Method, system and equipment for predicting ash deposition on heating surface of economizer and medium

Abstract

The invention discloses a method, a system and equipment for predicting ash deposition on a heating surface of a coal economizer and a medium, and relates to the technical field of ash deposition prediction, and the method comprises the steps: obtaining real-time operation data of the coal economizer; constructing a heat loss model of the heating surface of the economizer, and solving to obtain a soot blowing starting point and a corresponding soot blowing threshold value when the heat loss area is minimum; substituting the soot blowing starting point when the heat loss area is minimum and the corresponding soot blowing threshold value into the original sequence, and intercepting to obtain a target soot deposition section of interval prediction; decomposing the target ash deposition section by adopting an ensemble empirical mode decomposition method to obtain a high-frequency component, a low-frequency component and a trend term; taking the result as the input of a bidirectional gating recurrent neural network based on quantile regression to obtain a prediction interval of ash deposition; according to the method, the uncertainty of prediction points is better described for the target sequence by using interval prediction, and more response time can be provided for boiler operators to perform soot blowing operation and formulate soot blowing strategies in practical application.

Description

Technical Field

[0001] The invention relates to the technical field of ash accumulation prediction, and in particular to a soot blowing prediction method, system, equipment and medium for a heating surface of an economizer in a coal-fired power station. Background Art

[0002] Economizer is one of the important equipment on the boiler side of coal-fired power plants. Its main heat exchange mode is convection heat exchange. Ash deposition on the heating surface is a key factor affecting the efficiency of convection heat transfer. Long-term ash deposition will have an adverse effect on the performance of the economizer and the overall operation of the boiler, including the following aspects: 1. Decreased heat transfer efficiency: Ash deposition will form an insulation layer, reduce the heat transfer coefficient, and reduce the heat exchange efficiency of the economizer. This means that more heat will be lost with the flue gas instead of being absorbed by water or feed water, thus reducing the thermal efficiency of the boiler; 2. Increased energy consumption: Due to the decrease in heat transfer efficiency, more fuel consumption is required to maintain the same heat output, which increases operating costs. Ash pollution will also increase the resistance of flue gas flow, forcing the boiler fan to increase output to overcome the additional resistance, which will also lead to an increase in electricity consumption; 3. Safety risks: Severe ash pollution deposition may cause flue blockage, increase flue gas back pressure, and may cause furnace pressure to increase, posing a threat to the safe operation of the boiler. High concentrations of ash pollution may also contain corrosive components, causing corrosion to metal surfaces and accelerating aging and damage to equipment; 4. Increased maintenance costs: Cleaning the ash pollution on the economizer needs to be done regularly, whether it is mechanical soot blowing, chemical cleaning or other methods, it will increase maintenance costs and downtime. At present, the soot blowing operation of the heating surface of the economizer of a coal-fired power plant mostly adopts a soot blowing method with a fixed time and fixed mode. This time-based soot blowing method has many disadvantages in the process of variable operating conditions of the boiler, which is not conducive to the efficient operation of the boiler. Therefore, changing the time-based sootblowing method to a state-based sootblowing method is of great significance for improving the heat exchange efficiency of the economizer heating surface and the safety, energy conservation and emission reduction of the entire unit.

[0003] For example, patent announcement number CN115759279A discloses a "coal-fired power plant boiler ash accumulation prediction method", which mainly predicts ash accumulation and combines the introduction of CEEMDAN decomposition algorithm and KPCA kernel density dimensionality reduction algorithm to improve prediction accuracy. However, prediction alone cannot help operators determine "when to blow soot" and "how long to blow". Therefore, the soot blowing problem is solved by establishing a model that can effectively characterize the convection heat transfer efficiency of the boiler heating surface and comprehensively considering the actual operation of the boiler to formulate a soot blowing optimization strategy. Patent announcement number CN109654518B discloses a method for establishing the soot blowing heat transfer capacity Q per unit time of an economizer p The model is obtained by p The optimal soot accumulation time and soot blowing time at the maximum can be used to optimize the soot blowing cycle, and the approximate range of the soot blowing nodes and soot blowing time can be estimated.

[0004] However, traditional methods cannot perform quantitative analysis in economizer energy efficiency optimization, and it is difficult to determine the uncertainty of the prediction points, and cannot provide boiler operators with more reaction time for soot blowing operations and formulating soot blowing strategies. Summary of the invention

[0005] In view of the shortcomings of the prior art that it is impossible to conduct quantitative analysis and it is difficult to determine the prediction points, the present invention proposes a method for predicting ash accumulation on the heating surface of an economizer. By solving the heat loss model of the heating surface of an economizer with the minimum heat loss area as the objective function, interval prediction is used for the target sequence to better characterize the uncertainty of the prediction points, thereby solving the problems existing in the prior art.

