Thermal power generating unit cold end system operation optimization method based on machine learning

Abstract

The invention discloses a thermal power generating unit cold end system operation optimization method based on machine learning. The method comprises the following steps: collecting and preprocessing unit multi-source operation data; calculating intermediate parameters and carrying out time alignment and feature fusion; establishing an operation boundary set and a control constraint set, and performing windowing and screening; four types of sub-models are trained, and an R-GCN network is adopted to carry out multi-relation feature aggregation and parameter uniformization processing; calling the cold end system unified prediction model, and outputting a cold end state quantity and a performance index; building a multivariable optimization problem based on the unified prediction model of the cold end system; solving a multivariable optimization problem by adopting a near-end strategy optimization method; and updating the unified prediction model of the cold end system on line based on real-time data, periodically retraining and deploying, and outputting an optimal control instruction. The machine learning and multi-model collaborative optimization method is adopted, intelligent operation control over the thermal power generating unit cold end system is achieved, and the method has the advantages of being high in prediction precision, low in energy consumption and high in self-adaption.

Description

Technical Field

[0001] This invention relates to the field of thermal power unit operation optimization technology, and in particular to a machine learning-based method for optimizing the operation of the cold end system of thermal power units. Background Technology

[0002] Existing thermal power unit cold-end systems typically employ operation optimization methods based on empirical models or steady-state thermodynamic equations. These methods utilize pre-set operating curves or linear models to control the start-up, shutdown, and load regulation of circulating water pumps and cooling tower fans. While these methods are feasible under static or near-steady-state conditions, they fail to fully utilize multi-source operating data from the unit, making it difficult to accurately reflect the dynamic changes under complex coupling relationships. This results in insufficient prediction accuracy and significant deviations between optimization results and actual operation. Furthermore, traditional optimization methods often rely on single-variable control, neglecting the multi-factor interrelationships in energy transfer, fluid circulation, and ventilation processes within the cold-end system.

[0003] In recent years, some studies have attempted to introduce neural network models to model and optimize cold-end systems. However, most methods remain at the stage of single network structure or fixed weight learning, lacking multi-model fusion and feature consistency mechanisms, making it difficult to achieve unified modeling of multi-dimensional time-series data. In addition, existing methods generally do not consider mechanistic constraints and adaptive updates of real-time data, leading to prediction drift and control lag problems during long-term model operation, which cannot meet the dual requirements of energy saving and safe operation of thermal power units under variable load and complex environments.

[0004] Therefore, how to provide a machine learning-based method for optimizing the operation of the cold-end system of thermal power units is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a machine learning-based method for optimizing the operation of the cold-end system of a thermal power unit. This invention employs machine learning and multi-model collaborative optimization to achieve intelligent operation control of the cold-end system of a thermal power unit, which has the advantages of high prediction accuracy, low energy consumption, and strong adaptability.

[0006] According to an embodiment of the present invention, a machine learning-based method for optimizing the operation of a thermal power unit's cold-end system includes the following steps: Collect multi-source operation data from the unit, preprocess it, and generate a standardized operation dataset; Intermediate parameters are calculated based on a standardized running dataset, and time alignment and feature fusion are performed to generate an input feature set. Based on the input feature set, establish the running boundary set and the control constraint set, and perform windowing and filtering according to the preset time window length and sliding step size to construct the time series training sample set and label set; Four types of sub-models are trained based on time series training sample sets and label sets, and R-GCN network is used to perform multi-relation feature aggregation and parameter consistency processing to generate a unified prediction model for cold-end systems. The unified prediction model of the cold-end system is invoked based on the set of operating boundaries and the set of real-time input features, and the output cold-end state variables and performance indicators are output. A multivariate optimization problem is constructed based on a unified prediction model for cold-end systems. The decision variables are set as the speed and number of circulating water pumps and the speed and number of cooling tower fans. A set of control constraints is applied to limit the feasible solution space. The near-end strategy optimization method is used to solve the multivariate optimization problem, obtain the optimal setpoints of circulating cooling water flow rate and cooling tower ventilation volume, and output the optimal control command set. The unified prediction model of the cold-end system is updated online based on real-time operating data. Retraining and deployment replacement are completed within the preset update cycle, and the corresponding optimal control commands are re-output based on the updated model.

[0007] Optionally, the multi-source operating data of the unit includes turbine inlet steam parameters, regenerative system adjusted and non-adjusted extraction steam parameters, condenser parameters, circulating cooling water system parameters, cooling tower and cooling tower fan parameters, and atmospheric environment parameters. The preprocessing includes data cleaning, noise filtering, feature alignment, data normalization, and feature filtering.

[0008] Optionally, the generation of the input feature set specifically includes: The condenser cooling tube cleanliness coefficient is calculated based on a standardized operating dataset. The condenser cooling tube cleanliness coefficient is determined by the ratio of the actual heat transfer coefficient to the reference heat transfer coefficient according to the heat transfer balance relationship. The steam extraction rate of each stage of the low-pressure heater is calculated based on the energy conservation of the regenerative system. The steam extraction rate of the low-pressure heater is determined so that the heat required to meet the temperature rise of the feedwater is equal to the heat that can be provided by the enthalpy difference of the extraction steam ratio. The equivalent resistance coefficient of the circulating cooling water pipeline is calculated based on the hydraulic relationship of the circulating cooling water pipeline. The equivalent resistance coefficient of the circulating pipeline is determined according to the criterion of the relationship between friction and local pressure drop with respect to the square of volumetric flow rate. The cleaning coefficient of condenser cooling tubes, the steam extraction rate of each stage of low-pressure heaters and the equivalent resistance coefficient of circulation pipeline are aligned in time and the sampling window is unified. A fixed-length sliding time window with the current time as the starting point is adopted. The intermediate parameter sequence within the same time window is aligned with the direct operating parameter sequence within the same time window in time order, and the intermediate parameter sequence set under the unified time scale is output. The intermediate parameter sequence set and the direct operating parameter execution features are fused. The features are spliced ​​in the order of priority of equipment association, second priority of process sequence, and supplementation of environmental disturbance. The condenser-related features are placed first, the circulating cooling water system-related features are placed in the middle, and the cooling tower and cooling tower fan-related features and atmospheric environmental parameters are placed last. The feature dimensions within each time window are kept consistent to generate the input feature set.

[0009] Optionally, the construction of the time series training sample set and label set specifically includes: Extract turbine inlet steam parameters, regenerative system adjusted and non-adjusted extraction steam parameters, and atmospheric environmental parameters from the input feature set. Determine the upper and lower limit functions of each parameter value based on the unit load and atmospheric environmental conditions, and establish the operating boundary set. Extract parameters of the circulating cooling water system, cooling tower and cooling tower fan, condenser and atmospheric environment from the input feature set, determine the control constraint range based on the equipment operating boundary and system characteristic curve, and establish a control constraint set. Determine the time window length and sliding step size of the time series samples, extract the input sequence from the input feature set in order of time window, call the running boundary set for each time step to perform boundary checks, and mark the time windows that meet the boundary conditions as candidate samples. For each time window corresponding to the candidate sample, the control constraint set is further called to perform constraint verification. It is required that the speed and number of circulating water pumps, the speed and number of cooling tower fans, the condenser vacuum, the volume flow rate of circulating cooling water and the pipeline pressure drop, and the ventilation volume of cooling tower all meet their respective feasible regions or allowable intervals. Time windows that do not meet the constraints are eliminated, and the start and end times and boundary snapshots of time windows that meet the constraints are recorded. Using the time window that passes the test as a sample, the corresponding input sequence is used as the time series training sample. The cold end state variables and performance indicators within the same time window are used as labels, and they are organized according to the sample number and time step number to construct the time series training sample set and label set.

