A deep learning-based automatic control method and system for coal-fired power generation
Patent Information
- Application Number
- CN202611009438.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-08
- Publication Date
- 2026-09-15
Smart Images
Figure CN122755480A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of automatic control technology for coal-fired power generation, and specifically discloses an automatic control method and system for coal-fired power generation based on deep learning. Background Technology
[0002] Against the backdrop of dual carbon targets, the development of new energy sources is gaining momentum, while traditional coal-fired power generating units face transformation challenges. In the process of reducing coal consumption and carbon emissions, coal-fired power generation is gradually shifting from the main source of electricity supply to a supporting and regulating power source. Improving the operational flexibility and energy efficiency of units has become the most pressing technical requirement in the field of coal-fired power generation.
[0003] Coal-fired power generating units undertake peak shaving and frequency regulation tasks and operate under frequent load changes over long periods. The amount of fuel required by the unit changes rapidly and with large fluctuations. However, due to its large delay and inertia, the pulverizing system has a long response time to changes in pulverized coal output, making it difficult to meet the demand for rapid changes in pulverizing output when the unit changes load quickly. This severely limits the unit's ability to change load quickly and restricts its operational flexibility.
[0004] Currently, the load change rate of supercritical coal-fired power generating units in my country is typically 1.0%-2.0% of rated load per minute, which cannot meet the requirements of new power systems for the rapid load change capability of coal-fired power generating units. Due to the processes of coal pulverization and coal conveying, heat release from fuel combustion and heat absorption of the working fluid in the boiler, and heat storage in the metal components of the boiler's heating surface, the thermal inertia and hysteresis of the boiler result in a load change response rate of more than 10 minutes.
[0005] Therefore, existing coal-fired power generation control suffers from problems such as load response lag and significant delays in the pulverizing system. Summary of the Invention
[0006] The technical problem to be solved by this invention is the lag in load response and the large delay in the pulverizing system of existing coal-fired power generation control.
[0007] The technical solution of the present invention to solve the above-mentioned technical problems is as follows: A deep learning-based automatic control method for coal-fired power generation includes: Real-time acquisition of multi-dimensional operational data from coal-fired power generating units; After preprocessing the operational data, principal component features are extracted using principal component analysis. Based on these features, the predicted load sequence within a preset time window is obtained by improving the TPA-LSTM model. Based on the predicted load sequence, the low-frequency change trend and short-term fluctuation components are obtained through time series decomposition. Based on the low-frequency variation trend, the optimal number of coal mills to be put into operation and the coal mill output strategy are generated to adjust the operating status of the pulverizing system. Based on short-term fluctuation components, adjust the powder storage depth and powder feeding rate of the small powder hopper in advance; The coal feed from the pulverizer and small pulverizer bins is combined and transported to the boiler furnace, coordinating the coal-fired power generation units to participate in primary frequency regulation and AGC response.
[0008] Furthermore, after preprocessing the operational data, principal component features are extracted using principal component analysis, including: Principal component analysis is used to calculate the covariance matrix based on the preprocessed running data, solve for the eigenvalues and eigenvectors, and select the first preset number of eigenvectors whose cumulative contribution rate exceeds a set percentage as principal components to obtain the principal component features.
[0009] Furthermore, based on the predicted load sequence, the low-frequency variation trend and short-term fluctuation components are obtained through time series decomposition, including: The low-frequency variation trend is extracted from the predicted load sequence by passing it through a low-pass filter, and the short-term fluctuation component is obtained by subtracting the low-frequency variation trend from the predicted load sequence.
[0010] Furthermore, based on low-frequency variation trends, an optimal strategy for the number of coal mills in operation and their output is generated to adjust the operating status of the pulverizing system, including: The variable load rate is calculated based on the low-frequency change trend. If the variable load rate exceeds the preset overshoot threshold, the overdrive mode of the coal mill is activated to generate the optimal number of coal mills in operation and the coal mill output strategy. The number of coal mills in operation is increased to the optimal number of coal mills in operation, and the coal mill output is increased based on the coal mill output strategy. If the variable load rate does not exceed the preset overshoot threshold, the pulverizer override mode will not be activated. The objective function is to minimize the energy consumption of the pulverizing system, and the upper and lower limits of pulverizer output and the number of pulverizers in operation are used as constraints to generate the optimal number of pulverizers in operation and the pulverizer output strategy. Based on the pulverizer output strategy, the expected output of the pulverizer is obtained, and the actual output of the pulverizer is acquired. According to the deviation between the actual output of the pulverizer and the expected output of the pulverizer, the speed of the feeder and the air volume of the primary air fan are adjusted.