[0006] A method for predicting ash accumulation on a heating surface of an economizer comprises the following steps:

[0007] Acquire the real-time operation data of the economizer, analyze the real-time operation data, and obtain a curve of the clean factor data set that can characterize the cross-flow heat transfer efficiency of the economizer heating surface over time according to the analysis results;

[0008] The soot accumulation section y1, soot blowing section y2 and soot accumulation section y3 of the economizer heating surface are intercepted from the curve of the clean factor data set changing with time; based on the soot blowing starting point, soot blowing end point and the fitting curve of the clean factor data set, the heat loss model of the economizer heating surface is constructed with the minimum heat loss area as the objective function; the soot blowing starting point is the intersection of the soot accumulation section y1 and the soot blowing section y2, and the soot blowing end point is the intersection of the soot blowing section y2 and the soot accumulation section y3;

[0009] The heat loss model of the economizer heating surface is solved by an improved subtraction average algorithm to obtain the soot blowing starting point and end point when the heat loss area of ​​the economizer heating surface is the smallest; specifically, a piecewise chaotic mapping function is introduced to initialize the positions of the soot blowing starting point and the soot blowing end point, and the heat loss area is solved as the fitness function, and the fitness corresponding to each soot blowing starting point and the soot blowing end point after initialization is evaluated. A global search is performed by introducing a golden section search, and a roulette strategy is introduced. In each round of iteration, different probabilities are assigned to each soot blowing starting point and the soot blowing end point according to their corresponding fitness, and the soot blowing starting point and the soot blowing end point corresponding to the highest probability are selected, so as to obtain the soot blowing starting point and the soot blowing end point corresponding to the minimum heat loss area;

[0010] According to the soot blowing starting point and end point corresponding to the minimum heat loss area, the target ash accumulation section of interval prediction is intercepted in the curve of the cleaning factor data set changing with time; according to the target ash accumulation section of interval prediction, the ash accumulation interval of the economizer heating surface is predicted.

[0011] Furthermore, the ash accumulation section y1, the soot blowing section y2 and the ash accumulation section y3 of the economizer heating surface are intercepted from the curve of the clean factor data set changing with time, specifically including: smoothing the clean factor data set by a polynomial fitting method, and the fitting curve is expressed as:

[0012] y=a0+a1x+a2x 2 +a3x 3 +a4x 4

[0013] The fitting curves of the ash accumulation section y1, the soot blowing section y2, and the ash accumulation section y3 are respectively expressed as:

[0014] y1=0.78499-1.05678×10 -5 x 1 +8.63269×10 -10 x 2 -4.37152×10 -14 x 3 +6.99596×10 -19 x 4

[0015] y2=1416.15676-0.16281x 1 +7.02413×10 -6 x 2 -1.34737×10- 10 x 3 +9.69771×10 - 16 x 4

[0016] y3=161.41513-0.01616x 1 +6.09207×10 -7 x 2 -1.0198×10 -11 x 3 +6.39346×10 -17 x 4

[0017] Wherein, a0, a1, a2, a3, and a4 represent coefficients of different terms of the polynomial; x and y represent time and the corresponding cleaning factor, respectively.

[0018] Furthermore, the target ash accumulation section predicted according to the interval prediction is used to predict the ash accumulation section of the economizer heating surface, including the following steps:

[0019] The target dust accumulation section is decomposed by using the ensemble empirical mode decomposition method to obtain high-frequency components, low-frequency components and trend items.

[0020] The high-frequency component, low-frequency component, trend term and target ash accumulation section after wavelet threshold denoising are used as the input of a bidirectional gated recurrent neural network based on quantile regression to obtain the ash accumulation range of the economizer heating surface.

[0021] Furthermore, the target dust accumulation section is decomposed by using the ensemble empirical mode decomposition method, wherein the ratio of the standard deviation of the decomposed additional white noise to the standard deviation of the signal is 0.2, and the number of times the signal is averaged is 100.

