Optimal control method and system for flexibility reconstruction and deep peak regulation of thermal power plant

Abstract

The application provides a thermal power plant flexibility reconstruction and deep peak regulation optimization control method and system, relates to the technical field of control engineering, and comprises the following steps: collecting all-around operation data of the thermal power plant to construct a multilevel dynamic nonlinear model of the thermal power plant, performing real-time identification on model parameters, generating an identification result, and establishing a multidimensional operation characteristic database of the thermal power plant; based on the operation characteristic database of the thermal power plant, designing a reconstruction scheme, performing visual simulation and optimization to obtain flexibility reconstruction results; constructing a multi-objective optimization function, formulating an adaptive deep peak regulation optimization control strategy, performing short-term and medium-term prediction on power grid load, generating an initial control strategy based on the prediction result, performing online optimization and adjustment, and continuously evaluating and optimizing the control strategy.

Description

Technical Field

[0001] This invention relates to the field of control engineering technology, and in particular to a method and system for the flexible retrofitting and deep peak shaving optimization control of thermal power plants. Background Technology

[0002] As the proportion of renewable energy in the power system continues to increase, traditional thermal power plants are facing unprecedented challenges. On the one hand, thermal power plants need to be more flexible to cope with rapid fluctuations in grid load. On the other hand, thermal power plants also need to meet increasingly stringent environmental protection requirements while ensuring safe and economical operation.

[0003] Existing thermal power plant control systems and operating strategies suffer from limited peak-shaving capacity, low operating efficiency, and poor environmental performance when dealing with these challenges. At the same time, frequent peak shaving exacerbates equipment fatigue damage, reduces equipment lifespan, and thus increases maintenance costs.

[0004] Therefore, a solution is urgently needed to address the problems existing in the current technology. Summary of the Invention

[0005] This invention provides a method and system for the flexible retrofitting and deep peak shaving optimization control of thermal power plants, which can at least solve some of the problems existing in the prior art.

[0006] A first aspect of this invention provides a method for the flexibility retrofitting and deep peak-shaving optimization control of thermal power plants, comprising:

[0007] A distributed sensor network is used to collect comprehensive operational data from a thermal power plant. Edge computing technology is used to preprocess the collected data, including outlier detection, data smoothing, and compression. Based on the preprocessed data, a multi-level dynamic nonlinear model of the thermal power plant is constructed, including the fuel delivery system, boiler system, turbine system, generator system, and pollutant control system. A hybrid deep learning algorithm, combining long short-term memory networks and convolutional neural networks, is used to train and optimize the models of each subsystem. An improved adaptive quantum behavior particle swarm optimization algorithm is used to identify the model parameters in real time, generate identification results, and establish a multi-dimensional thermal power plant operation characteristic database based on the identification results, including efficiency curves, emission characteristics, and dynamic responses during start-up and shutdown processes under different loads.

[0008] Based on a database of thermal power plant operating characteristics, data mining techniques are used to analyze unit operating bottlenecks, including load change rate limiting factors and minimum technical output constraints. Multi-objective and multi-scheme flexible retrofitting plans are designed, including the design of a boiler anti-ash accumulation intelligent ash cleaning system, intelligent optimization of the turbine bypass system, frequency conversion retrofitting of the feedwater system, and combustion optimization system retrofitting. Virtual reality and augmented reality technologies are used to visualize and optimize the retrofitting plans. During the retrofitting process, edge computing-based digital twin technology is used to construct a high-fidelity digital model of the thermal power plant. Through retrofitting, the start-up and shutdown speed, load change rate, and minimum technical output of the units are improved, resulting in flexible retrofitting outcomes.

[0009] Based on the dynamic nonlinear model and the results of flexibility retrofitting, a multi-objective optimization function is constructed, with objectives including maximizing unit efficiency, minimizing emissions, maximizing peak-shaving capacity, and maximizing equipment lifespan. An improved model predictive control algorithm is adopted, combined with deep reinforcement learning methods, to formulate an adaptive deep peak-shaving optimization control strategy. Short-term and medium-term forecasts of grid load are performed. Based on the forecast results, a deep deterministic strategy gradient algorithm is used to generate an initial control strategy, which is then optimized and adjusted online. A hierarchical distributed control architecture is designed to generate a global optimization objective, which is decomposed into local objectives for each subsystem. The interactions between subsystems are coordinated, and precise control of each subsystem is achieved through intelligent actuators to realize coordinated control of the units. A real-time performance evaluation mechanism based on big data analysis is established, and a sliding time window method is used to continuously evaluate and optimize the control strategy. Based on the evaluation results, an adaptive learning algorithm is used to continuously optimize the control strategy.

[0010] In one alternative implementation,

[0011] A distributed sensor network is used to collect comprehensive operational data from a thermal power plant. Edge computing technology is used to preprocess the collected data, including outlier detection, data smoothing, and compression. Based on the preprocessed data, a multi-level dynamic nonlinear model of the thermal power plant is constructed, including the fuel delivery system, boiler system, turbine system, generator system, and pollutant control system. A hybrid deep learning algorithm, combining long short-term memory networks and convolutional neural networks, is used to train and optimize the models of each subsystem. An improved adaptive quantum behavior particle swarm optimization algorithm is used to identify the model parameters in real time, generating identification results. Based on these results, a multi-dimensional database of thermal power plant operating characteristics is established, including efficiency curves under different loads, emission characteristics, and dynamic responses during start-up and shutdown processes.

[0012] High-precision sensors are deployed on the steam turbines, boilers, and generators of thermal power plants to collect real-time operating parameters. Edge computing nodes are deployed on-site, equipped with ARM Cortex-A72 processors and 4GB RAM, and running a lightweight Linux operating system.

[0013] The edge computing nodes are used to preprocess the collected data. The preprocessing includes outlier detection using an improved local anomaly factor algorithm, which reduces computational complexity by combining a sliding time window technique; data smoothing using an adaptive Savitzky-Gore filter, which automatically adjusts the filter window size and polynomial order according to the local characteristics of the data; and data compression using a hybrid compression algorithm based on wavelet transform and deep autoencoder, to obtain the preprocessed data.

[0014] Based on the preprocessed data, a multi-level dynamic nonlinear model of a thermal power plant is constructed. The model includes five subsystems: fuel delivery system, boiler system, turbine system, generator system, and pollutant control system.

[0015] A hybrid deep learning algorithm is used to train and optimize the model of each subsystem. The hybrid deep learning algorithm combines a long short-term memory network and a convolutional neural network. The long short-term memory network adopts a bidirectional structure, and the convolutional neural network adopts a multi-scale convolutional kernel. The outputs of the long short-term memory network and the convolutional neural network are fused through an attention mechanism.

[0016] An improved adaptive quantum behavior particle swarm optimization algorithm is used to achieve real-time identification of model parameters. Quantum states are used to represent particle positions and update phase angles. Quantum rotation gates are used to realize particle motion to obtain model identification results. Based on the model identification results, a multi-dimensional thermal power plant operation characteristic database is constructed. The database adopts a distributed storage architecture, uses Apache Cassandra as the underlying storage system, develops a RESTful API-based data access interface, supports multi-parameter joint queries and time series slicing, uses AES-256 encryption to store data, implements role-based access control, and records audit logs for data access operations.

[0017] In one alternative implementation,

[0018] The particle position is represented by a quantum state, and the phase angle is updated using the following formula:

[0019] Δθ=ωΔθ prev +c1r1(θ pbest -θ)+c2r2(θ gbest -θ)+α·(rand()-0.5)·π;

[0020] Where Δθ is the change in phase angle, representing the update magnitude of the current phase angle, and ω represents the inertia weight. prevc1 represents the phase angle change in the previous step, r1 represents a random number between 0 and 1, and θ represents the individual learning factor. pbest Let c1 represent the individual's optimal phase angle, c2 represent the global learning factor, r2 represent another random number between 0 and 1, and θ represent the global learning factor. gbest α represents the global optimal phase angle, α represents the interference intensity, and rand() represents the uncertainty function, which is used to randomly generate a random number between 0 and 1.

[0021] In one alternative implementation,

[0022] Based on a database of thermal power plant operating characteristics, data mining techniques are used to analyze unit operating bottlenecks, including load change rate limiting factors and minimum technical output constraints. Multi-objective, multi-scheme flexible retrofitting plans are designed, including the design of an intelligent ash removal system for boilers to prevent ash accumulation, intelligent optimization of the turbine bypass system, frequency conversion retrofitting of the feedwater system, and combustion optimization system retrofitting. Virtual reality and augmented reality technologies are used to visualize and optimize the retrofitting plans. During the retrofitting implementation process, edge computing-based digital twin technology is used to construct a high-fidelity digital model of the thermal power plant. Through retrofitting, the unit's start-up and shutdown speed, load change rate, and minimum technical output are improved, resulting in flexible retrofitting outcomes including:

[0023] A dynamic nonlinear model of the boiler system is constructed, and an intelligent combustion optimization system is developed based on the dynamic nonlinear model of the boiler system. The intelligent combustion optimization system includes using a long short-term memory network to predict future load changes, using an improved nonlinear model predictive control algorithm to optimize the coal feed rate, primary air volume and secondary air volume in real time, using a multi-objective optimization strategy to construct an optimization objective function and using a sequential quadratic programming algorithm combined with the trust region method to solve the nonlinear optimization problem.