[0011] Furthermore, based on short-term fluctuation components, the powder storage depth and powder dispensing rate of the small powder hopper are adjusted in advance, including: Obtain the current maximum output of the coal mill, the current actual output of the coal mill, and the current real-time output of the coal mill; and obtain the peak value of the short-term fluctuation component based on the short-term fluctuation component. Calculate the fuel shortage based on the current maximum output of the coal mill, the current actual output of the coal mill, and the peak value of the short-term fluctuation component. If the fuel shortage is greater than zero, the powder storage control unit will be activated to increase the powder storage depth of the small powder bin in advance, so that the powder storage amount is at least a set multiple of the fuel shortage. If the fuel shortage is less than or equal to zero, the pulverized coal feeding control unit is activated. The pulverized coal feeding signal of the small pulverized coal bin is determined based on the difference between the peak value of the short-term fluctuation component and the real-time output of the current coal mill. The pulverized coal feeding signal of the small pulverized coal bin is corrected by multiple factors to obtain the pulverized coal feeding rate command of the small pulverized coal bin. The pulverized coal feeding rate of the small pulverized coal bin is adjusted based on the pulverized coal feeding rate command of the small pulverized coal bin.
[0012] Furthermore, the formula for calculating the powder feeding rate command of the small powder hopper is as follows: , , in, This is the powder dispensing rate instruction for the small powder container. This is a correction factor for small powder bin burners. This is the correction factor for the condition of the powder feeder. This is a correction factor for boiler combustion conditions. To provide an advance investment coefficient, This is a signal to send fans to the small fan warehouse. This is the time difference between the current time and the predicted start time of the load change. The time constant required for the powder feeding system of the small powder hopper to reach a stable powder supply from startup. This is an adjustable coefficient.
[0013] Furthermore, it also includes steps for incrementally training or retraining the improved TPA-LSTM model using periodically collected runtime data, including: The regularly collected operational data is divided into training set, validation set and test set according to a preset ratio; Based on the training set, training samples are constructed using a sliding window method; The prediction error of the improved TPA-LSTM model under the current working conditions is evaluated using the sliding window residual analysis method. When the prediction error exceeds the preset error threshold, retraining is automatically triggered, and the improved TPA-LSTM model is retrained using training samples. When the prediction error does not exceed the preset error threshold, incremental training is automatically triggered, and the improved TPA-LSTM model is incrementally trained using training samples. The improved TPA-LSTM model was validated using a validation set; The validated improved TPA-LSTM model was tested using a test set.
[0014] This invention also relates to a deep learning-based automatic control system for coal-fired power generation, used in the aforementioned deep learning-based automatic control method for coal-fired power generation, comprising: The data acquisition module is used to collect multi-dimensional operating data of coal-fired power generating units in real time; The deep learning load prediction module is used to preprocess the running data and extract principal component features using principal component analysis. Based on the principal component features, an improved TPA-LSTM model is used to obtain the predicted load sequence within a preset time window. The deep learning load prediction module includes a PCA unit and an improved TPA-LSTM model. The predicted load sequence decomposition module is used to obtain the low-frequency variation trend and short-term fluctuation components based on the predicted load sequence through time series decomposition methods. The coal mill pulverizing coordination control module is used to generate the optimal number of coal mills in operation and the expected output of the coal mills based on low-frequency variation trends, and to adjust the operating status of the pulverizing system. The powder storage control module for the small powder silo is used to adjust the powder storage depth and powder feeding rate of the small powder silo in advance based on short-term fluctuation components; the powder storage control module includes a powder storage control unit and a powder feeding control unit; The coordination module is used to superimpose the coal mill and small coal silo supply data and deliver them to the boiler furnace, and coordinate the coal-fired power generation units to participate in primary frequency regulation and AGC response. The online model update module is used to incrementally train or retrain the improved TPA-LSTM model using the latest running data collected periodically.
[0015] Furthermore, the improved TPA-LSTM model includes: an input layer, an LSTM layer, a temporal pattern attention layer, and a fully connected output layer connected in sequence; The input layer receives principal component features, the LSTM layer captures long-term dependencies in the load time series, the time-series pattern attention layer automatically calculates the attention weights for each historical time step and performs a weighted summation of the hidden states in key time periods to generate a context vector, and the fully connected output layer maps the output of the time-series pattern attention layer to multi-step load forecast values to generate a predicted load sequence.
[0016] Furthermore, the powder control module for the small powder silo also includes a powder level safety protection unit, which is used to monitor the powder level height and powder silo temperature in real time. When the powder level height is higher than the maximum warning value or lower than the minimum warning value, or when the powder silo temperature exceeds the safety threshold, an alarm signal is issued and the speed of the powder feeder is limited.