[0022] Furthermore, the high-frequency component, the low-frequency component, the trend term and the target ash accumulation section after wavelet threshold denoising are used as inputs of a bidirectional gated recurrent neural network based on quantile regression to obtain the ash accumulation interval of the economizer heating surface, which specifically includes the following steps:

[0023] The input data is divided into multiple specific quantiles; the quantile absolute deviation loss is set as the loss function, and its loss function is expressed as:

[0024]

[0025] Where τ is the target quantile, y k is the actual observed value, is the predicted value, is the indicator function;

[0026] Substitute the data divided into specific quantiles into the bidirectional gated recurrent neural network for prediction, and obtain the ash accumulation range of the economizer heating surface, including:

[0027] A BiGRU model based on the quantile regression loss function is established, and the conditional quantization value of its output response variable Y at the quantile point τ is:

[0028]

[0029] Where J is the number of units in the hidden state, f is the activation function of the output layer, and h j (τ) is the output of the BiGRU hidden state, w j (τ) and b j (τ) are the weight and bias of the output layer respectively; the model parameter l is updated according to the gradient descent algorithm Adam to obtain the BiGRU model with different weights and biases; after a series of forward propagation and backward learning, the predicted value of the response variable at each magnitude at time t+h is obtained:

[0030]

[0031] Estimate Y t+h|t And confidence intervals for discrete conditional quartiles:

[0032]

[0033] in, is the prediction interval at the β significance level, and τ are the upper and lower bounds of the prediction interval respectively. When , the confidence level of the interval is 1-β.

[0034] Furthermore, a distributed control system DCS is used to obtain real-time operating data of the economizer, with a sampling period of once every 5 seconds, for a total of 24 hours.

[0035] The present invention also includes a system for predicting ash accumulation on a heating surface of an economizer, comprising:

[0036] An acquisition module is used to acquire the real-time operation data of the economizer, analyze the real-time operation data, and obtain a curve of the clean factor data set that can characterize the cross-flow heat transfer efficiency of the economizer heating surface over time according to the analysis result;

[0037] A model building module is used to intercept the ash accumulation section y1, the soot blowing section y2 and the ash accumulation section y3 of the economizer heating surface from the curve of the clean factor data set changing with time; based on the soot blowing starting point, the soot blowing end point and the fitting curve of the clean factor data set, the heat loss model of the economizer heating surface is constructed with the minimum heat loss area as the objective function; the soot blowing starting point is the intersection of the soot accumulation section y1 and the soot blowing section y2, and the soot blowing end point is the intersection of the soot blowing section y2 and the ash accumulation section y3;

[0038] A calculation module is used to solve the heat loss model of the economizer heating surface by an improved subtraction average algorithm, and obtain the soot blowing starting point and the soot blowing end point when the heat loss area of ​​the economizer heating surface is the smallest; specifically, it includes introducing a piecewise chaotic mapping function, initializing the positions of the soot blowing starting point and the soot blowing end point, solving the heat loss area as a fitness function, evaluating the fitness corresponding to each soot blowing starting point and the soot blowing end point after initialization, and performing a global search by introducing a golden section search. At the same time, a roulette strategy is introduced in each round of iteration, and different probabilities are assigned to each soot blowing starting point and the soot blowing end point according to their corresponding fitness, and the soot blowing starting point and the soot blowing end point corresponding to the highest probability are screened out, so as to obtain the soot blowing starting point and the soot blowing end point corresponding to the minimum heat loss area;

[0039] The prediction module is used to obtain the target ash accumulation section of interval prediction in the curve of the cleaning factor data set changing with time according to the soot blowing starting point and soot blowing end point corresponding to the minimum heat loss area; and predict the ash accumulation interval of the economizer heating surface according to the target ash accumulation section of interval prediction.

[0040] The present invention also includes a computer device for predicting ash accumulation on the heating surface of an economizer, comprising: a memory, a processor, and a computer program stored in the memory, and when the processor executes the computer program, the steps of the method for predicting ash accumulation on the heating surface of an economizer are implemented.

[0041] Furthermore, a readable storage medium is provided, wherein the readable storage medium stores a computer program, wherein the computer program includes program instructions, and when the program instructions are executed by a processor, the steps of the economizer heating surface ash deposition prediction method are executed.

[0042] The present invention provides a method, system, device and storage medium for predicting ash accumulation on a heating surface of an economizer, which have the following beneficial effects:

[0043] The present invention analyzes and calculates the acquired real-time operation data of the economizer, and obtains a heat loss model of the economizer heating surface with the minimum heat loss area as the objective function; by adding a piecewise chaotic mapping, a golden section interval search strategy and a roulette strategy to the traditional subtraction average algorithm, the algorithm has higher convergence accuracy and faster convergence speed when solving the target problem; its piecewise chaotic mapping initializes the initial solutions of the soot blowing starting point and the soot blowing end point, and uses the fitness function to evaluate the fitness corresponding to each candidate solution of the soot blowing starting point and the soot blowing end point after initialization, thereby improving the randomness of population initialization, and introducing the golden section interval search strategy enables the algorithm to achieve a good balance between global search and local development capabilities, which is helpful in the search process It can not only fully explore the entire search space, but also conduct detailed search near the current optimal solution, thereby increasing the probability of finding the global optimal solution. At the same time, the roulette strategy is used to assign different probabilities according to the fitness of the individual, and the individual with the lowest fitness is eliminated, so that the probability of the individual represented by the optimal solution being selected will be greatly increased, thereby screening out the optimal sootblowing starting point and end point. According to this method, the difficulty of traditional methods in quantitative analysis in economizer energy efficiency optimization is overcome, and the speed and accuracy of solving the target problem are further improved. The use of interval prediction for the target sequence better characterizes the uncertainty of the prediction point. In practical applications, it can provide more reaction time for boiler operators to perform sootblowing operations and formulate sootblowing strategies. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 A flow chart of a method for predicting ash accumulation on a heating surface of an economizer provided in an embodiment of the present invention;

