent to each other before and after the missing data point from multiple sources. , , These represent the collection times for the corresponding multi-source information data points; Min-max normalization is used to map the multi-source information after outlier removal and imputation to the [0,1] interval to obtain a standardized dataset, as shown in the following formula: ; In the formula, For the normalized first The original data from multiple sources The minimum value of the data in the original dataset of multi-source information. This represents the maximum value of the data in the original dataset containing multi-source information.

[0008] Furthermore, the initial weights and thresholds of the optimized BP neural network are obtained in the following manner: The initial weights and thresholds of the BP neural network are optimized using a particle swarm optimization algorithm. The initial weights and thresholds of the BP neural network are encoded as particle position vectors. The root mean square value of the prediction error of the BP neural network is used as the fitness function value. Through global and local search of the particle swarm, the particle velocity and position are iteratively updated to obtain the optimal weights and thresholds corresponding to the minimum fitness function value.

[0009] Furthermore, the process of training a reinforcement learning agent based on the convergence of the reward function is as follows: Define the state space of a reinforcement learning agent. : ; in, The outlet temperature of the constant temperature device is the sample water temperature. , These represent the outlet temperature deviation and the rate of change of the temperature deviation of the thermostat, respectively. The outlet pressure of the pressure reducing device. , These are the outlet pressure deviation and the rate of change of the pressure deviation of the pressure reducing device, respectively. This refers to the inlet temperature of the cooling water in the cooling device. This refers to the cooling water flow rate of the cooling device. This represents the total actual traffic. These are the total actual flow deviation and the rate of change of the total actual flow deviation, respectively. The actual traffic flow of each branch road For the flow deviation of each branch, The rate of change of flow deviation for each branch; Action space Defined as follows: ; In the formula, , , These represent the correction amounts for the proportional, integral, and derivative coefficients of the PID parameters, respectively. This is a correction amount for the cooling water flow rate of the cooling device. This is the correction amount for the opening degree of the pressure reducing valve in the pressure reducing device. This is the correction amount for the opening of the regulating valve for the total flow rate of the sample water at the outlet of the constant temperature device. For the correction amount of the opening of the flow regulating valve in each branch; Reward function: ; In the formula, This represents the reward value of the reinforcement learning agent; These are the weighting coefficients. This represents the fault warning coefficient, which is 0 when there is no fault and 1 when a potential fault is detected. The predicted values ​​output by the particle swarm optimization algorithm-backpropagation neural network are used as the initial input to the reinforcement learning agent. Experience data is generated through interaction with the sampling device's operating environment, and the experience playback mechanism of the reinforcement learning agent is used to update the data. The network parameters are adjusted until the reward function converges, thus obtaining the optimal control decision strategy and completing the construction of the reinforcement learning neural network model.

[0010] Further, the step of inputting data from the preprocessed standardized dataset into a reinforcement learning neural network model to obtain the initial optimized parameters and correction values ​​of the PID controller, and outputting control signals to adjust the opening degree of each valve based on the fused optimized PID parameters and real-time data deviation, includes: The data from the standardized dataset obtained after real-time acquisition of multi-source information and preprocessing is input into the constructed reinforcement learning neural network model to obtain the predicted values ​​of the temperature and pressure of the sample outlet water, as well as the predicted value of the data deviation. At the same time, the initial optimization parameters of the PID controller are output. , , ; The reward function, calculated using a quantitative method, evaluates the current state of the sampling device and corrects the output PID parameters accordingly. , , ; Based on the initial optimized parameters of the PID controller and the correction values ​​of the PID parameters, the proportional coefficient of the final PID parameters is obtained. Integral coefficient Differential coefficients : ; ; ; The control output formula of the PID controller is: , In the formula, for The control output of the PID controller at any given time. for The deviation between the actual parameters and the target parameters at any given time; Control output based on PID controller , adjust the opening of each valve of the sampling device.

[0011] Further, judging the operation state of the sampling device based on the fault warning coefficient includes: Judging the operation state of the sampling device through the fault warning coefficient output by the particle swarm optimization algorithm - BP neural network , setting the warning threshold F1 and the fault threshold F2; When <F1, it is determined that the sampling device is operating normally; when F1 ≤ <F2, a fault warning is issued; when ≥F2, it is determined that the sampling device has a fault.

