A method for rapid load coordinated regulation of a coal-fired boiler for deep peak regulation

Abstract

The application discloses a kind of quick load collaborative regulation method for depth peak shaving of coal-fired boiler, and relates to the technical field of coal-fired boiler.The method steps include: collecting coal-fired boiler load step response data, and inputting to improved physical information time sequence neural network model, output to obtain thermal inertia coefficient;Based on thermal inertia coefficient, theoretical boiler load response path is constructed using quantile regression method;Based on theoretical boiler load response path, load response optimization is carried out on coal-fired boiler by improved particle swarm optimization algorithm, and the coal-fired boiler coal feeding and exhaust parameter sequence is obtained;According to coal-fired boiler coal feeding and exhaust parameter sequence, coal-fired boiler is heated and combusted, and the actual boiler load response path of coal-fired boiler is collected, the actual boiler load response path is compared with the theoretical boiler load response path, and the load path deviation is obtained, and the coal-fired boiler coal feeding and exhaust parameter sequence is adjusted based on the load path deviation, to obtain the corrected coal-fired boiler coal feeding and exhaust parameter sequence.

Description

Technical Field

[0001] This invention relates to the field of coal-fired boiler technology, specifically to a rapid load coordination regulation method for coal-fired boilers oriented towards deep peak shaving. Background Technology

[0002] With the large-scale grid connection of new energy power generation, the power system's requirements for the deep peak-shaving flexibility of coal-fired boilers are becoming increasingly stringent. They need to achieve rapid and accurate load response across a wide load range to balance grid supply and demand fluctuations. The internal thermal dynamics of a coal-fired boiler furnace are complex, and the thermal inertia and temperature response characteristics at different temperature measurement points vary significantly. Optimizing the matching of coal feeding and exhaust parameters directly affects the efficiency and stability of load regulation.

[0003] In the field of load regulation of coal-fired boilers, traditional methods often adopt strategies based on proportional-integral-derivative (PID) control or model predictive control (MPC), treating the coal-fired boiler as a unified whole, and following the regulation command of the target load, synchronously adjusting the fuel supply and air distribution of all burners in the coal-fired boiler to achieve load changes.

[0004] However, traditional methods treat the furnace as a whole thermal system, lacking consideration of the differences in thermal inertia of internal temperature measuring points. This makes it impossible to formulate differentiated coal feeding and exhaust parameter adjustment sequences based on the thermal inertia of different temperature measuring points, resulting in a mismatch between fuel air distribution and local heat demand in the furnace. Ultimately, this leads to problems such as response lag and temperature imbalance at temperature measuring points during load regulation, making it difficult to meet the requirements of deep peak shaving for the accuracy and speed of load regulation. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a rapid load coordination regulation method for coal-fired boilers oriented towards deep peak shaving, thereby solving the problems existing in the background technology.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for rapid load coordination regulation of coal-fired boilers for deep peak shaving, comprising the following steps: Step S1: Collect the load step response data of the coal-fired boiler and input it into the improved physical information time-series neural network model to output the thermal inertia coefficient; Step S2: Based on the thermal inertia coefficient, construct the theoretical boiler load response path using the quantile regression method; Step S3: Based on the theoretical boiler load response path, the load response of the coal-fired boiler is optimized using an improved particle swarm optimization algorithm to obtain the coal feeding and exhaust parameter sequence of the coal-fired boiler. Step S4: Heating and combustion of the coal-fired boiler according to the coal feeding and exhaust parameter sequence, and collecting the actual boiler load response path of the coal-fired boiler. The actual boiler load response path is compared with the theoretical boiler load response path to obtain the load path deviation. The coal feeding and exhaust parameter sequence is adjusted based on the load path deviation to obtain the corrected coal feeding and exhaust parameter sequence, thereby realizing the load regulation of the coal-fired boiler.

[0007] Preferably, the collection of load step response data of coal-fired boilers includes the following specific steps: Collect operating data of a coal-fired boiler at preset temperature measurement points during periods of a step change in load, i.e., before the step change in load. - , After step change + + , + + + Within the time window, the standard deviation of the load command is less than 0.5% of the rated load. This represents the length of the steady-state time window. The time required to complete the main transition process, This represents the starting moment of the step change. Indicates the duration from the start to the end of the load command; Calculate the values ​​before the step change respectively. - , [Time period and step change after] + + , + + + The arithmetic mean of u(t) over the time period is used to obtain the coal feed rate before the step jump. Coal feed rate after step jump Let u(t) be the total coal feed rate at time t, and finally calculate the step coal feed rate change. , = - ; Calculate the steady-state window before a step change [ - , The average temperature of the i-th temperature measuring point within the range Then the temperature relative change sequence at that temperature measuring point , , Let be the temperature value of the i-th temperature measuring point at time t; the sequence of relative changes of all temperature measuring points is the step response data of the coal-fired boiler load.

[0008] Preferably, the improved physical information temporal neural network model includes the following steps: An improved physical information temporal neural network model is constructed, which includes an input layer, a feature extraction layer, a temporal dependency layer, a parameter regression layer, and an output layer. The input layer of the physical information temporal neural network model is a pair of time-series data for each temperature measurement point. , =[ ; ],in This represents the step change in coal feed rate. Let be the sequence of relative temperature changes at the i-th temperature measurement point; The feature extraction layer is implemented using a two-layer one-dimensional convolutional neural network. The first layer contains 64 convolutional kernels of size 5, and the second layer contains 128 convolutional kernels of size 3. Both layers use Same Padding with a stride of 1 to maintain the temporal length and use the ReLU activation function for nonlinear transformation. The temporal dependency layer consists of two layers of gated recurrent units. The first gated recurrent unit contains 64 hidden units, and the second gated recurrent unit contains 32 hidden units. A dropout rate of 0.2 is set between layers to prevent overfitting. The parameter regression layer contains two fully connected layers. The first fully connected layer has 16 neurons, and the second fully connected layer has 8 neurons. The final output layer contains 1 neuron and outputs the thermal inertia coefficient.

