A soot blowing method for air preheater adapted to deep peak load regulation mode of thermal power units
By combining physical mechanisms with mathematical models, the CNN-LSTM prediction method was used to achieve automated control of air preheater ash accumulation, which solved the problems of flue blockage and low-temperature corrosion caused by air preheater ash accumulation in the existing technology, and improved the operational stability and economy of thermal power units.
Patent Information
- Application Number
- CN202211198384.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-09-29
- Publication Date
- 2024-11-29
- Estimated Expiration
- 2042-09-29
AI Technical Summary
Existing soot blowing technology cannot remove ash accumulation in the air preheater in a timely, thorough and intelligent manner under the deep peak shaving mode of thermal power units, resulting in flue blockage and low-temperature corrosion, which affects the normal operation of the unit.
Combining physical mechanisms and mathematical drivers, a CNN-LSTM model is used to predict the air preheater pressure difference. By determining whether the pressure difference exceeds the critical value and the maximum number of soot blowing cycles is reached, automated soot blowing control is achieved, and the optimal time for soot blowing is selected.
It achieves timeliness, thoroughness, safety, and economy in air preheater soot blowing, reduces unnecessary soot blowing fluid consumption and heat exchange surface wear, and improves the stability and efficiency of unit operation.
Smart Images

Figure CN115562012B_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the field of soot blowing of a boiler tail heating surface, and in particular relates to a soot blowing method for an air preheater adapted to a deep peak regulation mode of a thermal power unit. Background Art
[0002] Domestic thermal power units have been gradually transformed for flexibility. However, at the beginning of their design, most thermal power units did not consider that the units would operate in a deep peak-shaving mode for a long time. Therefore, the thermal power units in this deep peak-shaving operation mode will inevitably affect the operation safety of the boiler side, the turbine side and the auxiliary equipment. Among them, the long-term low-load operation of the unit makes the exhaust gas temperature low, which brings about the problems of ammonia escape and catalyst deactivation in the denitrification system. As the flue gas temperature drops, sulfur trioxide in the flue gas condenses with water vapor to form sulfuric acid vapor. This inevitably leads to dust accumulation and low-temperature corrosion at the cold end of the air preheater, and the dust accumulation in the flue gradually increases over time. Under the combined effect of low-temperature corrosion, the flue flow area is reduced or even blocked, resulting in an increase in the pressure difference between the inlet and outlet of the air preheater and an increase in the fan output. Once the flue is blocked, it can only be descaled by water flushing during the shutdown inspection. There is no effective means to solve it during operation, which affects the normal operation of the unit.
[0003] Sootblowing is an effective way to remove dust from the heating surface. Existing sootblowing methods are basically achieved through strong flushing of some other gases or liquids, including steam sootblower, high-energy gas pulse sootblower and sonic sootblower, as well as high-pressure water flushing after shutdown. Existing sootblowing control strategies are mostly manual sootblowing or timed and quantitative sootblowing. During the operation of the unit, if the sootblower is used too many times, that is, the sootblowing cycle is too short, it will cause unnecessary sootblowing fluid consumption and wear and corrosion of the heating surface boiler tubes; if the cycle is too long, the sootblowing effect is not good, and the loss caused by dust on the heating surface is too large. Considering the safety and economy of boiler operation, it is particularly important to select the best sootblowing time.
[0004] Existing soot blowing control strategies are mostly manual soot blowing, timed and quantitative soot blowing, and soot blowing based on the degree of soot accumulation determined by the soot thermal resistance.
[0005] 1. Manual soot blowing, which requires the unit to be shut down for operation, has a long soot blowing downtime period, which seriously affects the normal operation of the unit. In addition, this method requires manpower to disassemble and assemble the equipment, which is not smart enough.
[0006] 2. Timed and quantitative sootblowing. This method cannot be adjusted according to the actual blockage of the equipment. If the sootblower is used too many times, that is, the sootblowing cycle is too short, it will cause unnecessary sootblowing fluid consumption and wear and corrosion of the heating surface furnace tubes; if the cycle is too long, the sootblowing effect is not good, the loss caused by soot contamination on the heating surface is too great, and the sootblowing is not thorough.
[0007] 3. Using thermal resistance to reflect the degree of ash accumulation for soot blowing. This method introduces an ash accumulation factor and uses the ash accumulation thermal resistance of the heated surface to determine whether to blow soot. This calculation method involves many heat sources, has poor anti-interference ability, and is easily interfered by other thermal resistances, resulting in inaccurate ash accumulation judgment. In addition, this method is usually applicable to boiler heating surfaces and low-low economizers, and is not universal.