[0012] Further, training the reinforcement learning neural network model and updating the parameters of the reinforcement learning neural network model adopt the following methods: Hidden layer Sigmoid activation function: ; In the formula, is the input of the hidden layer neurons; ; In the formula, is the weight from the th neuron of the input layer to the th neuron of the hidden layer, is the input value of the th neuron of the input layer, is the threshold of the th neuron of the hidden layer, is the total number of hidden layer neurons; Output layer linear activation function: ; In the formula, is the input of the output layer neurons; ; In the formula, is the weight from the th neuron of the hidden layer to the th neuron of the output layer, is the output value of the th neuron of the hidden layer, represents the total number of output layer neurons, is the threshold of the th neuron of the output layer; The output values of the output layer neurons respectively correspond to the proportional coefficient, integral coefficient, and differential coefficient of the optimized PID parameters; Gradient descent method weight update formula: ; In the formula, for Weights at time points for Weights at time points For learning rate, For the error function value to the weight The partial derivatives; Error function value : ; In the formula, These are the actual sample water parameter values. For the target sample water parameter values, For the sample size, As a reward weighting coefficient, Real-time reward value; The error magnitude is quantified based on the error function value to guide the correction of the reinforcement learning neural network model parameters.

[0013] In a second aspect, the present invention provides an intelligent water and steam sampling control system for thermal power plants, which executes the intelligent water and steam sampling control method for thermal power plants as described above. The system includes: The information acquisition module collects multi-source information in real time throughout the entire sampling process; The control module preprocesses the collected multi-source information, constructs a reinforcement learning neural network model, and outputs the control output of the PID controller. The execution module receives the control output from the control module and adjusts the opening of each regulating valve of the sampling device to complete water vapor sampling.

[0014] Compared with existing technologies, the beneficial effects of this invention are as follows: The reinforcement learning neural network model proposed in this invention fully utilizes the PSO algorithm-BP neural network to optimize PID parameters, improving the accuracy and efficiency of PID control. Furthermore, by combining reinforcement learning algorithms, it achieves dynamic optimization of PID parameters and precise control of the sampling process. This invention can automatically adjust water and steam sampling parameters according to the load of the thermal power plant unit, ensuring the representativeness of water and steam sampling while meeting the requirements of online chemical instruments, thus achieving precise monitoring of water and steam quality. By integrating the control module and execution module of the control system, it achieves unmanned operation of water and steam sampling, effectively solving the data inaccuracies and safety risks to personnel and equipment caused by manual operation in the existing water and steam detection process of thermal power plants, thereby improving the intelligence level, operational reliability, and safety of thermal power plants. Attached Figure Description

[0015] Figure 1 This is an overall flowchart of the intelligent water and steam sampling control method for thermal power plants, as exemplified by the present invention. Figure 2 This is a flowchart of the reinforcement learning neural network model algorithm of the present invention. Detailed Implementation

[0016] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0017] like Figure 1 As shown, this invention provides an intelligent water and steam sampling control method for thermal power plants, comprising the following steps: Step S1: Collect multi-source information on the operating status of the sampling device. The collected multi-source information includes various data such as unit load. Sampling inlet water vapor temperature Sampling inlet water vapor pressure Water vapor flow rate at the sampling inlet Cooling water inlet temperature of the cooling device Cooling water outlet temperature of the cooling device Cooling water flow rate of cooling device Cooling sample water outlet temperature Pressure reducing device inlet pressure outlet pressure of pressure reducing device The outlet temperature of the constant temperature device for sample water Total actual flow ; Step S2: Preprocess the multi-source information. The preprocessing methods include using... Outliers were removed using the criteria, missing values ​​were filled in using linear interpolation, and the data was mapped to the [0,1] interval using min-max normalization to obtain a standardized dataset. The standardized dataset was then divided into a training set and a validation set.

[0018] The specific calculation formulas for the above preprocessing method are as follows: (1) Outlier Judgment Formula: ; In the formula, For the first The original data from multiple sources The mean of the original dataset containing multi-source information. is the standard deviation of the original dataset of multi-source information; when the data of the multi-source information to be processed meets this formula, the data is determined to be an outlier and removed.

[0019] (2) Linear interpolation method to complete the formula: ; In the formula, For the first The imputation value for missing data points from multiple sources. , These are the valid data adjacent to each other before and after the missing data point from multiple sources. , , These represent the collection times for the corresponding multi-source information data points.

[0020] (3) min-max normalization formula: ; In the formula, The first normalized multi-source information One data point, For the first The original data from multiple sources The minimum value of the data in the original dataset of multi-source information. This represents the maximum value of the data in the original dataset containing multi-source information.