[0010] Optionally, the generation of the unified prediction model for the cold-end system specifically includes: Based on the time series training sample set and label set, the training set and validation set are divided according to the time window and sliding step size, and the sequence alignment operation is performed. The time feature sequence within the same time window is used as the input of the sub-model, and the cold end state quantity and performance index in the historical running data are used as supervision labels. The condenser sub-model is trained by using an LSTM-PINN network for sequence-to-sequence learning. The time feature sequence is encoded, decoded, predicted, and the parameters are optimized until the network converges and the converged condenser sub-model is output. The circulating water pump sub-model is trained by using a ConvGRU network for sequence-to-sequence learning. With prediction accuracy as the optimization objective, the network parameters are iteratively updated using a gradient descent optimization method, and the converged circulating water pump sub-model is output. The cooling tower sub-model was trained using a Bi-TCN network for sequence-to-sequence learning, and a batch training and early stopping strategy was employed to obtain the converged cooling tower model. The cooling tower fan sub-model was trained using a DGCN network for sequence-to-sequence learning. A graph structure was constructed based on the fan nodes and their connections. The objective function was the sum of the mean squared errors between the predicted values ​​and the supervision labels. Batch training and parameter regularization were used to obtain the converged cooling tower fan sub-model. The four types of sub-models are coupled with parameters. An interface mapping between the sub-models is established according to the cold-end process correlation. The R-GCN network is used to aggregate and unify the features of shared intermediate quantities. The circulating cooling water volume flow rate, condenser vacuum, cooling tower ventilation volume and cooling tower fan speed are dynamically correlated among the sub-models to generate a unified prediction model for the cold-end system.

[0011] Optionally, the feature aggregation and consistency processing of the shared intermediate quantity specifically includes: The converged condenser sub-model, circulating water pump sub-model, cooling tower model, and cooling tower fan sub-model are input into the parameter coupling module. An interface mapping relationship is established based on the cold end process flow. The circulating cooling water volume flow rate, condenser vacuum, cooling tower ventilation volume, and cooling tower fan speed are set as shared intermediate quantities. Parameter transfer and consistency constraints are performed between the sub-models. The circulating cooling water volume flow rate output by the circulating water pump sub-model is used as the input of the condenser sub-model, the vacuum parameters output by the condenser sub-model are used as the input of the cooling tower sub-model, and the ventilation volume output by the cooling tower model is used as the input of the cooling tower fan sub-model, thus forming a parameter interaction path. The R-GCN network is used to perform multi-relation feature aggregation and parameter unification on the output features of each sub-model. A relation graph structure is constructed with each sub-model as a node and energy transfer, fluid coupling and ventilation as multi-type edges. Node initialization and relation embedding are performed to map the feature vectors output by each sub-model to a unified feature space and establish a relation type weight matrix. In the relational convolutional layer, feature aggregation is performed for each relation type. The features of adjacent nodes are weighted and summed according to the corresponding relation weights and then linearly transformed. In the node update layer, non-linear activation function is performed on the aggregated features to generate updated node features. The layer-by-layer propagation and fusion of multi-relation information is achieved through inter-layer stacking and residual connections. The features aggregated through multiple layers of relationships are input into the fusion mapping layer, where global feature mapping and parameter consistency optimization are performed. Shared intermediate quantities are subject to unified constraints and forward propagation verification, and a unified prediction model for the cold-end system is output.

[0012] Optionally, the output of the cold junction state variables and performance indicators specifically includes: Receive the running boundary set and the real-time input feature set, organize the inference samples according to the inference time window and sliding step size, and perform consistency verification, missing value filling and normalization processing on the real-time input feature set; Based on the set of operational boundaries, the feasible region of each time window is filtered, requiring all parameters to take values ​​between the corresponding lower and upper bound functions. Those that do not meet the requirements are eliminated, and the samples that pass the test are retained for inference. The unified prediction model for the cold-end system is invoked to perform forward calculations, and a preliminary prediction sequence of cold-end state variables and performance indicators within the corresponding time window is output. The preliminary predicted sequence is subjected to inverse normalization and unit conversion, and out-of-bounds pruning is performed according to the running boundary set; The pruned prediction sequence is subjected to physical consistency verification. The residual value of each time step is calculated based on the energy conservation relationship and the heat transfer balance relationship. The prediction sequence that exceeds the physical constraint range is corrected and the cold end state quantity and performance index after physical verification are output.

[0013] Optionally, the construction of the multivariate optimization problem specifically includes: calling the unified prediction model of the cold-end system within the inference time window, receiving the real-time input feature set and the operating boundary set, extracting the circulating water pump speed, the number of circulating water pumps, the cooling tower fan speed, and the number of cooling tower fans as optimization decision variables, determining the value range of each variable based on the control constraint set, constructing an objective function with the sum of circulating water pump power consumption and cooling tower fan power consumption as the objective function, and combining the feasible domain constraints of condenser vacuum, circulating cooling water volume flow rate and pipeline pressure drop, and cooling tower ventilation volume to form constraint terms, calling the unified prediction model of the cold-end system to perform forward calculation on the candidate decision variable combination, calculating the corresponding cold-end state variables and performance index prediction values, establishing the mapping relationship between the objective function and the constraint conditions, and obtaining the multivariate optimization problem.

[0014] Optionally, the optimal control instruction set output specifically includes: Set the optimization time window and rolling step size, initialize the initial values ​​of circulating water pump speed, number of circulating water pumps, cooling tower fan speed and number of cooling tower fans, load the control constraint set and set the cutoff constant, discount factor and maximum iteration round; Under the current combination of decision variables, a sequence of candidate decision variables is generated. The speed and number of circulating water pumps and the speed and number of cooling tower fans are constrained according to the allowable value range. The unified prediction model of the cold end system is called to calculate the cold end state variables and performance indicators of the corresponding time series, forming a prediction result set corresponding to the time window. Calculate the immediate reward and add penalties for condenser vacuum, circulating cooling water volume flow rate and pipeline pressure drop, and cooling tower ventilation exceeding the feasible region. Then, sum the time series according to the discount factor to obtain the total reward value. An improved objective function is constructed based on the difference between the current return value and the previous return value. Candidate decision variables are then pruned and updated, with constraints ensuring the variation range remains within a set range. The improved objective function is as follows: ; in, The objective function value for the current iteration. The objective function value from the previous iteration. This is a magnification factor for the difference in returns. This represents the packaging performance return value for the current round. This is the return on packaging performance from the previous round. Here, n represents the penalty coefficient for variable constraints, and n is the total number of control parameters. For the first One candidate control parameter, The change in the candidate control parameter. This represents the sum of the absolute values ​​of all changes in the control parameters; The decision variables are iteratively optimized in a mini-batch manner. In each iteration, the objective function value is updated and a convergence judgment is performed. The optimization is terminated when the objective function gain is less than a preset threshold or the maximum number of iterations is reached. Extract the optimal decision variable set at convergence, calculate the corresponding optimal setpoints for circulating cooling water flow and cooling tower ventilation, and convert them into control commands for circulating water pump speed and number of units, and cooling tower fan speed and number of units based on the control constraint set, and output the optimal control command set.

