A method for deploying computing power of a coal-fired power plant computing engine and a DCS controller in coordination

Abstract

This invention discloses a method for the collaborative deployment of computing power between a computing engine and a DCS controller in a coal-fired power plant. It relates to the field of collaborative computing power deployment technology for coal-fired power plants. First, it constructs an integrated computing power fusion architecture for the computing engine and DCS controller, breaking down physical isolation and data interaction barriers between the two. All DCS controllers are incorporated into the real-time computing scheduling domain of the computing engine, clearly defining the rules for computing power allocation. The computing engine undertakes high-intensity computing tasks such as modeling, statistical calculations, and knowledge computation, while the DCS controller undertakes real-time control execution and simple logic operations, achieving deep integration of computing power and control. Second, it dynamically adjusts the computing power allocation ratio between the two based on the real-time operating conditions of the unit. This invention effectively solves the problems of fragmented computing power, supply-demand mismatch, high operating latency, and easy interruption of computing power due to single-point failures in existing technologies. It significantly improves computing power utilization and greatly accelerates control response speed without requiring large-scale hardware modifications to existing DCS systems.

Description

Technical Field

[0001] This invention relates to the field of collaborative deployment technology of computing power in coal-fired power plants, and in particular to a method for collaborative deployment of computing power between a computing engine and a DCS controller in a coal-fired power plant. Background Technology

[0002] As coal-fired power plants gradually transition from baseload power to peak-shaving power, they frequently need to participate in flexible operation scenarios such as deep peak shaving and rapid load changes. To achieve safe, efficient, and low-emission control under all operating conditions, the demand for intelligent algorithms such as reinforcement learning and mechanism modeling in coal-fired power plants continues to increase. Computing resources have become a core factor restricting the large-scale application of intelligent control technologies. Currently, the control core of coal-fired power plants is the distributed control system (DCS), and the computing power deployment scheme for intelligent computing functions and DCS control functions is a core research direction in the industry.

[0003] One of the current mainstream deployment solutions is an external, independent computing engine deployment. This solution deploys the intelligent computing engine on an industrial server independent of the DCS (Distributed Control System). It interacts with the DCS through common communication interfaces such as Modbus and OPC. After independently completing complex calculations such as modeling, optimization, and diagnosis, the computing engine sends control commands to the DCS for execution. This solution does not require modification to the DCS's native system, has a low modification threshold, and can flexibly configure high-performance hardware to support the operation of complex algorithms. It has already been applied in non-core control scenarios such as combustion optimization and denitrification optimization in some coal-fired power plants. However, in this solution, the computing engine and the DCS computing power are completely separated, making it impossible to achieve cross-system computing power scheduling. Data transmission across devices is prone to control delays. Under complex operating conditions, insufficient computing power and disconnect between commands and execution are likely to occur. Furthermore, a single server failure can easily lead to the overall paralysis of intelligent control functions, making it unsuitable for core control loops and unable to support the unit's requirement for 168 hours of continuous operation without human intervention.

[0004] Another mainstream solution is the full local deployment of intelligent algorithms within the DCS (Distributed Control System). This solution directly integrates simplified intelligent control logic into the native program of the DCS controller, with all calculations performed locally on the DCS controller, eliminating the need for external computing devices. This solution boasts extremely low data transmission latency, deep integration of control logic with the DCS's native protection system, and high reliability, and has already been applied in simple logic optimization scenarios across a small number of units. However, this solution is limited by the computing power ceiling of the DCS controller itself, making it unable to support highly complex intelligent algorithms. Furthermore, differences in development environments and hardware architectures among different DCS vendors result in significant algorithm porting difficulties and high modification costs. The fixed computing power allocation rules cannot be dynamically adjusted according to operating conditions, leading to wasted computing power in steady-state conditions and insufficient computing power in dynamic conditions, thus failing to meet the needs of intelligent control across all scenarios. Summary of the Invention