[0021] Step S3: Optimize the initial weights and thresholds of the BP neural network to obtain the optimal weights and thresholds. Train the reinforcement learning agent based on the convergence of the reward function to construct the reinforcement learning neural network model. (Refer to...) Figure 2 The process is as follows: The initial weights and thresholds of the BP neural network are optimized using the particle swarm optimization (PSO) algorithm. The BP neural network is trained using the training set and its performance is verified using the validation set. By encoding the initial weights and thresholds of the BP neural network into particle position vectors, and using the root mean square value of the BP neural network prediction error as the fitness function value, the optimal weights and thresholds corresponding to the minimum fitness function value are obtained by initializing particle parameters, performing global and local searches of the particle swarm, and iteratively updating particle positions and velocities. This improves the training efficiency and prediction accuracy of the BP neural network, avoids local optima, and circumvents the defects of the BP neural network, thereby providing more accurate and stable initial control parameters for the PID controller.

[0022] (1) Particle velocity update formula: ; In the formula, For the first The particle in the first Wei Shang The speed of time For inertial weights, For the first The particle in the first Wei Shang The speed of time , As a learning factor, , A random number within the interval [0,1]. For the first The particle in the first The optimal position of an individual in the dimension. For the first The particle in the first Wei Shang Location at any given moment For the entire particle swarm in the th The globally optimal position in the dimension.

[0023] (2) Particle position update formula: ; In the formula For the first The particle in the first Wei Shang The position at time, and the meanings of the other parameters are the same as those in the particle velocity update formula.

[0024] (3) Inertia weight linearly decreasing update formula: ; in, Let be the inertia weight at time t. The initial inertia weight can be preferably 0.9. For the final inertia weight, a value of 0.4 is preferred. This represents the current iteration number. The maximum number of iterations. (4) Fitness function (root mean square error): ; In the formula, The root mean square error, For the sample size, For the first The actual value of each sample, i.e., the actual water parameter value. For the first The BP neural network prediction value of each sample, i.e. the parameter value of the target water sample.

[0025] If the fitness function value is not the minimum, then return to update the particle velocity and position until the minimum fitness function value is obtained, output the global optimal particle position, and based on the fitness function, obtain the optimal weights and thresholds corresponding to its minimum value. Use the optimal weights and thresholds corresponding to the minimum value to update the initial weights and thresholds of the BP neural network to improve the training efficiency and prediction accuracy of the BP neural network and avoid local optima.

[0026] Furthermore, based on the convergence of the reward function and the optimization strategy for the sampling device's operating environment using reinforcement learning algorithms, a reinforcement learning agent with a reinforcement learning neural network model is created. This agent integrates reinforcement learning algorithms with the PSO algorithm-BP neural network to achieve coordinated control of the sampling device's PID controller, specifically including: (1) State space Definition: ; in, The outlet temperature of the constant temperature device is the sample water temperature. , These represent the outlet temperature deviation and the rate of change of the temperature deviation of the thermostat, respectively. ; The outlet pressure of the pressure reducing device. , These are the outlet pressure deviation and the rate of change of the pressure deviation of the pressure reducing device, respectively. Pressure deviation change rate; This refers to the inlet temperature of the cooling water in the cooling device. This refers to the cooling water flow rate of the cooling device. This represents the total actual traffic. These are the total actual flow deviation and the rate of change of the total actual flow deviation, respectively. ; For the flow rate of each branch road, For the flow deviation of each branch, The rate of change of flow deviation for each branch, .

[0027] (2) Action space Definition: ; in, , , These are the proportional coefficients corresponding to the PID parameters. Integral coefficient Differential coefficients The correction amount, This is a correction amount for the cooling water flow rate of the cooling device. This is the correction amount for the opening degree of the pressure reducing valve in the pressure reducing device. This is the correction amount for the opening of the regulating valve for the total flow rate of the sample water at the outlet of the thermostat. This is the correction amount for the opening of the flow regulating valves in each branch.

[0028] (3) Reward function design: ; In the formula, This represents the reward value of the reinforcement learning agent, used to evaluate the quality of the current policy; These are all weighting coefficients, which can be adjusted according to control priorities and energy consumption requirements. Indicates the total number of branches; This refers to the temperature deviation at the outlet of the thermostat. , The target temperature value; The outlet pressure deviation of the pressure reducing device , This is the target pressure value; The deviation of the total actual flow rate, , The target value for total flow; For branch flow deviation, , The target value for branch flow; For cooling water flow rate correction, the larger the absolute value, the greater the energy consumption adjustment, and the higher the weight of the corresponding penalty item; This is the fault warning coefficient, which is 0 when there is no fault and 1 when a potential fault is detected. Fault states are avoided by introducing a weighting coefficient.