[0015] The beneficial effects of this invention are: This invention introduces a machine learning-based cold-end system operation optimization method, enabling full-process modeling, prediction, and control optimization of the cold-end operation of thermal power units. Unlike traditional operation optimization methods that rely on empirical models or steady-state thermodynamic formulas, this invention fully utilizes multi-source operating data from the unit, combining the operating characteristics of subsystems such as the circulating cooling water system, cooling tower, and condenser, to establish a multi-dimensional data system encompassing time-series features, energy transfer characteristics, and fluid coupling relationships. Through cleaning, normalization, feature fusion, and windowing of the raw operating data, this invention achieves high-precision data representation of the cold-end system's operating status, providing a unified data foundation for subsequent model training and optimization, and effectively improving the observability and modeling accuracy of complex coupled systems.

[0016] In the modeling stage, this invention employs a multi-sub-model hierarchical modeling strategy and a parameter coupling strategy using a relational graph convolutional network. Independent sub-models are established for four core components: condenser, circulating water pump, cooling tower, and cooling tower fan. Multiple relational features are aggregated and standardized using a relational graph convolutional network. This method maps three types of relationships—energy transfer, fluid circulation, and ventilation—to multi-type edge structures, achieving structured coupling and parameter sharing among multi-source features, thereby generating a unified cold-end system prediction model. Compared to traditional single-model or static feature fusion methods, this invention maintains the physical correlation and parameter consistency between sub-models under dynamic conditions, significantly improving the accuracy and stability of cold-end state variables and performance index predictions. The model prediction output undergoes physical consistency verification based on energy conservation and heat transfer balance, effectively suppressing the accumulation of biases in the data-driven model under unsteady-state conditions, ensuring the physical rationality and engineering usability of the prediction results.

[0017] In the optimization control phase, this invention constructs a multivariate optimization problem based on a unified prediction model for the cold-end system and employs a near-end strategy optimization algorithm for decision-making. By setting the rotational speed and number of circulating water pumps and cooling tower fans as optimization decision variables, and combining this with a set of control constraints to limit the feasible domain of parameters, this invention achieves optimal overall energy efficiency of the cold-end system while ensuring operational safety and equipment constraints. This method can quickly calculate the optimal setpoints for circulating cooling water flow and cooling tower ventilation under dynamic operating conditions and output corresponding pump and fan control commands, thereby achieving energy-saving operation under different loads and environmental conditions. Furthermore, this invention introduces a real-time data-driven online update mechanism, completing model retraining and replacement deployment within a preset update cycle. This adapts to model drift caused by changes in operating conditions and equipment aging, continuously maintaining the effectiveness of prediction and optimization. In summary, this invention achieves significant improvements over existing technologies in terms of data accuracy, modeling reliability, energy consumption optimization, and adaptability, possessing high engineering application value and promotional significance. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0019] Figure 1 This is a flowchart of a machine learning-based method for optimizing the operation of a thermal power unit's cold-end system, as proposed in this invention. Figure 2 This is a schematic diagram of the structure of the unified prediction model for the cold-end system of a machine learning-based optimization method for the operation of the cold-end system of a thermal power unit proposed in this invention. Detailed Implementation

[0020] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0021] refer to Figure 1-2 A machine learning-based method for optimizing the operation of cold-end systems in thermal power units includes the following steps: Collect multi-source operation data from the unit, preprocess it, and generate a standardized operation dataset; Intermediate parameters are calculated based on a standardized running dataset, and time alignment and feature fusion are performed to generate an input feature set. Based on the input feature set, establish the running boundary set and the control constraint set, and perform windowing and filtering according to the preset time window length and sliding step size to construct the time series training sample set and label set; Four types of sub-models are trained based on time series training sample sets and label sets, and R-GCN network is used to perform multi-relation feature aggregation and parameter consistency processing to generate a unified prediction model for cold-end systems. The unified prediction model of the cold-end system is invoked based on the set of operating boundaries and the set of real-time input features, and the output cold-end state variables and performance indicators are output. A multivariate optimization problem is constructed based on a unified prediction model for cold-end systems. The decision variables are set as the speed and number of circulating water pumps and the speed and number of cooling tower fans. A set of control constraints is applied to limit the feasible solution space. The near-end strategy optimization method is used to solve the multivariate optimization problem, obtain the optimal setpoints of circulating cooling water flow rate and cooling tower ventilation volume, and output the optimal control command set. The unified prediction model of the cold-end system is updated online based on real-time operating data. Retraining and deployment replacement are completed within the preset update cycle, and the corresponding optimal control commands are re-output based on the updated model.

[0022] In this embodiment, the multi-source operating data of the unit includes turbine inlet steam parameters, regenerative system adjusted and non-adjusted extraction steam parameters, condenser parameters, circulating cooling water system parameters, cooling tower and cooling tower fan parameters, and atmospheric environment parameters. The preprocessing includes data cleaning, noise filtering, feature alignment, data normalization, and feature filtering.