[0005] This invention proposes a method for the collaborative deployment of computing power between a computing engine and a DCS controller in a coal-fired power plant, in order to solve the problems mentioned in the prior art, such as computing power fragmentation, supply and demand mismatch, high response latency, insufficient operational reliability, and high cost of implementation and transformation in existing coal-fired power plant computing power deployments.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a method for collaborative deployment of computing power between a coal-fired power plant computing engine and a DCS controller, comprising the following steps: The first step is to construct an integrated computing power fusion architecture for the computing engine and DCS controller. By connecting to the computing power scheduling bus of the intelligent computing engine through the OPC UA unified real-time communication interface opened by the DCS controller, the physical isolation and data interaction barriers between the computing engine and the DCS controller are broken down. All DCS controllers are incorporated into the real-time computing scheduling domain of the computing engine. The computing power allocation rules are determined based on the task computing power requirement threshold and response time requirement threshold. The computing engine undertakes high-intensity computing power tasks with computing power requirements higher than 1 TOPS and response time requirements higher than 100ms, while the DCS controller undertakes real-time control execution tasks and simple logic operation tasks with response time requirements lower than 50ms. This achieves deep binding of computing power resources and control logic and eliminates the risk of decoupling between computing power and control execution. The second step is to collect the operating parameters of the coal-fired power plant units in real time, identify the current operating conditions based on the active power fluctuation rate, main steam pressure deviation, and unit operating status flag, and dynamically adjust the computing power allocation ratio of the computing engine and DCS controller according to the computing power demand priority of different operating conditions to match the differentiated computing power demand under each operating condition and maximize the overall computing power utilization. The third step is to encapsulate the intelligent algorithms required for coal-fired power plant control into standardized algorithm operators according to the rules of unified input and output format and unified resource consumption limit. Based on the embedded running environment of the DCS controller, cross-compile to generate a compatible lightweight algorithm program package. The algorithm container is embedded in the DCS controller for local operation through the algorithm container compatible with the DCS native system. The algorithm operation can be completed without relying on an external computing server. At the same time, a custom algorithm secondary encapsulation interface is provided to support the rapid deployment and adaptation of new algorithms. The fourth step involves the computing engine adopting a multi-server distributed cluster deployment mode, setting up a primary and backup computing node synchronization mechanism and a computing power overload protection mechanism. When the primary node is running normally, it is responsible for the computing power scheduling and task distribution of the entire system. In case of failure, the backup node quickly takes over the computing power tasks. When the computing power is overloaded, redundant resources are automatically started to take over the overflow tasks, ensuring continuous and stable output of computing power in all scenarios.

[0007] Preferably, the high-intensity computing tasks specifically include modeling calculation tasks such as combustion optimization modeling, steam temperature characteristic modeling, and denitrification reaction characteristic modeling; statistical calculation tasks such as hourly coal consumption statistics, pollutant emission statistics, and equipment runtime statistics; and knowledge calculation tasks such as fault diagnosis knowledge reasoning and optimal value reasoning of operating parameters. The real-time control execution and simple logic operation tasks specifically include PID control operation tasks such as water supply regulation, steam temperature regulation, and combustion regulation; switch logic judgment tasks such as valve opening and closing judgment, pump start and stop judgment, and protection trigger judgment; and analog quantity acquisition and preprocessing tasks such as temperature, pressure, and flow signal filtering, range conversion, and outlier removal.

[0008] Preferably, the real-time operating conditions of the coal-fired power plant unit are specifically divided into four categories: steady-state operation, rapid load change, deep peak shaving, and equipment start-up and shutdown. The criteria for judging steady-state operation is that the active power fluctuation rate of the unit is less than one percent of the rated power per minute, and the duration of this state exceeds thirty minutes. The criteria for judging rapid load change is that the active power fluctuation rate of the unit is greater than or equal to two percent of the rated power per minute. The criteria for judging deep peak shaving is that the active power of the unit is less than forty percent of the rated power. The criteria for judging equipment start-up and shutdown are that the unit is in the load increase stage before grid connection, the load decrease stage before disconnection, or the start-up and shutdown operation stage of a single core auxiliary machine.

[0009] Preferably, when the unit is in steady-state operation, 30% of the total computing power of the computing engine is allocated to the statistical computing and knowledge computing modules for optimizing unit operating parameters, predicting equipment degradation trends, and conducting energy efficiency benchmarking analysis. Of the remaining 70% of computing power, 40% is allocated to the real-time modeling computing module, and 30% is reserved for redundancy. This maximizes the added value of computing power without affecting steady-state control requirements.

[0010] Preferably, when the unit is in a rapid load change or deep peak shaving condition, 70% of the total computing power of the computing engine is allocated to the real-time computing module for real-time combustion optimization calculations, steam temperature advance control calculations, and dynamic adjustment calculations of denitrification and ammonia injection, ensuring that the delay in generating control commands is less than 50 milliseconds. Of the remaining 30% of computing power, 10% is allocated to the equipment status monitoring module, and 20% is reserved for redundancy, to avoid control lag problems caused by insufficient computing power during the change of operating conditions.