[0029] During the training phase of a reinforcement learning agent, the goal is to achieve convergence of the reward function and obtain the optimal control decision strategy. This enables the reinforcement learning agent to adaptively optimize PID parameters online, thereby achieving the intelligence of the reinforcement learning neural network model through training.

[0030] Parameter collaborative update: The predicted value output by the PSO (Particle Swarm Optimization) algorithm-BP neural network (i.e., the proportional coefficient of the PID controller) is updated. Integral coefficient Differential coefficients As the initial input to the reinforcement learning agent, empirical data is generated through interaction with the operating environment of the sampling device. This empirical data includes operating parameters collected during the operation of the sampling device and optimized initial state parameters of the reinforcement learning agent. The sampling device includes a cooling device, a pressure reducing device, and a temperature controlling device. The empirical data is updated using the experience playback mechanism within the reinforcement learning agent. Network parameters (depth) Network, Deep Algorithm The network parameters are processed until the reward function converges, outputting the optimal control parameters to obtain the optimal control decision strategy. The optimal control decision strategy is the correction amount of the PID parameters, thus completing the construction of the reinforcement learning neural network model. The control decision strategy is the decision rule obtained by the reinforcement learning agent through training, which is used to guide online inference.

[0031] Step S4: Input the data from the preprocessed standardized dataset into the constructed reinforcement learning neural network model, namely the PSO algorithm-BP neural network and reinforcement learning agent, to obtain the initial optimization parameters of the PID controller and the correction amount of the PID parameters; then, based on the fused optimized PID parameters and real-time data deviation, output control signals to adjust the opening degree of each valve of the sampling device. The control signals are directly applied to the actuators of each valve of the sampling device when the sampling device system is running online.

[0032] Specifically, data from the standardized dataset obtained after real-time acquisition of multi-source information and preprocessing is input into the constructed reinforcement learning neural network model to obtain predicted values ​​of the temperature and pressure of the sampled outlet water, as well as predicted values ​​of data deviation. At the same time, the initial optimization parameters of the PID controller are output. , , ; The reward function, calculated using a quantitative method, evaluates the current state of the sampling device and corrects the output PID parameters accordingly. , , ; The reward function, as a core indicator for evaluating the performance of a current policy in a reinforcement learning neural network model, outputs a real-time reward value by quantizing the current state of the sampling device and control bias. This is used to determine the suitability of the initial optimization parameters of the current PID controller. The specific steps are as follows: Obtain the current status of the sampling device: The current status consists of two parts, one is the real-time operating parameters of the sampling device, and the other is a standardized dataset of multi-source information after preprocessing, such as the actual temperature of the sample water at the outlet of the constant temperature device. Temperature deviation Pressure deviation at the outlet of the pressure reducing device Secondly, the initial state parameters of the PID parameters output by the PSO algorithm-BP neural network. , , .

[0033] Substituting into the reward function to calculate the real-time reward value: Combining the control objective of the sampling device, the reward function adopts a quantitative calculation method. The core logic is that the smaller the deviation and the more stable the adjustment, the higher the reward value; the larger the deviation and the more drastic the fluctuation, the lower the reward value. The specific calculation is as follows: First, the calculation method for the comprehensive control deviation is as follows: integrating key operating parameters of the sampling device such as temperature, pressure, and flow rate, the result is obtained through the following steps: 1. First, obtain the deviation of each individual data item: the deviation of the outlet sample water temperature of the constant temperature device. The difference between the actual temperature and the target temperature; the outlet pressure deviation of the pressure reducing device. The difference between actual pressure and target pressure; total actual flow deviation. The difference between the total actual flow and the total target flow; 2. Normalize the deviation of each individual data item to obtain the normalized data deviation. , , The values ​​are all in the range of [0,1]. 3. Combine PID parameter control priority settings with weights, prioritizing temperature > pressure > flow rate, and assigning weights of α, β, and γ respectively. These weights are set according to the sampling device's control priority, with α + β + γ = 1; and the overall control deviation is considered. The calculation formula is: ; Comprehensive control deviation It is a weighted sum of the three types of deviations: temperature, pressure, and flow rate, with a value range of [0,1]. The closer it is to 0, the better the control effect and the higher the reward value; The closer it is to 1, the greater the control deviation and the lower the reward value.