[0009] Preferably, the output yields the thermal inertia coefficient, comprising the following steps: The total loss function of the physical information temporal neural network model is , ,in, To fit the loss function to the data, The physical constraint loss function; Minimize the total loss function using the optimizer. The network weights and biases of the physical information temporal neural network model are optimized, and the thermal inertia coefficient of the i-th temperature measurement point is finally output. .

[0010] Preferably, the step of constructing the theoretical boiler load response path based on the thermal inertia coefficient using the quantile regression method includes the following steps: The response priorities of temperature measuring points are determined based on their thermal inertia coefficients; the activation time proportion of the p-th response priority is calculated using quantile regression. The proportion of time required to reach the target of the p-th response priority Scaling factor for the response time constant of the p-th response priority ; Startup time ratio based on the priority of the p-th response The proportion of time required to reach the target of the p-th response priority Scaling factor for the response time constant of the p-th response priority Calculate the temperature reference value and the overall load reference value at the temperature measuring point; The theoretical boiler load response path is finally obtained. , =[ , ],in, This represents the reference temperature value at time t for the i-th temperature measuring point. Reference value for the overall boiler load at time t.

[0011] Preferably, the step of prioritizing the response of temperature measuring points based on the thermal inertia coefficient of each temperature measuring point includes the following steps: Based on the thermal inertia coefficient and steady-state gain coefficient of each temperature measuring point, a dual feature vector is constructed. The dual feature vector is then clustered using the K-means clustering method to obtain P clusters. Based on the P clusters, the temperature measuring points are divided into P response priorities.

[0012] Preferably, the calculation of the temperature reference value and the overall load reference value at the temperature measuring point includes the following specific steps: For each temperature measurement point, a reference temperature trajectory is generated according to its response priority: ; in, This represents the reference temperature value at time t for the i-th temperature measuring point. The steady-state temperature reference value of the i-th temperature measurement point at the start of the step. To correspond to the boiler target load The expected steady-state temperature value of the i-th temperature measurement point. The start time of temperature regulation Temperature response time constant, where t is the time index; Calculate the reference value for the overall boiler load: ; in, Reference value for the overall boiler load at time t. The initial steady-state load of the boiler at the moment of step start. Boiler target load, This represents the overall load inertia time constant of the boiler.

[0013] Preferably, the improved particle swarm optimization algorithm includes the following specific steps: Construct an improved fitness function for the particle swarm optimization algorithm; the update formulas for the position and velocity of each particle in the improved particle swarm optimization algorithm are as follows: ; ; in, This represents the velocity of the b-th particle in the (g+1)-th iteration. This represents the velocity of the b-th particle in the g-th iteration. This represents the position of the b-th particle in the (g+1)-th iteration. This represents the position of the b-th particle in the g-th iteration. For inertial weights, The maximum number of iterations, and The learning factor is rand1() and rand2(), which are uniformly distributed random numbers in the range [0,1]. This represents the historical best position of the b-th particle. This represents the globally optimal position for all particles.

[0014] Preferably, the construction of the fitness function for the improved particle swarm optimization algorithm includes the following specific steps: ; Where F is the fitness function value, and H is the future time domain. Indexing future moments For load weighting, Temperature weighting, express The predicted load value, express The load value of the theoretical boiler load response path at any given time. express The predicted temperature value of the i-th temperature measuring point. express Temperature reference value for the theoretical boiler load response path at any given time.

[0015] Preferably, the specific steps for comparing the actual boiler load response path with the theoretical boiler load response path to obtain the load path deviation are as follows: ; ; in, This represents the load deviation at time t. This represents the temperature deviation at the i-th temperature measurement point. This represents the actual load at time t. This represents the actual temperature of the i-th temperature measurement point at time t; The final result is the load path deviation, which includes load deviation and temperature deviation.

[0016] This invention provides a rapid load coordination regulation method for coal-fired boilers oriented towards deep peak shaving, involving coal-fired and power load control system technology, which has the following beneficial effects: (1) The significance of adopting the physical information time-series neural network model lies in its effective integration of data-driven methods and prior physical knowledge by embedding traditional physical laws (such as first-order inertia plus pure time delay model) into the neural network architecture. This overcomes the limitations of traditional curve fitting methods, which are sensitive to initial values ​​and prone to getting trapped in local optima. At the same time, it overcomes the deficiency of pure data-driven models in lacking physical interpretability. By using a physical constraint loss function, it ensures that the predicted value of the thermal inertia coefficient conforms to actual physical laws (such as non-negativity and reasonable range), thereby accurately extracting the deep features of the thermal inertia coefficient from boiler operation data. This provides a reliable quantitative basis for the subsequent realization of precise load regulation through zone coordination, and significantly improves the accuracy and robustness of thermal inertia spatial mapping.

[0017] (2) The significance of adopting the improved physical information temporal neural network model lies in its enhanced temporal feature extraction capability and generalization performance through optimized model architecture (such as convolutional neural networks and gated recurrent units) and training strategies (such as dynamic learning rate adjustment). The improved model can capture the dynamic characteristics in boiler load step response data more precisely, and balance data fitting and physical constraints through an adaptive loss function, making the prediction of thermal inertia coefficient more consistent with actual operating conditions. This not only improves the preprocessing accuracy of the load regulation front end, but also provides high-quality input for subsequent quantile regression and particle swarm optimization, thereby improving the overall speed, stability and adaptability of load regulation.

[0018] (3) An improved particle swarm optimization algorithm is used to efficiently solve the complex optimization problem of boiler load response, which is characterized by high dimension and multiple constraints. This algorithm encodes the coal feeding and air distribution parameter sequences of each burner into particle positions and uses swarm intelligence to perform parallel search in the solution space. Its dynamically adjusted inertia weights and learning factors effectively balance global exploration and local development capabilities, thereby quickly finding the optimal operating instructions that enable the actual boiler state to accurately track the theoretical load response path. This closed-loop strategy based on rolling optimization gives the system a strong real-time adaptive capability, which can effectively compensate for uncertainties such as coal quality fluctuations. Under the premise of ensuring safety constraints, it achieves precise and efficient coordinated regulation of fuel and air distribution, and significantly improves the speed, stability and robustness of load regulation. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating the steps of a rapid load coordination regulation method for coal-fired boilers oriented towards deep peak shaving proposed in this invention. Figure 2 This is a structural diagram of the physical information temporal neural network model in the rapid load coordination regulation method for coal-fired boilers oriented towards deep peak shaving proposed in this invention; Figure 3 This is a step hierarchy diagram of obtaining the theoretical boiler load response path in a rapid load coordination regulation method for coal-fired boilers oriented towards deep peak shaving proposed in this invention. Detailed Implementation

[0021] 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.