[0024] A steam turbine system model is constructed, and an adaptive sliding pressure operation control strategy is developed based on the steam turbine system model. The adaptive sliding pressure operation control strategy includes training the control strategy using a deep deterministic policy gradient algorithm, defining a state space including main steam pressure, temperature, flow rate, and generator output power, defining the action space as the parameters of the sliding pressure curve and designing a reward function, deploying edge computing nodes on the feedwater pump, induced draft fan, and forced draft fan, running a lightweight model predictive control algorithm on each edge computing node, and designing a distributed coordination mechanism based on federated learning. Each edge node maintains a local model, updates the model parameters using local data, and sends the updated model parameters to a central server. The central server aggregates all updates and sends the aggregated updates back to each edge node, and the edge nodes use the received updates to update their local models.

[0025] Develop an integrated intelligent scheduling system that treats the boiler system, turbine system and auxiliary system as independent intelligent agents. Use an improved multi-agent soft actor-commentator algorithm to achieve overall optimal control. Each agent has its own agent policy network and value function network. Introduce a global value function to evaluate the value of joint actions and make decisions based on the local value function and the global value function.

[0026] A high-fidelity digital model of a thermal power plant is constructed through visualization simulation. By modifying the unit's start-up and shutdown speed, load change rate, and minimum technical output, flexible modification results are obtained.

[0027] In one alternative implementation,

[0028] The update objective of the value function network is shown in the following formula:

[0029]

[0030] Where E represents the expectation, s t Indicates the current state, a t Indicates the current action, r t s represents the current reward value. t+1 Denotes the next state, D represents the experience replay buffer, and Qi(s) represents the next state. t a t ) represents the value function network for the current state s t and current action a t The estimated value, where γ represents the discount factor. Indicates according to strategy π i From the next state s t+1 Randomly select action a t+1 The expected value, Q i (s t+1 a t+1 ) represents the value function network for the next state s t+1和 Next action a t+1 The estimated value, β represents the policy entropy adjustment factor, and P represents the current policy π. i In state s t+1 Choose action a t+1 The logarithmic probability.

[0031] In one alternative implementation,

[0032] Based on the dynamic nonlinear model and the results of flexibility retrofitting, a multi-objective optimization function is constructed, with objectives including maximizing unit efficiency, minimizing emissions, maximizing peak-shaving capacity, and maximizing equipment lifespan. An improved model predictive control algorithm, combined with deep reinforcement learning, is employed to formulate an adaptive deep peak-shaving optimization control strategy. Short-term and medium-term grid load forecasts are performed. Based on the forecast results, a deep deterministic strategy gradient algorithm is used to generate an initial control strategy, which is then optimized and adjusted online. A hierarchical distributed control architecture is designed to generate a global optimization objective, which is decomposed into local objectives for each subsystem. The interactions between subsystems are coordinated, and precise control of each subsystem is achieved through intelligent actuators to realize coordinated unit control. A real-time performance evaluation mechanism based on big data analysis is established, employing a sliding time window method to continuously evaluate and optimize the control strategy. Based on the evaluation results, an adaptive learning algorithm is used to continuously optimize the control strategy, including:

[0033] Based on the dynamic nonlinear model and the results of the flexibility modification, a multi-objective optimization function is constructed, with objectives including maximizing unit efficiency, minimizing emissions, maximizing peak-shaving capacity, and maximizing equipment lifespan. An improved model predictive control algorithm is designed, which includes a state estimator, a constraint processor, and an optimization solver.

[0034] A deep reinforcement learning framework is constructed, which includes a deep policy network and a value network. The deep policy network is used to generate control actions, and the value network is used to evaluate state values. The improved model predictive control algorithm is combined with the deep reinforcement learning framework to form an adaptive deep peak shaving optimization control strategy.

[0035] A method combining long short-term memory networks and convolutional neural networks is used to perform short-term and medium-term forecasts of power grid load, obtaining load forecast results. Based on the load forecast results, an initial control strategy is generated according to the improved model predictive control algorithm and the deep deterministic policy gradient algorithm. The deep deterministic policy gradient algorithm includes an actor network and a critic network, both of which adopt a four-layer fully connected neural network structure.

[0036] An online optimization module is designed to optimize and adjust the initial control strategy in real time. The online optimization module adopts the model predictive control method to solve the optimal control sequence in the rolling time domain and constructs a hierarchical distributed control architecture, which includes a global coordination layer and a local control layer. The global coordination layer is responsible for generating the global optimization objective and uses a decomposition coordination algorithm to decompose the global objective into local objectives of each subsystem. The local control layer is responsible for executing the precise control of each subsystem and uses an adaptive PID control algorithm for tracking control of the local objectives. An intelligent actuator is designed, which includes an advanced actuator and an embedded controller to achieve precise control of each subsystem. A real-time performance evaluation mechanism based on big data analysis is established, and a sliding time window method is used to continuously evaluate and optimize the adaptive deep peak shaving optimization control strategy.

[0037] Based on the evaluation results of the real-time performance evaluation mechanism, an adaptive learning algorithm is used to continuously optimize the adaptive deep peak shaving optimization control strategy. The adaptive learning algorithm uses the gradient descent method to update the control strategy parameters. A safety monitoring module is set up to monitor key operating parameters in real time. When an anomaly is detected or a constraint is about to be violated, a preset safety protection strategy is triggered.

[0038] A second aspect of this invention provides a control system for the flexibility retrofitting and deep peak shaving optimization of thermal power plants, comprising:

[0039] The first unit is used to collect comprehensive operational data of thermal power plants using a distributed sensor network. Edge computing technology is used to preprocess the collected data, including outlier detection, data smoothing and compression. Based on the preprocessed data, a multi-level dynamic nonlinear model of the thermal power plant is constructed, including the fuel delivery system, boiler system, turbine system, generator system and pollutant control system. A hybrid deep learning algorithm, combining long short-term memory network and convolutional neural network, is used to train and optimize the models of each subsystem. An improved adaptive quantum behavior particle swarm optimization algorithm is used to identify the model parameters in real time, generate identification results, and based on the identification results, establish a multi-dimensional thermal power plant operation characteristic database, including efficiency curves, emission characteristics and dynamic response of start-up and shutdown processes under different loads.

[0040] The second unit is used to analyze unit operation bottlenecks based on the thermal power plant operation characteristic database and using data mining technology, including load change rate limiting factors and minimum technical output constraints. It designs multi-objective and multi-scheme flexible transformation schemes, including the design of boiler anti-ash accumulation intelligent ash cleaning system, intelligent optimization of turbine bypass system, frequency conversion transformation of feedwater system and combustion optimization system transformation. It uses virtual reality and augmented reality technology to visualize and optimize the transformation schemes. During the transformation implementation process, it adopts edge computing-based digital twin technology to build a high-fidelity digital model of the thermal power plant. Through transformation, it improves the unit's start-up and shutdown speed, load change rate and minimum technical output, and obtains flexible transformation results.

[0041] The third unit is used to construct a multi-objective optimization function based on the dynamic nonlinear model and the results of flexibility modifications. The objectives include maximizing unit efficiency, minimizing emissions, maximizing peak-shaving capacity, and maximizing equipment lifespan. An improved model predictive control algorithm is adopted, combined with deep reinforcement learning methods, to formulate an adaptive deep peak-shaving optimization control strategy. Short-term and medium-term forecasts of grid load are performed. Based on the forecast results, a deep deterministic strategy gradient algorithm is used to generate an initial control strategy. The initial control strategy is optimized and adjusted online. A hierarchical distributed control architecture is designed to generate a global optimization objective. The global objective is decomposed into local objectives of each subsystem, and the interaction between subsystems is coordinated. Intelligent actuators are used to achieve precise control of each subsystem to realize coordinated control of the unit. A real-time performance evaluation mechanism based on big data analysis is established. A sliding time window method is used to continuously evaluate and optimize the control strategy. Based on the evaluation results, an adaptive learning algorithm is used to continuously optimize the control strategy.

[0042] A third aspect of the embodiments of the present invention,

[0043] An electronic device is provided, comprising:

[0044] processor;

[0045] Memory used to store processor-executable instructions;

[0046] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0047] Fourth aspect of the present invention,

[0048] A computer-readable storage medium is provided, having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0049] This invention significantly improves the load change rate of generating units, reduces minimum technical output, and enhances start-up and shutdown speed through multi-objective optimization and deep reinforcement learning. This greatly enhances the peak-shaving capacity and system flexibility of thermal power units. By employing improved model predictive control algorithms and deep learning technology, optimized operation of units under different loads is achieved, significantly improving unit efficiency, reducing power generation costs and fuel consumption. By incorporating maximizing equipment lifespan into the optimization objective and using an evaluation method based on fatigue damage models, equipment wear is effectively reduced, and the expected lifespan of major equipment is extended. Through the cooperation of online optimization modules and intelligent actuators, the system response speed is significantly improved, which is crucial for rapid peak shaving and frequency regulation. In summary, this invention not only optimizes the performance of individual generating units but also provides strong support for the flexibility, reliability, and cleanliness of the entire power system, which is of great significance for promoting energy transition and sustainable development. Attached Figure Description

[0050] Figure 1 This is a flowchart illustrating the method for flexible retrofitting and deep peak-shaving optimization control of thermal power plants according to an embodiment of the present invention.

