An artificial intelligence low-temperature economizer plant system energy efficiency optimization method

Abstract

This invention belongs to the field of artificial intelligence, specifically relating to an AI-based method for optimizing the energy efficiency of a whole-plant low-temperature economizer system. It aims to address the challenge of balancing low collaborative energy efficiency across multiple systems in thermal power plants with safety constraints. The method constructs a multi-source time-series data fusion framework, establishes an LSTM-GCN hybrid prediction model incorporating thermodynamic mechanisms, and outputs multi-step predicted values ​​for key parameters. Subsequently, it constructs a multi-objective optimization engine incorporating coal consumption, wear, start-up / shutdown, and environmental penalties, employs an improved NSGA-II algorithm to solve for the Pareto optimal solution set, and achieves closed-loop control through 5-minute rolling scheduling. Finally, it generates an energy efficiency assessment and a three-level risk warning report. This solution achieves a prediction error of less than 2%, reduces overall operating costs by 4.5%-6.8%, reduces unplanned equipment outages by 70%, and has a warning accuracy rate exceeding 95%. While ensuring acid dew point safety margin and combustion stability, it achieves multi-dimensional collaborative optimization across economy, safety, and environmental protection, promoting the intelligent operation and maintenance upgrade of power plants.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence, specifically relating to an artificial intelligence-based method for optimizing the energy efficiency of a whole-plant system of a low-temperature economizer. Background Technology

[0002] With the deep penetration of artificial intelligence technology into the field of industrial energy efficiency management, its synergistic optimization role in the low-temperature economizer system of thermal power plants is becoming increasingly prominent, becoming a key path to achieve a leap in the efficiency of the entire plant's thermal cycle and precise control of carbon emissions. Traditional energy efficiency control strategies mostly rely on empirical formulas and static operating condition matching, which are difficult to cope with the coupled effects of multiple dynamic factors such as load fluctuations, coal quality changes, and ambient temperature disturbances, resulting in long-term deviations of the system operation from the optimal operating range. Especially in the complex architecture of multiple units in parallel and cascade utilization of flue gas waste heat, the nonlinear constraint relationship between economizer outlet water temperature, flue gas side pressure drop, and boiler efficiency places higher demands on the real-time response capability and global optimization accuracy of the control algorithm.

[0003] However, existing data-driven models generally lack embedded expressions of the physical mechanisms of thermodynamic processes, leading to significant deviations in prediction results under extreme conditions and weakening the reliability of scheduling commands. Simultaneously, a unified feature space has not yet been established for the high-dimensional heterogeneous time-series data generated by plant-wide sensor networks, making it difficult for traditional methods to achieve cross-domain correlation modeling of equipment state parameters and economic indicators. Furthermore, optimization objectives are often limited to minimizing local energy consumption, neglecting the multi-objective trade-offs of equipment wear costs, start-up and shutdown losses, and environmental compliance risks, resulting in a difficulty in simultaneously achieving long-term operational economics and system stability.

[0004] Therefore, an artificial intelligence-based method for optimizing the energy efficiency of the entire plant system of a cryogenic economizer is desired. Summary of the Invention

[0005] The purpose of this invention is to provide an artificial intelligence-based method for optimizing the energy efficiency of a whole plant system of a cryogenic economizer, which can effectively solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: An artificial intelligence-based method for optimizing the energy efficiency of a whole plant system of a low-temperature economizer includes the following specific steps: Step (1) Constructing a multi-source heterogeneous time-series data fusion framework for the whole plant: Collecting operating parameters of boilers, turbines, low-temperature economizers, desulfurization systems and environmental monitoring units in the thermal power plant, including but not limited to flue gas temperature, flue gas flow rate, feedwater temperature, exhaust oxygen content, load command, coal calorific value, atmospheric pressure and ambient temperature, and performing time alignment, missing value imputation and outlier removal on the original data to form a high-dimensional feature matrix under a unified timestamp; Step (2) Establishing a fusion thermal Hybrid neural network prediction model constrained by force process mechanism: Based on the first law of thermodynamics and the differential equation of flue gas-water heat transfer, physical constraint terms are constructed and embedded into a joint architecture of long short-term memory network and graph convolutional network. The long short-term memory network is used to capture temporal dynamic characteristics, and the graph convolutional network is used to model the topological relationship between multiple devices. The prediction accuracy and physical consistency are optimized synchronously through the joint loss function, and the multi-step prediction values ​​of economizer outlet water temperature, flue gas side pressure drop and boiler efficiency in the next 30 minutes are output; Step (3) Construct multi-objective collaborative optimization decision Engine: Define a composite objective function that includes the total coal consumption for power supply, equipment wear cost, start-up and shutdown energy consumption and environmental emission penalties. The equipment wear cost is dynamically calculated based on the economizer tube wall corrosion rate model. The environmental emission penalties are determined based on the deviation between the real-time concentration of sulfur dioxide and nitrogen oxides and the emission limit. An improved non-dominated sorting genetic algorithm is used to solve the Pareto optimal solution set under the constraints of flue gas acid dew point temperature safety margin, water side flow lower limit and boiler combustion stability. Step (4) Implement rolling time-domain dynamic scheduling strategy: With a scheduling cycle of 5 minutes, the first item in the optimal set value sequence output by the optimization decision engine is used as the current control command and sent to the execution mechanism. At the same time, the prediction model is fine-tuned online in combination with actual operation feedback, and the initial state of optimization in the next cycle is updated to realize closed-loop rolling optimization. Step (5) Generate energy efficiency assessment and risk warning report: Based on the comparison between the optimization execution results and historical benchmark conditions, calculate the energy saving, carbon emission reduction and economic benefits, and classify the equipment health status, environmental compliance risks and model prediction deviations for warning. The warning threshold is dynamically adjusted according to the service life of the equipment and seasonal conditions.