[0023] In this embodiment, the generation of the input feature set specifically includes: The condenser cooling tube cleanliness coefficient is calculated based on a standardized operating dataset. The condenser cooling tube cleanliness coefficient is determined by the ratio of the actual heat transfer coefficient to the reference heat transfer coefficient according to the heat transfer balance relationship. Specifically, the inlet temperature, outlet temperature, condensing pressure, condensing temperature, and circulating cooling water flow rate of the standardized operating dataset are extracted. The heat transfer of the cooling tube is determined according to the heat balance relationship. The heat transfer is divided by the average temperature difference of the circulating cooling water and the heat transfer area of ​​the condenser to obtain the actual heat transfer coefficient. The unit's reference operating condition is used as the reference state. The circulating cooling water temperature difference, heat transfer, and heat transfer area under the corresponding state are extracted to calculate the reference heat transfer coefficient. The ratio of the actual heat transfer coefficient to the reference heat transfer coefficient is calculated, and the resulting ratio is the condenser cooling tube cleanliness coefficient. At the same time, a time sliding window is used to smooth the data and remove outliers during the calculation process. The steam extraction rate of each stage of the low-pressure heater is calculated based on the energy conservation principle of the regenerative system. The steam extraction rate of the low-pressure heater is determined so that the heat required to meet the feedwater temperature rise is equal to the heat provided by the enthalpy difference of the extraction steam. Specifically, the following steps are taken: extract the inlet feedwater temperature, outlet feedwater temperature, feedwater flow rate, extraction steam pressure and temperature, and heater condensate discharge status parameters of each stage of the low-pressure heater from the standardized operating data. Calculate the heat absorption of the feedwater based on the energy conservation principle of the regenerative system. Determine the heat absorption power of the heater corresponding to the feedwater flow rate and the inlet and outlet temperature differences. Calculate the steam extraction heating capacity by extracting the enthalpy of the corresponding stage of extraction steam at the extraction point and the enthalpy of the heater condensate discharge point. Divide the heat absorption power of the heater by the steam extraction heating capacity to obtain the steam extraction rate of the corresponding stage of the low-pressure heater. Calculate the steam extraction rate of all low-pressure heaters sequentially to form a sequence of steam extraction rates for each stage of the low-pressure heater. At the same time, the steam extraction rate sequence is smoothed by a fixed-length sliding time window and outlier removal is performed. The equivalent resistance coefficient of the circulating cooling water pipeline is calculated based on the hydraulic relationship of the circulating cooling water pipeline. The equivalent resistance coefficient of the circulating pipeline is determined according to the criterion of the relationship between frictional resistance and local pressure drop with respect to the square of volumetric flow rate. Specifically, the following steps are taken: extract the circulating cooling water volumetric flow rate, pipeline inlet pressure, pipeline outlet pressure, circulating water density, pipe diameter, elbow local resistance coefficient, valve local resistance coefficient, and pipeline length from the standardized operating data. Calculate the total pressure drop of the pipeline based on the hydraulic balance relationship of the circulating cooling water system. Take the pressure difference between the inlet and outlet as the total pressure drop. Determine the flow kinetic energy term using the circulating cooling water volumetric flow rate and density. Calculate the comprehensive pressure drop according to the sum of frictional resistance and local resistance. Then divide the total pressure drop by the product of the square of the volumetric flow rate and the density to obtain the equivalent resistance coefficient of the circulating pipeline. Use a fixed-length sliding time window to smooth the continuous operating data and remove outliers. The cleaning coefficient of condenser cooling tubes, the steam extraction rate of each stage of low-pressure heaters and the equivalent resistance coefficient of circulation pipeline are aligned in time and the sampling window is unified. A fixed-length sliding time window with the current time as the starting point is adopted. The intermediate parameter sequence within the same time window is aligned with the direct operating parameter sequence within the same time window in time order, and the intermediate parameter sequence set under the unified time scale is output. The intermediate parameter sequence set and the direct operating parameter execution features are fused. The features are spliced ​​in the order of priority of equipment association, second priority of process sequence, and supplementation of environmental disturbance. The condenser-related features are placed first, the circulating cooling water system-related features are placed in the middle, and the cooling tower and cooling tower fan-related features and atmospheric environmental parameters are placed last. The feature dimensions within each time window are kept consistent to generate the input feature set.

[0024] In this embodiment, the construction of the time series training sample set and label set specifically includes: Extract turbine inlet steam parameters, regenerator system adjusted and non-adjusted extraction steam parameters, and atmospheric environmental parameters from the input feature set. Determine the upper and lower limit functions of each parameter value based on the unit load and atmospheric environmental conditions, and establish an operating boundary set. The operating boundary set is defined as including the inlet steam boundary, extraction steam boundary, and environmental boundary. Parameters of the circulating cooling water system, cooling tower and cooling tower fan, condenser and atmospheric environment are extracted from the input feature set. The control constraint range is determined based on the equipment operating boundary and system characteristic curve, and a control constraint set is established. The control constraint set includes constraints on the speed and number of circulating water pumps, the speed and number of cooling tower fans, the vacuum allowable range, the feasible region constraints on the flow rate and pressure drop of circulating cooling water, and the feasible region constraints on the ventilation volume of cooling tower. Determine the time window length and sliding step size of the time series sample, extract the input sequence from the input feature set according to the time window order, call the running boundary set for each time step to perform boundary checks, and mark the time window that meets the boundary conditions as a candidate sample. The boundary checks ensure that the values ​​of each parameter are not less than the corresponding lower limit function value and not greater than the corresponding upper limit function value. For each time window corresponding to the candidate sample, the control constraint set is further called to perform constraint verification. It is required that the speed and number of circulating water pumps, the speed and number of cooling tower fans, the condenser vacuum, the volume flow rate of circulating cooling water and the pipeline pressure drop, and the ventilation volume of cooling tower all meet their respective feasible regions or allowable intervals. Time windows that do not meet the constraints are eliminated, and the start and end times and boundary snapshots of time windows that meet the constraints are recorded. Using the time window that passes the test as a sample, the corresponding input sequence is used as the time series training sample. The cold end state variables and performance indicators within the same time window are used as labels, and they are organized according to the sample number and time step number to construct the time series training sample set and label set.

[0025] In this embodiment, the generation of the unified prediction model for the cold-end system specifically includes: The training and validation sets are divided according to time windows and sliding step sizes based on the time series training sample set and label set, and sequence alignment is performed. The time feature sequence within the same time window is used as the input of the sub-model. The cold end state variables and performance indicators in the historical operation data are used as supervision labels. The cold end state variables include condenser vacuum, condensate outlet temperature, circulating cooling water volume flow rate, circulating cooling water inlet and outlet temperature difference, cooling tower inlet water temperature, cooling tower outlet water temperature, and cooling tower ventilation. The performance indicators include condenser heat exchange efficiency, circulating cooling efficiency, unit cooling load energy consumption, vacuum loss rate, cooling tower heat balance deviation coefficient, and unit comprehensive thermal efficiency correction coefficient. The condenser sub-model is trained by using an LSTM-PINN network for sequence-to-sequence learning. The time feature sequence is encoded, decoded, predicted, and the parameters are optimized until the network converges and the converged condenser sub-model is output. The training of the condenser sub-model specifically includes: inputting the time feature sequence into the LSTM-PINN network; the input layer receiving the input features at each time step; the long short-term memory encoding layer extracting the temporal dependencies of the time feature sequence through the gating unit and generating the hidden state sequence; the long short-term memory decoding layer generating the prediction sequence step by step with the terminal hidden state as the initial state; obtaining the prediction output through fully connected mapping at each time step; the physical constraint layer applying the energy conservation relationship and heat transfer balance relationship to the prediction output for physical consistency constraints during the forward propagation process; passing the constraint residual to the loss calculation unit; and the output layer outputting the final prediction sequence. During the training process, the error is calculated based on the prediction output and the supervision label, and the loss value is generated by combining the physical constraint residual. Backpropagation and parameter iterative updates are performed based on the loss value until the network converges. The criteria for model convergence are that the mean square error between the predicted value and the supervision label is lower than a preset error threshold, and the average absolute value of the residuals of the energy conservation relationship and the residuals of the heat transfer balance relationship are both lower than a preset physical constraint threshold. The circulating water pump sub-model is trained by using a ConvGRU network for sequence-to-sequence learning. With prediction accuracy as the optimization objective, the network parameters are iteratively updated using a gradient descent optimization method, and the converged circulating water pump sub-model is output. The training of the circulating water pump sub-model specifically includes: inputting the time feature sequence into the ConvGRU network; the input layer receives the time feature sequence and reconstructs its shape to match the convolution calculation dimension; the convolution gated recurrent unit layer performs convolution operations at each time step to extract spatial features, and controls the state transfer through update gates and reset gates to achieve spatiotemporal feature fusion and memory retention; the decoding layer receives the hidden state sequence of the convolution gated recurrent unit layer, generates the output of each time step of the prediction sequence in sequence, and obtains the prediction value through convolution mapping; the output layer outputs the complete prediction sequence; during the training process, the mean square error is calculated based on the prediction value and the supervision label; gradient backpropagation and network parameter iterative updates are performed based on the error until the prediction error meets the preset convergence condition. The cooling tower sub-model was trained using a Bi-TCN network for sequence-to-sequence learning, and a batch training and early stopping strategy was employed to obtain the converged cooling tower model. The training of the cooling tower sub-model specifically includes: inputting the time feature sequence into the Bi-TCN network; the input layer receives the time feature sequence and performs normalization processing; the forward temporal convolutional layer performs one-dimensional dilated convolution in the forward order of the time sequence to extract forward temporal features; the backward temporal convolutional layer performs one-dimensional dilated convolution in the reverse order of the time sequence to extract backward temporal features; the fusion layer concatenates the outputs of the forward and backward convolutional layers and performs nonlinear activation to obtain joint temporal features; the output layer generates a prediction sequence through fully connected mapping; during training, forward propagation is performed on batch time series samples to calculate the prediction value; the mean square error is calculated based on the prediction value and the supervision label; the network parameters are updated using gradient backpropagation; and the verification error is monitored in consecutive training rounds. When the verification error no longer decreases within a preset number of rounds, an early stopping strategy is triggered, and the converged cooling tower sub-model is output. The cooling tower fan sub-model was trained using a DGCN network for sequence-to-sequence learning. A graph structure was constructed based on the fan nodes and their connections. The objective function was the sum of the mean squared errors between the predicted values ​​and the supervision labels. Batch training and parameter regularization were used to obtain the converged cooling tower fan sub-model. The training of the cooling tower fan sub-model specifically includes: inputting the time feature sequence into the DGCN network; the input layer receiving the cooling tower fan operation data and standardizing the features; the dynamic graph construction layer using the fan unit as the graph node and the airflow coupling strength and spatial adjacency relationship as the graph edge weights to generate a dynamically changing adjacency matrix; the temporal graph convolutional layer performing graph convolution calculation based on the current adjacency matrix at each time step to realize feature propagation between nodes and temporal feature extraction; the aggregation layer performing weighted aggregation of node features at each time step and generating global temporal features; and the output layer generating a prediction sequence through a mapping function. During training, time series samples are input in batches, forward propagation is performed to calculate the predicted value, the mean square error is calculated based on the predicted value and the supervision label, and a parameter regularization term is added to the error function to limit the model complexity. Gradient backpropagation and parameter iterative updates are performed based on the total loss value until the loss converges, and the converged cooling tower fan sub-model is output. The four types of sub-models are coupled with parameters. An interface mapping between the sub-models is established according to the cold-end process correlation. The R-GCN network is used to aggregate and unify the features of shared intermediate quantities. The circulating cooling water volume flow rate, condenser vacuum, cooling tower ventilation volume and cooling tower fan speed are dynamically correlated among the sub-models to generate a unified prediction model for the cold-end system.