[0034] Based on the above comprehensive control deviation The specific calculation formula for the reward function using the quantitative calculation method is as follows: ; This is the real-time reward value, ranging from [20, 100]. When the sampling device is operating normally, the overall control deviation approaches 0, there is no overshoot, and the adjustment speed is fast, Approaching 100 (maximum reward); when the sampling device exhibits a slight deviation and is adjusted smoothly... A score between 60 and 99 indicates a high reward; however, if the sampling device has a large deviation, fluctuates, or approaches the warning threshold, a negative result may occur. Between 30 and 59 (low reward); when the sampling device deviation exceeds the allowable range or an abnormality occurs. <30 (extremely low reward).

[0035] The current state of the sampling device is evaluated based on real-time reward values: (This is achieved through real-time reward values.) The system determines whether the initial state parameters of the current PID controller are suitable for the current operating conditions. A high reward indicates that the current initial state parameters are suitable and do not require significant correction. A low reward indicates that the current initial state parameters are poorly adapted and require targeted correction. An extremely low reward indicates that the parameters are severely mismatched and require significant correction.

[0036] Based on the real-time reward value output by the reward function, the corresponding correction amount for the PID parameters is output to achieve online adaptive optimization of the initial state parameters. The specific steps are as follows: Based on the current state and real-time reward value, call the depth of DQN. The network queries the optimal action and depth in the current state. The network utilizes the experience data from previous offline training and the established mapping relationship between state, reward, and optimal action.

[0037] The hierarchical logic of the correction output is aligned with actual control, avoiding overcorrection that could cause fluctuations in the sampling device's operation. The selection of the reward value range and the determination of its suitability can be formulated based on actual production needs. This implementation only provides one example.

[0038] 1. High-reward scenarios ( ≥60: This indicates that the initial state parameters of the current PID controller are well adapted, and the output has a small correction value. , , (Absolute value ≤ 5% of initial state parameters), only fine-tuning is required to accommodate minor fluctuations in operating conditions; 2. Low-reward scenarios (30≤ <60): The initial state parameter adaptation of the current PID controller is average, and the output has a moderate correction amount. , , (The absolute value is 5% to 15% of the initial state parameters), and the deviation is corrected accordingly. 3. Extremely low reward scenarios ( <30): The initial state parameters of the current PID controller are severely mismatched, resulting in a large output correction ( , , (The absolute value is 15% to 30% of the initial state parameters), and it is quickly adjusted to suppress the expansion of the deviation and restore the normal operation of the sampling device.

[0039] Correction Quantity Verification and Output: After outputting the correction quantity of the PID parameters, combine it with the current state. Predict the effect of the corrected control (through depth) Network prediction correction prediction reward value ),like > If this indicates an improvement in performance after correction, then the correction amount should be output directly; otherwise... ≤ If the correction does not improve the effect, then adjust the correction amount and reduce the correction ratio until the predicted reward value improves, and finally output the correction amount of the optimal PID parameters.

[0040] Based on the correction amount of the PID parameters for optimal output, and combining the initial optimized parameters of the PID controller output by the PSO algorithm - BP neural network and the correction amount of the PID parameters, the proportional coefficient of the final PID parameters is obtained , integral coefficient , differential coefficient : ; ; ; The control output formula of the PID controller is: ; In the formula, is the control output quantity of the PID controller at time , that is, the output control signal adjusts the opening degrees of the valves of the sampling device, is the deviation value between the actual parameter and the target parameter at time

[0041] The state of the sampling device refers to whether the sampling device operates normally, whether there are abnormalities or faults under the current working conditions, including three categories: normal state, early warning state, and fault state, including but not limited to the opening and closing operation states of the valves of the sampling device

[0042] The fault early warning coefficient output by the PSO algorithm - BP neural network is used to judge the operation state of the sampling device; the early warning threshold F1 and the fault threshold F2 (F2 > F1) are set. When < F1, it is determined that the sampling device operates normally; when F1 ≤ < F2, a fault early warning is issued; when ≥ F2, it is determined that the sampling device has a fault; based on the output characteristics of the hidden layer of the BP neural network and combined with historical fault data, the fault type is located. The fault types include filter blockage, decline in the efficiency of the cooling device, sensor fault, etc