[0051] Figure 2 This is a schematic diagram of the structure of the thermal power plant flexibility retrofit and deep peak shaving optimization control system according to an embodiment of the present invention. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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.

[0053] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0054] Figure 1 This is a flowchart illustrating the method of an embodiment of the present invention, as shown below. Figure 1 As shown, the method includes:

[0055] S1. A distributed sensor network is used to collect comprehensive operational data of the thermal power plant. Edge computing technology is used to preprocess the collected data, including outlier detection, data smoothing and compression. Based on the preprocessed data, a multi-level dynamic nonlinear model of the thermal power plant is constructed, including the fuel delivery system, boiler system, turbine system, generator system and pollutant control system. A hybrid deep learning algorithm, combining long short-term memory network and convolutional neural network, is used to train and optimize the models of each subsystem. An improved adaptive quantum behavior particle swarm optimization algorithm is used to identify the model parameters in real time, generate identification results, and based on the identification results, establish a multi-dimensional thermal power plant operation characteristic database, including efficiency curves, emission characteristics and dynamic response of start-up and shutdown processes under different loads.

[0056] The hybrid deep learning algorithm combines multiple deep learning techniques (such as convolutional neural networks and recurrent neural networks) to improve model performance and generalization ability, and can more effectively capture complex data features. The improved adaptive quantum behavior particle swarm optimization algorithm is a variant of particle swarm optimization, which incorporates the idea of ​​quantum computing. By introducing qubits and quantum rotation gates, the search behavior of particles is more flexible, thereby improving global search ability and convergence speed. The multi-dimensional thermal power plant operation characteristic database is a data management system that integrates multiple operating parameters and performance indicators, and is designed to support the optimization and decision-making of thermal power plants.

[0057] In one alternative implementation,

[0058] A distributed sensor network is used to collect comprehensive operational data from a thermal power plant. Edge computing technology is used to preprocess the collected data, including outlier detection, data smoothing, and compression. Based on the preprocessed data, a multi-level dynamic nonlinear model of the thermal power plant is constructed, including the fuel delivery system, boiler system, turbine system, generator system, and pollutant control system. A hybrid deep learning algorithm, combining long short-term memory networks and convolutional neural networks, is used to train and optimize the models of each subsystem. An improved adaptive quantum behavior particle swarm optimization algorithm is used to identify the model parameters in real time, generating identification results. Based on these results, a multi-dimensional database of thermal power plant operating characteristics is established, including efficiency curves under different loads, emission characteristics, and dynamic responses during start-up and shutdown processes.

[0059] High-precision sensors are deployed on the steam turbines, boilers, and generators of thermal power plants to collect real-time operating parameters. Edge computing nodes are deployed on-site, equipped with ARM Cortex-A72 processors and 4GB RAM, and running a lightweight Linux operating system.

[0060] The edge computing nodes are used to preprocess the collected data. The preprocessing includes outlier detection using an improved local anomaly factor algorithm, which reduces computational complexity by combining a sliding time window technique; data smoothing using an adaptive Savitzky-Gore filter, which automatically adjusts the filter window size and polynomial order according to the local characteristics of the data; and data compression using a hybrid compression algorithm based on wavelet transform and deep autoencoder, to obtain the preprocessed data.

[0061] Based on the preprocessed data, a multi-level dynamic nonlinear model of a thermal power plant is constructed. The model includes five subsystems: fuel delivery system, boiler system, turbine system, generator system, and pollutant control system.

[0062] A hybrid deep learning algorithm is used to train and optimize the model of each subsystem. The hybrid deep learning algorithm combines a long short-term memory network and a convolutional neural network. The long short-term memory network adopts a bidirectional structure, and the convolutional neural network adopts a multi-scale convolutional kernel. The outputs of the long short-term memory network and the convolutional neural network are fused through an attention mechanism.

[0063] An improved adaptive quantum behavior particle swarm optimization algorithm is used to achieve real-time identification of model parameters. Quantum states are used to represent particle positions and update phase angles. Quantum rotation gates are used to realize particle motion to obtain model identification results. Based on the model identification results, a multi-dimensional thermal power plant operation characteristic database is constructed. The database adopts a distributed storage architecture, uses Apache Cassandra as the underlying storage system, develops a RESTful API-based data access interface, supports multi-parameter joint queries and time series slicing, uses AES-256 encryption to store data, implements role-based access control, and records audit logs for data access operations.

[0064] The adaptive Savitzky-Gore filter is a signal processing technique suitable for noise reduction and data smoothing. The hybrid compression algorithm based on wavelet transform and deep autoencoder combines the time-frequency analysis capabilities of wavelet transform with the feature extraction capabilities of deep autoencoder, aiming to improve data compression efficiency, especially when processing complex signals (such as images or audio). The quantum state refers to the state of a quantum system, described by a wave function or density matrix, reflecting the physical properties of the system. The phase angle is an important parameter in quantum computing, where a quantum state can be represented in complex form, determining its position in the complex plane and affecting the outcome of quantum operations. The quantum rotation gate is one of the fundamental operations in quantum computing, manipulating qubits by rotating the phase angle of the quantum state. Common gates include RX, RY, and RZ gates, corresponding to rotations along different axes.

[0065] High-precision sensors are installed on key equipment in thermal power plants. Vibration sensors, temperature sensors, and pressure sensors are installed on steam turbines to monitor parameters such as bearing vibration, steam temperature, and pressure. Oxygen content sensors, temperature sensors, and flow sensors are installed on boilers to monitor parameters such as combustion efficiency, superheated steam temperature, and feedwater flow. Voltage sensors, current sensors, and power factor sensors are installed on generators to monitor parameters such as output voltage, current, and power factor. The sampling frequency of these sensors is set to 100Hz to ensure that rapid changes in the operating status of the equipment are captured.

[0066] Next, edge computing nodes were deployed on-site, with one edge computing node deployed in each major equipment area (such as the turbine room, boiler room, and generator room). Each node was equipped with an ARM Cortex-A72 processor with a main frequency of 2.5GHz, 4GB DDR4 RAM, and 128GB SSD storage. The nodes ran a customized lightweight Linux operating system, based on Ubuntu 20.04LTS and trimmed down to retain only the necessary system components and drivers to reduce system overhead and improve operating efficiency.

[0067] Edge computing nodes preprocess the collected data and use an improved local anomaly factor algorithm to detect outliers. A sliding time window technique is used, with the window size set to 1 minute and sliding once every 10 seconds. For each data point, the local density ratio of it to other data points in the window is calculated. If the ratio exceeds a preset threshold (e.g., 3.0), it is marked as an outlier.

[0068] The data smoothing was performed using an adaptive Savitzky-Gore filter with an initial window size of 21 data points and a polynomial order of 3. The filter dynamically adjusted the window size by calculating the local variance of the data within the window. If the local variance increased, the window size was reduced to preserve details; if the local variance decreased, the window size was increased to enhance the smoothing effect. The polynomial order was also adjusted according to the complexity of the data, ranging from 2 to 5.

[0069] Data compression is performed using a hybrid compression algorithm based on wavelet transform and deep autoencoder. First, the data is subjected to discrete wavelet transform using the Daubechies 4 wavelet basis with a decomposition layer of 3. The wavelet coefficients are then input into the deep autoencoder. The encoder part of the autoencoder contains 3 hidden layers with 256, 128, and 64 neurons, respectively. The decoder part is symmetrical and uses ReLU activation function and batch normalization layer to improve training efficiency and generalization ability. In this way, the original data can be compressed to 1 / 10 of its original size while keeping the reconstruction error below 1%.

[0070] Based on the preprocessed data, a multi-level dynamic nonlinear model of a thermal power plant is constructed. The model includes five subsystems: fuel delivery system, boiler system, turbine system, generator system, and pollutant control system. Each subsystem model considers the nonlinear relationship and dynamic characteristics between key parameters. For example, the boiler system model includes the combustion process, water-cooled wall heat transfer, superheater and reheater, and considers the influence of factors such as fuel composition, feedwater flow rate, and steam parameters.

[0071] A hybrid deep learning algorithm was used to train and optimize the model of each subsystem. This algorithm combines a long short-term memory network and a convolutional neural network. The long short-term memory network adopts a bidirectional structure with 128 hidden layer neurons to capture long-term dependencies in time series data. The convolutional neural network uses multi-scale convolutional kernels, including 1x1, 3x3 and 5x5 kernels, with 32 filters for each size to extract multi-scale features. The outputs of the long short-term memory network and the convolutional neural network are fused through an attention mechanism. The attention weights are calculated using the softmax function. The model is trained using the Adam optimizer with a learning rate of 0.001, a batch size of 64, and 100 training epochs.

[0072] An improved adaptive quantum behavior particle swarm optimization algorithm is used to identify model parameters in real time. The particle swarm size is set to 50, the maximum number of iterations is 100, each particle uses a qubit to represent its position, the initial phase angle is randomly generated, and particle motion is realized through a quantum rotation gate. The rotation angle is calculated based on the particle's current position, the individual optimal position, and the global optimal position. An adaptive inertia weight is introduced, with an initial value of 0.9, which decreases linearly to 0.4 as the number of iterations increases, in order to balance global search and local search capabilities.