[0022] Please see Figures 1-3 The present invention provides a technical solution: a method for rapid load coordination regulation of coal-fired boilers for deep peak shaving.

[0023] Step S1: Collect the load step response data of the coal-fired boiler and input it into the improved physical information time-series neural network model to output the thermal inertia coefficient.

[0024] Based on the pre-arrangement of temperature measuring points at key temperature measuring points in the boiler furnace (such as burner temperature measuring points, furnace outlet, side walls and front and rear walls), the operating data of each temperature measuring point of the coal-fired boiler during the period of load step change are collected, that is, the load step response data of the coal-fired boiler.

[0025] It should be noted that the selection of temperature measurement points must comprehensively cover the key temperature measurement points in the coal-fired boiler furnace to accurately construct a spatial distribution mapping of thermal inertia. Specifically, temperature measurement points should be arranged at locations representing different heat exchange intensities, such as burner temperature measurement points, furnace outlet, side walls, and front and rear walls, ensuring spatial uniformity and representativeness in the height and width directions of the furnace. This allows for the capture of the dynamic characteristics of each local temperature measurement point throughout the entire thermodynamic process from fuel entry to flue gas exit. High-precision, high-response-speed T-type or S-type thermocouples are used for temperature acquisition. Their measuring ends must be securely installed inside the furnace wall or penetrate deep into the furnace through a sleeve to ensure direct contact with the flue gas or flame temperature measurement points. The data acquisition system must have high-frequency sampling capability (e.g., 1-10Hz) and be equipped with signal filtering functions to accurately record the rapid temperature changes during load steps, providing a reliable, low-hysteresis, and high signal-to-noise ratio raw data foundation for subsequent parameter identification of the first-order inertial plus pure time-delay model.

[0026] The load step change must meet the following conditions: the steady-state change of the load command is greater than 5% of the rated load, and the duration from the start to the end of the load command change is... The duration must be less than 60 seconds, and the average rate of change during this period must be greater than 1% / second of the rated load. Before and after a load change event, the boiler must have a sufficiently long period of stable operation to accurately calculate the excitation amplitude and response baseline. Specifically, this is defined as the period before the change... - , ] and after the change [ + + , + + + Within the time window, the standard deviation of the load command is less than 0.5% of the rated load, where The time window length representing the steady state is typically 300 seconds. The estimated time required to complete the main transition process is typically taken as 600 seconds. This represents the starting moment of the step change.

[0027] Calculate the values ​​before the step change respectively. - , [Time period and step change after] + + , + + + The arithmetic mean of u(t) over the time period is denoted as the coal feed rate before the step jump. Coal feed rate after step jump Let u(t) be the total coal feed rate at time t, and finally calculate the step coal feed rate change. , = - .

[0028] Calculate the steady-state window before a step change [ - , The average temperature of the i-th temperature measuring point within the range (The steady-state window of the i-th temperature measurement point before the step change [ - , The arithmetic mean of the temperatures collected at all times in the data is used to determine the relative temperature change sequence at that temperature measurement point. , , Let be the temperature value of the i-th temperature measuring point at time t, where t ranges from... To T, T is from The total time from the start to the re-entry into a new steady state. The sequence of relative changes at all temperature measurement points constitutes the step response data of the coal-fired boiler load.

[0029] An improved physical information temporal neural network model is constructed, which adopts a layered architecture including an input layer, a feature extraction layer, a temporal dependency layer, and a parameter regression layer. The input layer is a pair of temporal data for each temperature measurement point. , =[ ; ],in This represents the step change in coal feed rate. For the sequence of relative temperature changes at the i-th temperature measurement point, the scalar... with sequence The features at each time step are concatenated to form an input tensor of dimension (T,2). The feature extraction layer is implemented using a two-layer one-dimensional convolutional neural network. The first layer contains 64 convolutional kernels of size 5, and the second layer contains 128 convolutional kernels of size 3. Both layers use SamePadding with a stride of 1 to maintain the temporal length and use the ReLU activation function for nonlinear transformation. The temporal dependency layer consists of two layers of gated recurrent units. The first gated recurrent unit contains 64 hidden units, and the second gated recurrent unit contains 32 hidden units. A dropout rate of 0.2 is set between layers to prevent overfitting. This layer only outputs the feature representation of the last time step. The parameter regression layer contains two fully connected layers. The first fully connected layer has 16 neurons, and the second fully connected layer has 8 neurons. The final output layer contains 1 neuron that outputs the thermal inertia coefficient. The Softplus function is used as the activation function in the output layer to ensure that the thermal inertia coefficient satisfies the positive constraint.

[0030] The model training adopted a mini-batch gradient descent strategy with a batch size of 32, the initial learning rate was set to 0.001 with an exponential decay mechanism, the Adam optimizer was used for parameter updates, the training process lasted for 200 to 500 rounds, and an early stopping strategy was set to terminate training when the validation set loss did not improve for 20 consecutive rounds.

[0031] The total loss function of a physical information temporal neural network: ,in The total loss function of the physical information temporal neural network is... To fit the loss function to the data, This is the physical constraint loss function.

[0032] The data fitting loss function uses the mean squared error form to measure the difference between the thermal inertia coefficient predicted by the network and the reference value of the thermal inertia coefficient obtained through traditional parameter identification methods (such as a first-order inertia plus pure time delay model). The physical constraint loss function mainly penalizes predicted values ​​of the thermal inertia coefficient that violate physical laws, such as negative values ​​or values ​​that exceed the reasonable range.