[0026] In this embodiment, the feature aggregation and consistency processing of the shared intermediate quantity specifically includes: The converged condenser sub-model, circulating water pump sub-model, cooling tower model, and cooling tower fan sub-model are input into the parameter coupling module. An interface mapping relationship is established based on the cold end process flow. The circulating cooling water volume flow rate, condenser vacuum, cooling tower ventilation volume, and cooling tower fan speed are set as shared intermediate quantities. Parameter transfer and consistency constraints are performed between the sub-models. The circulating cooling water volume flow rate output by the circulating water pump sub-model is used as the input of the condenser sub-model, the vacuum parameters output by the condenser sub-model are used as the input of the cooling tower sub-model, and the ventilation volume output by the cooling tower model is used as the input of the cooling tower fan sub-model, thus forming a parameter interaction path. The R-GCN network is used to perform multi-relation feature aggregation and parameter unification on the output features of each sub-model. A relation graph structure is constructed with each sub-model as a node and energy transfer, fluid coupling and ventilation as multi-type edges. Node initialization and relation embedding are performed to map the feature vectors output by each sub-model to a unified feature space and establish a relation type weight matrix. In the relational convolutional layer, feature aggregation is performed for each relation type. The features of adjacent nodes are weighted and summed according to the corresponding relation weights and then linearly transformed. In the node update layer, non-linear activation function is performed on the aggregated features to generate updated node features. The layer-by-layer propagation and fusion of multi-relation information is achieved through inter-layer stacking and residual connections. The features aggregated through multiple layers of relationships are input into the fusion mapping layer, where global feature mapping and parameter consistency optimization are performed. Shared intermediate quantities are subject to unified constraints and forward propagation verification, and a unified prediction model for the cold-end system is output.

[0027] In this embodiment, the output of the cold junction state quantity and performance index specifically includes: Receive the running boundary set and the real-time input feature set, organize the inference samples according to the inference time window and sliding step size, and perform consistency verification, missing value filling and normalization processing on the real-time input feature set; Based on the set of operational boundaries, the feasible region of each time window is filtered, requiring all parameters to take values ​​between the corresponding lower and upper bound functions. Those that do not meet the requirements are eliminated, and the samples that pass the test are retained for inference. The unified prediction model for the cold-end system is invoked to perform forward calculations, and a preliminary prediction sequence of cold-end state variables and performance indicators within the corresponding time window is output. The preliminary predicted sequence is subjected to inverse normalization and unit conversion, and out-of-bounds pruning is performed according to the running boundary set; The pruned prediction sequence is subjected to physical consistency verification. The residual value of each time step is calculated based on the energy conservation relationship and the heat transfer balance relationship. The prediction sequence that exceeds the physical constraint range is corrected. The cold end state quantity and performance index after physical verification are output. The energy conservation relationship is the time average absolute value of the difference between the heat absorbed by the feed water and the heat supplied by multiplying the enthalpy difference of the extraction steam by the extraction steam quantity. The residual of the heat transfer balance relationship is the time average absolute value of the difference between the heat exchange and the actual heat transfer coefficient multiplied by the average temperature difference of the circulating cooling water and the heat exchange area of ​​the condenser.

[0028] In this embodiment, the construction of the multivariate optimization problem specifically includes: calling the unified prediction model of the cold-end system within the inference time window, receiving the real-time input feature set and the operating boundary set, extracting the circulating water pump speed, the number of circulating water pumps, the cooling tower fan speed, and the number of cooling tower fans as optimization decision variables, determining the value range of each variable based on the control constraint set, constructing an objective function with the sum of circulating water pump power consumption and cooling tower fan power consumption as the objective function, and combining the feasible domain constraints of condenser vacuum, circulating cooling water volume flow rate and pipeline pressure drop, and cooling tower ventilation volume to form constraint terms, calling the unified prediction model of the cold-end system to perform forward calculation on the candidate decision variable combination, calculating the corresponding cold-end state variables and performance index prediction values, establishing the mapping relationship between the objective function and the constraint conditions, and obtaining the multivariate optimization problem.