[0011] Preferably, when the unit is in the equipment start-up and shutdown mode, more than 60% of the total computing power of the computing engine and DCS controller is allocated to the auxiliary machine control algorithm module for starting and stopping logic calculations, real-time status monitoring, and parameter anomaly early warning for the induced draft fan, forced draft fan, and coal mill. The remaining computing power is allocated to the parameter trend prediction module during the start-up and shutdown process to ensure that there are no protection malfunctions and no significant fluctuations in parameters during the equipment start-up and shutdown process.

[0012] Preferably, the intelligent algorithm includes reinforcement learning algorithm, fuzzy neural network algorithm, and genetic optimization algorithm. The lightweight algorithm package is developed based on JAVA or PYTHON language and is cross-compiled to adapt to the ARM or X86 architecture processor of the DCS controller. The memory usage limit of each lightweight algorithm operator does not exceed 128 megabytes, and the CPU utilization rate does not exceed 20% of the single core of the DCS controller. The secondary encapsulation interface provides unified parameter configuration, input and output signal mounting, and resource usage configuration functions, supporting power plant operation and maintenance personnel to customize and add fault diagnosis and optimization control algorithms without modifying the underlying DCS system.

[0013] Preferably, the primary and backup computing node synchronization mechanism specifically involves setting up one primary computing node and two backup computing nodes, and using the Raft distributed consensus protocol to achieve real-time status synchronization between the primary and backup nodes. The synchronization content includes the computing power scheduling queue, intermediate computation results, and algorithm model parameters. The data synchronization latency is less than ten milliseconds. When the primary node fails to send a heartbeat signal three times consecutively, it is determined to be a primary node failure. The backup nodes complete the switchover within fifty milliseconds based on the consensus voting results and take over all computing power scheduling tasks. Control commands are issued without interruption during the switchover process.

[0014] Preferably, the computing power overload protection mechanism specifically collects the CPU, memory, and GPU utilization rates of all nodes in the computing engine cluster every 100 milliseconds. When the total computing power utilization rate of the cluster exceeds 90% for three consecutive collections, the redundant computing power nodes in low-power standby state are automatically woken up, and statistical computing and knowledge computing tasks with lower priority than real-time control are migrated to the redundant nodes for execution. When the total computing power utilization rate of the cluster drops below 70%, the migrated tasks are gradually moved back to the original nodes and the redundant nodes are shut down to avoid computing delay problems caused by computing power overload.

[0015] Preferably, the integrated computing power fusion architecture, dynamic computing power allocation rules, lightweight algorithm deployment mode, and cluster-based fault tolerance mechanism form a complete computing power collaboration system, which collaboratively solves four core problems: computing power fragmentation, supply and demand mismatch, operation delay, and single point of failure. This fully meets the computing power requirements of coal-fired power plant units under all operating conditions and supports the units to achieve 168 hours of continuous and stable operation without human intervention.

[0016] Compared with existing technologies, the beneficial effects of this invention are: This invention breaks down the physical isolation and computing resource barriers between the computing engine and the DCS controller by constructing an integrated computing power fusion architecture. It divides computing power according to task computing power requirements and response time requirements. High-intensity and complex computing tasks are undertaken by the computing engine, while low-latency control tasks are executed locally by the DCS controller. This effectively solves the problems of completely fragmented computing power, high latency in cross-device transmission of control commands, and disconnect between control calculation and execution in existing external deployment solutions. At the same time, it avoids the shortcomings of the fully DCS local deployment solution, such as insufficient computing power limit and inability to support highly complex intelligent algorithms.

[0017] This invention utilizes a dynamic computing power allocation mechanism based on the unit's operating conditions to identify the current operating scenario of the unit and flexibly adjust the computing power allocation ratio according to the priority of computing power demand in different scenarios. High-priority tasks such as real-time control are given priority to computing power resources, while non-urgent tasks such as statistical analysis and status prediction are allocated computing power as needed. This effectively solves the problems of fixed computing power allocation rules, redundant and wasteful computing power in steady-state operating conditions, and insufficient computing power supply in dynamic operating conditions in existing solutions, and significantly improves the overall utilization rate of computing power resources.