[0043] Step S5: Iterative optimization: Based on the fault early warning coefficient, judge the operation state of the sampling device, and combine historical data to locate the fault type. Regularly collect the operation data and fault records of the sampling device, train the reinforcement learning neural network model, update the parameters of the reinforcement learning neural network model, and improve the adaptability and prediction accuracy of the reinforcement learning neural network model to different working conditions of the sampling device based on the feedback of the sampling device operation environment

[0044] The activation function of the BP neural network: The hidden layer uses the Sigmoid function, and the output layer uses the linear function; the training algorithm uses the gradient descent method, and the formula is as follows: (1) Hidden layer Sigmoid activation function: ; In the formula, This serves as the input to the neurons in the hidden layer. ; For the input layer The first neuron is connected to the hidden layer. The weights of each neuron, For the input layer The input values ​​of each neuron (including the deviation between the actual parameters and the target parameters) and the rate of change of deviation). For the hidden layer Threshold of a neuron This represents the total number of neurons in the hidden layer.

[0045] (2) Output layer linear activation function: ; In the formula, For the input of the output layer neurons; ; In the formula, For the hidden layer The nth neuron to the output layer The weights of each neuron, For the hidden layer The output value of each neuron This represents the total number of neurons in the output layer; For the output layer The threshold for each neuron does not need to be manually fixed. In the initial stage, a random value method is used, and the value range is close to the normal training range of the BP neural network. Subsequently, the PSO algorithm and BP neural network training are used for collaborative optimization, and finally converge to the optimal threshold.

[0046] This embodiment provides a reference value: the total number of neurons in the output layer. =3 (corresponding to the proportional coefficient of the PID parameters) Integral coefficient Differential coefficients ), initial thresholds of the three output layer neurons , , The values ​​can be selected as 0.08, 0.12, and 0.05, with a suggested range of [-0.2, 0.2]. It should be understood that the above example values ​​are for reference only and will be optimized to the optimal value through the PSO algorithm-BP neural network.

[0047] The output values ​​of the output layer neurons correspond to the optimized proportional coefficients of the PID controller. Integral coefficient Differential coefficients (3) To minimize the prediction error of the reinforcement learning neural network model and optimize the learning effect, the gradient descent method is used for weight update: ; In the formula, for Weights at time (including) , ), for Weights at time points For learning rate, For the error function value to the weight The partial derivatives of .

[0048] To quantify the bias in predictive control using a reinforcement learning neural network model, guide model optimization and parameter correction, and ensure optimal control performance, an error function is used to quantify the bias and correct the reinforcement learning neural network model to achieve optimal control performance. The error function value is as follows: ; In the formula, These are the actual sample water parameter values. For the target sample water parameter values, For the sample size, As a reward weighting coefficient, This represents the real-time reward value. The error function value is then calculated. Based on error function value The magnitude of the error is quantized as deviation (i.e., the deviation between the actual input parameters of the hidden layer neurons and the target parameters). (and the rate of change of deviation), thereby guiding the correction of the parameters of the reinforcement learning neural network model to achieve the optimal control effect.

[0049] In another scenario of this embodiment, a smart water and steam sampling system for thermal power plants based on reinforcement learning neural networks is proposed to realize the above-mentioned smart water and steam sampling control method for thermal power plants. The system includes: Information acquisition module: Real-time acquisition of multi-source information throughout the sampling process, including unit load. Sampling inlet water vapor temperature Sampling inlet water vapor pressure Water vapor flow rate at the sampling inlet Cooling water inlet temperature of the cooling device Cooling water outlet temperature of the cooling device Cooling water flow rate of cooling device Cooling sample water outlet temperature Pressure reducing device inlet pressure outlet pressure of pressure reducing device The outlet temperature of the constant temperature device for sample water and the flow of each branch and total actual traffic ; Control module: Preprocesses the collected multi-source information, constructs a reinforcement learning neural network model, and uses this model to infer the results from the preprocessed multi-source data, combining it with empirical data generated through interaction with the sampling device's operating environment. The resulting output is the control signal from the PID controller. ; It should be noted that the empirical data generated through interaction with the operating environment of the sampling device refers to basic operational data: real-time operational data of the sampling device (such as unit load, temperature, pressure, flow rate, etc. at various points); and agent action data: action-related data output by the reinforcement learning agent (such as output correction data). , , Control effect data: Control feedback data generated corresponding to the above data, such as changes in the operating status of the sampling device and changes in control deviation after parameter correction.