[0033] The data fitting loss function is: ; in, Error loss value, Let be the predicted thermal inertia coefficient at the i-th temperature measurement point. Here is the reference value for the thermal inertia coefficient of the i-th temperature measuring point, and N is the number of temperature measuring points.

[0034] The physical constraint loss function is: ; in, , , , This is a hyperparameter, with a value range from 0.1 to 1. The preset maximum value of the thermal inertia coefficient is typically set to 600-1200 seconds, depending on the boiler characteristics. ) is the sign function. This represents the step change in coal feed rate. The pure time delay of the temperature response at the i-th temperature measuring point, where c is an empirical coefficient with a value range of 2 to 3.

[0035] Minimize using an optimizer (such as Adam) The network weights W and bias b are optimized on the training set. The final output... The thermal inertia coefficient of the i-th temperature measuring point is used to complete the spatial mapping of the furnace thermal inertia.

[0036] It should be noted that the first-order inertial plus pure time-delay model serves as the theoretical basis for describing the dynamic characteristics of each temperature measurement point i, and its mathematical expression is: ; in, The steady-state gain of the temperature response at the i-th temperature measurement point It is the pure time lag of the temperature response at the i-th temperature measuring point. This represents the step change in coal feed rate. The least squares method and other parameter estimation algorithms are applied to the response data at each temperature measuring point i. Curve fitting is performed to optimize and solve for a set of optimal model parameters. , , This minimizes the sum of squared errors between the model output and the measured data. The thermal inertia coefficient obtained through this process is the reference value for the thermal inertia coefficient in the data fitting loss function of the physical information temporal neural network model.

[0037] It should be noted that the improved physical information time-series neural network model, by embedding the traditional physical law of first-order inertia plus pure time delay model into the neural network architecture, overcomes the limitations of traditional curve fitting methods, such as sensitivity to initial values ​​and susceptibility to local optima, and also overcomes the lack of physical interpretability in purely data-driven models. By using the physical model fitting results as a supervision signal, the network is guided to accurately extract deep features of the thermal inertia coefficient from boiler operation data, laying a solid quantitative foundation for subsequent precise load regulation through zone coordination.

[0038] Step S2: Based on the thermal inertia coefficient, the theoretical boiler load response path is constructed using the quantile regression method.

[0039] Based on the furnace thermal inertia space mapping results obtained in step S1 (thermal inertia coefficients at each temperature measuring point) ), and the K-means adaptive clustering partitioning method is used to prioritize response, specifically: (1) Construct a dual feature vector: the thermal inertia coefficient Normalize to [0,1], and simultaneously normalize the steady-state gain coefficients of the first-order inertial plus pure time-delay model output. Normalized to [0,1], forming a double eigenvector. , (2) Determine the optimal number of clusters: Use the elbow rule to calculate the sum of squared errors of the number of clusters (ranging from 2 to 5), thus obtaining the optimal number of clusters P for the current working condition; (3) K-means clustering: With the goal of minimizing the intra-class variance and maximizing the inter-class variance, cluster the dual feature vectors of all temperature measurement points to obtain (4) Priority partitioning: [The text appears to be incomplete and contains several clusters.] The clusters are divided into 1 to P response priority partitions based on the average thermal inertia coefficient within the cluster, from smallest to largest, where p = 1, 2, 3, ... p is the response priority index. The smaller the average thermal inertia, the smaller the priority index p, and the earlier and faster the corresponding coal feeding and exhaust regulation starts.

[0040] From the boiler's historical deep peak-shaving optimal operating conditions, the boiler's stability characteristics are extracted: the average thermal inertia coefficient of each priority temperature measuring point. (p is the response priority index), global maximum thermal inertia coefficient Intensity of load variation , This is the rated load value. This represents the difference between the target load (the target load of the load adjustment command) and the current load. The corresponding tag data is: the proportion of activation time for each priority temperature measurement point. Target achievement time ratio Response time constant scaling factor Among them, the proportion of startup time The time when the temperature trajectory first deviates from the initial steady-state bandwidth (the fluctuation range of the temperature measuring point during stable operation of the boiler before load adjustment is ±1% to ±3% of the rated temperature), is obtained by normalizing it relative to the total adjustment time; target achievement time ratio. The earliest moment when the temperature trajectory enters and remains within the target steady-state bandwidth (the allowable fluctuation range where the temperature measuring point remains stable near the target value after load adjustment, such as ±1% to ±2% of the rated temperature) is normalized relative to the total adjustment time; the response time constant scaling factor is used. The temperature response trajectory under optimal operating conditions was obtained by least-squares fitting using a first-order inertial plus pure lag model.

[0041] The optimal operating condition for peak shaving in the boiler's historical data is that the load regulation rate reaches 1.5%-2.5% / minute of the rated load, the maximum deviation between the actual load value and the set value is less than ±1.5% of the rated load, the fluctuation of key parameters such as main steam temperature and pressure is within the allowable range (temperature ±5℃, pressure ±0.3MPa), and the boiler efficiency decrease does not exceed 1.5 percentage points of the benchmark value.

[0042] A multivariate quantile regression model was constructed using the quantile regression method: WD represents the boiler stability characteristics. To select the target quantile, prioritize =0.5 (median quantile), balancing robustness and central tendency representation ability. For target quantiles The regression coefficient matrix is ​​given below. In practical applications, the boiler stability characteristics calculated based on the current operating conditions are substituted into the pre-trained quantile regression model. The model directly outputs three adaptive core parameters: the start-up time ratio of the p-th response priority. The proportion of time required to reach the target of the p-th response priority Scaling factor for the response time constant of the p-th response priority The smaller the priority index p, the shorter the startup time ratio. The smaller the ratio of time to target achievement The smaller the scaling factor, the better. The smaller the value, the earlier the adjustment starts, the faster the load response is completed, and the faster the temperature / load response speed.

[0043] From the current load to target load Total transition time : ; in, The total transition time, This is the scaling factor, which defaults to 2. For a safe buffer period, a range of 60 to 120 seconds is preferred. The average thermal inertia coefficient for the p-th response priority.