[0018] This invention combines lightweight algorithm embedding deployment with a clustered computing power fault tolerance mechanism to encapsulate intelligent algorithms into standardized operators adapted to the DCS operating environment. Deployment can be completed without modifying the underlying DCS system. At the same time, it adopts a cluster deployment mode with real-time synchronization of primary and backup nodes and automatic protection against computing power overload. When a single node fails, it can quickly switch over to take over computing power tasks, and automatically activate redundant resources when computing power is overloaded. This effectively solves the problems of existing solutions, such as high difficulty in algorithm porting and adaptation, high investment in modification, easy interruption of computing power due to single node failure, and insufficient operational reliability. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the overall process of collaborative deployment of computing power between the computing engine and DCS controller in a coal-fired power plant, as proposed in this invention. Figure 2 This is a flowchart illustrating the integrated architecture construction and computing power division of this invention. Figure 3 This is a flowchart illustrating the dynamic computing power allocation based on real-time operating conditions of the present invention. Figure 4 This is a flowchart illustrating the lightweight deployment and cluster fault tolerance management of the algorithm of this invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Reference Figures 1 to 4 This invention discloses a method for the collaborative deployment of computing power between a computing engine and a DCS controller in a coal-fired power plant. The specific implementation process is as follows: The first step is the integration of the unified computing power architecture: First, obtain the hardware parameters of all NARI Ruilan DCS controllers in the target coal-fired power plant, including processor architecture, memory capacity, and open communication port numbers. Configure the X.509 security certificate for OPC UA communication, set the data transmission cycle to 10ms, and upload the computing resource tags and available task types of all DCS controllers to the computing engine's computing resource pool via the computing power scheduling bus using a Kafka low-latency message queue. This breaks down the physical isolation and data interaction barriers between the two, bringing all DCS controllers into the computing engine's real-time computing scheduling domain. Pre-calibrate the computing power and response time of all tasks to be executed, and assign execution sides using a task attribution determination formula. ; in This is a test of computing power requirements for a single task. The computing power threshold is fixed at 1 TOPS. This represents the maximum allowable response latency for a single task. The response time threshold is fixed at 50ms.

[0022] Based on this rule, the division of computing power is determined: the computing engine undertakes high-intensity computing tasks, while the DCS controller undertakes real-time control execution and simple logic operation tasks, thereby achieving a deep binding between computing resources and control logic and eliminating the risk of disconnection between computing power and control execution.

[0023] The second step is dynamic and elastic allocation of computing power: Real-time collection of parameters such as active power, main steam pressure, and unit operating status words of the coal-fired power plant units. Active power is sampled once per minute, and the power fluctuation rate is calculated by dividing the difference between two consecutive sample values ​​by the rated power. Based on preset rules, the current operating condition is identified, and the computing power allocation ratio between the computing engine and the DCS controller is dynamically adjusted according to the priority of computing power demand under different operating conditions. This matches the differentiated computing power requirements of each operating condition and maximizes the overall computing power utilization. The core computing power allocation logic code snippet is as follows: def allocate_power(working_condition, total_computing_power): # Input parameters: current working condition, total computing power of the cluster; Output parameter: computing power allocated to each module if working_condition == "steady": stat_knowledge = total_computing_power * 0.3 model_calc = total_computing_power * 0.28 redundant = total_computing_power * 0.42 return [stat_knowledge, model_calc, redundant] elif working_condition in ["fast_load_change", "deep_peaking"]: realtime_control = total_computing_power * 0.7 device_monitor = total_computing_power * 0.03 redundant = total_computing_power * 0.27 return [realtime_control, device_monitor, redundant] elif working_condition == "equipment_start_stop": aux_control = total_computing_power * 0.6 trend_pred = total_computing_power * 0.2 redundant = total_computing_power * 0.2 return [aux_control, trend_pred, redundant] The third step is the lightweight algorithm embedding and deployment: The intelligent algorithms required for coal-fired power plant control are encapsulated into standardized algorithm operators according to unified input / output formats and unified resource consumption limits. Based on the embedded runtime environment of the DCS controller, a corresponding cross-compilation toolchain is used to generate adapted lightweight algorithm packages. These packages are then embedded into the DCS controller locally using lightweight Docker containers compatible with the DCS native system, enabling algorithm computation without relying on external computing servers. Simultaneously, a secondary encapsulation interface based on the RESTful specification is provided, along with accompanying Swagger interface documentation, supporting rapid deployment and adaptation of new algorithms.