[0050] Execution module: Receives control signals output from the control module, adjusts the opening of each regulating valve in the sampling device, and completes water vapor sampling.

[0051] It should be noted that any parts not covered in this invention are the same as or can be implemented using existing technology. The above description is merely a preferred embodiment of this invention, but the scope of protection of this invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in this invention, based on the technical solution and inventive concept of this invention, should be covered within the scope of protection of this invention.

Claims

1. A method for intelligent water and steam sampling and control in a thermal power plant, characterized in that, include: Collect information from multiple sources, preprocess the information from multiple sources, obtain a standardized dataset, and divide it into a training set and a validation set; Optimize the initial weights and thresholds of the BP neural network to obtain the optimal weights and thresholds, train the reinforcement learning agent based on the convergence of the reward function, and construct the reinforcement learning neural network model. The data in the preprocessed standardized dataset is input into the reinforcement learning neural network model to obtain the initial optimization parameters of the PID controller and the correction amount of the PID parameters. Based on the fused optimized PID parameters and the real-time data deviation, the control signal is output to adjust the opening degree of each valve. The operating status of the sampling device is determined based on the fault warning coefficient, and the fault type is located by combining historical data. Operating data and fault records are collected regularly to train the reinforcement learning neural network model and update the parameters of the reinforcement learning neural network model.

2. The intelligent water and steam sampling control method for thermal power plants according to claim 1, characterized in that, The multi-source information includes unit load, sampling inlet water vapor temperature, sampling inlet water vapor pressure, sampling inlet water vapor flow rate, cooling water inlet temperature of the cooling device, cooling water outlet temperature of the cooling device, cooling water flow rate of the cooling device, cooling sample water outlet temperature, pressure reducing device inlet pressure, pressure reducing device outlet pressure, constant temperature device outlet sample water temperature, and total actual flow rate.

3. The intelligent water and steam sampling control method for thermal power plants according to claim 1, characterized in that, The multi-source information is preprocessed using the following method: use The criterion for removing outliers from multi-source information is as follows: ； In the formula, For the first The original data from multiple sources The mean of the original dataset containing multi-source information. The standard deviation of the original dataset of multi-source information is used. When the data of the multi-source information to be processed meets this formula, it is judged as an outlier and removed. The missing values ​​are filled using linear interpolation, as shown in the following formula: ； In the formula, For the first The imputation value for missing data points from multiple sources. , These are the valid data adjacent to each other before and after the missing data point from multiple sources. , , These represent the collection times for the corresponding multi-source information data points; Min-max normalization is used to map the multi-source information after outlier removal and imputation to the [0,1] interval to obtain a standardized dataset, as shown in the following formula: ； In the formula, For the normalized first Data from multiple sources, The minimum value of the data in the original dataset of multi-source information. This represents the maximum value of the data in the original dataset containing multi-source information.

4. The intelligent water and steam sampling control method for thermal power plants according to claim 1, characterized in that, The method for optimizing the initial weights and thresholds of the BP neural network to obtain the optimal weights and thresholds is as follows: The initial weights and thresholds of the BP neural network are optimized using a particle swarm optimization algorithm. The initial weights and thresholds of the BP neural network are encoded as particle position vectors. The root mean square value of the prediction error of the BP neural network is used as the fitness function value. Through global and local search of the particle swarm, the particle velocity and position are iteratively updated to obtain the optimal weights and thresholds corresponding to the minimum fitness function value.

5. The intelligent water and steam sampling control method for thermal power plants according to claim 4, characterized in that, The process of training a reinforcement learning agent based on the convergence of the reward function is as follows: Define the state space of a reinforcement learning agent. : ； in, The outlet temperature of the constant temperature device is the sample water temperature. , These represent the outlet temperature deviation and the rate of change of the temperature deviation of the thermostat, respectively. The outlet pressure of the pressure reducing device. , These are the outlet pressure deviation and the rate of change of the pressure deviation of the pressure reducing device, respectively. This refers to the inlet temperature of the cooling water in the cooling device. This refers to the cooling water flow rate of the cooling device. This represents the total actual traffic. These are the total actual flow deviation and the rate of change of the total actual flow deviation, respectively. The actual traffic flow of each branch road For the flow deviation of each branch, The rate of change of flow deviation for each branch; Action space Defined as follows: ； In the formula, , , These represent the correction amounts for the proportional, integral, and derivative coefficients of the PID parameters, respectively. This is a correction amount for the cooling water flow rate of the cooling device. This is the correction amount for the opening degree of the pressure reducing valve in the pressure reducing device. This is the correction amount for the opening of the regulating valve for the total flow rate of the sample water at the outlet of the constant temperature device. For the correction amount of the opening of the flow regulating valve in each branch; Reward function: ； In the formula, To enhance the reward value of the learning agent; These are the weighting coefficients. Indicates the total number of branches. This is the fault warning coefficient, which is 0 when there is no fault and 1 when a potential fault is detected. The predicted values ​​output by the particle swarm optimization algorithm-backpropagation neural network are used as the initial input to the reinforcement learning agent. Experience data is generated through interaction with the sampling device's operating environment, and the experience playback mechanism of the reinforcement learning agent is used to update the data. The network parameters are adjusted until the reward function converges, thus obtaining the optimal control decision strategy and completing the construction of the reinforcement learning neural network model.