[0044] Assigning different start-up times and response speeds to each priority temperature measurement point, the start-up time of the p-th priority temperature measurement point is... : = × Target completion time : = × Response time constant : = × .

[0045] It should be noted that by setting differentiated start-up times, response time constants, and target arrival times for priority regions, peak shifting and coordination in the load regulation process are achieved. Specifically, these parameters are adaptively determined through a quantile regression model, enabling the low thermal inertia region to start up immediately and respond quickly to rapidly provide the initial load change rate; the medium thermal inertia region to start up later and change at a uniform rate to receive and maintain the load change process; and the high thermal inertia region to start last and change slowly to smoothly transition and stabilize to the target load. This parameter system together constitutes a theoretically optimal load response path, which essentially guides the sequential and coordinated release of energy input in the furnace space and time dimensions, thereby fundamentally avoiding local shocks and temperature imbalances caused by synchronous regulation, and ultimately achieving the overall goal of rapid and stable load regulation.

[0046] For each temperature measurement point i, a reference temperature trajectory is generated according to its priority p: ; in, This represents the reference temperature value at time t for the i-th temperature measuring point. The steady-state temperature reference value of the i-th temperature measurement point at the start of the step. To correspond to the boiler target load The expected steady-state temperature value of the i-th temperature measurement point. The start time of temperature regulation Temperature response time constant.

[0047] It should be noted that this corresponds to the boiler target load. The expected steady-state temperature value of the i-th temperature measurement point This is not a preset constant, but is determined through a boiler steady-state thermodynamic calculation model or a load-temperature mapping relationship established based on historical operating data. Specifically, when the boiler is under load... During stable operation, a steady-state distribution exists at each temperature measuring point. When the boiler switches to the target load During stable operation, each measuring point will correspond to a new desired steady-state temperature value. This mapping relationship → First, a historical operating database needs to be collected, where each record represents a steady-state operating point, containing the stable load value and the corresponding stable temperature values ​​of all temperature measuring points within that period. For each temperature measuring point i, a nonlinear function is established using a support vector regression algorithm, such that any target load is input to the nonlinear function. This allows us to predict the expected steady-state temperature value of temperature measuring point i when operating stably under that load. .

[0048] Boiler overall load reference trajectory formula: ; in, Reference value for the overall boiler load at time t. The initial steady-state load of the boiler at the moment of step start. Boiler target load, This represents the overall load inertia time constant of the boiler.

[0049] It should be noted that the overall load inertia time constant of the boiler is... Based on historical load step data, a first-order inertial model is used for fitting to identify the load time constant corresponding to each step. The boiler's rated load is divided into several typical deep peak-shaving load intervals, for example, 30%~50%, 50%~70%, 70%~90%, and 90%~100% of the rated load. The arithmetic mean of the time constants identified in each load interval is then taken as the overall load inertial time constant of the boiler in the corresponding load interval. The nominal value is matched and called according to the current operating load range of the boiler in actual application. The advantage of this method is that its conclusions are directly derived from the unit's own historical operating data, and by using nominal values ​​for different load ranges, it avoids the problem of simple averaging masking the differences in dynamic characteristics of different load segments. It can better match the actual thermal inertia characteristics of the boiler in different load ranges, effectively avoiding systematic deviations in the theoretical load reference trajectory during deep adjustment of low or high load segments, and providing a more reliable key basis for generating accurate, smooth, and reasonable theoretical load response paths.

[0050] The theoretical boiler load response path is finally obtained. , =[ , ].

[0051] It should be noted that the overall boiler load reference trajectory It provides a globally optimal transition path from the current load to the target load. This path is determined by the overall inertial characteristics of the boiler. This decision ensured both the smoothness and speed of load changes. Reference temperature trajectories for each temperature measuring point. This is a localized combustion strategy designed to achieve this global load target. Different start-up times are assigned to temperature measurement points with different thermal inertia (based on the response priority of that temperature measurement point). ) and response speed ( This guides the decoupling and coordinated delivery of fuel and air in the furnace space and time dimensions. Theoretical boiler load response path. An ideal path is defined: when the temperature at each measuring point strictly follows its differentiated trajectory. When changes occur, the overall load of the boiler will naturally track the global target trajectory. This enables rapid and stable load regulation while avoiding temperature imbalance caused by differences in local dynamic characteristics.

[0052] Step S3: Based on the theoretical boiler load response path, the load response of the coal-fired boiler is optimized using an improved particle swarm optimization algorithm to obtain the coal feeding and exhaust parameter sequence of the coal-fired boiler.

[0053] This step uses the graded load response path generated in step S2. To achieve rigid tracking of the target, an improved particle swarm optimization algorithm is used to solve a sequence of coal feeding and exhaust parameters in a short time domain H. This drives the actual state of the boiler to accurately conform to the preset reference trajectory, realizing closed-loop, rolling, and adaptive tracking of the planned path in step S2.

[0054] Define the future H time domain (the value of the time domain needs to strike a balance between control performance and computational efficiency, and its corresponding physical duration should cover the main dynamic process of the boiler load step response, which is consistent with the overall load inertia time constant calculated in step S2). Closely related, based on the step response of a first-order system, it takes approximately [time period] to reach steady state. Given that the predicted time domain typically covers 60% to 80% of the total duration (4 to 5 times longer), the physical duration of H is approximately 2.4 to 4.0 times longer. Therefore, the physical duration corresponding to H can be empirically taken as... Three to five times, taking three to five times is a slightly conservative engineering experience value that includes this range, so when If the interval is 300 seconds, then H (which could be on the order of 900 to 1500 seconds) is the sequence of coal feeding and exhaust parameters to be optimized. Among them, control increment u(t+ |t)=u(t+ ∣t)-u(t+ -1∣t), u(t+ |t) represents the prediction of the future t+ at the current time t. The control increment at time step, u(t+ |t) represents the prediction of the future t+ at time t. Control vector at time, Indexing future moments.