[0024] The fourth step is cluster-based computing power fault tolerance: The computing engine adopts a multi-server distributed cluster deployment mode. All cluster nodes are connected to a 10 Gigabit industrial Ethernet network, with one primary computing power node, two backup computing power nodes, and two redundant computing power nodes. A real-time synchronization mechanism between the primary and backup nodes and a computing power overload protection mechanism are configured. When the primary node is running normally, it is responsible for the computing power scheduling and task distribution of the entire system. In case of failure, the backup node quickly takes over the computing power tasks. When the computing power is overloaded, redundant resources are automatically activated to take over the overflow tasks, ensuring continuous and stable computing power output in all scenarios.

[0025] This invention also discloses the specific task division rules between the computing engine and the DCS controller: In high-intensity computing tasks, modeling tasks specifically include offline modeling based on more than one year of historical operating data to train combustion characteristic models, steam temperature characteristic models, and denitrification reaction characteristic models, as well as online modeling based on real-time data with model parameters updated hourly; statistical calculation tasks are executed once per hour, outputting energy efficiency and environmental protection reports such as unit power supply coal consumption, plant power consumption rate, and pollutant emissions; knowledge calculation tasks are based on a pre-set fault diagnosis knowledge base of 2000+ entries, using forward reasoning logic to identify early signs of equipment degradation.

[0026] In the real-time control and simple calculation tasks on the DCS side, the execution cycle of PID control calculation is 10ms, covering core control loops such as water supply regulation, steam temperature regulation, and combustion regulation; the execution cycle of switch logic judgment is 5ms, covering logic such as valve opening / closing judgment, pump start / stop judgment, and protection trigger judgment; the analog signal acquisition and preprocessing task is executed once every 2ms, performing amplitude limiting filtering and outlier removal on the acquired 4-20mA current signal and PT100 temperature signal. The filtering adopts a first-order inertial filtering algorithm, and the formula is as follows: ;in This is the filtered output value at the current moment. This represents the raw signal value collected at the current moment. This is the filtered output value from the previous moment. With the filter coefficient fixed at 0.2, signal spike interference can be effectively filtered out, avoiding control fluctuations caused by signal abrupt changes.

[0027] This invention also discloses specific rules for determining the operating conditions of the generating unit: The criteria for judging steady-state operating conditions are: the active power of the unit is maintained within 50%-100% of the rated power, the power fluctuation rate per minute is less than 1% of the rated power, and no AGC load change command or auxiliary machine start / stop command is received for 30 consecutive minutes. The criteria for determining the rapid load change condition are: the received AGC command load change rate is greater than or equal to 2% of the rated power per minute, and the condition continues until the active power reaches the target value and ends 3 minutes later; The criteria for determining deep peak shaving conditions are: the active power of the unit is less than 40% of the rated power, and this state lasts for more than 10 minutes. The criteria for determining equipment start-up and shutdown conditions are as follows: when the unit is in the load increase stage before grid connection, the load decrease stage before disconnection, or the individual start-up and shutdown operation stage of core auxiliary equipment such as induced draft fan, forced draft fan, coal mill, and feedwater pump, the determination mark is the time period from when the DCS system issues the corresponding start-up and shutdown command to when the equipment operating parameters stabilize.

[0028] This invention also discloses the computing power allocation rules under steady-state operating conditions: Under steady-state conditions, the unit's operating parameters fluctuate little, and the basic control requirements are low. Therefore, 30% of the computing power of the computing engine is allocated to the statistical computing and knowledge computing modules. The statistical computing module collects the unit's power supply coal consumption, plant power consumption rate, NOx emission concentration, and other indicators hourly, compares them with historical best values ​​and industry benchmark values, and outputs operation optimization suggestions. The knowledge computing module collects parameters such as stator temperature and bearing vibration of the equipment in real time, compares them with the deterioration characteristics in the fault knowledge base, and can predict potential equipment failures 72 hours in advance. Of the remaining 70% of computing power, 40% is allocated to the real-time modeling computing module to update the parameters of control models such as combustion and steam temperature, and 30% is reserved as redundancy to cope with sudden fluctuations in computing power demand.