[0017] The beneficial effects of this invention are: This invention, based on real-time acquired operational data, obtains high-precision predicted load by improving the deep learning capabilities of the TPA-LSTM model. It decomposes the predicted load into low-frequency variation trends and short-term fluctuation components using a time-series decomposition method. Based on the low-frequency variation trends, it generates the optimal number of coal mills to be put into operation and the optimal coal mill output strategy, adjusting the operating status of the pulverizing system. Based on the short-term fluctuation components, it adjusts the powder storage depth and feed rate of the small powder silo in advance. The obtained powder supply from the coal mills and small powder silos is superimposed and delivered to the boiler furnace, coordinating the coal-fired power generation units to participate in primary frequency regulation and AGC response. This significantly improves the variable load response rate, greatly reduces the delay and energy consumption of the pulverizing system, and deeply integrates load prediction with pulverizing and powder storage control, achieving a fundamental shift from passive response-based deviation compensation to active predictive feedforward control. It also enhances control stability and is applicable to multiple platforms. Attached Figure Description
[0018] Figure 1 This is a flowchart of an automatic control method for coal-fired power generation based on deep learning, as described in an embodiment of the present invention. Figure 2 This is a block diagram of an automatic control system for coal-fired power generation based on deep learning, as described in an embodiment of the present invention. Detailed Implementation
[0019] The principles and features of the present invention are described below. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.
[0020] A pulverizing system is a combination of all equipment and pipelines that dry and grind raw coal into pulverized coal that meets the requirements for combustion and then transport it to the boiler burner. A pulverizing system typically includes: coal feeder, coal mill, coarse powder separator, primary air fan, pulverized coal pipeline, burner, fine powder separator, small powder silo, pulverizer, etc.
[0021] A deep learning-based automatic control method for coal-fired power generation, such as... Figure 1 As shown, it includes the following steps: S1. Real-time acquisition of multi-dimensional operating data of coal-fired power generating units; Multi-dimensional operating data of coal-fired power generating units are collected in real time from the distributed control system (DCS). The operating data includes unit load, main steam pressure, boiler oxygen content, AGC load command and peak shaving command in the power grid dispatch command data, coal quality parameters and ambient temperature, etc. The coal quality parameters include volatile matter, calorific value and ash content.
[0022] In the power grid dispatch instruction data, the AGC load instruction is an active power target value instruction automatically issued by the power grid dispatch center (EMS system) to the grid-connected power plant (or energy storage station). It is used to adjust the output of the AGC unit in real time to maintain the system frequency and tie-line exchange power according to plan.
[0023] Incremental training or retraining of the improved TPA-LSTM model is performed using regularly collected runtime data, specifically including: S11. Divide the latest operational data collected periodically into training set, validation set and test set according to the preset ratio; S12. Based on the training set, construct training samples using a sliding window method; A fixed-size window is used to slide sequentially across the ordered sequences (such as time series or text tokens) in the training set. The data within each window is used as the input feature and the data immediately following it is used as the target label, generating a series of overlapping input-label pairs to improve the training of the TPA-LSTM model.
[0024] S13. Use the sliding window residual analysis method to evaluate the prediction error of the improved TPA-LSTM model under the current working conditions. When the prediction error exceeds the preset error threshold, retraining is automatically triggered, and the improved TPA-LSTM model is retrained using training samples. When the prediction error does not exceed the preset error threshold, incremental training is automatically triggered, and the improved TPA-LSTM model is incrementally trained using training samples. Retraining involves discarding old weights and training from scratch with the full dataset. It is costly and time-consuming, but it can achieve architectural leaps and qualitative changes in capabilities, making it suitable for major version upgrades or dramatic changes in data distribution.
[0025] Incremental training retains the original model parameters and fine-tunes them using only new data. It is low-cost and fast, and suitable for small-scale data updates or vertical capability enhancement. The improved TPA-LSTM model uses mean squared error (MSE) as the loss function and the Adam optimizer is used for optimization training.
[0026] S14. Validate the trained improved TPA-LSTM model using the validation set; Adjust the learning rate during the validation of the improved TPA-LSTM model.
[0027] S15. Test the validated improved TPA-LSTM model using the test set.
[0028] The performance metrics of the improved TPA-LSTM model were tested and validated using a test set.
[0029] S2. After preprocessing the operational data, principal component features are extracted using principal component analysis. Based on the principal component features, the predicted load sequence within the preset time window is obtained by improving the TPA-LSTM model. Preprocessing of the running data includes: filling missing values with cubic spline interpolation; removing outliers after detection using the 3σ criterion; and normalizing all running data to the [0,1] interval using the maximum-minimum normalization method.
[0030] Extracting principal component features using principal component analysis involves: calculating the covariance matrix based on normalized operating data, solving for eigenvalues and eigenvectors, and selecting a preset number of eigenvectors with a cumulative contribution rate exceeding a set percentage as principal components. The set percentage can be 90%, and the preset number can be 5 (i.e., obtaining the cumulative contribution rate based on the eigenvalues, and selecting the top 5 eigenvectors with a cumulative contribution rate exceeding 90% as principal components), thus obtaining the principal component features.
[0031] The principal component features are input into the improved TPA-LSTM model to perform multi-step prediction of the load demand of coal-fired power generating units within a preset time window, and output the predicted load sequence F_pred within the preset time window.