[0008] 4. Soot blowing is performed based on the unit operation big data to reflect the degree of blockage of air preheater ash accumulation. This soot blowing method is purely mathematically driven, lacks integration with physical models, and does not consider the physical mechanism behind the blockage of air preheater ash accumulation. Based on the mathematically driven model, the amount of data calculation is too complex, the calculation time is too long, the prediction results under different working conditions have large errors, and the reliability is affected.
[0009] In summary, since the existing sootblowing technology strategy is imperfect, the sootblowing is not timely, thorough and intelligent, the present invention proposes an air preheater sootblowing method that is adapted to the deep peak regulation mode of thermal power units. Summary of the invention
[0010] In the context of flexible power generation, the unit load often fluctuates over a large range, making the air preheater ash accumulation and blockage irregular. The present invention combines physical mechanism with mathematical drive, deeply analyzes the mechanism of air preheater ash accumulation and blockage, and comprehensively considers relevant factors such as unit operating conditions and flue gas parameters to select the most suitable input parameters; by using the CNN model to accurately extract the potential features of the time series data of the input parameters and integrate them, and then through the LSTM neural network module, the integrated features are memorized and screened, and the pressure difference is fitted and predicted, and finally the prediction result is output through the fully connected layer; by judging whether the pressure difference is greater than the critical value and whether the maximum number of soot blowing is reached, the soot blowing action is performed, so as to automatically control the soot blowing system to achieve the purpose of timely and thorough soot blowing, which is safer, energy-saving and environmentally friendly. Through the above control ideas, the physical mechanism drive and the mathematical model drive are combined to form an air preheater soot blowing method that adapts to the deep peak-shaving mode of thermal power units, grasp the best time for soot blowing, and achieve the timeliness, thoroughness, economy and safety of the unit soot blowing system.
[0011] The invention discloses an air preheater soot blowing method adapted to the deep peak regulation mode of a thermal power unit, comprising three parts: operation parameter selection, pressure difference prediction and soot blowing control.
[0012] Preferably, in the above-mentioned air preheater sootblowing method, the operating parameters include flue gas operating parameters, denitrification parameters, air preheater operating parameters, flue gas acid dew point, and actual operating load of the unit.
[0013] Preferably, the flue gas operating parameters specifically include exhaust temperature, sulfur content of coal, moisture content of flue gas and sulfur trioxide content in flue gas; the denitrification parameters specifically include SCR ammonia injection amount, SCR flue gas oxygen content, SCR inlet ammonia content and SCR outlet ammonia content; the air preheater operating parameters specifically include air leakage coefficient, cold air temperature, inlet flue gas temperature and outlet flue gas temperature.
[0014] Preferably, the calculation method of the flue gas acid dew point includes the following three methods:
[0015] (a) Standard method for thermometers of boiler units
[0016]
[0017] In the above formula, t ld Water dew point, ℃, S zs Converted sulfur content, %, A zs Estimated ash content, %, α fh The fly ash share, β, is related to the excess air coefficient and has a value of 125;
[0018] (b) Mueller curve fitting formula
[0019]
[0020] In the above formula, is the volume parts per million of SO3;
[0021] (c)AGOkkes formula
[0022]
[0023] In the above formula, is the water vapor partial pressure, is the partial pressure of SO3.
[0024] The flue gas acid dew point and denitrification parameters need to be corrected based on theoretical values and on-site measurement values.
[0025] Preferably, in the sootblowing method, the pressure difference prediction is divided into two steps: data cleaning preprocessing and CNN-LSTM prediction model construction, which are as follows:
[0026] Step 1: Data cleaning and preprocessing
[0027] For the specific data of the selected parameters, the 3δ method is used to eliminate the data with large errors, and the standard deviation is calculated according to the Bessel formula. Suppose the sample data is x 1 ,x 2 ,x 3 ,…,x n , the average value is deviation but:
[0028]
[0029] In order to facilitate calculation and make the specific data of the selected parameters in the same order of magnitude, they are normalized and the value range of the processed data is [0, 1];
[0030] Step 2: CNN-LSTM prediction model construction
[0031] The CNN module is responsible for feature extraction and integration of input data. The specific calculation formula is as follows:
[0032] Convolution calculation:
[0033]
[0034] In the above formula, is the i-th input feature value of the l-th layer, is the weight matrix of the i-th convolution kernel in the l-th layer, x l -1 is the output feature of the lth layer, is the i-th bias term of the l-th layer, and f is the activation function;
[0035] The pooling layer uses the maximum pooling formula and the average pooling formula:
[0036]
[0037] In the above formula, is the value in the i-th feature matrix of the l+1th layer after pooling, is the value of the i-th feature matrix of the l-th layer within the pooling range, D j is the pooling range;
[0038] The LSTM module is responsible for memorizing and filtering the integrated features, performing fitting predictions, and finally outputting the prediction results through the fully connected layer. The specific calculation formula is as follows:
[0039] f t =σ(x t W xf +h t-1 W hf +b f )
[0040] i t =σ(x t W xi +h t-1 W hi +b i )
[0041]
[0042] o t = σ(x t W xo + h t-1 W ho + b o )
[0043]
[0044] h t = o t * tanh(c t )
[0045] Among them, f t , i t , o t are the forget gate, input gate, candidate state memory unit, and output gate at time step t respectively; W represents the weight; b is the bias term; σ(·) represents the sigmoid activation function.