[0029] In this embodiment, the optimal control instruction set output specifically includes: Set the optimization time window and rolling step size, initialize the initial values ​​of circulating water pump speed, number of circulating water pumps, cooling tower fan speed and number of cooling tower fans, load the control constraint set and set the cutoff constant, discount factor and maximum iteration round; Under the current combination of decision variables, a sequence of candidate decision variables is generated. The speed and number of circulating water pumps and the speed and number of cooling tower fans are constrained according to the allowable value range. The unified prediction model of the cold end system is called to calculate the cold end state variables and performance indicators of the corresponding time series, forming a prediction result set corresponding to the time window. The instant reward is calculated and the penalties for condenser vacuum, circulating cooling water volume flow rate and pipeline pressure drop, and cooling tower ventilation volume exceeding the feasible region are added. The time series is summed according to the discount factor to obtain the total reward value. The instant reward is the negative value of the sum of the power consumption of the circulating water pump and the power consumption of the cooling tower fan. An improved objective function is constructed based on the difference between the current return value and the previous return value. Candidate decision variables are then pruned and updated, with constraints ensuring the variation range remains within a set range. The improved objective function is as follows: ; in, The objective function value for the current iteration. The objective function value from the previous iteration. This is a magnification factor for the difference in returns. This represents the packaging performance return value for the current round. This is the return on packaging performance from the previous round. Here, n represents the penalty coefficient for variable constraints, and n is the total number of control parameters. For the first One candidate control parameter, The change in the candidate control parameter. This represents the sum of the absolute values ​​of all changes in the control parameters; The decision variables are iteratively optimized in a mini-batch manner. In each iteration, the objective function value is updated and a convergence judgment is performed. The optimization is terminated when the objective function gain is less than a preset threshold or the maximum number of iterations is reached. Extract the optimal decision variable set at convergence, calculate the corresponding optimal set values ​​of circulating cooling water flow rate and cooling tower ventilation volume, and convert them into control commands for circulating water pump speed and number of units, and cooling tower fan speed and number of units based on the control constraint set. Output the optimal control command set, which includes circulating water pump speed and number of units setting commands, cooling tower fan speed and number of units setting commands, and time synchronization signals.

[0030] Example 1: To verify the feasibility of this invention in practice, it was applied to the cold-end system of a megawatt-class generator unit in a large inland thermal power plant. This unit employs a closed-loop, wet-cooling system. During operation, it is affected by seasonal temperature and load changes, resulting in significant condenser vacuum fluctuations and frequent adjustments to the circulating water pumps and cooling tower fans. This leads to a decrease in the unit's cold-end operating efficiency and significant fluctuations in energy consumption. Under high temperature and humidity conditions in summer, if the cooling tower ventilation is not adjusted in a timely manner, the cold-end heat transfer capacity weakens, and the condensing pressure increases, thereby causing a decrease in turbine efficiency and affecting overall power supply performance. Traditional control methods often rely on manual experience to set equipment operating parameters, lacking global optimization methods based on multi-source data, leading to problems such as delayed adjustment response, inaccurate energy consumption prediction, and fragmented control strategies.

[0031] In practical application, the multi-source data acquisition system proposed in this invention was first deployed to simultaneously collect and standardize data on turbine steam inlet, condenser heat exchange, circulating water pump operation, cooling tower fan operation, and environmental meteorological data, forming a standardized dataset suitable for machine learning. Subsequently, based on this dataset, intermediate parameters such as the condenser cooling tube cleanliness coefficient, low-pressure heater extraction steam rate, and equivalent resistance coefficient of the circulating pipeline were calculated, and time alignment and feature fusion were performed to generate an input feature set. In the training sample construction stage, an operational boundary set and a control constraint set were established based on the input feature set. Historical operational data were divided into time windows, and samples satisfying the boundary and constraint conditions were selected for model training, forming a time-series training sample set and a label set.

[0032] In the model training phase, sub-models for the condenser, circulating water pump, cooling tower, and cooling tower fan were constructed. Sequence-to-sequence learning was performed using a combination of long short-term memory networks (LSTM) with physical constraints, convolutional gated recurrent unit networks (GNU), bidirectional temporal convolutional networks, and dynamic graph convolutional networks. After training, the parameters of the four sub-models were coupled using a relational graph convolutional network. A relational graph structure with energy transfer, fluid coupling, and ventilation effects as multi-type edges was established to achieve feature aggregation and parameter consistency, resulting in a unified prediction model for the cold-end system. This model can predict cold-end state variables and performance indicators, including condenser vacuum, circulating cooling water volumetric flow rate, cooling tower ventilation volume, and circulating cooling efficiency, based on inputs such as unit load, cooling water temperature, and ambient temperature during real-time operation.

[0033] During unit operation optimization, a multivariate optimization problem is constructed based on a predictive model. The rotational speed and number of circulating water pumps and cooling tower fans are used as decision variables, and a feasible solution space is formed by combining the set of control constraints. A near-end strategy optimization method is employed to solve the optimization problem, dynamically generating optimal setpoints for the circulating cooling water flow rate and cooling tower ventilation volume. These setpoints are then converted into control commands for the circulating water pumps and cooling tower fans and input into the control system, achieving automated closed-loop regulation. Through real-time data feedback and updates, the system can periodically perform online retraining and model replacement to ensure that the model parameters maintain accuracy and adaptability under different seasons and operating conditions.

[0034] During continuous operation, monitoring results showed that the cold-end system operated more stably, the condenser vacuum stability significantly improved, the power coordination between the circulating water pump and the cooling tower fan improved, and the cold-end heat exchange efficiency was effectively maintained. The model's output state prediction values ​​matched the field measurements well, indicating that the unified prediction model has high dynamic response capability and generalization performance. Compared with traditional experience-based adjustment methods, this invention can achieve coordinated operation of the cooling water system under the same load conditions, effectively avoiding excessive energy consumption and inefficient cooling.

[0035] The results show that the method of the present invention improves the overall energy efficiency and control accuracy of the cold end system while ensuring the safe and stable operation of the unit, and has practical application value in the cold end system of large thermal power plants.

[0036] Table 1. Performance Comparison of the Invention and Traditional Cold-End System Operation Optimization Methods for Thermal Power Units

[0037] As can be clearly seen from Table 1, the method of the present invention is superior to the traditional method in many indicators.

[0038] Regarding condenser vacuum stability, when operating boundary parameters (turbine load, ambient temperature, ambient humidity, etc.) change, the pressure drop is 1.05 kPa / h, while that of this invention is only 0.52 kPa / h. This indicates that through machine learning modeling and multi-relation feature aggregation, this invention can maintain vacuum stability under complex load changes and environmental disturbances, reducing energy loss in the thermal system. This stability stems from the parameter unification mechanism of the unified prediction model for the cold-end system, which creates a dynamic coupling feedback between the circulating water pump, cooling tower, and fan.

[0039] The response time for adjusting the cooling tower outlet water temperature was reduced from 585 seconds to 350 seconds, demonstrating a significant improvement in response speed due to control strategy optimization. This invention utilizes near-end strategy optimization to solve multivariate control problems, enabling the cooling tower fan to quickly adjust the ventilation volume based on predicted thermal balance deviations. This allows the circulating water temperature to adapt to rapid changes in the operating boundary of the cold-end system, achieving an instantaneous response in heat and mass transfer of the cooling tower and avoiding the accumulation of thermal deviations caused by excessive delays.

[0040] Regarding the stability of the circulating cooling water volumetric flow rate, within a condenser vacuum variation range of 1 kPa, the conventional control method achieves a stability of 26.8%, while the present invention reduces this to 12.1%, significantly improving the stability of flow control. This result reflects that in the method of this invention, the circulating water pump sub-model can automatically adjust its speed and number of pumps based on real-time characteristics, ensuring a stable cooling water supply. Furthermore, the feature aggregation of the R-GCN network effectively suppresses the propagation of hydraulic fluctuations, thereby avoiding localized flow imbalances.