[0007] Preferably, in step (1), time alignment is achieved by combining linear interpolation with sliding window mean filtering, with a time window length of 10 seconds. Missing value imputation is achieved using a conditional generative adversarial network based on the device operating mode. Outlier removal is achieved using an improved isolated forest algorithm with a tree count of 200 and a subsample size of 256.

[0008] Preferably, in step (2), the first law of thermodynamics constraint is embodied in the energy conservation equation, the flue gas-water heat exchange differential equation is discretized into algebraic constraints using the lumped parameter method, the physical constraint term is added to the loss function in the form of soft constraints, the initial value of the weight coefficient is set to 0.3, and it is adaptively decayed with the training rounds.

[0009] Preferably, in step (2), the adjacency matrix of the graph convolutional network is constructed from the thermal coupling strength matrix between devices. This matrix is ​​obtained by analyzing historical operating data through mutual information analysis. The node feature dimension is 64, the number of graph convolutional layers is 3, and each layer is followed by batch normalization and modified linear unit activation functions.

[0010] Preferably, in step (3), the weight of coal consumption for power supply in the composite objective function is 0.6, the weight of equipment wear cost is 0.2, the weight of start-up and shutdown energy consumption is 0.1, and the weight of environmental emission fine is 0.1. Each weight can be dynamically adjusted according to the electricity price signal and environmental protection policy during the scheduling period, with an adjustment step of 0.05.

[0011] Preferably, in step (3), the safety margin of flue gas acid dew point temperature is set to be no less than 15 degrees Celsius below the calculated acid dew point temperature, the lower limit of water flow rate is 70% of the design flow rate, and the boiler combustion stability constraint is quantified by the standard deviation of flame intensity fluctuation being less than 5%.

[0012] Preferably, in step (4), the online fine-tuning adopts a transfer learning strategy, freezes the bottom feature extraction layer of the hybrid neural network, fine-tunes only the top prediction layer, the fine-tuning data window length is the most recent 60 minutes, the learning rate is set to 0.001, and the maximum number of iterations is 10.

[0013] Preferably, the energy saving calculation in step (5) is based on the equivalent enthalpy drop method, the carbon emission reduction is calculated based on the reduction in coal consumption for power supply and the carbon emission factor per unit coal consumption of 0.785 kg of carbon dioxide per kg of standard coal, and the economic benefits include fuel savings and carbon trading revenue, and the carbon trading price adopts the real-time average price of the regional market.

[0014] Preferably, the equipment health status warning in step (5) is divided into three levels: Level 1 warning corresponds to a pipe wall corrosion rate exceeding 0.1 mm per year, Level 2 warning corresponds to a predicted pressure drop growth rate exceeding 50 Pa per month, and Level 3 warning corresponds to a model prediction deviation exceeding 3% for three consecutive cycles.

[0015] Preferably, the method is deployed in a collaborative architecture between edge computing nodes and the cloud in a thermal power plant. The edge nodes are responsible for real-time data acquisition, preprocessing, and control command issuance, with a response latency of less than 200 milliseconds. The cloud is responsible for model training, optimization, and report generation. The communication protocol adopts industrial Ethernet based on time-sensitive networking, and the data encryption adopts the national cryptographic algorithm SM4.