[0046] Preferably, in terms of soot blowing control, according to the prediction result of the CNN-LSTM model, a differential pressure change curve of the air preheater is constructed, and a soot blowing threshold C lj and the daily soot blowing times N are set. When the change rate dp / dt of the differential pressure change curve at time t > C lj and the current soot blowing times n < N, the soot blowing action is executed. Description of the Drawings
[0047] Figure 1 is the overall logic diagram of an air preheater soot blowing method adapted to the deep peak shaving mode of a thermal power unit;
[0048] Figure 2 is the schematic diagram of air preheater ash accumulation and blockage;
[0049] Figure 3 is the diagram for selecting input parameters of air preheater differential pressure prediction;
[0050] Figure 4 is the basic structure diagram of the LSTM model. Detailed Implementation Manner
[0051] To make the purpose, technical solution, and advantages of the implementation of the present invention clearer, the technical solutions in the embodiments of the present invention will be described in more detail below with reference to the drawings in the embodiments of the present invention. The described embodiments are some, but not all, of the embodiments of the present invention. The embodiments described below are intended to explain the present invention and should not be construed as a limitation on the implementation of the present invention.
[0052] The present invention uses the CNN-LSTM model to predict the critical pressure difference of flue ash blockage, judge the actual blockage of the air preheater and select the appropriate soot blowing time, so as to automatically control the soot blowing system to achieve the purpose of timely and thorough soot blowing, which is safer, energy-saving and environmentally friendly. The overall logic is as follows Figure 1 shown.
[0053] The present invention is generally divided into three parts: operation parameter selection, pressure difference prediction and action soot blowing.
[0054] Run parameter selection
[0055] The degree of ash accumulation and blockage of the on-site air preheater can be directly reflected by the pressure difference between the inlet and outlet of the air preheater. The pressure difference between the inlet and outlet of the air preheater is mainly affected by the unit load, coal and flue gas parameters, SCR denitrification system operating parameters and air preheater operating parameters. The basis for parameter selection is as follows:
[0056] The blockage of the on-site air preheater is mainly caused by the adsorption of fly ash in the flue gas by ammonium bisulfate, which then adheres to the heated wall surface. Figure 2 As shown. The reason is that under the background of flexible power generation, the units frequently perform deep peak regulation, which leads to increased ammonia escape from the SCR denitrification system. The escaped ammonia reacts chemically with SO3 in the flue gas to form ammonium bisulfate. The melting point of ammonium bisulfate is 147°C. Below this temperature, ammonium bisulfate will gradually condense into a liquid state, and its solidification process has a strong gelling property. However, the temperature of the medium temperature section and the cold section of the air preheater is about 120°C to 230°C. Within this temperature range, ammonium bisulfate is dispersed in the flue gas in the form of mist or accumulates in liquid on the heated surface of the air preheater. Due to the strong hygroscopicity and adhesion of ammonium bisulfate, a large amount of fly ash adheres to the heat exchange surface of the air preheater or is stuck between the two layers. In the long run, the flow area on the flue gas side is reduced, which leads to blockage of the air preheater.
[0057] When the air preheater is clogged by dust accumulation, it is often accompanied by low-temperature corrosion. Therefore, the effect of flue gas acid dew point on the air preheater pressure difference needs to be considered. The dew point temperature of flue gas is mainly affected by SO 3 The present invention combines the actual conditions of different units and adopts the following acid dew point calculation method:
[0058] (a) Standard method for thermometers of boiler units
[0059]
[0060] Where: t ld Water dew point, ℃; S zs Converted sulfur content, %; A zs Equivalent ash content, %; α fh Fly ash share; β is related to the excess air coefficient and has a value of 125.