[0073] Based on the model identification results, a multi-dimensional database of thermal power plant operation characteristics was constructed. The database adopts a distributed storage architecture, using Apache Cassandra as the underlying storage system, and sets up 3 data nodes. Each node is equipped with an 8-core CPU, 32GB RAM and 1TB SSD storage. The data is partitioned according to timestamp and device ID to optimize query performance.

[0074] Develop a data access interface based on a RESTful API, supporting multi-parameter join queries and time series slicing. The API is implemented using the Python Flask framework and deployed on a separate application server.

[0075] The stored data is encrypted using the AES-256 algorithm. The key length is 256 bits, encrypted using CBC mode and padded with PKCS7. Key management is protected by a Hardware Security Module (HSM), implementing role-based access control. Three roles are defined: administrator, engineer, and operator, each with different data access permissions. For example, operators can only view real-time operational data, while engineers can access historical data and perform advanced analysis.

[0076] In this embodiment, by deploying high-precision sensors on key equipment, the core operating parameters of the thermal power plant are accurately captured, providing high-quality raw data for subsequent data analysis and model building. Preliminary data preprocessing on-site significantly reduces data transmission volume, lowers network bandwidth requirements, and reduces the computational burden on the central server. The distributed computing architecture enables the system to respond more quickly to changes in equipment status, providing the possibility for real-time control and fault early warning. The hybrid compression algorithm based on wavelet transform and deep autoencoder significantly reduces storage and transmission costs while maintaining low reconstruction errors. The dynamic nonlinear model can better capture the transient characteristics and nonlinear behavior of the system, providing a more reliable theoretical basis for optimized control and fault diagnosis. In summary, this embodiment achieves efficient acquisition, processing, storage, and analysis of thermal power plant operating data, providing strong technical support for the intelligent and refined management of thermal power plants and helping to improve the overall operating efficiency and economic benefits of thermal power plants.

[0077] In one alternative implementation,

[0078] The particle position is represented by a quantum state, and the phase angle is updated using the following formula:

[0079] Δθ=ωΔθ prev +c1r1(θ pbest -θ)+c2r2(θ gbest -θ)+α·(rand()-0.5)·π;

[0080] Where Δθ is the change in phase angle, representing the update magnitude of the current phase angle, and ω represents the inertia weight. prev c1 represents the phase angle change in the previous step, r1 represents a random number between 0 and 1, and θ represents the individual learning factor. pbest Let c1 represent the individual's optimal phase angle, c2 represent the global learning factor, r2 represent another random number between 0 and 1, and θ represent the global learning factor. gbest α represents the global optimal phase angle, α represents the interference intensity, and rand() represents the uncertainty function, which is used to randomly generate a random number between 0 and 1.

[0081] In this embodiment, quantum states are used to represent particle positions, allowing particles to explore multiple possible solutions simultaneously. This enables the algorithm to conduct a broader and deeper search in the solution space, significantly increasing the probability of finding the global optimum. The superposition property of quantum states allows particles to evaluate multiple potential solutions simultaneously in each iteration, thereby accelerating the search for the optimal solution and achieving better optimization results in fewer iterations. The inertia term maintains the particle's motion trend, which is helpful for global search. The individual cognition term and the social cognition term guide the particle to move towards known good solutions, promoting local search. The random perturbation term increases the diversity of the search and helps to escape local optima. The adaptive mechanism enables the algorithm to maintain good performance at different stages of the optimization process, allowing for both extensive global search in the early stages and refined local search in the later stages. In summary, this embodiment provides a more efficient and accurate solution for parameter identification of dynamic nonlinear models of thermal power plants, thus laying a solid foundation for the intelligent operation and optimized control of thermal power plants.

[0082] S2. Based on the thermal power plant operation characteristic database, data mining technology is used to analyze the unit operation bottlenecks, including load change rate limiting factors and minimum technical output constraints. Multi-objective and multi-scheme flexible transformation schemes are designed, including the design of a boiler anti-ash accumulation intelligent ash cleaning system, intelligent optimization of the turbine bypass system, frequency conversion transformation of the feedwater system, and combustion optimization system transformation. Virtual reality and augmented reality technologies are used to visualize and optimize the transformation schemes. During the transformation implementation process, digital twin technology based on edge computing is used to construct a high-fidelity digital model of the thermal power plant. Through transformation, the start-up and shutdown speed, load change rate, and minimum technical output of the unit are improved, and the results of flexible transformation are obtained.

[0083] Data mining is the process of extracting valuable information from large amounts of data. Commonly used techniques include classification, clustering, association rule mining, and regression analysis. The minimum technical output constraint refers to setting the minimum output level that power generation equipment must reach when performing power generation scheduling or resource allocation to ensure the stability and reliability of the system. The edge computing-based digital twin technology refers to a virtual model that uses edge computing capabilities to monitor and optimize physical systems in real time. Digital twins create digital copies of physical devices or systems, collect and analyze data in real time, and thus reflect their status and behavior. Edge computing processes data at the data generation location (such as next to the device), reducing latency and improving response speed.

[0084] In one alternative implementation,

[0085] Based on a database of thermal power plant operating characteristics, data mining techniques are used to analyze unit operating bottlenecks, including load change rate limiting factors and minimum technical output constraints. Multi-objective, multi-scheme flexible retrofitting plans are designed, including the design of an intelligent ash removal system for boilers to prevent ash accumulation, intelligent optimization of the turbine bypass system, frequency conversion retrofitting of the feedwater system, and combustion optimization system retrofitting. Virtual reality and augmented reality technologies are used to visualize and optimize the retrofitting plans. During the retrofitting implementation process, edge computing-based digital twin technology is used to construct a high-fidelity digital model of the thermal power plant. Through retrofitting, the unit's start-up and shutdown speed, load change rate, and minimum technical output are improved, resulting in flexible retrofitting outcomes including:

[0086] A dynamic nonlinear model of the boiler system is constructed, and an intelligent combustion optimization system is developed based on the dynamic nonlinear model of the boiler system. The intelligent combustion optimization system includes using a long short-term memory network to predict future load changes, using an improved nonlinear model predictive control algorithm to optimize the coal feed rate, primary air volume and secondary air volume in real time, using a multi-objective optimization strategy to construct an optimization objective function and using a sequential quadratic programming algorithm combined with the trust region method to solve the nonlinear optimization problem.

[0087] A steam turbine system model is constructed, and an adaptive sliding pressure operation control strategy is developed based on the steam turbine system model. The adaptive sliding pressure operation control strategy includes training the control strategy using a deep deterministic policy gradient algorithm, defining a state space including main steam pressure, temperature, flow rate, and generator output power, defining the action space as the parameters of the sliding pressure curve and designing a reward function, deploying edge computing nodes on the feedwater pump, induced draft fan, and forced draft fan, running a lightweight model predictive control algorithm on each edge computing node, and designing a distributed coordination mechanism based on federated learning. Each edge node maintains a local model, updates the model parameters using local data, and sends the updated model parameters to a central server. The central server aggregates all updates and sends the aggregated updates back to each edge node, and the edge nodes use the received updates to update their local models.

[0088] Develop an integrated intelligent scheduling system that treats the boiler system, turbine system and auxiliary system as independent intelligent agents. Use an improved multi-agent soft actor-commentator algorithm to achieve overall optimal control. Each agent has its own agent policy network and value function network. Introduce a global value function to evaluate the value of joint actions and make decisions based on the local value function and the global value function.

[0089] A high-fidelity digital model of a thermal power plant is constructed through visualization simulation. By modifying the unit's start-up and shutdown speed, load change rate, and minimum technical output, flexible modification results are obtained.

[0090] The dynamic nonlinear model of the boiler system describes the dynamic behavior of the boiler under different operating conditions, including the combustion process, heat exchange, and pressure changes. The intelligent combustion optimization system refers to optimizing the combustion process using data-driven methods and intelligent algorithms to achieve the goals of improving combustion efficiency and reducing emissions. The coal feed rate refers to the amount of fuel (such as coal) supplied in the boiler system, which is an important parameter affecting combustion efficiency and boiler operation. The sequential quadratic programming algorithm is an optimization algorithm commonly used to solve optimization problems with quadratic objective functions and linear constraints. The adaptive sliding pressure operation control strategy is a control strategy that optimizes operating efficiency and stability by adjusting the system's operating points (such as pressure and flow rate). The improved multi-agent soft actor-commentator algorithm is an algorithm in reinforcement learning that optimizes the decision-making process through cooperation and competition among multiple agents.

[0091] A dynamic nonlinear model of the boiler system is constructed, taking into account the complex thermodynamic processes and chemical reactions inside the boiler, including the combustion process, water circulation system, steam system, etc. The model adopts the partitioning method to divide the boiler into several control volumes. Mass balance, energy balance and momentum balance equations are established in each control volume. Considering the nonlinear characteristics of the boiler system, a neural network is introduced to describe the parts that are difficult to express by the mechanistic model, such as the complex relationship between combustion efficiency and multiple factors.

[0092] Based on the constructed dynamic nonlinear model of the boiler system, an intelligent combustion optimization system was developed. A long short-term memory network was used to predict future load changes. The input of the long short-term memory network includes historical load data, weather forecast information, holiday information, etc., and the output is the load prediction for the next 24 hours. The network structure includes a long short-term memory network layer with 128 neurons, a fully connected layer with 64 neurons, and an output layer. The network was trained using the Adam optimizer with a learning rate of 0.001 and a batch size of 32.