[0045] Figure 2 It is a schematic diagram of the cleaning factors of the original dust accumulation section 1, the dust blowing section, and the dust accumulation section 2 changing with time in an embodiment of the present invention.

[0046] Figure 3It is a schematic diagram of the fitting and modeling process of the original ash accumulation section 1, the ash blowing section, and the ash accumulation section 2 in an embodiment of the present invention;

[0047] Figure 4 A box plot comparing various optimization algorithms in an embodiment of the present invention;

[0048] Figure 5 Filtering effect diagrams of four different wavelet threshold functions in the embodiments of the present invention;

[0049] Figure 6 This is a comparison chart of the prediction results of four neural networks with quantile algorithm added in the embodiment of the present invention;

[0050] Figure 7 Schematic diagram of four different evaluation indicators in an embodiment of the present invention. DETAILED DESCRIPTION

[0051] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0052] The present invention proposes a method for predicting ash accumulation on the heating surface of an economizer, such as Figure 1 As shown, the specific steps include:

[0053] S1. The real-time operation data of the economizer is obtained through the distributed control system DCS of the coal-fired power plant. The obtained data is subjected to a series of numerical calculations to obtain the clean factor data set (CF) that can characterize the heat transfer efficiency of the heated surface of the economizer. After analysis, the soot blowing cycle and related rules are intercepted to prepare for modeling, soot accumulation section 1, soot blowing section, and soot accumulation section 2. Among them, the sampling period of the distributed control system (DCS) data collection is once every 5 seconds, and the total sampling hours are 24 hours.

[0054] S2. Use polynomial fitting method to smooth the cleaning factor data set and convert the original sequence into a linear sequence to facilitate the calculation of heat loss area. If the nonlinear non-stationary sequence is directly integrated, it will lead to large errors. The expression of the fitting curve is as follows:

[0055] y=a0+a1x+a2x 2 +a3x 3 +a4x 4

[0056] Wherein, a0, a1, a2, a3, and a4 represent coefficients of different terms of the polynomial; x and y represent time and the corresponding cleaning factor, respectively.

[0057] The fitted expressions of dust accumulation section 1 (y1), soot blowing section (y2), and dust accumulation section 2 (y3) are:

[0058] y1=0.78499-1.05678×10 -5 x 1 +8.63269×10 -10 x 2 -4.37152×10 -14 x 3 +6.99596×10 -19 x 4

[0059] y2=1416.15676-0.16281x 1 +7.02413×10 -6 x 2 -1.34737×10 -10 x 3 +9.69771×10 - 16 x 4

[0060] y3=161.41513-0.01616x 1 +6.09207×10 -7 x 2 -1.0198×10 -11 x 3 +6.39346×10 -17 x 4

[0061] like Figure 2 The figure shows the change of the cleaning factor of the original sooting section 1, the soot blowing section, and the sooting section 2 over time. The area enclosed by the straight line where the original soot blowing starting point x1=32725 and the soot blowing end point x2=36660, the fitting curves y1, y2, y3, and the boundary x=0, x=44690, y=1 is defined as the heat loss area A of the economizer heating surface. g , where the soot blowing start point and soot blowing end point refer to the time nodes of the soot accumulation-soot blowing-soot accumulation process, where the intersection of soot accumulation section 1 and soot blowing section is the soot blowing start point, and the intersection of soot blowing section and soot accumulation section 2 is the soot blowing end point, such as Figure 3 The figure shows the schematic diagram of the fitting and modeling process of the original ash accumulation section 1, the ash blowing section, and the ash accumulation section 2.