[0055] Control vector u(t+ |t)=[ (t+ |t), (t+ ∣t)], where (t+ |t) represents the j-th burner in the future t+ Coal feed rate at any given time (t+ |t) represents the j-th burner in the future t+ The air distribution volume at any given time, with a dimension of 2J×1, includes the coal feed and air distribution volume commands for all J-layer burners, therefore The dimension is D = 2J × H.

[0056] Control vector magnitude constraints: , To minimize the control vector, Maximum value of the control vector; control increment constraint , To control the minimum increment, To control the maximum increment. Wherein, =[ , ], =[ , ], and These are the minimum and maximum values ​​of the coal feed increment, respectively. and These are the minimum and maximum values ​​for the exhaust gas increment, respectively. The coal feed increment is ±2 to ±5 tons / hour, and the exhaust gas increment is ±5% to ±15% of the air volume command. The specific values ​​are determined by the characteristics of the coal-fired boiler. At the same time, the following safety constraints must be met: the main steam temperature fluctuation does not exceed ±5℃ of the rated value, the furnace negative pressure is maintained within the range of -100Pa to +200Pa, and the oxygen content at the furnace outlet is controlled within 3% to 6%.

[0057] The position vector of each particle b in the improved particle swarm optimization algorithm Directly represents a candidate coal feeding and exhaust parameter sequence ,Right now = The particle flies within a D-dimensional search space. Within the feasible region (satisfying the control increment constraint)... ≤Δu≤ Randomly initialize the positions of M particles (population size) and speed .

[0058] The core of improving the particle swarm optimization algorithm is to iteratively update the position and velocity of the particles. The position vector of each particle b at the g-th iteration... This represents a candidate sequence of coal feeding and exhaust parameters. The update formulas for particle position and velocity are: ; ; in, The inertial weights are dynamically adjusted during iteration. , and Take values ​​of 0.88 and 0.18 respectively. The maximum number of iterations, and As a learning factor, As the number of iterations decreases linearly from the initial value of 2.0 to 0.5, The value increases linearly from 0.5 to 2.0 with the number of iterations, and rand1() and rand2() generate uniformly distributed random numbers in the range [0,1]. This represents the historical optimal position of particle b. This represents the globally optimal position for all particles.

[0059] The fitness value of each particle is calculated by the following function, where a smaller value indicates a better solution: ; Where F is the fitness function, Indexing future moments For load weighting, Temperature weighting, + =1, in one embodiment and satisfy + =1, preferably 0.5 and 0.5 respectively; when load tracking is emphasized more, it can be appropriately increased. When temperature stability is emphasized, the temperature can be appropriately increased. When high temperature response requirements are needed, it can be improved. , express The predicted load value, express The load value of the theoretical boiler load response path at any given time. express The predicted temperature value of the i-th temperature measuring point. express Temperature reference value for the theoretical boiler load response path at any given time.

[0060] It should be noted that the above and The values ​​can be calculated using a first-order inertial plus pure time delay model: First, the actual measured value at the current time t is used as the initial state. Then, the coal feeding and exhaust parameters in the candidate control sequence are used as the model input excitation. The load prediction adopts the overall first-order inertial model of the boiler, with the total coal feed change as the input and the load change as the output. The prediction of each temperature measuring point adopts the distributed first-order inertial plus pure time delay model identified in step S1, with the coal feed change and air distribution change of the combustion layer corresponding to the temperature measuring point and the adjacent combustion layer (the upper combustion layer and the lower combustion layer) as the input and the local temperature as the output. This adapts to the actual combustion characteristics of the furnace temperature being strongly affected by the multi-layer coal feeding and air distribution. By substituting the coal feeding and exhaust parameter sequence in the future time domain into these parallel local models, the predicted trajectory of the load and temperature at each point in the entire prediction time domain can be gradually obtained, thus providing a basis for fitness evaluation for particle swarm optimization.

[0061] During algorithm initialization, the positions and velocities of M particles are randomly generated (M is typically between 50 and 100), and assigned values ​​within constraints. In each iteration, the fitness of each particle is calculated and updated. and The termination condition is... The iteration is complete or the fitness change is less than the threshold ε. Final output. The globally optimal solution is the optimal sequence of coal feeding and ventilation parameters.

[0062] The obtained optimal coal feeding and exhaust parameter sequence in the H time domain is used to heat and burn coal-fired boilers.

[0063] The significance of employing the improved particle swarm optimization algorithm in step S3 lies in its ability to efficiently solve the complex, high-dimensional, multi-constraint optimization problem of boiler load response. This algorithm encodes the coal feed and air distribution parameter sequences of each burner into particle positions, simulating parallel search in the solution space through swarm intelligence. Its dynamically adjusted inertia weights and learning factors effectively balance global exploration and local development capabilities, thereby quickly finding the optimal operating command that allows the boiler's actual state to accurately track the theoretical path planned in S2. This closed-loop strategy based on rolling optimization endows the system with strong real-time adaptive capabilities, effectively compensating for uncertainties such as coal quality fluctuations. Ultimately, under the premise of ensuring safety constraints, it achieves precise and efficient coordinated regulation of fuel and air distribution.

[0064] Step S4: Heating and combustion of the coal-fired boiler according to the coal feeding and exhaust parameter sequence, and collecting the actual boiler load response path of the coal-fired boiler. The actual boiler load response path is compared with the theoretical boiler load response path to obtain the load path deviation. The coal feeding and exhaust parameter sequence is adjusted based on the load path deviation to obtain the corrected coal feeding and exhaust parameter sequence, thereby realizing the load regulation of the coal-fired boiler.

[0065] This step is the core of the entire control method, ensuring its safety and adaptability. During the execution of the preset optimal graded load response path in step S3, the matching degree between the actual thermal state of the boiler and the theoretical path is monitored in real time. The system intelligently diagnoses path deviations caused by uncertainties such as coal quality fluctuations, environmental changes, and equipment characteristic drift, and dynamically triggers targeted correction strategies. This provides online guidance and replanning for the optimization process in S3, ensuring that the load change process always proceeds safely and smoothly around the predetermined thermal inertia decoupling objective. This significantly improves the robustness and practicality of the entire control method under complex operating conditions.