[0032] S3. Based on the predicted load sequence, the low-frequency change trend and short-term fluctuation components are obtained through time series decomposition. The time series decomposition method is used to decompose the predicted load sequence into low-frequency variation trend and short-term fluctuation component from the 1st to the mth minute in the future. This includes: extracting the low-frequency variation trend F_low from the predicted load sequence through a low-pass filter (a first-order low-pass digital filter), and then subtracting the low-frequency variation trend from the predicted load sequence to obtain the short-term fluctuation component F_high; setting the time window width n of the low-frequency variation trend, that is, taking the moving average of the low-frequency variation trend from the 1st to the nth minute, and using the short-term fluctuation component of the nm minute as the fast compensation target.
[0033] The low-frequency trend is transmitted to the coal mill pulverizing coordination control module to adjust the base load, and the short-term fluctuation component is transmitted to the small powder storage control module to respond to rapid load fluctuations.
[0034] S4. Based on the low-frequency change trend, generate the optimal number of coal mills to be put into operation and the coal mill output strategy, and adjust the operating status of the pulverizing system. The load change rate is calculated based on the low-frequency change trend. If the load change rate exceeds the preset overshoot threshold, the overdrive mode of the coal mill is activated to generate the optimal number of coal mills in operation and the coal mill output strategy. The number of coal mills in operation is increased to the optimal number of coal mills in operation, and the output of a single coal mill is increased to above the rated value so that the coal mill output meets the coal mill output strategy. If the variable load rate does not exceed the preset overshoot threshold, the pulverizer override mode will not be activated. The objective function is to minimize the energy consumption of the pulverizing system, and the upper and lower limits of pulverizer output and the number of pulverizers in operation are used as constraints to generate the optimal number of pulverizers in operation and the pulverizer output strategy. Based on the pulverizer output strategy, the expected output of the pulverizer is obtained, and the actual output of the pulverizer is acquired. According to the deviation between the actual output of the pulverizer and the expected output of the pulverizer, the speed of the feeder and the air volume of the primary air fan are adjusted to achieve dynamic tracking of the pulverizer output, thereby achieving basic load tracking.
[0035] Based on high-precision load prediction, this invention optimizes the number of coal mills in operation and the output distribution of coal mills in advance, avoiding peak power impact and energy waste caused by temporarily starting multiple coal mills due to sudden load increases, and greatly reducing the delay and energy consumption of the pulverizing system.
[0036] S5. Based on short-term fluctuation components, adjust the powder storage depth and powder feeding rate of the small powder hopper in advance; Obtain the current maximum output of the coal mill B_max, the current actual output of the coal mill B_current, and the current real-time output of the coal mill B(t). Based on the short-term fluctuation component, obtain the peak value of the short-term fluctuation component F_high_peak (i.e., the fuel demand of the short-term fluctuation component per unit time).
[0037] Calculate the fuel shortage based on the current maximum output of the coal mill, the current actual output of the coal mill, and the peak value of the short-term fluctuation component. .
[0038] The formula for calculating the fuel deficit is: .
[0039] If the fuel deficit is greater than zero ( If the powder storage control unit is activated, the powder storage depth of the small powder bin will be increased in advance so that the powder storage amount reaches at least a set multiple of the fuel shortage amount, which can be 1.2 times. If the fuel deficit is less than or equal to zero ( If the short-term fluctuation component peak value is less than the current real-time output of the coal mill, then the feed control unit will be activated. Determine the powder supply signal from the small powder container. The toner feeding rate command for the small toner hopper is obtained by performing multi-factor correction on the toner feeding signal of the small toner hopper. The powder feeding rate of the small powder bin is adjusted based on the powder feeding rate command of the small powder bin.
[0040] The formula for calculating the difference between the peak value of the short-term fluctuation component and the current real-time output of the coal mill is as follows: , The formula for calculating the powder feeding rate command of the small powder hopper is: , , in, This is a correction factor for small powder bin burners. This is the correction factor for the condition of the powder feeder. This is a correction factor for boiler combustion conditions. To provide an advance investment coefficient, This is the time difference between the current time and the predicted start time of the load change. The time constant required for the small powder silo feeding system to reach a stable powder supply from startup; This is an adjustable coefficient, with a value range of 0.5-1.5.
[0041] This invention utilizes deep learning based on an improved TPA-LSTM model to achieve high-precision prediction of the load demand of coal-fired power generating units in the next 5-30 minutes. It adjusts the pulverized coal storage depth and feeding rate of the small pulverized coal silo in advance, realizing coordinated pulverized coal supply between the small pulverized coal silo and the coal mill pulverizing system. This transforms the small pulverized coal silo from an end-of-line compensator to an active reserver, advancing the timing of fuel input from after the load gap occurs to before the gap occurs. This significantly improves the variable load response rate, enabling it to meet the requirement of 3%-5% of rated load per minute, and greatly reduces the delay of the pulverizing system.