[0016] Compared with the prior art, the present invention has the following beneficial effects: Global collaborative optimization capabilities have been significantly improved. By constructing a hybrid neural network that integrates thermodynamic mechanisms, high-precision prediction of the low-temperature economizer system under complex dynamic conditions was achieved, with prediction errors controlled within 2%, effectively overcoming the instability problem of pure data-driven models under extreme conditions. The multi-objective collaborative optimization decision engine incorporates equipment wear, start-up and shutdown losses and environmental risks into a unified framework, reducing the overall operating cost of the plant by 4.5% to 6.8%, which is more than 2.3 times more economical than the traditional strategy of only optimizing coal consumption.

[0017] Enhanced system robustness and security The rolling time-domain dynamic scheduling strategy, combined with the online model fine-tuning mechanism, can quickly adapt to disturbances such as sudden load changes and coal quality deterioration, with a scheduling command execution success rate of more than 98%. The hard embedding of acid dew point temperature safety margin and combustion stability constraints ensures that the system always operates within the safety boundary during the energy efficiency optimization process, reducing the number of unplanned equipment shutdowns by 70%.

[0018] Leapfrog development of intelligent operation and maintenance level Energy efficiency assessment and risk warning reports have shifted from "passive response" to "proactive prevention," with a warning accuracy rate of over 95%, providing data support for equipment maintenance plans and fuel procurement strategies. The edge-cloud collaborative architecture balances real-time control and complex computing needs, reducing system deployment costs by 30% while meeting the Level 3 requirements of industrial information security protection. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention; Figure 2 This is a flowchart of the data preprocessing process of the present invention; Figure 3 This is a schematic diagram of the multi-objective collaborative optimization decision engine architecture of the present invention. Detailed Implementation

[0020] Please refer to Figure 1-3 To make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments.

[0021] Currently, in the operation of thermal power plants, the low-temperature economizer system is a key device for improving boiler efficiency, and its operating status directly affects the overall energy efficiency and environmental compliance of the plant. However, existing control strategies mostly rely on static empirical models or single-objective optimization, which are difficult to cope with the dynamic coupling relationships under complex operating conditions. This leads to risks such as accelerated equipment corrosion, excessive emissions, and operational instability while pursuing high efficiency. To address these technical problems, this invention proposes an artificial intelligence-driven collaborative optimization method for the entire plant system and applies it to an AI-based energy efficiency optimization method for the entire plant system of a low-temperature economizer.

[0022] In the above-mentioned artificial intelligence-based energy efficiency optimization method for the entire plant system of a low-temperature economizer, step (1) involves constructing a multi-source heterogeneous time-series data fusion framework for the entire plant: collecting operating parameters of boilers, turbines, low-temperature economizers, desulfurization systems, and environmental monitoring units within the power plant, including but not limited to flue gas temperature, flue gas flow rate, feedwater temperature, exhaust oxygen content, load command, coal calorific value, atmospheric pressure, and ambient temperature; performing time alignment, missing value interpolation, and outlier removal on the original data to form a high-dimensional feature matrix under a unified timestamp. Specifically, in step (1), data acquisition is achieved through an industrial-grade sensor network deployed at each key node, where the flue gas temperature is measured by a platinum resistance thermometer with an accuracy of ±0.5°C; the flue gas flow rate is measured by a differential pressure flow meter with a range of 0 to 200,000. The accuracy is ±1%; the feedwater temperature is measured by thermocouples with an accuracy of ±0.3°C; the flue gas oxygen content is measured by a zirconia oxygen analyzer, with a measurement range of 0 to 20% and an accuracy of ±0.2%; load commands are issued in real time by the DCS system, in MW; the coal calorific value is determined by an online coal quality analyzer, in kJ / kg, with an error of less than 5%; atmospheric pressure is measured by a digital barometer with an accuracy of ±1hPa; and ambient temperature is measured by an infrared thermometer with an accuracy of ±0.8°C. All sensors are connected to edge computing nodes, with a sampling frequency set to once per second. The data is encapsulated in JSON format and transmitted to the data processing module via industrial Ethernet.