[0066] The actual boiler load response path is acquired in real time using a high-frequency sampling method, mainly including: the measured temperature values ​​at each temperature measuring point. Actual load value It then calculates and compares the result with the setpoint corresponding to the theoretical boiler load response path generated in step S2: ; ; in, This represents the load deviation at time t. This represents the temperature deviation at the i-th temperature measurement point. This represents the actual load at time t. This represents the actual temperature at the i-th temperature measurement point at time t; finally, the load path deviation is obtained, which includes the load deviation. and temperature deviation .

[0067] The coal feeding and exhaust parameters are corrected based on the load path deviation to obtain the corrected coal feeding and exhaust parameters: ; ; in, This represents the corrected coal feed rate for the j-th burner at time t. This represents the corrected exhaust volume of the j-th burner at time t. This represents the theoretical coal feed rate of the j-th burner at time t. The theoretical exhaust volume of the j-th burner at time t is Let be the change in coal feed for the j-th burner at time t. The change in exhaust air volume corrected for the j-th burner at time t.

[0068] and The calculation formula is: ; ; in, and These are the proportional gain and integral gain for coal feed rate control. and These are the proportional gain and integral gain for exhaust volume control. The proportional gain is 0.1~1.0, and the integral gain is 0.01~0.1, but the specific values ​​depend on the dynamic characteristics of the boiler.

[0069] It should be noted that, , , and The specific values ​​are not fixed constants, but need to be tuned on-site according to the dynamic characteristics of the specific boiler. Typically, a trial-and-error method based on historical operating data or engineering tuning methods such as the Ziegler-Nichols approach are used: First, a step disturbance is applied within the boiler's safe operating range, and the system response curve is observed to initially determine the critical gain and oscillation period; then, adjustments are made gradually from smaller values ​​(e.g., proportional gain of 0.3-0.5 and integral gain of 0.03-0.05), prioritizing the tuning of the proportional gain to ensure response speed, followed by fine-tuning of the integral gain to eliminate steady-state errors, ultimately achieving the optimal balance between speed, stability, and robustness in load and temperature tracking. The entire tuning process must follow the "proportional first, integral second" principle and verify parameter adaptability under actual peak-shaving conditions. If necessary, online fine-tuning should be performed based on factors such as coal quality changes and load ranges.

[0070] After the parameter correction calculation is completed, the corrected coal feeding and exhaust parameter sequence will be sent to the underlying actuators through the boiler's distributed control system or programmable logic controller.

[0071] Specifically, a dual-loop control is adopted to achieve real-time adjustment: the outer loop is PSO (Particle Swarm Optimization) rolling re-optimization, and the inner loop is PI (Proportional-Integral) feedback correction. The two work together according to the following logic, with the following frequency allocation: the PSO rolling re-optimization cycle is relatively long, preferably updating the optimal candidate control sequence once every 30 seconds, for periodically updating the global optimal control benchmark; the PI correction cycle is synchronized with the high-frequency sampling interval of the control system, preferably executed once every 1 second, to achieve real-time disturbance suppression.

[0072] Within each PSO cycle, the algorithm outputs the optimal coal feeding and ventilation parameter sequence for the future period, serving as the baseline trajectory for PI correction. In each subsequent PI control cycle, the controller combines the deviation between the current actual operating data and the theoretical path to perform real-time corrections on this baseline trajectory, generating the final actual control commands.

[0073] Within each control cycle, the corrected coal feed rate and corrected exhaust rate corresponding to the current time t are converted into actual control commands, driving the corresponding coal feeder and damper actuators to achieve real-time adjustment of the combustion rate. After correcting the coal feed rate and exhaust rate at time t, the newly acquired actual operating data at time t+1 is calculated (including the measured temperature values ​​of each temperature measuring point at time t+1). Actual load value ) and theoretical path parameters (theoretical temperature values ​​at each temperature measuring point) Theoretical load value The deviation is calculated, and the coal feeding and exhaust parameters at time t+1 are corrected based on this deviation (the calculation method for the correction is the same as the correction method for the coal feeding and exhaust at time t). The above steps are repeated to continuously update the coal feeding and exhaust parameter sequence for the remaining time domain (from t+2 to t+H-1), thereby ensuring that the actual load dynamic process of the boiler always closely follows the theoretical optimal path.

[0074] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, the phrase "comprising an element defined as..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0075] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for rapid load coordination regulation of coal-fired boilers for deep peak shaving, characterized in that: Includes the following steps: Step S1: Collect the load step response data of the coal-fired boiler and input it into the improved physical information time-series neural network model to output the thermal inertia coefficient; Step S2: Based on the thermal inertia coefficient, construct the theoretical boiler load response path using the quantile regression method; Step S3: Based on the theoretical boiler load response path, the load response of the coal-fired boiler is optimized using an improved particle swarm optimization algorithm to obtain the coal feeding and exhaust parameter sequence of the coal-fired boiler. Step S4: Heating and combustion of the coal-fired boiler according to the coal feeding and exhaust parameter sequence, and collecting the actual boiler load response path of the coal-fired boiler. The actual boiler load response path is compared with the theoretical boiler load response path to obtain the load path deviation. The coal feeding and exhaust parameter sequence is adjusted based on the load path deviation to obtain the corrected coal feeding and exhaust parameter sequence, thereby realizing the load regulation of the coal-fired boiler.

2. The rapid load coordination regulation method for coal-fired boilers oriented towards deep peak shaving as described in claim 1, characterized in that: The collection of load step response data from coal-fired boilers includes the following specific steps: Collect operating data of a coal-fired boiler at preset temperature measurement points during periods of a step change in load, i.e., before the step change in load. - , After step change + + , + + + Within the time window, the standard deviation of the load command is less than 0.5% of the rated load. This represents the length of the steady-state time window. The time required to complete the main transition process, This represents the starting moment of the step change. Indicates the duration from the start to the end of the load command; Calculate the values ​​before the step change respectively. - , [Time period and step change after] + + , + + + The arithmetic mean of u(t) over the time period is used to obtain the coal feed rate before the step jump. Coal feed rate after step jump Let u(t) be the total coal feed rate at time t, and finally calculate the step coal feed rate change. , = - ; Calculate the steady-state window before a step change [ - , The average temperature of the i-th temperature measuring point within the range Then the temperature relative change sequence at that temperature measuring point , , Let be the temperature value of the i-th temperature measuring point at time t; the sequence of relative changes of all temperature measuring points is the step response data of the coal-fired boiler load.