[0042] S6. The obtained coal mill and small coal silo supply quantities are superimposed and transported to the boiler furnace to coordinate the coal-fired power generation unit to participate in primary frequency regulation and AGC response.
[0043] The coal supply from the pulverizer and the small pulverizer silo is combined and delivered to the boiler furnace so that the total coal supply is equal to the fuel required for the peak load of the predicted load, and the coal-fired power generation units are coordinated to participate in primary frequency regulation and AGC response.
[0044] This invention improves the deep learning capabilities of the TPA-LSTM model to achieve high-precision load prediction. Based on the predicted load, it enables advance adjustment of the pulverizing system's operating status, as well as the pulverizing depth and feeding rate of the small pulverizing silo, significantly improving the variable load response rate and greatly reducing the pulverizing system's delay and energy consumption. It deeply integrates load prediction with pulverizing and pulverizing storage control. Compared with traditional PID control based on deviation feedback, the predictive feedforward control of this invention significantly reduces the fluctuation amplitude and overshoot of the main steam pressure, enhancing control stability. Furthermore, the data-driven improved TPA-LSTM model does not rely on a precise physical mechanism model, can be deployed in 300MW to 1000MW coal-fired units, has low migration costs, and is suitable for multiple platforms.
[0045] This invention also relates to a deep learning-based automatic control system for coal-fired power generation, such as... Figure 2 As shown, it includes: The data acquisition module is used to collect multi-dimensional operating data of coal-fired power generating units in real time. The operating data includes unit load, main steam pressure, boiler oxygen content, AGC load instructions and peak shaving instructions from the power grid dispatching instructions, coal quality parameters, and ambient temperature, etc. The coal quality parameters include volatile matter, calorific value, and ash content.
[0046] The deep learning-based load forecasting module preprocesses the runtime data and extracts principal component features using Principal Component Analysis (PCA). Based on these features, an improved TPA-LSTM model is used to obtain the predicted load sequence within a preset time window. The module comprises a PCA unit and an improved TPA-LSTM model. The PCA unit preprocesses the runtime data and extracts principal component features using PCA. The improved TPA-LSTM model is a neural network model used to obtain the predicted load sequence within a preset time window based on the principal component features. The improved TPA-LSTM model consists of a sequentially connected input layer, an LSTM layer, a time-series pattern attention layer, and a fully connected output layer. The input layer receives the principal component features, the LSTM layer captures long-term dependencies in the load time series, the time-series pattern attention layer automatically calculates the attention weights for each historical time step and performs a weighted summation of the hidden states in key time periods to generate a context vector, and the fully connected output layer maps the output of the time-series pattern attention layer to multi-step load forecasts for the next 5-30 minutes to generate the predicted load sequence.
[0047] The predicted load sequence decomposition module is used to obtain the low-frequency variation trend and short-term fluctuation components based on the predicted load sequence through time series decomposition method; the low-frequency variation trend is sent to the coal mill pulverizing coordination control module, and the short-term fluctuation components are sent to the small powder storage control module.
[0048] The coal mill pulverizing coordination control module is used to generate the optimal number of coal mills to be put into operation and the expected output of the coal mills based on low-frequency variation trends, and to adjust the operating status of the pulverizing system.
[0049] The small powder hopper storage control module is used to adjust the powder storage depth and powder feeding rate of the small powder hopper in advance based on short-term fluctuation components. The small powder hopper storage control module includes a powder storage control unit, a powder feeding control unit, and a powder level safety protection unit. The powder storage control unit is used to adjust the powder storage depth of the small powder hopper and control the powder storage amount. The powder feeding control unit is used to control the powder feeding rate of the small powder hopper. The powder level safety protection unit is used to monitor the powder level height and powder hopper temperature in real time. When the powder level height is higher than the maximum warning value or lower than the minimum warning value, or when the powder hopper temperature exceeds the safety threshold, an alarm signal is issued and the powder feeder speed is limited.
[0050] The coordination module is used to superimpose the coal supply from the coal mill and small coal silo and deliver it to the boiler furnace, and coordinate the coal-fired power generation unit to participate in primary frequency regulation and AGC response.
[0051] The online model update module is used to incrementally train or retrain the improved TPA-LSTM model using the latest running data collected periodically.
[0052] Example 1 Taking a 350MW coal-fired power plant unit as an example, a deep learning-based automatic control method for coal-fired power generation includes the following steps: S1. Real-time acquisition of multi-dimensional operating data of coal-fired power generating units; Operational data from the coal-fired power generating units over the past 12 months was collected from the distributed control system (DCS), totaling approximately 31.5 million records. The operational data was divided into training, validation, and test sets at a ratio of 80%, 10%, and 10%, respectively.
[0053] The training set is constructed into training samples using a sliding window approach. Each training sample contains the running data of the past 30 time steps, and the target label is the load value of the next 20 time steps (one step corresponds to 1 minute).