[0062] S3, set the originally known soot blowing starting point x1 and soot blowing end point x2 as unknown, and take the heat loss area A gAs the objective function, x1 and x2 are solved by the improved subtraction average algorithm (GRSABO), and x1 and x2 are moved on the x-axis under the constraints of two nonlinear constraints: x1 ≥ 28800, 4500 ≤ x2-x1 ≤ 5400 (unit: seconds), so that A g The value of is the smallest. Get A g The minimum soot blowing starting point x1 and the corresponding soot blowing threshold. The algorithm is improved according to the actual problem, and the segmented chaotic mapping, golden section interval search strategy and roulette strategy are added to the traditional subtraction average algorithm. This makes the algorithm converge more accurately and faster when solving the target problem. The heat loss model of the economizer heating surface is solved by the improved subtraction average algorithm, and the soot blowing starting point and end point when the heat loss area of ​​the economizer heating surface is the minimum are obtained, including the following steps:

[0063] (1) Initialize the algorithm parameters, set the initial number of search agents to 30, the maximum number of iterations to 100, the heat loss area as the fitness function, the upper bound to 44690, the lower bound to 28800, and the output dimension to 2 (the soot blowing start point and the soot blowing end point).

[0064] (2) Piecewise chaotic mapping is introduced to initialize the search agent position (the position of the initial solution of the soot blowing starting point and the soot blowing end point), and the fitness function is used to evaluate the fitness of each individual after initialization. The individual refers to the candidate solution of the soot blowing starting point and the soot blowing end point.

[0065] (3) The golden sine algorithm is introduced. First, the golden section coefficient is introduced to enable the algorithm to achieve a good balance between global search and local development capabilities. This balance helps the algorithm to fully explore the entire search space during the search process, and to conduct detailed searches near the current optimal solution, thereby increasing the probability of finding the global optimal solution. Secondly, the periodic fluctuation characteristics of the sine function can be used to quickly traverse different areas in the search space. This characteristic helps the algorithm avoid falling into the local optimal solution during the search process, thereby improving the global search capability. By quickly searching for feasible solutions of the unit circle near the optimal solution, the golden sine strategy can accelerate the convergence of the algorithm.

[0066] (4) The roulette strategy is introduced to increase the probability of individuals with high fitness being selected, thereby avoiding the difficulty of selecting the optimal solution at the boundary due to the insensitivity of the traditional optimization algorithm to the boundary. In each iteration of the improved subtraction average algorithm, the roulette algorithm assigns different probabilities according to the fitness of the individual. The individual with high fitness is more likely to be selected, and then the individual with the lowest fitness is eliminated. After several cycles, the probability of the individual representing the optimal solution being selected will be greatly increased.

[0067] (5) After reaching the program termination condition (reaching the maximum number of iterations or the fitness value meets the requirements), compare whether the heat loss area corresponding to each current search agent is the minimum, run the improved subtraction average algorithm multiple times to compare the optimal solution and the corresponding fitness value each time, and select the soot blowing start point and soot blowing end point corresponding to the minimum heat loss area as the output.

[0068] S4. Substitute the soot blowing starting point and the corresponding soot blowing threshold obtained in step 3 into the original data set and intercept 1000 sets of data to obtain the target soot accumulation section for interval prediction. The target soot accumulation section is denoised by wavelet threshold method and ensemble empirical decomposition (EEMD). The intrinsic mode components obtained by decomposition are classified using t-test to obtain three types of features: high-frequency component, low-frequency component and trend item. The ratio of the standard deviation of the additional white noise decomposed by EEMD to the standard deviation of the signal is 0.2 and the number of times the signal is averaged is 100.

[0069] The test indicators of the t-test are divided as follows: IMF1 is indicator 1, IMF1+IMF2 is indicator 2, and so on. It is determined whether the t values ​​of indicator 1 and indicator 2 are significantly different from 0, whether indicators 2 and indicators 3 are significantly different from 0, and so on.

[0070] S5. Divide the ratio of training set and test set into 8:2, bring the three types of features and the denoised target sequence into the bidirectional gated recurrent neural network (BiGRU) based on quantile regression (QR) for interval prediction, and judge the accuracy of prediction based on evaluation indicators such as mean absolute error (MAE) and mean square error (MSE). A bidirectional information flow mechanism is introduced to the traditional gated recurrent neural network (GRU), which improves the model's context perception and ability to process long sequence data; the addition of quantile regression algorithm and confidence interval can better characterize the uncertainty of interval prediction. The bidirectional gated recurrent neural network divides the training set and test set into a ratio of 8:2. On the premise of having 1,000 sets of original data, the ash accumulation trend can be known about 15 minutes in advance, which is convenient for boiler operation and maintenance personnel and technicians to prepare for soot blowing operations and formulate reasonable soot blowing strategies.

[0071] The high-frequency component, low-frequency component, trend term and target dust accumulation segment after wavelet threshold denoising are used as input, and a bidirectional gated recurrent neural network (BiGRU) based on quantile regression (QR) is used for interval prediction. The method includes the following steps:

[0072] (1) Use quantile regression to estimate the uncertainty of the prediction results and provide quantitative information (confidence interval).