6. The intelligent water and steam sampling control method for thermal power plants according to claim 1, characterized in that, The process involves inputting data from the preprocessed standardized dataset into a reinforcement learning neural network model to obtain the initial optimized parameters and correction values ​​for the PID controller. Based on the fused optimized PID parameters and real-time data deviations, a control signal is output to adjust the opening degree of each valve, including: The data from the standardized dataset obtained after real-time acquisition of multi-source information and preprocessing is input into the constructed reinforcement learning neural network model to obtain the predicted values ​​of the temperature and pressure of the sample outlet water, as well as the predicted value of the data deviation. At the same time, the initial optimization parameters of the PID controller are output. , , ; The reward function, calculated using a quantitative method, evaluates the current state of the sampling device and corrects the output PID parameters accordingly. , , ; Based on the initial optimized parameters of the PID controller and the correction values ​​of the PID parameters, the proportional coefficient of the final PID parameters is obtained. Integral coefficient Differential coefficients : ； ； ； The control output formula of the PID controller is: ， In the formula, for The control output of the PID controller at any given time. for The deviation between the actual parameters and the target parameters at any given time; Control output based on PID controller Adjust the opening degree of each valve in the sampling device.

7. The intelligent water and steam sampling control method for thermal power plants according to claim 5, characterized in that, The method of determining the operating status of the sampling device based on the fault early warning coefficient includes: Fault warning coefficient output by particle swarm optimization algorithm-backpropagation neural network Determine the operating status of the sampling device and set the early warning threshold F1 and the fault threshold F2; When <When F1, it is determined that the sampling device is operating normally; when F1 ≤ <F2, a fault warning is issued; when ≥ F2, it is determined that the sampling device has a fault.

8. The intelligent water and steam sampling control method for thermal power plants according to claim 1, characterized in that, The reinforcement learning neural network model is trained and its parameters are updated in the following manner: Hidden layer Sigmoid activation function: In the formula, This serves as the input to the neurons in the hidden layer. ； In the formula, For the input layer The first neuron is connected to the hidden layer. The weights of each neuron, For the input layer The input value of each neuron. For the hidden layer Threshold of a neuron The total number of neurons in the hidden layer; the linear activation function of the output layer: In the formula, For the input of the output layer neurons; ； In the formula, For the hidden layer The nth neuron to the output layer The weights of each neuron, For the hidden layer The output value of each neuron This represents the total number of neurons in the output layer. For the output layer The threshold values ​​of each neuron; the output values ​​of the output layer neurons correspond to the proportional coefficient, integral coefficient, and derivative coefficient of the optimized PID parameters, respectively; the gradient descent weight update formula: ； In the formula, for Weights at time points for Weights at time points For learning rate, For the error function value to the weight The partial derivatives; Error function value : ； In the formula, These are the actual sample water parameter values. For the target sample water parameter values, For the sample size, As a reward weighting coefficient, Real-time reward value; The error magnitude is quantified based on the error function value to guide the correction of the reinforcement learning neural network model parameters.

9. A smart water and steam sampling control system for a thermal power plant, executing the smart water and steam sampling control method for a thermal power plant as described in any one of claims 1-8, characterized in that, The system includes: The information acquisition module collects multi-source information in real time throughout the entire sampling process; The control module preprocesses the collected multi-source information, constructs a reinforcement learning neural network model, and outputs the control output of the PID controller. The execution module receives the control output from the control module and adjusts the opening of each regulating valve of the sampling device to complete water vapor sampling.

Images