[0054] The improved TPA-LSTM model was trained using training samples, validated using a validation set, and tested using a test set. The testing process included: using mean squared error (MSE) as the loss function, optimizing the training with the Adam optimizer, an initial learning rate of 0.001, a batch size of 64, and a pause interval of 10. After 120 training epochs, the improved TPA-LSTM model achieved the following performance metrics on the test set: a mean absolute percentage error (MAPE) of 1.45% and a root mean square error (RMSE) of 2.31 MW for predicting the load in the next 10 minutes, with a prediction response time of less than 50 milliseconds.
[0055] After training, the improved TPA-LSTM model and its parameters will be deployed to the inference engine of the industrial control server to achieve real-time inference.
[0056] S2. After preprocessing the operational data, principal component features are extracted using principal component analysis. Based on the principal component features, the predicted load sequence within the preset time window is obtained by improving the TPA-LSTM model. Principal component analysis (PCA) was performed on the normalized 8-dimensional running data to calculate the covariance matrix, solve for the eigenvalues and eigenvectors, and select the top 5 principal components with a cumulative contribution rate of over 90% as the model input features.
[0057] An improved TPA-LSTM model is constructed, including: an input layer receiving 5-dimensional principal component features with a time step of 30; an LSTM layer with 128 hidden units and a dropout rate of 0.2; a temporal pattern attention layer (TPA Layer) with an attention window length of 10 steps and an attention dimension of 64, employing an additive attention mechanism to calculate the attention weights at each historical time step, and weighted summing the weights with the LSTM hidden state vectors to generate a context vector; and a fully connected output layer containing three fully connected layers with 64, 32, and 20 neurons respectively, all using ReLU activation functions, and outputting a load prediction sequence of length 20 (corresponding to the next 20 minutes).
[0058] Real-time prediction includes: assuming the system completes a load sampling and update at time t=0, at which point the coal-fired generating unit is operating at a steady-state load of 100MW. The data acquisition module obtains real-time multi-dimensional operating data including: unit load 100MW, main steam pressure 12.5MPa, boiler oxygen content 3.2%, total instantaneous flow rate of coal feeder 35t / h, stable AGC load command, volatile matter 28%, calorific value 5200 kcal / kg, and ambient temperature 15℃. The deep learning load prediction module receives the above operating data at the current moment and the historical sequence of the past 30 time steps. After PCA preprocessing, LSTM layer extraction of temporal features, and invocation of the attention mechanism, it determines that the load will experience a rapid increase in the next 10 minutes, outputting the predicted load sequence F_pred as follows: minutes 1-3 maintain 100-105MW, minute 4 rises to 118MW, minute 5 rises to 135MW, minute 6 rises to 150MW, and minutes 7-10 maintain 150-148MW.
[0059] S3. Based on the predicted load sequence, the low-frequency change trend and short-term fluctuation components are obtained through time series decomposition. The predicted load sequence F_pred is processed through a first-order low-pass digital filter (cutoff frequency 0.05Hz) to extract the low-frequency trend F_low. Then, F_pred is subtracted from F_low to obtain the short-term fluctuation component F_high. Specifically, the time window width for the low-frequency trend is set to n=5 minutes, meaning the low-frequency trend from minutes 1 to 5 is taken as a moving average, and the short-term fluctuation component from minutes 6 to 10 is used as a rapid compensation target. After decomposition, the low-frequency trend F_low and the short-term fluctuation component F_high are obtained. F_low is sent to the coal mill pulverizing coordination control module, and F_high is sent to the small-powder storage control module.
[0060] S4. Based on the low-frequency change trend, generate the optimal number of coal mills to be put into operation and the coal mill output strategy, and adjust the operating status of the pulverizing system. The coal mill pulverizing coordination control module uses the low-frequency trend F_low as the feedforward input. Currently, there are two coal mills in operation: mill B and mill C. Each mill has a maximum output of 35MW, i.e., B_max = 70MW. The calculated load change rate is approximately (142-100)MW / 10min = 0.42MW / min, which does not exceed the preset overshoot threshold of 1.2MW / min. Therefore, the coal mill override mode is not activated. Instead, the optimal number of coal mills to be operated is determined with the goal of minimizing the energy consumption of the pulverizing system. Calculations show that during the load increase to 150MW, relying solely on mills B and C can increase the output to the maximum of 35MW per mill, for a total output of 70MW, without needing to start more mills in advance. Based on this, the coal mill pulverizing coordination control module generates a coal mill output strategy: mills B and C synchronously increase their output, adjusting the feeder speed and primary air volume to achieve base load tracking.