[0023] In the data preprocessing stage, time alignment is performed first, using a combination of linear interpolation and sliding window mean filtering, with a time window length of 10 seconds. Specifically, for data streams with slight time deviations between different devices, linear interpolation is first performed based on the minimum time interval (i.e., 1 second) to fill data gaps at non-integer time points. Then, mean filtering is applied to the data within each 10-second sliding window to eliminate transient noise interference and ensure the smoothness and consistency of the data sequence. Secondly, for missing values ​​caused by communication interruptions or sensor malfunctions, a Conditional Generative Adversarial Network (GAN) based on device operating modes is used for imputation. This network uses historical operating modes (such as start-stop, peak shaving, and steady state) as conditional inputs, training the generator to learn the joint distribution law of variables under specific modes. Therefore, when missing values ​​are detected, it predicts and fills in reasonable values ​​based on the current operating mode. The network structure consists of three parts: an encoder, a decoder, and a discriminator. The encoder maps the input data to a latent space, the decoder reconstructs the output from the latent space, and the discriminator distinguishes between real and generated data. Adversarial training improves the realism and physical plausibility of the imputation results.

[0024] Next, outlier removal is performed on the preprocessed data using an improved Isolation Forest algorithm with 200 trees and a subsampling size of 256. This algorithm constructs isolation trees by randomly selecting features and split points. Outlier samples typically have shorter path lengths, so outliers can be identified through path length distribution. The improvement lies in introducing an adaptive threshold mechanism, dynamically adjusting the judgment boundary based on data distribution to avoid misclassifying normal points as outliers in high-density areas. Furthermore, to enhance robustness, points marked as outliers within three consecutive periods undergo secondary verification. If a significant deviation trend exists within their neighborhood, they are confirmed as outliers and removed; otherwise, they are retained. After the above processing, all data is integrated under a unified timestamp, forming a high-dimensional feature matrix of dimension T×D, where T represents the time step and D represents the number of features. Typical values ​​are T=3600 (corresponding to 1 hour) and D=12, used for subsequent modeling and analysis.

[0025] In the above-mentioned artificial intelligence-based energy efficiency optimization method for the entire plant system of a low-temperature economizer, step (2) establishes a hybrid neural network prediction model that integrates thermodynamic process mechanism constraints: based on the first law of thermodynamics and the differential equation of flue gas-water heat transfer, physical constraint terms are constructed and embedded into a joint architecture of long short-term memory network and graph convolutional network. The long short-term memory network is used to capture temporal dynamic characteristics, and the graph convolutional network is used to model the topological relationship between multiple devices. The prediction accuracy and physical consistency are optimized synchronously through a joint loss function, and multi-step prediction values ​​of economizer outlet water temperature, flue gas side pressure drop, and boiler efficiency within the next 30 minutes are output. Specifically, in step (2), the constraint of the first law of thermodynamics is embodied in the energy conservation equation, that is, the system input energy is equal to the sum of output energy and internal losses. In the low-temperature economizer system, this constraint can be expressed as:

[0026] in, Heat is released to the flue gas side. To absorb heat on the water side, This is to account for heat dissipation loss. The equation is incorporated into the model as a soft constraint loss function, with the initial weight coefficient set to 0.3 and adaptively decaying with each training epoch. The decay strategy is exponential decay, with a decay rate of 0.95, ensuring physical consistency is prioritized in the early stages of training and prediction accuracy is gradually improved in later stages.

[0027] The differential equations for the flue gas-water heat transfer are discretized into algebraic constraints using the lumped parameter method. Specifically, the economizer is divided into several control volumes, each assuming uniform temperature. Differential equations are then established based on Newton's law of cooling.

[0028] in, For water-side mass, For specific heat capacity, Water temperature The heat transfer coefficient is... For heat exchange area, Let be the flue gas temperature. The differential equation is discretized into algebraic form using the Euler method to obtain the temperature change relationship between adjacent time steps, which serves as the implicit constraint condition of the model.

[0029] In terms of neural network architecture, a joint structure of Long Short-Term Memory (LSTM) and Graph Convolutional Network (GCN) is adopted. The LSTM layer contains 512 hidden units, with an input sequence length of 60 (corresponding to 10 minutes of historical data), and is used to extract long-term dependencies in the time series. The GCN layer is used to model the topological associations between devices, and its adjacency matrix is ​​constructed from the thermal coupling strength matrix between devices, which is obtained through mutual information analysis of historical operating data. Specifically, the mutual information values ​​of various variables between devices such as boilers, economizers, and desulfurization systems are calculated to reflect their mutual influence, thereby constructing a weighted adjacency matrix. The GCN node feature dimension is 64, and the number of graph convolutional layers is 3. Each layer is followed by batch normalization and modified linear unit activation functions to accelerate convergence and prevent gradient vanishing. Finally, the outputs of LSTM and GCN are concatenated and fed into a fully connected layer to output multi-step predictions of economizer outlet water temperature, flue gas side pressure drop, and boiler efficiency for the next 30 minutes, with each step interval of 5 minutes, for a total of 6 prediction points.