[0029] This invention also discloses computing power allocation rules under rapid load changes and deep peak shaving conditions: Under these two operating conditions, the unit parameters fluctuate greatly and the control precision requirements are high. Therefore, 70% of the computing power of the computing engine is allocated to the real-time computing module. The real-time computing module runs reinforcement learning combustion optimization algorithm and fuzzy neural network steam temperature control algorithm. The calculation cycle is 50ms. The output control commands are directly sent to the DCS controller for execution, ensuring that the main steam pressure deviation does not exceed ±0.2MPa and the superheated steam temperature deviation does not exceed ±5℃ during the changing operating conditions. Of the remaining 30% of the computing power, 10% is allocated to the equipment status monitoring module to monitor the current, vibration and temperature parameters of the auxiliary equipment in real time to avoid overload of the auxiliary equipment during the changing operating conditions, and 20% is reserved for redundancy.

[0030] This invention also discloses the computing power allocation rules under equipment start-up and shutdown conditions: The control logic during equipment start-up and shutdown is complex and carries a high risk of misoperation. Therefore, more than 60% of the computing power of the computing engine and DCS controller is allocated to the auxiliary machine control algorithm module. The auxiliary machine control algorithm module runs the start-up and shutdown timing control of the coal mill and the induced draft fan blade adjustment logic, and monitors the current, vibration, and temperature parameters of the auxiliary machine in real time. Once the parameters exceed the warning threshold, an adjustment command is immediately triggered. The remaining computing power is allocated to the parameter trend prediction module during start-up and shutdown. The LSTM time series prediction algorithm is used to predict the main steam pressure and steam temperature change trends in the next 10 minutes and provide adjustment suggestions in advance to ensure that the parameters do not fluctuate significantly during start-up and shutdown and that no protection malfunctions are triggered.

[0031] This invention also discloses specific encapsulation and deployment rules for lightweight algorithms: The intelligent algorithms include three categories: deep reinforcement learning algorithms, fuzzy neural network algorithms, and genetic optimization algorithms. When encapsulating them, the types and lengths of input and output parameters are uniformly set, and the memory usage limit for a single operator is set to 128MB, and the CPU utilization limit is set to 20% of a single core. During development, they are written in JAVA or PYTHON language, and the corresponding cross-compilation toolchain is used to compile and generate executable packages for the ARM or X86 architecture of the DCS controller. The size of the generated single operator package does not exceed 50MB. The algorithm container uses a lightweight Docker runtime adapted to industrial embedded environments, which can run without installing a full Docker environment; the secondary encapsulation interface provides unified parameter configuration, signal mounting, and testing and verification functions. After the operation and maintenance personnel upload the code package of the custom algorithm, the system automatically completes compatibility testing and resource usage testing. After the test is passed, it can be deployed to the DCS controller for operation without modifying the underlying DCS system code.

[0032] This invention also discloses the specific synchronization and switching rules for primary and backup computing nodes: The computing engine cluster uses the Raft distributed consensus protocol to synchronize the status of primary and standby nodes. The primary node sends a heartbeat signal to the two standby nodes every 10ms, and simultaneously synchronizes the computing power scheduling queue, intermediate computation results, and algorithm model parameters. The synchronization adopts an incremental synchronization method, only synchronizing the content that has changed since the last synchronization, and the synchronization delay does not exceed 10ms. When a standby node fails to receive a heartbeat signal from the primary node for three consecutive times, an election process is triggered. The standby node that receives more than half of the node votes becomes the new primary node. The switchover process is completed within 50ms. During the switchover, control commands that have been issued continue to be executed, and control commands that have not been generated are processed by the new primary node, so there is no control interruption.

[0033] This invention also discloses the specific execution rules for computing power overload protection: The CPU, memory, and GPU utilization of all nodes in the cluster are collected every 100ms, and the total computing power utilization of the cluster is calculated using the following formula: ;in This represents the total computing power utilization rate of the cluster. Let be the CPU utilization of the i-th node. Let be the memory usage of the i-th node. Let i be the GPU utilization of the i-th node. The CPU weight is fixed at 0.4. The memory weight is fixed at 0.3. The GPU weight is fixed at 0.3. This represents the total number of online nodes in the cluster. When three consecutive data collections are completed... When the power consumption exceeds 90%, the redundant computing nodes in low-power standby state are automatically woken up, and statistical computing and knowledge computing tasks with lower priority than real-time control are migrated to the redundant nodes for execution. when When the energy consumption drops below 70%, gradually migrate the tasks back to the original nodes and shut down redundant nodes to reduce energy consumption.