[0061] S5. Based on short-term fluctuation components, adjust the powder storage depth and powder feeding rate of the small powder hopper in advance; The small-scale coal pulverizer storage control module uses the short-term fluctuation component F_high as the feedforward input. It calculates the current maximum output of the operating coal mill as B_max = 70MW, the current actual output of the coal mill as B_current = 35MW (at t=0), the peak value of the short-term fluctuation component is approximately 8MW, and the fuel shortage is... =8-(70-35)=-27MW, there is actually no shortage. However, due to the rapid peak in the predicted overall load increase, the active pre-storage strategy of this invention is still implemented: the small powder silo feeding transmission mechanism is started 2 minutes before the predicted load increase starts at t=4 minutes, increasing the powder storage depth of the small powder silo to 80% of the maximum capacity. At the same time, the powder feeding rate command is calculated: t_settle=1.0 minute, t_lead=2.0 minute, =1.0, k=min(1, 1.0×2.0 / 1.0)=1.0, choose =0.96, =1.03, =1.02, thus obtaining =0.96×1.03×1.02×1×f(t)≈1.01f(t). When the load reaches its peak at t=6-7 minutes, the output of the coal mill has increased to 70MW. The additional 8MW required for the short-term fluctuation component is provided by the small pulverizer at the instructed feed rate.
[0062] S6. The obtained coal mill and small coal silo supply quantities are superimposed and transported to the boiler furnace to coordinate the coal-fired power generation unit to participate in primary frequency regulation and AGC response.
[0063] The coal feed from the pulverizer and the small pulverizer silo is combined and delivered to the boiler furnace. During the period t=4-7 minutes, the output of pulverizers B and C increases to the upper limit of 70MW (corresponding to a fuel consumption of approximately 84t / h), and the small pulverizer silo replenishes approximately 8MW of fuel (approximately 4.5t / h) according to the feed rate command. The total coal feed equals the fuel required for the predicted peak load of 150MW. The main steam pressure fluctuation is controlled within ±0.25MPa, and the coal-fired power generation unit successfully participates in primary frequency regulation and AGC response, with the actual load change rate reaching 3.2%Pe / min.
[0064] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.
Claims
1. A deep learning-based automatic control method for coal-fired power generation, characterized by, include: Real-time acquisition of multi-dimensional operational data from coal-fired power generating units; After preprocessing the operational data, principal component features are extracted using principal component analysis. Based on the principal component characteristics, the predicted load sequence within a preset time window is obtained by improving the TPA-LSTM model. Based on the predicted load sequence, the low-frequency change trend and short-term fluctuation components are obtained through time series decomposition. Based on the low-frequency variation trend, the optimal number of coal mills to be put into operation and the coal mill output strategy are generated to adjust the operating status of the pulverizing system. Based on short-term fluctuation components, adjust the powder storage depth and powder feeding rate of the small powder hopper in advance; The coal feed from the pulverizer and small pulverizer bins is combined and transported to the boiler furnace, coordinating the coal-fired power generation units to participate in primary frequency regulation and AGC response.
2. The method for automatic control of coal-fired power generation based on deep learning according to claim 1, characterized in that, After preprocessing the operational data, principal component features are extracted using principal component analysis, including: Principal component analysis is used to calculate the covariance matrix based on the preprocessed running data, solve for the eigenvalues and eigenvectors, and select the first preset number of eigenvectors whose cumulative contribution rate exceeds a set proportion as principal components to obtain the principal component features.
3. The automatic control method for coal-fired power generation based on deep learning according to claim 1, characterized in that, The method of obtaining low-frequency variation trends and short-term fluctuation components based on the predicted load sequence through time series decomposition includes: The low-frequency variation trend is extracted from the predicted load sequence by passing it through a low-pass filter, and the short-term fluctuation component is obtained by subtracting the low-frequency variation trend from the predicted load sequence.
4. The method for automatic control of coal-fired power generation based on deep learning according to claim 1, characterized in that, The process of generating the optimal number of coal mills in operation and the coal mill output strategy based on low-frequency variation trends, and adjusting the operating status of the pulverizing system, includes: The variable load rate is calculated based on the low-frequency change trend. If the variable load rate exceeds the preset overshoot threshold, the overdrive mode of the coal mill is activated to generate the optimal number of coal mills in operation and the coal mill output strategy. The number of coal mills in operation is increased to the optimal number of coal mills in operation, and the coal mill output is increased based on the coal mill output strategy. If the variable load rate does not exceed the preset overshoot threshold, the pulverizer override mode will not be activated. The objective function is to minimize the energy consumption of the pulverizing system, and the upper and lower limits of pulverizer output and the number of pulverizers in operation are used as constraints to generate the optimal number of pulverizers in operation and the pulverizer output strategy. Based on the pulverizer output strategy, the expected output of the pulverizer is obtained, and the actual output of the pulverizer is acquired. According to the deviation between the actual output of the pulverizer and the expected output of the pulverizer, the speed of the feeder and the air volume of the primary air fan are adjusted.