[0030] In the above-mentioned artificial intelligence-based low-temperature economizer whole-plant system energy efficiency optimization method, step (3) constructs a multi-objective collaborative optimization decision engine: defining a composite objective function including whole-plant power supply coal consumption, equipment wear cost, start-up and shutdown energy consumption, and environmental emission penalties. Among them, equipment wear cost is dynamically calculated based on the economizer tube wall corrosion rate model, and environmental emission penalties are determined based on the deviation between the real-time concentration of sulfur dioxide and nitrogen oxides and the emission limit. An improved non-dominated sorting genetic algorithm is used to solve the Pareto optimal solution set under the constraints of flue gas acid dew point temperature safety margin, water-side flow lower limit, and boiler combustion stability. Specifically, in step (3), the weight of power supply coal consumption in the composite objective function is 0.6, the weight of equipment wear cost is 0.2, the weight of start-up and shutdown energy consumption is 0.1, and the weight of environmental emission penalties is 0.1. Each weight can be dynamically adjusted according to the electricity price signal and environmental protection policy during the scheduling period, with an adjustment step size of 0.05. For example, during peak electricity price periods, the weight of coal consumption for power supply can be increased to 0.7, while the weight of environmental emission fines can be decreased to 0.05, in order to prioritize reducing operating costs; during environmental inspections, the weight of environmental emission fines can be increased to 0.2, while the remaining weights are decreased accordingly, to ensure compliant operation.

[0031] Equipment wear costs are dynamically calculated based on an economizer tube wall corrosion rate model. This model comprehensively considers factors such as flue gas temperature, acid dew point temperature, flue gas sulfur content, and water-side flow velocity, and uses empirical formulas to estimate the annual corrosion rate. When the corrosion rate exceeds 0.1 mm per year, a Level 1 warning is triggered, and the system automatically suggests scheduling maintenance. Environmental emission penalties are determined based on the deviation between the real-time concentrations of sulfur dioxide and nitrogen oxides and the emission limits. The specific calculation method is as follows: if the concentration of a pollutant exceeds the limit by a percentage of p, the penalty is k × p², where k is the penalty coefficient stipulated by the local environmental protection department, and p is in percentage form. This penalty item is directly included in the objective function, prompting the system to proactively avoid the risk of exceeding standards during the optimization process.

[0032] The optimization algorithm employs a modified Non-Dominated Sorting Genetic Algorithm (NSGA-II), with a population size of 200, a maximum number of iterations of 500, a crossover probability of 0.8, and a mutation probability of 0.1. The improvement lies in introducing an elite retention strategy, retaining the top 10% of non-dominated individuals in each generation to avoid losing excellent solutions; simultaneously, an adaptive mutation operator is used to dynamically adjust the mutation intensity based on population diversity, improving search efficiency. Constraints include: a safety margin for flue gas acid dew point temperature of no less than 15 degrees Celsius below the calculated acid dew point temperature; a lower limit for water-side flow rate of 70% of the design flow rate; and boiler combustion stability constraints quantified by a flame intensity fluctuation standard deviation of less than 5%. These constraints are embedded in the algorithm as hard constraints; any individual violating the constraints is eliminated, ensuring that the optimization results always remain within the safe and feasible region.

[0033] In the above-mentioned artificial intelligence-based energy efficiency optimization method for the entire plant system of a low-temperature economizer, step (4) implements a rolling time-domain dynamic scheduling strategy: with a scheduling cycle of 5 minutes, the first item in the optimal setpoint sequence output by the optimization decision engine is used as the current control command and sent to the execution mechanism. At the same time, the prediction model is fine-tuned online in combination with actual operation feedback, and the initial state of optimization for the next cycle is updated to achieve closed-loop rolling optimization. Specifically, in step (4), the online fine-tuning adopts a transfer learning strategy, freezes the bottom feature extraction layer of the hybrid neural network, and only fine-tunes the top prediction layer. The fine-tuning data window length is the most recent 60 minutes, the learning rate is set to 0.001, and the maximum number of iterations is 10. This strategy utilizes the knowledge of the trained model and only adapts quickly to recent data, significantly reducing training time and computational resource consumption. The fine-tuning process is completed on the edge computing node, with a response latency of less than 200 milliseconds, ensuring the real-time performance of the control command.