[0093] Based on load forecasting results, an improved nonlinear model predictive control algorithm is used to optimize coal feed rate, primary air volume, and secondary air volume in real time. This improved algorithm considers the strongly coupled nonlinear characteristics of the boiler system. The prediction time domain is set to 30 minutes, and the control time domain is 5 minutes. The constraints for the control variables include a coal feed rate between 30-100 t / h and a primary air volume between 100-300 kNm³. 3 The secondary air volume is between 50-150 kNm / h. 3 Between / h, the objective function adopts a multi-objective optimization strategy, which comprehensively considers boiler efficiency, NOx emissions and steam parameter stability.

[0094] The objective function is defined as a weighted sum of boiler efficiency, NOx emissions, and main steam temperature deviation. The weighting coefficients, determined through expert experience and historical data analysis, are 0.5, 0.3, and 0.2, respectively. A sequential quadratic programming algorithm combined with the trust region method is used to solve the nonlinear optimization problem. The initial trust region radius of the sequential quadratic programming algorithm is set to 1, the maximum number of iterations is 100, and the convergence tolerance is 1e-6.

[0095] A steam turbine system model is constructed. The model includes major components such as the high-pressure cylinder, intermediate-pressure cylinder, low-pressure cylinder, reheater, and condenser. Thermodynamic equations and energy balance equations are established for each component, taking into account the nonlinear characteristics of the steam turbine, such as valve characteristic curves and efficiency curves. The model inputs include main steam parameters, extraction steam rate, cooling water temperature, etc., and the outputs are performance indicators such as generator output power and heat rate.

[0096] Based on a steam turbine system model, an adaptive sliding pressure operation control strategy was developed. First, a deep deterministic strategy gradient algorithm was used to train the control strategy. The state space was defined to include main steam pressure, temperature, flow rate, and generator output power, with a sampling frequency of 1 Hz. The action space was defined as the parameters of the sliding pressure curve, including the coordinates of the control points of the cubic spline curve. A reward function was designed considering power generation efficiency, equipment lifespan, and operational stability; specifically, it was a weighted sum of the power generation efficiency improvement and the equipment stress penalty term. The weighting coefficients were determined experimentally to be 0.7 and 0.3.

[0097] The deep deterministic policy gradient algorithm uses two neural networks: an actor network and a critic network. The actor network consists of three fully connected layers with 256, 128, and 64 neurons, respectively, and uses the ReLU activation function. The critic network has a similar structure, but merges the action inputs after the second layer and uses the Adam optimizer with learning rates of 1e-4 and 1e-3, respectively. It uses an empirical replay buffer of size 100,000 and a batch size of 64. The soft update parameter τ of the target network is set to 0.001.

[0098] Edge computing nodes are deployed on the water pump, induced draft fan and forced draft fan. Each node is equipped with an ARM Cortex-A72 processor and 4GB RAM. A lightweight model predictive control algorithm is run on each edge computing node. The prediction time domain is 10 minutes and the control time domain is 1 minute. The control variables include equipment speed and valve opening. The constraints take into account equipment operating limitations and process parameter requirements.

[0099] Design a distributed coordination mechanism based on federated learning. Each edge node maintains a local model and updates its parameters using local data. The local model employs a three-layer fully connected neural network with 64 and 32 neurons in the hidden layers, using the ReLU activation function. Every 30 minutes, the edge nodes send the updated model parameters to the central server. The central server aggregates all updates using a federated averaging algorithm, with the aggregation weight proportional to the amount of data at each node. The aggregated updates are then sent back to each edge node, which uses the received updates to update its local model.

[0100] A comprehensive intelligent scheduling system was developed, treating the boiler system, turbine system, and auxiliary systems as independent intelligent agents. An improved multi-agent soft actor-commentator (MA-SAC) algorithm was used to achieve overall optimal control. Each agent corresponds to an agent policy network and a value function network. The agent policy network adopts a Gaussian policy and consists of a three-layer fully connected network with 128 and 64 neurons in the hidden layer. The value function network has a similar structure, but the output layer has only one neuron.

[0101] A global value function is introduced to evaluate the value of joint actions. The global value function network consists of four fully connected layers with 256, 128, and 64 hidden layer neurons. Decisions are made based on local and global value functions. A soft Q-learning update strategy is used. The initial value of the temperature parameter is set to 0.2 and dynamically adjusted through an automatic adjustment mechanism. Importance sampling technique is used to improve sample efficiency. The truncation parameter is set to 0.2.

[0102] Visual simulations were conducted to construct a high-fidelity digital model of the thermal power plant. The Unity 3D engine was used to create a 3D scene with centimeter-level modeling accuracy. By modifying the unit, the start-up and shutdown speed, load change rate, and minimum technical output were improved, resulting in flexible modification outcomes. Specifically, by optimizing the boiler water-cooled wall structure, improving the turbine regulation system, and upgrading the DCS control strategy, the unit start-up time was shortened from 4 hours to 2.5 hours, the load change rate was increased from 2% of rated load / minute to 5% of rated load / minute, and the minimum technical output was reduced from 40% of rated load to 25% of rated load.

[0103] In this embodiment, by constructing a dynamic nonlinear model including the boiler system and turbine system, and combining it with neural networks to handle complex relationships that are difficult to express using mechanistic models, the accuracy of system modeling is significantly improved. Through an adaptive sliding pressure operation control strategy, the control strategy is trained using the Deep Deterministic Policy Gradient (DDPG) algorithm, significantly improving the unit's load regulation capability. Edge computing nodes are deployed on key equipment to run a lightweight model predictive control algorithm, and global optimization is achieved through a distributed coordination mechanism based on federated learning, significantly improving the response speed and reliability of the control system. A high-fidelity digital model is constructed, achieving centimeter-level simulation accuracy, providing a highly realistic virtual environment for equipment modification and operation optimization, greatly reducing the risks and costs of actual modifications, and accelerating the verification and implementation of optimization schemes. Through a series of modification measures, the unit's start-up and shutdown speed, load change rate, and minimum technical output are significantly improved, enabling the thermal power unit to better adapt to grid fluctuations brought about by the large-scale integration of renewable energy, improving the unit's economy and environmental friendliness. In summary, this embodiment not only improves the operating efficiency and economic benefits of thermal power plants but also enhances the adaptability of thermal power units in the context of large-scale integration of new energy, providing strong support for the intelligent and green transformation of thermal power plants.

[0104] In one alternative implementation,

[0105] The update objective of the value function network is shown in the following formula:

[0106]

[0107] Where E represents the expectation, s t Indicates the current state, a t Indicates the current action, r t s represents the current reward value. t+1 Denotes the next state, D represents the experience replay buffer, and Qi(s) represents the next state. t a t ) represents the value function network for the current state s t and current action a t The estimated value, where γ represents the discount factor. Indicates according to strategy π i From the next state s t+1 Randomly select action a t+1 The expected value, Q i (s t+1 a t+1 ) represents the value function network for the next state s t+1和 Next action a t+1 The estimated value, β represents the policy entropy adjustment factor, and P represents the current policy π. i In state s t+1Choose action a t+1 The logarithmic probability.

[0108] In this embodiment, the strong correlation between consecutive states is broken by introducing an experience replay buffer, reducing the correlation between samples, effectively reducing the variance in the learning process, and improving the stability of the algorithm. By introducing a policy entropy adjustment factor and the log probability of the policy, it helps to avoid the policy getting stuck in local optima and increases the possibility of finding the global optimum. By using the difference between the target network and the main network to calculate the loss, this method implements a soft update mechanism, which can smooth the update process of the value function, reduce oscillations in the learning process, and improve the stability and convergence of the algorithm. The dynamic balancing mechanism enables the algorithm to adopt the most suitable policy at different learning stages, improving learning efficiency. In summary, this embodiment enables it to cope with complex and ever-changing operating environments more intelligently, efficiently, and flexibly, providing strong support for the intelligent transformation and green development of thermal power plants.

[0109] S3. Based on the dynamic nonlinear model and the results of flexibility modification, a multi-objective optimization function is constructed. The objectives include maximizing unit efficiency, minimizing emissions, maximizing peak-shaving capacity, and maximizing equipment lifespan. An improved model predictive control algorithm is adopted, combined with deep reinforcement learning methods, to formulate an adaptive deep peak-shaving optimization control strategy. Short-term and medium-term forecasts of grid load are performed. Based on the forecast results, a deep deterministic strategy gradient algorithm is used to generate an initial control strategy. The initial control strategy is optimized and adjusted online. A hierarchical distributed control architecture is designed to generate a global optimization objective. The global objective is decomposed into local objectives of each subsystem, and the interaction between subsystems is coordinated. Intelligent actuators are used to achieve precise control of each subsystem and coordinated control of the unit is realized. A real-time performance evaluation mechanism based on big data analysis is established. A sliding time window method is used to continuously evaluate and optimize the control strategy. Based on the evaluation results, an adaptive learning algorithm is used to continuously optimize the control strategy.