[0073] (2) Dealing with outliers and deviations: Divide the data into several specific quantiles (median, upper quartile, etc.) to enhance the capture of asymmetric effects and reduce the impact of outliers on the overall model.

[0074] (3) Set the quantile absolute deviation loss as the loss function and set τ as the target quantile. This loss function allows the model to make predictions at different quantiles, so that different parts of the prediction distribution can be estimated, such as the median, upper bound, and lower bound. The loss function formula is as follows:

[0075]

[0076] Where τ is the target quantile, usually in the interval (0, 1). k is the actual observed value, is the predicted value, is an indicator function, if Returns 1 if true, otherwise returns 0.

[0077] The parameters in the BiGRU model are the weight W and bias vector b of each neuron. The quantile regression method is introduced into the neural network, and a BiGRU model based on the quantile regression loss function is established. The conditional quantization value of the output response variable Y at the quantile point τ is:

[0078]

[0079] Where J is the number of units in the hidden state, f is the activation function of the output layer, and h j (τ) is the output of the BiGRU hidden state, w j (τ) and b j (τ) are the weight and bias of the output layer respectively.

[0080] By updating the network parameter l according to the gradient descent algorithm (Adam), a BiGRU model with different weights and biases can be obtained. After a series of forward propagation and backward learning, the predicted value of the response variable at each magnitude at time t+h can be obtained:

[0081]

[0082] This allows us to estimate Y t+h|t and confidence intervals for the discrete conditional quartiles.

[0083]

[0084] in, is the prediction interval at the β significance level, and τ are the upper and lower bounds of the prediction interval respectively. When , the confidence level of the interval is 1-β.

[0085] (4) Substitute the data divided into specific quantiles into the bidirectional gated recurrent neural network for prediction to obtain the prediction results.

[0086] The prediction results obtained in step 5 can allow boiler operators to obtain the soot accumulation trend about 15 minutes in advance, making it easier for them to prepare for soot blowing operations in advance and formulate more reasonable soot blowing strategies.

[0087] Finally, two conventional evaluation indicators: Mean Square Error (MSE), Mean Absolute Percentage Error (MAPE) and two interval prediction-specific evaluation indicators: Prediction Interval Coverage Probability (PICP) and Prediction Interval Average Width (PINAW), and a comparison of the prediction effects of multiple neural network combined with quantile algorithms verified the superiority and feasibility of the model proposed in this invention. This provides an effective and comprehensive method for improving the energy efficiency and economic benefits of coal-fired power plant boilers.

[0088] Under the premise of ensuring the safe operation of the equipment, the present invention uses a series of numerical calculations on the measured data of the distributed control system (DCS) of the coal-fired power plant and combines it with relevant thermodynamic laws and formulas to calculate the convective heat transfer efficiency of the economizer heating surface in real time; the whole process modeling of the economizer heat exchange surface can more intuitively analyze the changing trends and internal connections of multiple processes; the existing optimization algorithm is improved according to practical problems, overcoming the difficulty that traditional methods cannot perform quantitative analysis in economizer energy efficiency optimization, and the speed and accuracy of the algorithm in solving the target problem are further improved through improvements; the use of interval prediction for the target sequence better characterizes the uncertainty of the prediction point, which can provide more reaction time for boiler operators to perform soot blowing operations and formulate soot blowing strategies in practical applications.

[0089] Example: The economizer of the boiler of the No. 2 300MW unit of a coal-fired power plant is taken as the research object. The boiler is a subcritical, natural circulation, one-time intermediate reheat, double-arch single furnace, "W" flame combustion mode, tail double flue, flue gas damper temperature adjustment, balanced ventilation, solid slag discharge, open-air layout, full steel frame suspended drum boiler, and is a HG-1025 / 17.3-WM18 type produced by Harbin Boiler Plant. The main design parameters are shown in Table 1.

[0090] Table 1 Boiler design parameters

[0091]

[0092]

[0093] The type of economizer to be implemented is a cast iron economizer, which is a device installed at the bottom of the flue at the tail of the boiler to recover the waste heat of the flue gas. The main heat exchange mode of the economizer is convection heat exchange. After the boiler feed water is heated to saturated water, the heating surface under the pressure of the drum absorbs the heat of the high-temperature flue gas, reduces the flue gas exhaust temperature, saves energy, and improves efficiency, so it is called an economizer. Its main design parameters are shown in Table 2.