3. The rapid load coordination regulation method for coal-fired boilers oriented towards deep peak shaving, as described in claim 2, is characterized in that: The improved physical information temporal neural network model includes the following steps: An improved physical information temporal neural network model is constructed, which includes an input layer, a feature extraction layer, a temporal dependency layer, a parameter regression layer, and an output layer. The input layer of the physical information temporal neural network model is a pair of time-series data for each temperature measurement point. , =[ ; ],in This represents the step change in coal feed rate. Let be the sequence of relative temperature changes at the i-th temperature measurement point; The feature extraction layer is implemented using a two-layer one-dimensional convolutional neural network. The first layer contains 64 convolutional kernels of size 5, and the second layer contains 128 convolutional kernels of size 3. Both layers use Same Padding with a stride of 1 to maintain the temporal length and use the ReLU activation function for nonlinear transformation. The temporal dependency layer consists of two layers of gated recurrent units. The first gated recurrent unit contains 64 hidden units, and the second gated recurrent unit contains 32 hidden units. A dropout rate of 0.2 is set between layers to prevent overfitting. The parameter regression layer contains two fully connected layers. The first fully connected layer has 16 neurons, and the second fully connected layer has 8 neurons. The final output layer contains 1 neuron and outputs the thermal inertia coefficient.

4. The rapid load coordination regulation method for coal-fired boilers oriented towards deep peak shaving, as described in claim 3, is characterized in that: The output yields the thermal inertia coefficient, including the following steps: The total loss function of the physical information temporal neural network model is , ,in, To fit the loss function to the data, The physical constraint loss function; Minimize the total loss function using the optimizer. The network weights and biases of the physical information temporal neural network model are optimized, and the thermal inertia coefficient of the i-th temperature measurement point is finally output. .

5. A method for rapid load coordination regulation of coal-fired boilers for deep peak shaving according to claim 4, characterized in that: The method of constructing a theoretical boiler load response path based on the thermal inertia coefficient and using quantile regression includes the following steps: The response priorities of temperature measuring points are determined based on their thermal inertia coefficients; the activation time proportion of the p-th response priority is calculated using quantile regression. The proportion of time required to reach the target of the p-th response priority Scaling factor for the response time constant of the p-th response priority ; Startup time ratio based on the priority of the p-th response The proportion of time required to reach the target of the p-th response priority Scaling factor for the response time constant of the p-th response priority Calculate the temperature reference value and the overall load reference value at the temperature measuring point; The theoretical boiler load response path is finally obtained. , =[ , ],in, This represents the reference temperature value at time t for the i-th temperature measuring point. The reference value for the overall boiler load at time t.

6. A method for rapid load coordination regulation of coal-fired boilers for deep peak shaving according to claim 5, characterized in that: The method of prioritizing the response of temperature measuring points based on the thermal inertia coefficient of each temperature measuring point includes the following steps: Based on the thermal inertia coefficient and steady-state gain coefficient of each temperature measuring point, a dual feature vector is constructed. The dual feature vector is then clustered using the K-means clustering method to obtain P clusters. Based on the P clusters, the temperature measuring points are divided into P response priorities.

7. A method for rapid load coordination regulation of coal-fired boilers for deep peak shaving according to claim 6, characterized in that: The calculation of the temperature reference value and the overall load reference value at the temperature measurement point includes the following specific steps: For each temperature measurement point, a reference temperature trajectory is generated according to its response priority: ； in, This represents the reference temperature value at time t for the i-th temperature measuring point. The steady-state temperature reference value of the i-th temperature measurement point at the start of the step. To correspond to the boiler target load The expected steady-state temperature value of the i-th temperature measurement point. The start time of temperature regulation Temperature response time constant, where t is the time index; Calculate the reference value for the overall boiler load: ； in, Reference value for the overall boiler load at time t. The initial steady-state load of the boiler at the moment of step start. Boiler target load, This represents the overall load inertia time constant of the boiler.

8. A method for rapid load coordination regulation of coal-fired boilers for deep peak shaving according to claim 7, characterized in that: The improved particle swarm optimization algorithm includes the following specific steps: Construct an improved fitness function for the particle swarm optimization algorithm; the update formulas for the position and velocity of each particle in the improved particle swarm optimization algorithm are as follows: ； ； in, This represents the velocity of the b-th particle in the (g+1)-th iteration. This represents the velocity of the b-th particle in the g-th iteration. This represents the position of the b-th particle in the (g+1)-th iteration. This represents the position of the b-th particle in the g-th iteration. For inertial weights, The maximum number of iterations, and The learning factor is rand1() and rand2(), which are uniformly distributed random numbers in the range [0,1]. This represents the historical best position of the b-th particle. This represents the globally optimal position for all particles.

9. A method for rapid load coordination regulation of coal-fired boilers for deep peak shaving according to claim 8, characterized in that: The construction of the fitness function for the improved particle swarm optimization algorithm includes the following specific steps: ； Where F is the fitness function value, and H is the future time domain. Indexing future moments For load weighting, Temperature weighting, express The predicted load value, express The load value of the theoretical boiler load response path at any given time. express The predicted temperature value of the i-th temperature measuring point. express Temperature reference value for the theoretical boiler load response path at any given time.

10. A method for rapid load coordination regulation of coal-fired boilers for deep peak shaving according to claim 9, characterized in that: The specific steps for comparing the actual boiler load response path with the theoretical boiler load response path to obtain the load path deviation are as follows: ； ； in, This represents the load deviation at time t. This represents the temperature deviation at the i-th temperature measurement point. This represents the actual load at time t. This represents the actual temperature of the i-th temperature measurement point at time t; The final result is the load path deviation, which includes load deviation and temperature deviation.

Images