[0061] (b) Mueller curve fitting formula
[0062]
[0063] Where: It is the volume parts per million of SO3.
[0064] (c)AGOkkes formula
[0065]
[0066] Where: is the water vapor partial pressure, Pa; is the partial pressure of SO3, Pa.
[0067] In summary, the input parameters selected by the present invention, in addition to the unit load, also involve coal combustion and flue gas parameters, SCR denitrification system operating parameters and air preheater operating parameters. The specific input parameters are as follows: Figure 3 As shown, the comprehensive parameters are flue gas acid dew point and denitrification parameters, which are corrected according to the on-site measurement values.
[0068] Pressure difference prediction
[0069] The pressure difference prediction is divided into two steps: data cleaning preprocessing and CNN-LSTM prediction model construction.
[0070] 1) Data cleaning and preprocessing
[0071] According to the input parameters selected above, the 3δ method is used to eliminate the data with large errors, and the standard deviation is calculated according to the Bessel formula. Suppose the sample data is x 1 ,x 2 ,x 3 ,…,x n , the average value is deviation (i=1,2,…,n), then:
[0072]
[0073] In order to make all input parameters at the same order of magnitude for easy calculation, normalization processing is performed, and the value range of the processed data is [0, 1].
[0074] 2) CNN-LSTM prediction model construction
[0075] Convolutional neural network (CNN) is a feedforward neural network with deep structure and convolutional calculation. Its basic structure consists of input layer, convolution layer, pooling layer, fully connected layer and output layer. CNN model has powerful feature recognition and extraction capabilities, and can be applied to time series data classification and effectively predict the trend of time series data.
[0076] The convolution layer obtains the potential features of the data through the convolution operation of the convolution kernel and the input data. Among them, the convolution kernel acts as a filter to reduce the noise of the data and extract the potential key features in the input parameters. The convolution formula is as follows:
[0077]
[0078] In the formula, is the i-th input feature value of the l-th layer, is the weight matrix of the i-th convolution kernel in the l-th layer, x l-1 is the output feature of the lth layer, is the i-th bias term of the l-th layer, and f is the activation function.
[0079] Pooling layer: The pooling layer reduces the amount of calculation by reducing the parameters of the network, improves the calculation speed, and improves the robustness of the extracted features to prevent overfitting. The maximum pooling formula and the average pooling formula are as follows:
[0080]
[0081] In the formula, is the value in the i-th feature matrix of the l+1th layer after pooling, is the value of the i-th feature matrix of the l-th layer within the pooling range, D j is the pooling range.
[0082] Long short-term memory neural network (LSTM) is an improved recurrent neural network (RNN). Compared with the traditional recurrent neural network, it effectively solves the problems of gradient explosion and gradient disappearance in the training process. The basic structure of the LSTM model is as follows: Figure 4 As shown. The present invention combines the extracted feature vector xt, the state memory unit Ct-1 and the intermediate output ht-1 with the input forget gate to determine the part that needs to be forgotten in the state memory unit; the xt in the input gate is respectively activated by σ and tanh functions to jointly determine the vector that needs to be retained in the state memory unit; the intermediate output ht is jointly determined by the updated state memory unit Ct and the output gate Ot. The calculation formula involved is as follows:
[0083] f t =σ(x t W xf +h t-1 W hf +b f )
[0084] i t =σ(x t W xi +ht-1 W hi +b i )
[0085]
[0086] o t =σ(x t W xo +h t-1 W ho +b o )
[0087]
[0088] h t =o t *tanh(c t )
[0089] In the formula, f t , i t , o t are the forgetting gate, input gate, candidate state memory unit, and output gate at time step t, respectively; W represents the weight; b is the bias term; σ(·) represents the sigmoid activation function.
[0090] Among them, the main function of the CNN module is to extract and integrate the features of the historical unit load, coal combustion and flue gas parameters, SCR denitration system operation parameters, and air preheater operation data of each power plant; the LSTM neural network module is to memorize and screen the integrated features, and perform fitting prediction, and finally output the prediction result through the fully connected layer. In order to make the predicted value of the air preheater differential pressure of each power plant more in line with the actual situation, the present invention optimizes the prediction result. The CNN-LSTM hybrid model accurately extracts the potential features of the time series data by using the CNN model, provides more effective and accurate input data for the LSTM network model, and improves the accuracy of the air preheater differential pressure prediction.