[0034] This invention also discloses the overall operational performance verification rules for the computing power collaboration system: The integrated computing power fusion architecture solves the problem of the separation between traditional external computing engines and DCS computing power. The dynamic computing power allocation rules solve the problem of the mismatch between computing power supply and demand under different working conditions. The lightweight algorithm deployment mode solves the problems of high algorithm running latency and high transformation costs. The cluster-based fault tolerance mechanism solves the problem of computing power interruption caused by single node failure.

[0035] The four mechanisms work together and have been verified through 720 hours of continuous industrial field testing. The computing power utilization rate has been increased by more than 60%, the control command response latency is less than 50ms, and there are no problems of computing power interruption or control disconnection. It fully meets the computing power requirements of coal-fired power plant units under all operating conditions and supports the units to achieve 168 hours of continuous and stable operation without human intervention.

[0036] Reference Figure 1This diagram illustrates the complete steps and macro-level logic of the collaborative deployment of computing power between the computing engine and the distributed control system in a coal-fired power plant. First, the system breaks down the physical isolation between traditional devices, constructing an integrated converged computing architecture and establishing computing power allocation rules based on task type, achieving deep decoupling and rebinding of the underlying computing power and high-level control logic. Subsequently, the system collects unit operating data in real time and dynamically adjusts the computing power allocation weights between the two based on the specific operating conditions to maximize the utilization of global computing resources. At the algorithm application level, the system transforms complex intelligent algorithms into standardized operators and encapsulates them into lightweight program packages adapted to the local operating environment, directly deploying them within the controller for execution. Finally, to ensure industrial-grade high availability, the system employs a multi-server cluster deployment on the computing engine side, supplemented by a millisecond-level master / slave switching mechanism and overload protection response, thereby jointly supporting the long-term, uninterrupted, stable operation of the power plant units.

[0037] Reference Figure 2 This diagram details the construction mechanism and refined task offloading division of the computing power convergence architecture. In the initial construction phase, the system broke down the original device silos from both physical and logical communication perspectives, fully integrating the underlying controllers into the real-time scheduling domain of the upper-layer computing engine. When clarifying the division of computing power, the system implemented a heterogeneous task separation strategy. Leveraging its superior data throughput and parallel processing capabilities, the computing engine primarily undertakes high-intensity computing tasks such as complex modeling calculations, high-dimensional statistical analysis, and knowledge graph inference. Correspondingly, the controllers return to their core functions of high real-time performance and high reliability, focusing on tasks such as proportional-integral-derivative (PI) control of execution loops, determination of underlying switch logic states, and high-frequency acquisition and basic digital filtering preprocessing of front-end analog signals. This constructs a new type of industrial computing power foundation where each component performs its specific function and complements the others efficiently.

[0038] Reference Figure 3 This diagram illustrates the system's adaptive computing power routing and dynamic tilting mechanism for handling complex and ever-changing industrial scenarios. The system continuously monitors and identifies the current global operating status of the generating units, precisely dividing it into four core operating conditions: steady state, rapid load change, deep peak shaving, and equipment start-up and shutdown. When a steady state condition is identified, the system proactively and smoothly transfers approximately 30% of its computing power to the statistics and knowledge computing module to conduct in-depth equipment health assessments and optimize operating parameters. When facing high-frequency grid response demands such as rapid load change or deep peak shaving, the system rapidly concentrates up to 70% of its core computing power on the real-time computing module, ensuring the rapid generation and accurate issuance of millisecond-level control commands. If the generating units are in the equipment start-up and shutdown phase, the scheduling focus immediately shifts to the auxiliary machine control algorithm module, ensuring the safe and stable execution of complex auxiliary machine timing control logic.

[0039] Reference Figure 4This diagram focuses on analyzing the edge deployment path of intelligent algorithms and the security defense design of cloud clusters. In terms of algorithm engineering, the system lightweights and standardizes complex models such as cutting-edge reinforcement learning and fuzzy neural networks, encapsulating them into independently runnable packages. These packages are directly embedded into the underlying controller using secure container technology, enabling rapid deployment of computing power to the edge and providing open secondary development interfaces. Regarding hardware cluster security, multiple computing servers form a high-availability array. High-frequency heartbeats and model parameter synchronization are maintained between primary and backup nodes. In the event of a sudden hardware failure or network interruption of the primary node, the backup node can seamlessly take over all scheduling operations within a very short latency. Simultaneously, overload protection probes monitor the total system load in real time. When computing power consumption exceeds the safety threshold, the system automatically wakes up and connects redundant physical nodes, effectively mitigating computing power surges and ensuring absolute continuity of control output.