[0094] Table 2 Main design parameters of economizer

[0095] parameter unit Numeric Steam flow t\h 220 Inlet water temperature ℃ 274 Furnace tail width m 2.1 Rib tube specifications mm 51×3.5 Water flow rate m / s 1.04 Inlet flue gas temperature ℃ 400 Flue gas velocity m / s 8.13

[0096] Based on the above parameters, a comprehensive method for full-process modeling, optimization and prediction of soot blowing of the heating surface of the economizer of a coal-fired power plant in this embodiment is specifically divided into three modules: full-process modeling, optimization algorithm solution, and interval prediction, as follows:

[0097] (1) Full-process modeling of economizer heating surface:

[0098] Step 1: Obtain real-time data collected by sensors on the economizer side through the distributed control system (DCS) of Unit 2. The sampling period is 5 seconds, with a total sampling period of 24 hours and 17,280 sets of data.

[0099] Step 2: The data obtained in step 1 are analyzed and summarized through the basic laws of thermodynamics and engineering numerical calculations, with a total of 8938 groups of data that are less affected by load fluctuations and environmental factors, of which 0-6545 is the dust accumulation section 1, 6545-7332 is the soot blowing section, and 7332-8938 is the dust accumulation section 2;

[0100] Step 3: Perform polynomial fitting on the three segments of data divided in step 2 to obtain three fitting functions. The function expressions are shown below, and the chi-square statistic (chi-sqr tolerance value) of 1e-09 is used as the fitting convergence indicator.

[0101] Ash accumulation section 1:

[0102] y1=0.78499-1.05678×10 -5 x+8.63269×10 -10 x 2 -4.37152×10 -14 x 3 +6.99596×10- 19 x 4

[0103] Sootblowing section:

[0104] y2=1416.15676-0.16281x+7.02413×10 -6 x2 -1.34737×10 -10 x 3 +9.69771×10 -16 x 4

[0105] Dust accumulation section 2:

[0106] y3=161.41513-0.01616x+6.09207×10 -7 x 2 -1.0198×10 -11 x 3 +6.39346×10 -17 x 4

[0107] Step 4: Set the area enclosed by the fitting curve y1 of the ash accumulation section 1 in step 3 and the boundary x0=0, y0=1, and the original soot blowing starting point x1=32725 as the heat loss area A1 of the ash accumulation section 1; set the area enclosed by the fitting curve y2 of the soot blowing section and the boundary y0=1, the original soot blowing starting point x1=32725, and the original soot blowing end point x2=36660 as the heat loss area A2 of the soot blowing section; set the area enclosed by the fitting curve y3 of the ash accumulation section and the boundary y0=1, x3=44690, and the original soot blowing end point x2=36660 as the heat loss area A3 of the ash accumulation section 2, and the sum of the heat loss areas of the economizer heating surface A g The expression is:

[0108] A g =A1+A2+A3

[0109] Step 5: Set the soot blowing starting point x1 and the soot blowing end point x2 in step 4 as unknown numbers, so that x1 and x2 move on the x-axis under the constraints of two nonlinear constraints x1 ≥ 28800 and 4500 ≤ x2-x1 ≤ 5400 (unit: seconds), so that the heat loss area A of the economizer heating surface g This is both the modeling basis of the heat loss model of the economizer heating surface and the objective function of the following optimization problem. Here we calculate the cross-flow heat transfer efficiency of the economizer heating surface under the original soot blowing node:

[0110]

[0111] (2) Optimization algorithm solution:

[0112] Step 1: First, based on the defects of traditional optimization algorithms and the characteristics of the objective function solution, three areas that need improvement are summarized: (1) The population initialization is not random enough, resulting in a waste of time and space when the population transmits information to each other; (2) The interval search strategy is not efficient enough, and multiple functions are required to limit the search range, resulting in low convergence accuracy and convergence speed; (3) The algorithm is not sensitive to boundary values ​​and is easily pulled into the interval by penalty terms and feasibility conditions.

[0113] Step 2: Make three improvements based on the shortcomings in step 1:

[0114] (1) Add a piecewise chaotic mapping function to initialize the population;

[0115]

[0116] The value range of P is [0,0.5].

[0117] (2) Use the golden section strategy to replace the original interval search strategy;

[0118] c=a+(1-φ))*(ba)

[0119] d=a+φ*(ba)

[0120] φ≈1.618 This is a mathematically fixed ratio. Let f(c) and f(d) be the objective functions, and the expected goal is to find the minimum value on the interval [a,b]. When f(c)<f(d), the new search interval becomes [a,d], otherwise it becomes [c,b]. Repeat the above steps until the length of the search interval is less than the predetermined threshold or the upper limit of the number of iterations is reached.

[0121] (3) To address the problem that the optimization algorithm is insensitive to boundary values, a roulette strategy is used to make fitness inversely proportional to probability, that is, the smaller the fitness, the greater the probability of being selected.

[0122]

[0123] Add multiple processes so that the individual with the smallest probability is eliminated at the end of each process, thereby increasing the probability of the optimal solution being selected.