[0110] The model predictive control algorithm is a model-based control strategy that optimizes future control inputs by predicting the dynamic mathematical model of the system. The adaptive deep peak shaving optimization control strategy combines adaptive control and deep learning methods to optimize the system's load scheduling and peak management. The hierarchical distributed control architecture divides the control system into multiple levels and distributed units, typically including a field layer, a control layer, and a management layer. Each level is responsible for different functions, and the levels coordinate through communication. The real-time performance evaluation mechanism based on big data analysis utilizes big data analysis technology to monitor and analyze real-time data to evaluate the system's performance.

[0111] In one alternative implementation,

[0112] Based on the dynamic nonlinear model and the results of flexibility retrofitting, a multi-objective optimization function is constructed, with objectives including maximizing unit efficiency, minimizing emissions, maximizing peak-shaving capacity, and maximizing equipment lifespan. An improved model predictive control algorithm, combined with deep reinforcement learning, is employed to formulate an adaptive deep peak-shaving optimization control strategy. Short-term and medium-term grid load forecasts are performed. Based on the forecast results, a deep deterministic strategy gradient algorithm is used to generate an initial control strategy, which is then optimized and adjusted online. A hierarchical distributed control architecture is designed to generate a global optimization objective, which is decomposed into local objectives for each subsystem. The interactions between subsystems are coordinated, and precise control of each subsystem is achieved through intelligent actuators to realize coordinated unit control. A real-time performance evaluation mechanism based on big data analysis is established, employing a sliding time window method to continuously evaluate and optimize the control strategy. Based on the evaluation results, an adaptive learning algorithm is used to continuously optimize the control strategy, including:

[0113] Based on the dynamic nonlinear model and the results of the flexibility modification, a multi-objective optimization function is constructed, with objectives including maximizing unit efficiency, minimizing emissions, maximizing peak-shaving capacity, and maximizing equipment lifespan. An improved model predictive control algorithm is designed, which includes a state estimator, a constraint processor, and an optimization solver.

[0114] A deep reinforcement learning framework is constructed, which includes a deep policy network and a value network. The deep policy network is used to generate control actions, and the value network is used to evaluate state values. The improved model predictive control algorithm is combined with the deep reinforcement learning framework to form an adaptive deep peak shaving optimization control strategy.

[0115] A method combining long short-term memory networks and convolutional neural networks is used to perform short-term and medium-term forecasts of power grid load, obtaining load forecast results. Based on the load forecast results, an initial control strategy is generated according to the improved model predictive control algorithm and the deep deterministic policy gradient algorithm. The deep deterministic policy gradient algorithm includes an actor network and a critic network, both of which adopt a four-layer fully connected neural network structure.

[0116] An online optimization module is designed to optimize and adjust the initial control strategy in real time. The online optimization module adopts the model predictive control method to solve the optimal control sequence in the rolling time domain and constructs a hierarchical distributed control architecture, which includes a global coordination layer and a local control layer. The global coordination layer is responsible for generating the global optimization objective and uses a decomposition coordination algorithm to decompose the global objective into local objectives of each subsystem. The local control layer is responsible for executing the precise control of each subsystem and uses an adaptive PID control algorithm for tracking control of the local objectives. An intelligent actuator is designed, which includes an advanced actuator and an embedded controller to achieve precise control of each subsystem. A real-time performance evaluation mechanism based on big data analysis is established, and a sliding time window method is used to continuously evaluate and optimize the adaptive deep peak shaving optimization control strategy.

[0117] Based on the evaluation results of the real-time performance evaluation mechanism, an adaptive learning algorithm is used to continuously optimize the adaptive deep peak shaving optimization control strategy. The adaptive learning algorithm uses the gradient descent method to update the control strategy parameters. A safety monitoring module is set up to monitor key operating parameters in real time. When an anomaly is detected or a constraint is about to be violated, a preset safety protection strategy is triggered.

[0118] The rolling time domain refers to the dynamically updated time range during the control or prediction process. The sliding time window method is used for data analysis and signal processing. It collects and processes data within a fixed-length time window and continuously moves the window over time.

[0119] Based on the aforementioned dynamic nonlinear model and the results of the flexible modification, a multi-objective optimization function is constructed. This optimization function comprehensively considers four objectives: maximizing unit efficiency, minimizing emissions, maximizing peak-shaving capacity, and maximizing equipment life. Unit efficiency is measured by heat rate, emissions mainly consider NOx, SO2, and dust, peak-shaving capacity includes load change rate and minimum technical output, and equipment life is evaluated based on fatigue damage models of key components.

[0120] To achieve multi-objective optimization, an improved model predictive control algorithm is designed, comprising three key components: a state estimator, a constraint processor, and an optimization solver. The state estimator employs the extended Kalman filter method, fusing real-time data from various sensors to accurately estimate the current state of the system. The constraint processor considers various hard and soft constraints on equipment operation, such as boiler steam parameter limits and turbine vibration limits. The optimization solver uses a sequential quadratic programming method to solve for the optimal control sequence under the condition of satisfying the constraints.

[0121] A deep reinforcement learning framework is constructed, including a deep policy network and a value network. The deep policy network adopts a four-layer fully connected neural network structure. The number of nodes in the input layer is the system state dimension, the hidden layers have 128 and 64 nodes respectively, and the number of nodes in the output layer is the control action dimension. The value network also adopts a similar structure, but the output layer has only one node, which is used to estimate the state value. Both networks use the ReLU activation function and are trained using the Adam optimizer.

[0122] By combining the improved model predictive control algorithm with a deep reinforcement learning framework, an adaptive deep peak shaving optimization control strategy is formed. Specifically, the model predictive control algorithm is used to generate an initial control sequence, which is then used as the starting point for deep reinforcement learning. Through interaction with the environment, deep reinforcement learning continuously optimizes the control strategy and gradually improves peak shaving performance.

[0123] A method combining Long Short-Term Memory (LSTM) networks and Convolutional Neural Networks (CNNs) is used for short-term and medium-term load forecasting. The LSTM network consists of two layers, each with 128 hidden units, used to capture long-term dependencies in load sequences. The CNN network consists of two convolutional layers, using 32 and 64 3x3 convolutional kernels respectively, used to extract local features from the load data. The outputs of the LSTM network and the CNN are connected and passed through a fully connected layer to obtain the final load forecast result.

[0124] Based on the load forecast results, an initial control policy is generated using an improved model predictive control algorithm and a deep deterministic policy gradient algorithm. In the deep deterministic policy gradient algorithm, both the actor network and the critic network adopt a four-layer fully connected neural network structure. The input layer of the actor network corresponds to the system state, and the output layer corresponds to the continuous control actions. The input of the critic network includes the state and the action, and the output is the value estimate of the state-action pair. Both networks use batch normalization technology to accelerate the training process.

[0125] The design includes an online optimization module that optimizes and adjusts the initial control strategy in real time. This module uses model predictive control to solve for the optimal control sequence within a 10-minute rolling time domain and updates the control input every 30 seconds to ensure that the control strategy can respond to changes in system state in a timely manner.

[0126] A hierarchical distributed control architecture is constructed, including a global coordination layer and a local control layer. The global coordination layer is responsible for generating the global optimization objective and uses the Lagrange relaxation method to decompose the global objective into the local objectives of each subsystem. The local control layer is responsible for executing the precise control of each subsystem and uses an adaptive PID control algorithm to track and control the local objectives. The PID parameters are adjusted in real time through a fuzzy inference system to adapt to different operating conditions.

[0127] Design an intelligent actuator, including an advanced actuator and an embedded controller. The actuator uses a high-precision servo motor with a response time of less than 50 milliseconds. The embedded controller is based on an ARM Cortex-M4 processor with an operating frequency of 168MHz and is capable of implementing complex control algorithms.

[0128] Establish a real-time performance evaluation mechanism based on big data analysis, adopt a 5-minute sliding time window, continuously evaluate the system's efficiency, emissions, peak-shaving performance and other indicators, use principal component analysis to reduce the dimensionality of multidimensional performance indicators, and calculate the comprehensive performance score.

[0129] Based on the evaluation results of the real-time performance evaluation mechanism, an adaptive learning algorithm is used to continuously optimize the adaptive deep peak shaving optimization control strategy. The adaptive learning algorithm uses the stochastic gradient descent method, with the learning rate initially set to 0.001 and dynamically adjusted according to the degree of performance improvement. The control strategy parameters are updated every 100 time steps.

[0130] A safety monitoring module is set up to monitor key operating parameters in real time. When an anomaly is detected or a constraint is about to be violated, a preset safety protection strategy is triggered. The safety protection strategy includes a multi-level response mechanism such as rapid load reduction and emergency shutdown. The key parameters monitored include boiler steam temperature and pressure, turbine vibration, and bearing temperature, with a sampling frequency of 100Hz.

[0131] In a practical application of a 600MW subcritical unit, after adopting the above method, the unit efficiency increased by 1.8 percentage points, NOx emissions decreased by 22%, the load change rate increased from 2% of rated load / minute to 6% of rated load / minute, the minimum technical output decreased from 40% of rated load to 22% of rated load, and the expected life of the main equipment was extended by about 15%. Under the condition of frequent grid peak shaving, the unit's annual equivalent utilization hours increased by about 450 hours.