5. The automatic control method for coal-fired power generation based on deep learning according to claim 1, characterized in that, The method of adjusting the powder storage depth and powder feeding rate of the small powder hopper in advance based on short-term fluctuation components includes: Obtain the current maximum output of the coal mill, the current actual output of the coal mill, and the current real-time output of the coal mill; and obtain the peak value of the short-term fluctuation component based on the short-term fluctuation component. Calculate the fuel shortage based on the current maximum output of the coal mill, the current actual output of the coal mill, and the peak value of the short-term fluctuation component. If the fuel shortage is greater than zero, the powder storage control unit will be activated to increase the powder storage depth of the small powder bin in advance, so that the powder storage amount is at least a set multiple of the fuel shortage. If the fuel shortage is less than or equal to zero, the pulverized coal feeding control unit is activated. The pulverized coal feeding signal of the small pulverized coal bin is determined based on the difference between the peak value of the short-term fluctuation component and the real-time output of the current coal mill. The pulverized coal feeding signal of the small pulverized coal bin is corrected by multiple factors to obtain the pulverized coal feeding rate command of the small pulverized coal bin. The pulverized coal feeding rate of the small pulverized coal bin is adjusted based on the pulverized coal feeding rate command of the small pulverized coal bin.
6. The automatic control method for coal-fired power generation based on deep learning according to claim 5, characterized in that, The formula for calculating the powder feeding rate command of the small powder hopper is: ， ， in, This is the powder dispensing rate instruction for the small powder container. This is a correction factor for small powder bin burners. This is the correction factor for the condition of the powder feeder. This is a correction factor for boiler combustion conditions. To provide an advance investment coefficient, This is a signal to send fans to the small fan warehouse. This is the time difference between the current time and the predicted start time of the load change. The time constant required for the powder feeding system of the small powder hopper to reach a stable powder supply from startup. This is an adjustable coefficient.
7. The method for automatic control of coal-fired power generation based on deep learning according to claim 1, characterized in that, It also includes steps for incremental training or retraining the improved TPA-LSTM model using periodically collected runtime data, including: The regularly collected operational data is divided into training set, validation set and test set according to a preset ratio; Based on the training set, training samples are constructed using a sliding window method; The prediction error of the improved TPA-LSTM model under the current working conditions is evaluated using the sliding window residual analysis method. When the prediction error exceeds the preset error threshold, retraining is automatically triggered, and the improved TPA-LSTM model is retrained using training samples. When the prediction error does not exceed the preset error threshold, incremental training is automatically triggered, and the improved TPA-LSTM model is incrementally trained using training samples. The improved TPA-LSTM model was validated using a validation set; The validated improved TPA-LSTM model was tested using a test set.
8. A deep learning-based automatic control system for coal-fired power generation, characterized in that, A deep learning-based automatic control method for coal-fired power generation, used in any one of claims 1-7, comprises: The data acquisition module is used to collect multi-dimensional operating data of coal-fired power generating units in real time; The deep learning load prediction module is used to preprocess the running data and extract principal component features using principal component analysis. Based on the principal component features, an improved TPA-LSTM model is used to obtain the predicted load sequence within a preset time window. The deep learning load prediction module includes a PCA unit and an improved TPA-LSTM model. The predicted load sequence decomposition module is used to obtain the low-frequency variation trend and short-term fluctuation components based on the predicted load sequence through time series decomposition methods. The coal mill pulverizing coordination control module is used to generate the optimal number of coal mills in operation and the expected output of the coal mills based on low-frequency variation trends, and to adjust the operating status of the pulverizing system. The powder storage control module for the small powder silo is used to adjust the powder storage depth and powder feeding rate of the small powder silo in advance based on short-term fluctuation components; the powder storage control module includes a powder storage control unit and a powder feeding control unit; The coordination module is used to superimpose the coal mill and small coal silo supply data and deliver them to the boiler furnace, and coordinate the coal-fired power generation units to participate in primary frequency regulation and AGC response. The online model update module is used to incrementally train or retrain the improved TPA-LSTM model using the latest running data collected periodically.
9. The automatic control system for coal-fired power generation based on deep learning according to claim 8, characterized in that, The improved TPA-LSTM model includes: an input layer, an LSTM layer, a temporal pattern attention layer, and a fully connected output layer connected in sequence. The input layer is used to receive principal component features, the LSTM layer is used to capture the long-term dependencies of the load time series, the time series pattern attention layer is used to automatically calculate the attention weights of each historical time step and perform weighted summation of the hidden states of key time periods to generate a context vector, and the fully connected output layer is used to map the output of the time series pattern attention layer to multi-step load prediction values to generate a predicted load sequence.
10. The deep learning-based automatic control system for coal-fired power generation according to claim 8, characterized in that, The powder storage control module for the small powder silo also includes a powder level safety protection unit, which is used to monitor the powder level height and powder silo temperature in real time. When the powder level height is higher than the maximum warning value or lower than the minimum warning value, or when the powder silo temperature exceeds the safety threshold, an alarm signal is issued and the speed of the powder feeder is limited.