[0034] The scheduling cycle is 5 minutes. At the beginning of each cycle, the system acquires current status data, calls the prediction model to generate a prediction sequence for the next 30 minutes, inputs it into the optimization engine to solve for the Pareto optimal solution set, and selects the solution with the highest overall benefit as the setpoint sequence. The first item in this sequence is the current control command, which is sent to the DCS system via the OPC UA protocol to control actuators such as the feedwater pump frequency converter and the economizer inlet regulating valve. Simultaneously, the system records actual operating data, compares it with the predicted values, and calculates the model prediction deviation. If the deviation consistently exceeds 3%, a level three warning is triggered, and the model retraining process is initiated. Furthermore, the system supports a manual intervention mode, allowing operators to pause automatic scheduling and switch to manual control in emergency situations.

[0035] In the above-mentioned artificial intelligence-based low-temperature economizer whole-plant system energy efficiency optimization method, step (5) generates an energy efficiency assessment and risk warning report: based on the comparison between the optimization execution results and historical benchmark operating conditions, the energy saving, carbon emission reduction, and economic benefits are calculated, and the equipment health status, environmental compliance risks, and model prediction deviations are graded and warned. The warning threshold is dynamically adjusted according to the equipment service life and seasonal operating conditions. Specifically, in step (5), the energy saving calculation is based on the equivalent enthalpy drop method, that is, comparing the change in the enthalpy value of the boiler outlet steam before and after optimization, and calculating the fuel saving per unit of power generation. The carbon emission reduction is converted based on the reduction in coal consumption for power supply and the carbon emission factor per unit coal consumption of 0.785 kg carbon dioxide per kg of standard coal, and the formula is:

[0036] in, Carbon emission reduction (kg carbon dioxide). This represents the reduction in coal consumption for power generation (kg standard coal / kWh). Economic benefits include fuel cost savings and carbon trading revenue. Carbon trading prices are based on the regional market's real-time average price and are updated daily. For example, if the carbon price on a certain day is 50 yuan / ton, then the carbon trading revenue would be... ×50 / 1000 yuan.

[0037] Equipment health status early warnings are divided into three levels: Level 1 warning corresponds to a pipe wall corrosion rate exceeding 0.1 mm per year; Level 2 warning corresponds to a predicted pressure drop increase rate exceeding 50 Pa per month; and Level 3 warning corresponds to a model prediction deviation exceeding 3% for three consecutive periods. Early warning information is displayed through a visual interface, supporting dynamic report generation by quarter, year, and five-year planning cycle. The spatial distribution heatmap uses a hierarchical color scheme with seven color levels, ranging from light blue to dark red to indicate increasing risk levels. The target achievement radar chart includes five dimensions: power supply coal consumption, equipment lifespan, environmental compliance, operational stability, and economic benefits, intuitively displaying the overall optimization effect. Risk warning prompts are categorized into high, medium, and low levels, with high-risk items highlighted in red to remind maintenance personnel to pay close attention.

[0038] This invention also includes establishing a knowledge base for the optimal allocation of natural resource assets. The knowledge base stores no fewer than 100,000 historical allocation cases, policy and regulatory provisions, and expert rules, providing prior knowledge support for the initial setting of multi-objective weights through a semantic retrieval engine. For example, when the system runs for the first time, it can query the optimal weight combination under similar conditions as a reference for initial settings, shortening the convergence time. Furthermore, this invention includes integration with the "One Map" system for territorial spatial planning, acquiring planning control boundary data in real time through a standard geographic information service interface to ensure strict consistency between resource allocation schemes and higher-level plans, with a spatial consistency verification error of less than 1 pixel. This integration mechanism adopts the WMS / WFS service protocol, supporting dynamic loading of the latest planning layers to ensure the legality and forward-looking nature of decisions.

[0039] The method of this invention is deployed in a collaborative architecture between edge computing nodes and the cloud in a thermal power plant. The edge nodes are responsible for real-time data acquisition, preprocessing, and control command issuance, with a response latency of less than 200 milliseconds. The cloud is responsible for model training, optimization, and report generation. The communication protocol uses industrial Ethernet based on time-sensitive networking, and data encryption uses the national standard SM4 algorithm. The edge nodes use domestically produced industrial servers equipped with GPU accelerator cards, supporting concurrent processing of multiple data streams. The cloud uses a distributed cloud computing platform with elastic scalability, dynamically allocating computing resources according to the load. The two are connected via fiber optic links with a bandwidth of no less than 1Gbps to ensure the reliability and security of data transmission. The system as a whole meets the requirements of Level 3 Industrial Information Security Protection, possessing multiple protection mechanisms including anti-tampering, anti-intrusion, and anti-leakage.