[0132] In this embodiment, multi-objective optimization and deep reinforcement learning significantly improved the load change rate of the unit and reduced the minimum technical output, greatly enhancing the unit's peak-shaving flexibility and enabling it to better adapt to the large-scale grid connection of renewable energy. By incorporating equipment lifespan maximization into the optimization objective and employing an evaluation method based on a fatigue damage model, equipment wear was effectively reduced, not only lowering maintenance costs but also improving equipment reliability and availability. By combining model predictive control and deep reinforcement learning, the control strategy can be adaptively adjusted according to changes in system state and environment, better addressing uncertainties such as load fluctuations and fuel quality changes, thus improving system robustness. In summary, this embodiment not only improves the performance of individual units but also provides strong support for the flexibility and reliability of the entire power system, which is of great significance for promoting energy transition and sustainable development.

[0133] Figure 2 This is a schematic diagram of the system structure according to an embodiment of the present invention, such as... Figure 2 As shown, the system includes:

[0134] The first unit is used to collect comprehensive operational data of thermal power plants using a distributed sensor network. Edge computing technology is used to preprocess the collected data, including outlier detection, data smoothing and compression. Based on the preprocessed data, a multi-level dynamic nonlinear model of the thermal power plant is constructed, including the fuel delivery system, boiler system, turbine system, generator system and pollutant control system. A hybrid deep learning algorithm, combining long short-term memory network and convolutional neural network, is used to train and optimize the models of each subsystem. An improved adaptive quantum behavior particle swarm optimization algorithm is used to identify the model parameters in real time, generate identification results, and based on the identification results, establish a multi-dimensional thermal power plant operation characteristic database, including efficiency curves, emission characteristics and dynamic response of start-up and shutdown processes under different loads.

[0135] The second unit is used to analyze unit operation bottlenecks based on the thermal power plant operation characteristic database and using data mining technology, including load change rate limiting factors and minimum technical output constraints. It designs multi-objective and multi-scheme flexible transformation schemes, including the design of boiler anti-ash accumulation intelligent ash cleaning system, intelligent optimization of turbine bypass system, frequency conversion transformation of feedwater system and combustion optimization system transformation. It uses virtual reality and augmented reality technology to visualize and optimize the transformation schemes. During the transformation implementation process, it adopts edge computing-based digital twin technology to build a high-fidelity digital model of the thermal power plant. Through transformation, it improves the unit's start-up and shutdown speed, load change rate and minimum technical output, and obtains flexible transformation results.

[0136] The third unit is used to construct a multi-objective optimization function based on the dynamic nonlinear model and the results of flexibility modifications. The objectives include maximizing unit efficiency, minimizing emissions, maximizing peak-shaving capacity, and maximizing equipment lifespan. An improved model predictive control algorithm is adopted, combined with deep reinforcement learning methods, to formulate an adaptive deep peak-shaving optimization control strategy. Short-term and medium-term forecasts of grid load are performed. Based on the forecast results, a deep deterministic strategy gradient algorithm is used to generate an initial control strategy. The initial control strategy is optimized and adjusted online. A hierarchical distributed control architecture is designed to generate a global optimization objective. The global objective is decomposed into local objectives of each subsystem, and the interaction between subsystems is coordinated. Intelligent actuators are used to achieve precise control of each subsystem to realize coordinated control of the unit. A real-time performance evaluation mechanism based on big data analysis is established. A sliding time window method is used to continuously evaluate and optimize the control strategy. Based on the evaluation results, an adaptive learning algorithm is used to continuously optimize the control strategy.

[0137] A third aspect of the embodiments of the present invention,

[0138] An electronic device is provided, comprising:

[0139] processor;

[0140] Memory used to store processor-executable instructions;

[0141] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0142] Fourth aspect of the present invention,

[0143] A computer-readable storage medium is provided, having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0144] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0145] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for the flexible retrofitting and deep peak-shaving optimization control of thermal power plants, characterized in that, include: A distributed sensor network is used to collect comprehensive operational data from a thermal power plant. Edge computing technology is used to preprocess the collected data, including outlier detection, data smoothing, and compression. Based on the preprocessed data, a multi-level dynamic nonlinear model of the thermal power plant is constructed, including the fuel delivery system, boiler system, turbine system, generator system, and pollutant control system. A hybrid deep learning algorithm, combining long short-term memory networks and convolutional neural networks, is used to train and optimize the models of each subsystem. An improved adaptive quantum behavior particle swarm optimization algorithm is used to identify the model parameters in real time, generate identification results, and establish a multi-dimensional thermal power plant operation characteristic database based on the identification results, including efficiency curves, emission characteristics, and dynamic responses during start-up and shutdown processes under different loads. Based on a database of thermal power plant operating characteristics, data mining techniques are used to analyze unit operating bottlenecks, including load change rate limiting factors and minimum technical output constraints. Multi-objective and multi-scheme flexible retrofitting plans are designed, including the design of a boiler anti-ash accumulation intelligent ash cleaning system, intelligent optimization of the turbine bypass system, frequency conversion retrofitting of the feedwater system, and combustion optimization system retrofitting. Virtual reality and augmented reality technologies are used to visualize and optimize the retrofitting plans. During the retrofitting process, edge computing-based digital twin technology is used to construct a high-fidelity digital model of the thermal power plant. Through retrofitting, the start-up and shutdown speed, load change rate, and minimum technical output of the units are improved, resulting in flexible retrofitting outcomes. Based on the dynamic nonlinear model and the results of flexibility retrofitting, a multi-objective optimization function is constructed, with objectives including maximizing unit efficiency, minimizing emissions, maximizing peak-shaving capacity, and maximizing equipment lifespan. An improved model predictive control algorithm is adopted, combined with deep reinforcement learning methods, to formulate an adaptive deep peak-shaving optimization control strategy. Short-term and medium-term forecasts of grid load are performed. Based on the forecast results, a deep deterministic strategy gradient algorithm is used to generate an initial control strategy, which is then optimized and adjusted online. A hierarchical distributed control architecture is designed to generate a global optimization objective, which is decomposed into local objectives for each subsystem. The interactions between subsystems are coordinated, and precise control of each subsystem is achieved through intelligent actuators to realize coordinated control of the units. A real-time performance evaluation mechanism based on big data analysis is established, and a sliding time window method is used to continuously evaluate and optimize the control strategy. Based on the evaluation results, an adaptive learning algorithm is used to continuously optimize the control strategy.

2. The method according to claim 1, characterized in that, A distributed sensor network is used to collect comprehensive operational data from a thermal power plant. Edge computing technology is used to preprocess the collected data, including outlier detection, data smoothing, and compression. Based on the preprocessed data, a multi-level dynamic nonlinear model of the thermal power plant is constructed, including the fuel delivery system, boiler system, turbine system, generator system, and pollutant control system. A hybrid deep learning algorithm, combining long short-term memory networks and convolutional neural networks, is used to train and optimize the models of each subsystem. An improved adaptive quantum behavior particle swarm optimization algorithm is used to identify the model parameters in real time, generating identification results. Based on these results, a multi-dimensional database of thermal power plant operating characteristics is established, including efficiency curves under different loads, emission characteristics, and dynamic responses during start-up and shutdown processes. High-precision sensors are deployed on the steam turbines, boilers, and generators of thermal power plants to collect real-time operating parameters. Edge computing nodes are deployed on-site, equipped with ARM Cortex-A72 processors and 4GB RAM, and running a lightweight Linux operating system. The edge computing nodes are used to preprocess the collected data. The preprocessing includes outlier detection using an improved local anomaly factor algorithm, which reduces computational complexity by combining a sliding time window technique; data smoothing using an adaptive Savitzky-Gore filter, which automatically adjusts the filter window size and polynomial order according to the local characteristics of the data; and data compression using a hybrid compression algorithm based on wavelet transform and deep autoencoder, to obtain the preprocessed data. Based on the preprocessed data, a multi-level dynamic nonlinear model of a thermal power plant is constructed. The model includes five subsystems: fuel delivery system, boiler system, turbine system, generator system, and pollutant control system. A hybrid deep learning algorithm is used to train and optimize the model of each subsystem. The hybrid deep learning algorithm combines a long short-term memory network and a convolutional neural network. The long short-term memory network adopts a bidirectional structure, and the convolutional neural network adopts a multi-scale convolutional kernel. The outputs of the long short-term memory network and the convolutional neural network are fused through an attention mechanism. An improved adaptive quantum behavior particle swarm optimization algorithm is used to achieve real-time identification of model parameters. Quantum states are used to represent particle positions and update phase angles. A quantum rotation gate is used to realize particle motion to obtain model identification results. Based on the model identification results, a multi-dimensional thermal power plant operation characteristic database is constructed. The database adopts a distributed storage architecture, uses Apache Cassandra as the underlying storage system, develops a RESTful API-based data access interface, supports multi-parameter joint queries and time series slicing, uses AES-256 encryption to store data, implements role-based access control, and records audit logs for data access operations.

3. The method according to claim 2, characterized in that, The particle position is represented by a quantum state, and the phase angle is updated using the following formula: Δθ=ωΔθ prev +c1r1(θ pbest -θ)+c2r2(θ gbest -θ)+α·(rand()-0.5)·π; Where Δθ is the change in phase angle, representing the update magnitude of the current phase angle, and ω represents the inertia weight. prev c1 represents the phase angle change in the previous step, r1 represents a random number between 0 and 1, and θ represents the individual learning factor. pbest Let c1 represent the individual's optimal phase angle, c2 represent the global learning factor, r2 represent another random number between 0 and 1, and θ represent the global learning factor. gbest α represents the global optimal phase angle, α represents the interference intensity, and rand() represents the uncertainty function, which is used to randomly generate a random number between 0 and 1.