[0091] Soot Blowing Control
[0092] Set the soot blowing control loop, visualize the prediction result of the CNN-LSTM model, and obtain the change curve of the air preheater differential pressure. Set the soot blowing critical value C lj and the daily soot blowing times N. The growth rate of the air preheater differential pressure is intuitively reflected by the curve. When the growth rate of the differential pressure curve is greater than the set critical value (dp / dt > C lj ), and the current soot blowing times n < N, the soot blowing action is executed. To prevent frequent soot blowing actions, it is appropriate to set the daily soot blowing times N to 3 - 4 times.
[0093] In addition, considering the complexity of unit operation under the background of deep peak regulation, a soot blowing method combining intelligent soot blowing and timed soot blowing can be adopted. Since the unit has been in low-load operation mode for a long time, there is less soot accumulation in the air preheater, so that the pressure difference growth rate is always lower than the set critical value. In order to prevent the soot accumulation from hardening and scaling, it is recommended to set the soot blowing timed 1 to 2 times a day on the basis of intelligent soot blowing.
Claims
1. A soot blowing method for an air preheater adapted to the deep peak load regulation mode of a thermal power unit, characterized in that: It includes three parts: operating parameter selection, pressure difference prediction and soot blowing control; The operating parameters include flue gas operating parameters, denitration parameters, air preheater operating parameters, flue gas acid dew point, and actual operating load of the unit; The flue gas operation parameters specifically include exhaust gas temperature, sulfur content of coal, moisture content of flue gas and sulfur trioxide content in flue gas; the denitrification parameters specifically include SCR ammonia injection amount, SCR flue gas oxygen content, SCR inlet ammonia content and SCR outlet ammonia content; the air preheater operation parameters specifically include air leakage coefficient, cold air temperature, inlet flue gas temperature and outlet flue gas temperature; The calculation methods of flue gas acid dew point include the following three methods: (a) Standard method for thermometers of boiler units In the above formula, t ld Water dew point, ℃, S zs Converted sulfur content, %, A zs Estimated ash content, %, α fh The fly ash share, β, is related to the excess air coefficient and has a value of 125; (b) Mueller curve fitting formula In the above formula, is the volume parts per million of SO3; (c)AGOkkes formula In the above formula, is the water vapor partial pressure, is the partial pressure of SO3; The flue gas acid dew point and denitrification parameters need to be corrected based on the theoretical values and the on-site measured values; The pressure difference prediction is divided into two steps: data cleaning preprocessing and CNN-LSTM prediction model construction, as follows: Step 1: Data cleaning and preprocessing For the specific data of the selected parameters, the 3δ method is used to eliminate the data with large errors, and the standard deviation is calculated according to the Bessel formula. Suppose the sample data is x1, x2, x3, ..., x n , the average value is deviation but: In order to facilitate calculation, the specific data of the selected parameters are normalized to the same order of magnitude, and the value range of the processed data is [0, 1]; Step 2: CNN-LSTM prediction model construction The CNN module is responsible for feature extraction and integration of input data. The specific calculation formula is as follows: Convolution calculation: In the above formula, is the i-th input feature value of the l-th layer, is the weight matrix of the i-th convolution kernel in the l-th layer, x l-1 is the output feature of the lth layer, is the i-th bias term of the l-th layer, and f is the activation function; The pooling layer uses the maximum pooling formula and the average pooling formula: In the above formula, is the value in the i-th feature matrix of the l+1th layer after pooling, is the value of the i-th feature matrix of the l-th layer within the pooling range, D j is the pooling range; The LSTM module is responsible for memorizing and filtering the integrated features, performing fitting predictions, and finally outputting the prediction results through the fully connected layer. The specific calculation formula is as follows: f t =σ(x t W xf +h t-1 W hf +b f ) i t =σ(x t W xi +h t-1 W hi +b i ) o t =σ(x t W xo +h t-1 W ho +b o ) h t =o t *tanh(c t ) Among them, f t 、i t , o t are the forget gate, input gate, candidate state memory unit and output gate at time step t respectively; W represents the weight; b is the bias term; σ(·) represents the sigmoid activation function; In terms of soot blowing control, according to the prediction results of the CNN-LSTM model, a differential pressure change curve of the air preheater is constructed, and a soot blowing critical value C lj and the daily soot blowing times N are set. When the change rate d p / d t >C lj and the current soot blowing times n < N, the soot blowing action is executed.
Citation Information
Patent Citations
Method for establishing air pre-heater clean factor calculation model by using smoke pressure difference, and application
CN105069185A
Boiler ash-blowing optimization system and optimization method thereof
CN106773955A