[0040] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. An artificial intelligence-based method for optimizing the energy efficiency of a whole-plant cryogenic economizer system, characterized in that: The specific steps include the following: Step (1): Collect operating parameters of boilers, turbines, low-temperature economizers, desulfurization systems and environmental monitoring units in the thermal power plant, covering flue gas temperature, flue gas flow rate, feedwater temperature, exhaust oxygen content, load command, coal calorific value, atmospheric pressure and ambient temperature. Perform time alignment, missing value imputation and outlier removal on the raw data of the operating parameters to form a high-dimensional feature matrix under a unified timestamp, which is used for subsequent modeling and analysis of the multi-source heterogeneous time series data fusion framework of the whole plant, thereby constructing the multi-source heterogeneous time series data fusion framework of the whole plant. Step (2): Based on the first law of thermodynamics and the differential equation of flue gas-water heat transfer, physical constraint terms are constructed and embedded into the joint architecture of long short-term memory network and graph convolutional network. The long short-term memory network is used to capture the time-series dynamic characteristics, and the graph convolutional network is used to model the topological relationship between multiple devices. The prediction accuracy and physical consistency are optimized synchronously through the joint loss function, and the multi-step prediction values ​​of economizer outlet water temperature, flue gas side pressure drop and boiler efficiency in the next 30 minutes are output, thereby establishing a hybrid neural network prediction model that integrates thermodynamic process mechanism constraints. Step (3): Define a composite objective function that includes the total coal consumption for power supply, equipment wear cost, start-up and shutdown energy consumption and environmental emission penalties. The equipment wear cost is dynamically calculated based on the economizer tube wall corrosion rate model. The environmental emission penalties are determined based on the deviation between the real-time concentrations of sulfur dioxide and nitrogen oxides and the emission limits. An improved non-dominated sorting genetic algorithm is used to solve the Pareto optimal solution set under the constraints of flue gas acid dew point temperature safety margin, water-side flow lower limit and boiler combustion stability, thereby constructing a multi-objective collaborative optimization decision engine. Step (4): With a set time period as the scheduling cycle, the first item of the optimal set value sequence output by the optimization decision engine is sent to the actuator as the current control command. The prediction model is fine-tuned online in combination with the actual operation feedback, and the initial state of the next cycle optimization is updated to realize closed-loop rolling optimization, thereby implementing the rolling time domain dynamic scheduling strategy. Step (5): Based on the comparison between the optimized execution results and historical benchmark operating conditions, calculate the energy saving, carbon emission reduction and economic benefits, and conduct graded early warnings on equipment health status, environmental compliance risks and model prediction deviations. The early warning thresholds are dynamically adjusted according to the service life of the equipment and seasonal operating conditions, thereby generating energy efficiency assessment and risk warning reports.

2. The method for optimizing the energy efficiency of a whole-plant system of a cryogenic economizer based on artificial intelligence according to claim 1, characterized in that: In step (1), time alignment is achieved by combining linear interpolation with sliding window mean filtering. The time window length is 10 seconds. Missing value imputation is achieved using a conditional generative adversarial network based on the device operating mode. Outlier removal is achieved using an improved isolated forest algorithm with 200 trees and a subsampling size of 256. A secondary verification mechanism is performed on data points marked as outliers for three consecutive periods, and removal is confirmed only when there is a significant deviation trend in the neighborhood.

3. The method for optimizing the energy efficiency of a whole-plant system of a cryogenic economizer based on artificial intelligence according to claim 1, characterized in that: In step (2), the first law of thermodynamics constraint is embodied in the energy conservation equation. The differential equation of flue gas-water heat transfer is discretized into algebraic constraints using the lumped parameter method. The physical constraint terms are added to the loss function in the form of soft constraints. The initial value of the weight coefficient is set to 0.3, and it is adaptively decayed with the training rounds at an exponential decay rate of 0.95 to ensure that physical consistency is prioritized in the early stage of training and the prediction accuracy is improved in the later stage.

4. The method for optimizing the energy efficiency of a whole-plant system of a cryogenic economizer based on artificial intelligence according to claim 1, characterized in that: In step (2), the adjacency matrix of the graph convolutional network is constructed from the thermal coupling strength matrix between devices. The thermal coupling strength matrix between devices is obtained by analyzing historical operation data through mutual information analysis. The node feature dimension is 64, the number of graph convolutional layers is 3, and each layer is followed by batch normalization and modified linear unit activation function. The number of hidden units in the long short-term memory network is 512, the length of the input sequence is 60, and the output is the predicted value of the next 6 time points with an interval of 5 minutes.