[0041] In terms of energy efficiency, the power consumption rate of cold-end power-consuming equipment decreased from 2.8% to 2.1%, and the cooling energy consumption per unit of net power generation decreased from 26.2 kWh / MWh to 19.5 kWh / MWh, indicating that the overall heat exchange process of the system is more efficient. This improvement is mainly due to the coordinated scheduling of the cooling tower and circulating water system under a unified prediction model, which makes fuller use of the temperature difference in the heat exchange process and shortens the energy conversion chain, thereby effectively reducing energy consumption.

[0042] Regarding the condenser heat balance deviation coefficient, the traditional method yields 5.2%, while this invention reduces it to 1.8%. This index reflects the level of energy conservation deviation in the condenser heat transfer process; a smaller value indicates a higher degree of agreement between the heat transfer calculation and the actual thermodynamic process. This invention introduces physical constraints of energy conservation and heat transfer balance through the LSTM-PINN sub-model, making the prediction process not only dependent on data fitting but also consistent with the actual thermodynamic relationship, thereby significantly reducing the heat balance deviation.

[0043] The cooling tower heat balance deviation coefficient decreased from 4.8% to 1.7%, indicating that the heat exchange process of the cooling tower is closer to the theoretical heat balance state under different environmental and load conditions. This improvement comes from the deep integration of the cooling tower sub-model and the fan sub-model, which keeps the ventilation rate and water temperature difference consistent in prediction and control, reducing the phenomenon of uneven heat exchange.

[0044] The model prediction error (MAPE) decreased from 5.6% to 1.9%, demonstrating that the unified prediction model established in this invention has higher accuracy and stability across different time windows. Because the model incorporates physical constraints and multi-source data features during training, its prediction results better reflect the actual operating conditions of the cold-end system, exhibiting strong generalization ability and real-time reliability.

[0045] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present 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 the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for optimizing the operation of a thermal power unit's cold-end system based on machine learning, characterized in that, Includes the following steps: Collect multi-source operation data from the unit, preprocess it, and generate a standardized operation dataset; Intermediate parameters are calculated based on a standardized running dataset, and time alignment and feature fusion are performed to generate an input feature set. Based on the input feature set, establish the running boundary set and the control constraint set, and perform windowing and filtering according to the preset time window length and sliding step size to construct the time series training sample set and label set; Four types of sub-models are trained based on time series training sample sets and label sets, and R-GCN network is used to perform multi-relation feature aggregation and parameter consistency processing to generate a unified prediction model for cold-end systems. The unified prediction model of the cold-end system is invoked based on the set of operating boundaries and the set of real-time input features, and the output cold-end state variables and performance indicators are output. A multivariate optimization problem is constructed based on a unified prediction model for cold-end systems. The decision variables are set as the speed and number of circulating water pumps and the speed and number of cooling tower fans. A set of control constraints is applied to limit the feasible solution space. The near-end strategy optimization method is used to solve the multivariate optimization problem, obtain the optimal setpoints of circulating cooling water flow rate and cooling tower ventilation volume, and output the optimal control command set. The unified prediction model of the cold-end system is updated online based on real-time operating data. Retraining and deployment replacement are completed within the preset update cycle, and the corresponding optimal control commands are re-output based on the updated model.

2. The method for optimizing the operation of a thermal power unit's cold-end system based on machine learning according to claim 1, characterized in that, The multi-source operating data of the unit includes turbine inlet steam parameters, regenerative system adjusted and non-adjusted extraction steam parameters, condenser parameters, circulating cooling water system parameters, cooling tower and cooling tower fan parameters, and atmospheric environment parameters. The preprocessing includes data cleaning, noise filtering, feature alignment, data normalization, and feature filtering.

3. The method for optimizing the operation of a thermal power unit's cold-end system based on machine learning according to claim 1, characterized in that, The generation of the input feature set specifically includes: The condenser cooling tube cleanliness coefficient is calculated based on a standardized operating dataset. The condenser cooling tube cleanliness coefficient is determined by the ratio of the actual heat transfer coefficient to the reference heat transfer coefficient according to the heat transfer balance relationship. The steam extraction rate of each stage of the low-pressure heater is calculated based on the energy conservation of the regenerative system. The steam extraction rate of the low-pressure heater is determined so that the heat required to meet the temperature rise of the feedwater is equal to the heat that can be provided by the enthalpy difference of the extraction steam ratio. The equivalent resistance coefficient of the circulating cooling water pipeline is calculated based on the hydraulic relationship of the circulating cooling water pipeline. The equivalent resistance coefficient of the circulating pipeline is determined according to the criterion of the relationship between friction and local pressure drop with respect to the square of volumetric flow rate. The cleaning coefficient of condenser cooling tubes, the steam extraction rate of each stage of low-pressure heaters and the equivalent resistance coefficient of circulation pipeline are aligned in time and the sampling window is unified. A fixed-length sliding time window with the current time as the starting point is adopted. The intermediate parameter sequence within the same time window is aligned with the direct operating parameter sequence within the same time window in time order, and the intermediate parameter sequence set under the unified time scale is output. The intermediate parameter sequence set and the direct operating parameter execution features are fused. The features are spliced ​​in the order of priority of equipment association, second priority of process sequence, and supplementation of environmental disturbance. The condenser-related features are placed first, the circulating cooling water system-related features are placed in the middle, and the cooling tower and cooling tower fan-related features and atmospheric environmental parameters are placed last. The feature dimensions within each time window are kept consistent to generate the input feature set.

4. The method for optimizing the operation of a thermal power unit's cold-end system based on machine learning according to claim 1, characterized in that, The construction of the time series training sample set and label set specifically includes: Extract turbine inlet steam parameters, regenerative system adjusted and non-adjusted extraction steam parameters, and atmospheric environmental parameters from the input feature set. Determine the upper and lower limit functions of each parameter value based on the unit load and atmospheric environmental conditions, and establish the operating boundary set. Extract parameters of the circulating cooling water system, cooling tower and cooling tower fan, condenser and atmospheric environment from the input feature set, determine the control constraint range based on the equipment operating boundary and system characteristic curve, and establish a control constraint set. Determine the time window length and sliding step size of the time series samples, extract the input sequence from the input feature set in order of time window, call the running boundary set for each time step to perform boundary checks, and mark the time windows that meet the boundary conditions as candidate samples. For each time window corresponding to the candidate sample, the control constraint set is further called to perform constraint verification. It is required that the speed and number of circulating water pumps, the speed and number of cooling tower fans, the condenser vacuum, the volume flow rate of circulating cooling water and the pipeline pressure drop, and the ventilation volume of cooling tower all meet their respective feasible regions or allowable intervals. Time windows that do not meet the constraints are eliminated, and the start and end times and boundary snapshots of time windows that meet the constraints are recorded. Using the time window that passes the test as a sample, the corresponding input sequence is used as the time series training sample. The cold end state variables and performance indicators within the same time window are used as labels, and they are organized according to the sample number and time step number to construct the time series training sample set and label set.