[0040] The above are merely preferred embodiments 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 collaborative deployment of computing power between a computing engine and a DCS controller in a coal-fired power plant, characterized in that, The steps include the following: An integrated computing power fusion architecture is constructed for the computing engine and DCS controller, breaking the physical isolation between the computing engine and DCS controller. The DCS controller is incorporated into the real-time computing scheduling domain of the computing engine, and the computing power division rules are determined. The computing engine undertakes high-intensity computing tasks, while the DCS controller undertakes real-time control execution and simple logic operation tasks, thus achieving deep binding between computing power and control logic. Based on the real-time operating conditions of the coal-fired power plant units, the computing power allocation ratio of the computing engine and DCS controller is dynamically adjusted to match the computing power requirements under different operating conditions and maximize the utilization of computing power. The intelligent algorithm is encapsulated into standardized operators to generate a lightweight algorithm package that is adapted to the DCS controller's operating environment. The algorithm is then embedded into the DCS controller for local execution via an algorithm container, while also providing a custom algorithm secondary encapsulation interface. The computing engine adopts a multi-server cluster deployment mode, sets up primary and backup computing nodes and computing power overload protection mechanisms to ensure continuous and stable output of computing power.

2. The method for collaborative deployment of computing power between the computing engine and the DCS controller in a coal-fired power plant according to claim 1, characterized in that, The high-intensity computing tasks include modeling calculation tasks, statistical calculation tasks, and knowledge calculation tasks. The real-time control execution tasks and simple logic operation tasks include PID control calculation tasks, switch quantity logic judgment tasks, and analog quantity acquisition and preprocessing tasks.

3. The method for collaborative deployment of computing power between the computing engine and the DCS controller in a coal-fired power plant according to claim 1, characterized in that, The real-time operating conditions of the coal-fired power plant units include four categories: steady-state operating conditions, rapid load change operating conditions, deep peak shaving operating conditions, and equipment start-up and shutdown operating conditions.

4. The method for collaborative deployment of computing power between the computing engine and the DCS controller in a coal-fired power plant according to claim 3, characterized in that, When the unit is in steady-state operation, 30% of the total computing power of the computing engine is allocated to the statistical computing and knowledge computing modules for unit operation optimization and equipment status prediction.

5. The method for collaborative deployment of computing power between the computing engine and the DCS controller in a coal-fired power plant according to claim 3, characterized in that, When the unit is in a rapid load change or deep peak shaving condition, 70% of the total computing power of the computing engine is allocated to the real-time computing module for the rapid generation and issuance of control commands.

6. The method for collaborative deployment of computing power between the computing engine and the DCS controller in a coal-fired power plant according to claim 3, characterized in that, When the unit is in the equipment start-up and shutdown mode, computing power is prioritized to the auxiliary machine control algorithm module to ensure the stable execution of the auxiliary machine start-up and shutdown process.

7. The method for collaborative deployment of computing power between the computing engine and the DCS controller in a coal-fired power plant according to claim 1, characterized in that, The intelligent algorithms include reinforcement learning algorithms and fuzzy neural network algorithms. The lightweight algorithm package is developed based on JAVA or PYTHON language and is adapted to the local operating environment of DCS controller without relying on external computing servers.

8. The method for collaborative deployment of computing power between the computing engine and the DCS controller in a coal-fired power plant according to claim 1, characterized in that, In the primary and backup computing nodes, the backup node synchronizes the operating data and algorithm model of the primary node in real time. When the primary node fails, the backup node completes the switchover and takes over all computing power scheduling tasks within milliseconds.

9. The method for collaborative deployment of computing power between the computing engine and the DCS controller in a coal-fired power plant according to claim 1, characterized in that, The computing power overload protection mechanism specifically involves: real-time monitoring of the total computing power utilization rate of the computing engine; and when the computing power utilization rate exceeds a set threshold, automatically activating redundant computing power nodes to take over the overflow computing power tasks.

10. The method for collaborative deployment of computing power between the computing engine and the DCS controller in a coal-fired power plant according to claim 1, characterized in that, The integrated computing power fusion architecture, dynamic computing power allocation rules, lightweight algorithm deployment mode, and cluster-based fault tolerance mechanism jointly support the computing power requirements of coal-fired power plant units to achieve 168 hours of continuous operation without human intervention.

Images