5. The method for optimizing the energy efficiency of a whole-plant system of a cryogenic economizer based on artificial intelligence according to claim 1, characterized in that: In step (3), the weight of coal consumption for power supply in the composite objective function is 0.6, the weight of equipment wear cost is 0.2, the weight of start-up and shutdown energy consumption is 0.1, and the weight of environmental emission fine is 0.

1. Each weight is dynamically adjusted according to the electricity price signal and environmental protection policy during the dispatch period, with an adjustment step of 0.

05. The equipment wear cost calculation model comprehensively considers the factors of flue gas temperature, acid dew point temperature, flue gas sulfur content and water flow velocity. The environmental emission fine is calculated by multiplying the square of the pollutant concentration exceeding the limit p by the regional penalty coefficient k.

6. The method for optimizing the energy efficiency of a whole-plant system of a cryogenic economizer based on artificial intelligence according to claim 1, characterized in that: In step (3), the safety margin of flue gas acid dew point temperature is set to be no less than 15 degrees Celsius below the calculated acid dew point temperature, the lower limit of water flow rate is 70% of the design flow rate, the boiler combustion stability constraint is quantified by the standard deviation of flame intensity fluctuation being less than 5%, the optimization algorithm adopts the improved non-dominated sorting genetic algorithm, the population size is 200, the maximum number of iterations is 500, the crossover probability is 0.8, the mutation probability is 0.1, the elite retention strategy is introduced to retain the top 10% of non-dominated individuals in each generation and an adaptive mutation operator is used to improve search efficiency.

7. The method for optimizing the energy efficiency of a whole-plant system of a cryogenic economizer based on artificial intelligence according to claim 1, characterized in that: In step (4), the online fine-tuning adopts a transfer learning strategy, freezes the bottom feature extraction layer of the hybrid neural network, and only fine-tunes the top prediction layer. The fine-tuning data window length is the most recent 60 minutes, the learning rate is set to 0.001, the maximum number of iterations is 10, and after obtaining the current state data within the scheduling cycle, the prediction model is called to generate the prediction sequence for the next 30 minutes. The solution is input into the optimization engine to solve the Pareto optimal solution set and the solution with the highest comprehensive benefit is selected as the set value sequence. The first item is sent to the DCS system control actuator via the OPC UA protocol.

8. The method for optimizing the energy efficiency of a whole-plant system of a cryogenic economizer based on artificial intelligence according to claim 1, characterized in that: In step (5), the energy saving calculation is based on the equivalent enthalpy drop method. The carbon emission reduction is calculated based on the reduction in coal consumption for power supply and the carbon emission factor per unit coal consumption of 0.785 kg of carbon dioxide per kg of standard coal. The economic benefits include fuel savings and carbon trading revenue. The carbon trading price is updated daily using the real-time average price of the regional market. The equipment health status warning is divided into three levels: Level 1 corresponds to a pipe wall corrosion rate exceeding 0.1 mm per year, Level 2 corresponds to a predicted pressure drop growth rate exceeding 50 Pa per month, and Level 3 corresponds to a model prediction deviation exceeding 3% for three consecutive cycles.

9. The method for optimizing the energy efficiency of a whole-plant system of a cryogenic economizer based on artificial intelligence according to claim 1, characterized in that: It also includes establishing a knowledge base for the optimal allocation of natural resource assets, storing no less than 100,000 historical allocation cases, policy and regulatory provisions and expert rules. The knowledge base provides prior knowledge support for the initial setting of multi-objective weights through a semantic retrieval engine. When running for the first time, it can query the optimal weight combination of similar working conditions as a reference for the initial setting, thus shortening the convergence time.

10. The method for optimizing the energy efficiency of a whole-plant system of a cryogenic economizer based on artificial intelligence according to claim 1, characterized in that: Deployed in a collaborative architecture between edge computing nodes and the cloud in thermal power plants, the edge nodes are responsible for real-time data acquisition, preprocessing, and control command issuance with a response latency of less than 200 milliseconds. The cloud is responsible for model training, optimization, and report generation. The communication protocol adopts industrial Ethernet based on time-sensitive networking, and the data encryption adopts the national cryptographic SM4 algorithm. The edge nodes are equipped with GPU acceleration cards to support concurrent processing, and the cloud adopts a distributed cloud computing platform with elastic scalability. The system as a whole meets the requirements of Level 3 of the Industrial Information Security Protection System.

Images