5. The method for optimizing the operation of a thermal power unit's cold-end system based on machine learning according to claim 1, characterized in that, The generation of the unified prediction model for the cold-end system specifically includes: Based on the time series training sample set and label set, the training set and validation set are divided according to the time window and sliding step size, and the sequence alignment operation is performed. The time feature sequence within the same time window is used as the input of the sub-model, and the cold end state quantity and performance index in the historical running data are used as supervision labels. The condenser sub-model is trained by using an LSTM-PINN network for sequence-to-sequence learning. The time feature sequence is encoded, decoded, predicted, and the parameters are optimized until the network converges and the converged condenser sub-model is output. The circulating water pump sub-model is trained by using a ConvGRU network for sequence-to-sequence learning. With prediction accuracy as the optimization objective, the network parameters are iteratively updated using a gradient descent optimization method, and the converged circulating water pump sub-model is output. The cooling tower sub-model was trained using a Bi-TCN network for sequence-to-sequence learning, and a batch training and early stopping strategy was employed to obtain the converged cooling tower model. The cooling tower fan sub-model was trained using a DGCN network for sequence-to-sequence learning. A graph structure was constructed based on the fan nodes and their connections. The objective function was the sum of the mean squared errors between the predicted values ​​and the supervision labels. Batch training and parameter regularization were used to obtain the converged cooling tower fan sub-model. The four types of sub-models are coupled with parameters. An interface mapping between the sub-models is established according to the cold-end process correlation. The R-GCN network is used to aggregate and unify the features of shared intermediate quantities. The circulating cooling water volume flow rate, condenser vacuum, cooling tower ventilation volume and cooling tower fan speed are dynamically correlated among the sub-models to generate a unified prediction model for the cold-end system.

6. The method for optimizing the operation of a thermal power unit's cold-end system based on machine learning according to claim 5, characterized in that, The feature aggregation and consistency processing of the shared intermediate quantity specifically includes: The converged condenser sub-model, circulating water pump sub-model, cooling tower model, and cooling tower fan sub-model are input into the parameter coupling module. An interface mapping relationship is established based on the cold end process flow. The circulating cooling water volume flow rate, condenser vacuum, cooling tower ventilation volume, and cooling tower fan speed are set as shared intermediate quantities. Parameter transfer and consistency constraints are performed between the sub-models. The circulating cooling water volume flow rate output by the circulating water pump sub-model is used as the input of the condenser sub-model, the vacuum parameters output by the condenser sub-model are used as the input of the cooling tower sub-model, and the ventilation volume output by the cooling tower model is used as the input of the cooling tower fan sub-model, thus forming a parameter interaction path. The R-GCN network is used to perform multi-relation feature aggregation and parameter unification on the output features of each sub-model. A relation graph structure is constructed with each sub-model as a node and energy transfer, fluid coupling and ventilation as multi-type edges. Node initialization and relation embedding are performed to map the feature vectors output by each sub-model to a unified feature space and establish a relation type weight matrix. In the relational convolutional layer, feature aggregation is performed for each relation type. The features of adjacent nodes are weighted and summed according to the corresponding relation weights and then linearly transformed. In the node update layer, non-linear activation function is performed on the aggregated features to generate updated node features. The layer-by-layer propagation and fusion of multi-relation information is achieved through inter-layer stacking and residual connections. The features aggregated through multiple layers of relationships are input into the fusion mapping layer, where global feature mapping and parameter consistency optimization are performed. Shared intermediate quantities are subject to unified constraints and forward propagation verification, and a unified prediction model for the cold-end system is output.

7. The method for optimizing the operation of a thermal power unit's cold-end system based on machine learning according to claim 1, characterized in that, The output of the cold-end state variables and performance indicators specifically includes: Receive the running boundary set and the real-time input feature set, organize the inference samples according to the inference time window and sliding step size, and perform consistency verification, missing value filling and normalization processing on the real-time input feature set; Based on the set of operational boundaries, the feasible region of each time window is filtered, requiring all parameters to take values ​​between the corresponding lower and upper bound functions. Those that do not meet the requirements are eliminated, and the samples that pass the test are retained for inference. The unified prediction model for the cold-end system is invoked to perform forward calculations, and a preliminary prediction sequence of cold-end state variables and performance indicators within the corresponding time window is output. The preliminary predicted sequence is subjected to inverse normalization and unit conversion, and out-of-bounds pruning is performed according to the running boundary set; The pruned prediction sequence is subjected to physical consistency verification. The residual value of each time step is calculated based on the energy conservation relationship and the heat transfer balance relationship. The prediction sequence that exceeds the physical constraint range is corrected and the cold end state quantity and performance index after physical verification are output.

8. The method for optimizing the operation of a thermal power unit's cold-end system based on machine learning according to claim 1, characterized in that, The construction of the multivariate optimization problem specifically includes: calling the unified prediction model of the cold-end system within the inference time window, receiving the real-time input feature set and the operating boundary set, extracting the circulating water pump speed, the number of circulating water pumps, the cooling tower fan speed, and the number of cooling tower fans as optimization decision variables, determining the value range of each variable based on the control constraint set, constructing an objective function with the sum of circulating water pump power consumption and cooling tower fan power consumption as the objective function, and combining the feasible domain constraints of condenser vacuum, circulating cooling water volume flow rate and pipeline pressure drop, and cooling tower ventilation volume to form constraint terms, calling the unified prediction model of the cold-end system to perform forward calculation on the candidate decision variable combinations, calculating the corresponding cold-end state variables and performance index prediction values, establishing the mapping relationship between the objective function and the constraint conditions, and obtaining the multivariate optimization problem.

9. The method for optimizing the operation of a thermal power unit's cold-end system based on machine learning according to claim 1, characterized in that, The optimal control instruction set output specifically includes: Set the optimization time window and rolling step size, initialize the initial values ​​of circulating water pump speed, number of circulating water pumps, cooling tower fan speed and number of cooling tower fans, load the control constraint set and set the cutoff constant, discount factor and maximum iteration round; Under the current combination of decision variables, a sequence of candidate decision variables is generated. The speed and number of circulating water pumps and the speed and number of cooling tower fans are constrained according to the allowable value range. The unified prediction model of the cold end system is called to calculate the cold end state variables and performance indicators of the corresponding time series, forming a prediction result set corresponding to the time window. Calculate the immediate reward and add penalties for condenser vacuum, circulating cooling water volume flow rate and pipeline pressure drop, and cooling tower ventilation exceeding the feasible region. Then, sum the time series according to the discount factor to obtain the total reward value. An improved objective function is constructed based on the difference between the current return value and the previous return value. Candidate decision variables are then pruned and updated, with constraints ensuring the variation range remains within a set range. The improved objective function is as follows: ； in, The objective function value for the current iteration. The objective function value from the previous iteration. This is a magnification factor for the difference in returns. This represents the packaging performance return value for the current round. This is the return on packaging performance from the previous round. Here, n represents the penalty coefficient for variable constraints, and n is the total number of control parameters. For the first One candidate control parameter, The change in the candidate control parameter. This represents the sum of the absolute values ​​of all changes in the control parameters; The decision variables are iteratively optimized in a mini-batch manner. In each iteration, the objective function value is updated and a convergence judgment is performed. The optimization is terminated when the objective function gain is less than a preset threshold or the maximum number of iterations is reached. Extract the optimal decision variable set at convergence, calculate the corresponding optimal setpoints for circulating cooling water flow and cooling tower ventilation, and convert them into control commands for circulating water pump speed and number of units, and cooling tower fan speed and number of units based on the control constraint set, and output the optimal control command set.

Images