4. The method according to claim 1, characterized in that, Based on a database of thermal power plant operating characteristics, data mining techniques are used to analyze unit operating bottlenecks, including load change rate limiting factors and minimum technical output constraints. Multi-objective, multi-scheme flexible retrofitting plans are designed, including the design of an intelligent ash removal system for boilers to prevent ash accumulation, intelligent optimization of the turbine bypass system, frequency conversion retrofitting of the feedwater system, and combustion optimization system retrofitting. Virtual reality and augmented reality technologies are used to visualize and optimize the retrofitting plans. During the retrofitting implementation process, edge computing-based digital twin technology is used to construct a high-fidelity digital model of the thermal power plant. Through retrofitting, the unit's start-up and shutdown speed, load change rate, and minimum technical output are improved, resulting in flexible retrofitting outcomes including: A dynamic nonlinear model of the boiler system is constructed, and an intelligent combustion optimization system is developed based on the dynamic nonlinear model of the boiler system. The intelligent combustion optimization system includes using a long short-term memory network to predict future load changes, using an improved nonlinear model predictive control algorithm to optimize the coal feed rate, primary air volume and secondary air volume in real time, using a multi-objective optimization strategy to construct an optimization objective function and using a sequential quadratic programming algorithm combined with the trust region method to solve the nonlinear optimization problem. A steam turbine system model is constructed, and an adaptive sliding pressure operation control strategy is developed based on the steam turbine system model. The adaptive sliding pressure operation control strategy includes training the control strategy using a deep deterministic policy gradient algorithm, defining a state space including main steam pressure, temperature, flow rate, and generator output power, defining the action space as the parameters of the sliding pressure curve and designing a reward function, deploying edge computing nodes on the feedwater pump, induced draft fan, and forced draft fan, running a lightweight model predictive control algorithm on each edge computing node, and designing a distributed coordination mechanism based on federated learning. Each edge node maintains a local model, updates the model parameters using local data, and sends the updated model parameters to a central server. The central server aggregates all updates and sends the aggregated updates back to each edge node, and the edge nodes use the received updates to update their local models. Develop an integrated intelligent scheduling system that treats the boiler system, turbine system and auxiliary system as independent intelligent agents. Use an improved multi-agent soft actor-commentator algorithm to achieve overall optimal control. Each agent has its own agent policy network and value function network. Introduce a global value function to evaluate the value of joint actions and make decisions based on the local value function and the global value function. A high-fidelity digital model of a thermal power plant is constructed through visualization simulation. By modifying the unit's start-up and shutdown speed, load change rate, and minimum technical output, flexible modification results are obtained.

5. The method according to claim 4, characterized in that, The update objective of the value function network is shown in the following formula: Where E represents the expectation, s t Indicates the current state, a t Indicates the current action, r t s represents the current reward value. t+1 Denotes the next state, D represents the experience replay buffer, and Qi(s) represents the next state. t a t ) represents the value function network for the current state s t and current action a t The estimated value, where γ represents the discount factor. Indicates according to strategy π i From the next state s t+1 Randomly select action a t+1 The expected value, Q i (s t+1 a t+1 ) represents the value function network for the next state s t+1和 Next action a t+1 The estimated value, β represents the policy entropy adjustment factor, and P represents the current policy π. i In state s t+1 Choose action a t+1 The logarithmic probability.

6. The method according to claim 1, characterized in that, Based on the dynamic nonlinear model and the results of flexibility retrofitting, a multi-objective optimization function is constructed, with objectives including maximizing unit efficiency, minimizing emissions, maximizing peak-shaving capacity, and maximizing equipment lifespan. An improved model predictive control algorithm, combined with deep reinforcement learning, is employed to formulate an adaptive deep peak-shaving optimization control strategy. Short-term and medium-term grid load forecasts are performed. Based on the forecast results, a deep deterministic strategy gradient algorithm is used to generate an initial control strategy, which is then optimized and adjusted online. A hierarchical distributed control architecture is designed to generate a global optimization objective, which is decomposed into local objectives for each subsystem. The interactions between subsystems are coordinated, and precise control of each subsystem is achieved through intelligent actuators to realize coordinated unit control. A real-time performance evaluation mechanism based on big data analysis is established, employing a sliding time window method to continuously evaluate and optimize the control strategy. Based on the evaluation results, an adaptive learning algorithm is used to continuously optimize the control strategy, including: Based on the dynamic nonlinear model and the results of the flexibility modification, a multi-objective optimization function is constructed, with objectives including maximizing unit efficiency, minimizing emissions, maximizing peak-shaving capacity, and maximizing equipment lifespan. An improved model predictive control algorithm is designed, which includes a state estimator, a constraint processor, and an optimization solver. A deep reinforcement learning framework is constructed, which includes a deep policy network and a value network. The deep policy network is used to generate control actions, and the value network is used to evaluate state values. The improved model predictive control algorithm is combined with the deep reinforcement learning framework to form an adaptive deep peak shaving optimization control strategy. A method combining long short-term memory networks and convolutional neural networks is used to perform short-term and medium-term forecasts of power grid load, obtaining load forecast results. Based on the load forecast results, an initial control strategy is generated according to the improved model predictive control algorithm and the deep deterministic policy gradient algorithm. The deep deterministic policy gradient algorithm includes an actor network and a critic network, both of which adopt a four-layer fully connected neural network structure. An online optimization module is designed to optimize and adjust the initial control strategy in real time. The online optimization module adopts the model predictive control method to solve the optimal control sequence in the rolling time domain and constructs a hierarchical distributed control architecture, which includes a global coordination layer and a local control layer. The global coordination layer is responsible for generating the global optimization objective and uses a decomposition coordination algorithm to decompose the global objective into local objectives of each subsystem. The local control layer is responsible for executing the precise control of each subsystem and uses an adaptive PID control algorithm for tracking control of the local objectives. An intelligent actuator is designed, which includes an advanced actuator and an embedded controller to achieve precise control of each subsystem. A real-time performance evaluation mechanism based on big data analysis is established, and a sliding time window method is used to continuously evaluate and optimize the adaptive deep peak shaving optimization control strategy. Based on the evaluation results of the real-time performance evaluation mechanism, an adaptive learning algorithm is used to continuously optimize the adaptive deep peak shaving optimization control strategy. The adaptive learning algorithm uses the gradient descent method to update the control strategy parameters. A safety monitoring module is set up to monitor key operating parameters in real time. When an anomaly is detected or a constraint is about to be violated, a preset safety protection strategy is triggered.

7. A thermal power plant flexibility retrofit and deep peak-shaving optimization control system, used to implement the method of any one of claims 1-6, characterized in that, include: The first unit is used to collect comprehensive operational data of thermal power plants using a distributed sensor network. Edge computing technology is used to preprocess the collected data, including outlier detection, data smoothing and compression. Based on the preprocessed data, a multi-level dynamic nonlinear model of the thermal power plant is constructed, including the fuel delivery system, boiler system, turbine system, generator system and pollutant control system. A hybrid deep learning algorithm, combining long short-term memory network and convolutional neural network, is used to train and optimize the models of each subsystem. An improved adaptive quantum behavior particle swarm optimization algorithm is used to identify the model parameters in real time, generate identification results, and based on the identification results, establish a multi-dimensional thermal power plant operation characteristic database, including efficiency curves, emission characteristics and dynamic response of start-up and shutdown processes under different loads. The second unit is used to analyze unit operation bottlenecks based on the thermal power plant operation characteristic database and using data mining technology, including load change rate limiting factors and minimum technical output constraints. It designs multi-objective and multi-scheme flexible transformation schemes, including the design of boiler anti-ash accumulation intelligent ash cleaning system, intelligent optimization of turbine bypass system, frequency conversion transformation of feedwater system and combustion optimization system transformation. It uses virtual reality and augmented reality technology to visualize and optimize the transformation schemes. During the transformation implementation process, it adopts edge computing-based digital twin technology to build a high-fidelity digital model of the thermal power plant. Through transformation, it improves the unit's start-up and shutdown speed, load change rate and minimum technical output, and obtains flexible transformation results. The third unit is used to construct a multi-objective optimization function based on the dynamic nonlinear model and the results of flexibility modifications. The objectives include maximizing unit efficiency, minimizing emissions, maximizing peak-shaving capacity, and maximizing equipment lifespan. An improved model predictive control algorithm is adopted, combined with deep reinforcement learning methods, to formulate an adaptive deep peak-shaving optimization control strategy. Short-term and medium-term forecasts of grid load are performed. Based on the forecast results, a deep deterministic strategy gradient algorithm is used to generate an initial control strategy. The initial control strategy is optimized and adjusted online. A hierarchical distributed control architecture is designed to generate a global optimization objective. The global objective is decomposed into local objectives of each subsystem, and the interaction between subsystems is coordinated. Intelligent actuators are used to achieve precise control of each subsystem to realize coordinated control of the unit. A real-time performance evaluation mechanism based on big data analysis is established. A sliding time window method is used to continuously evaluate and optimize the control strategy. Based on the evaluation results, an adaptive learning algorithm is used to continuously optimize the control strategy.

8. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 6.

Images