[0124] (4) Figure 4 The box plots of various optimization algorithms are shown. These three improved methods are added to the subtraction average optimization algorithm, and 20 experiments are conducted to record the optimal solution x1 of the soot blowing starting point, the optimal solution x2 of the soot blowing end point, and the corresponding heat loss area A obtained by the algorithm. gAnd the frequency of the optimal solution being selected. Then, the solutions of the optimization algorithms such as Whale (WOA), Grey Wolf (GWO), Subtraction Average (SABO), Particle Swarm (PSO), Seagull (SOA), and Sparrow (SSA) were compared. Each optimization algorithm was tested 20 times independently. The comparison results are shown in Table 3.

[0125] Table 3 Comparison results

[0126] Optimization Algorithm (frequency) (frequency) (frequency) WOA 29000(5％) 33500(5％) 11324.74(5％) GWO 28890(5％) 33432(5％) 11187.51(5％) SABO 28800(10％) 33300(5％) 11057.36(5％) PSO 28800(5％) 33308(5％) 11058.48(5％) SOA 28800(5％) 33400(5％) 11071.07(5％) SSA 28800(5％) 33378(5％) 11068.12(5％) GRSABO 28800(85％) 33300(85％) 11057.36(85％)

[0127] The optimal sootblowing starting point obtained by the optimization algorithm can already guide boiler operation and maintenance personnel to perform sootblowing operations at the corresponding time. However, knowing only the specific sootblowing time is still not enough. If there is an emergency, the operation and maintenance personnel need enough time to do preparation work. Next, the present invention will bring the optimal sootblowing starting point and the corresponding sootblowing threshold obtained by the optimization module into the original data, and intercept the 1000 groups of data before the sootblowing threshold as the objective function of interval prediction. It truly transforms the time-based sootblowing strategy into a state-based sootblowing strategy. Calculate the front-to-back heat transfer efficiency of the economizer after optimizing the sootblowing node:

[0128]

[0129] The optimization scheme improves the cross-flow heat transfer efficiency of the economizer heating surface:

[0130] 28.7239%-24.7423%=3.9816%;

[0131] (3) Interval prediction based on quantile regression:

[0132] Step 1: De-noise the target function in the summary of module 2 using the wavelet unbiased risk estimation threshold method. Since the optimized data is still mixed with some noise and burrs, it will have a certain impact on the prediction accuracy. Therefore, denoising the target sequence is an indispensable process. The design of different threshold functions also has a great impact on the denoising results. Figure 5 The denoising results of unbiased risk estimation threshold function, heuristic threshold function, fixed threshold function and maximum and minimum threshold function are shown.

[0133] Step 2: Use ensemble empirical mode decomposition (EEMD) on the denoised data, set the ratio of the standard deviation of the additional noise to the standard deviation of the original signal to 0.2, average the signal 100 times, and adaptively decompose the denoised signal into 9 IMF components, including 8 signal components and 1 trend term.

[0134] Step 3: Use the t-test to divide the 9 IMF components in step 2 into three important features: high-frequency components, low-frequency components, and trend items.

[0135] Step 4: Take these three types of features and denoised data as input, use the quantile algorithm to divide this set of input into several specific quantiles, and then divide the ratio of training set and test set into 8:2, and use the bidirectional gated recurrent neural network to perform interval prediction on the denoised data to compare the predicted results with the actual values. Then use the other three neural networks (long-short-term neural network, gated recurrent neural network, bidirectional long-short-term neural network) with quantile regression to compare with the bidirectional gated recurrent neural network, such as Figure 6 As shown. The evaluation indicators (MSE, MAPE, PICP, PINAM) verify that the prediction results of the proposed model QRBiGRU have high accuracy, as shown in Figure 7 The evaluation indicators are shown in Table 4:

[0136] Table 4 Evaluation indicators

[0137]

[0138] In the prediction module, the 1,000 sets of data of the objective function are divided into a training set and a test set in a ratio of 8:2. This ratio allows boiler operation and maintenance personnel to know the ash accumulation trend of the economizer about 15 minutes in advance, thereby gaining more preparation time and formulating a more reasonable sootblowing strategy in the state-based sootblowing strategy.

[0139] In summary, through theoretical calculation and case analysis, the comprehensive method of modeling, prediction and optimization of the whole process of economizer proposed in this invention transforms the traditional time-based fixed sootblowing mode into a state-based intelligent sootblowing model, which not only improves the convection heat transfer efficiency of the economizer heating surface by about 4%, but also provides boiler operation and maintenance personnel with about 15 minutes to prepare for sootblowing operations and specify a more reasonable sootblowing plan. The application of this comprehensive method provides important theoretical and technical support for the improvement of energy efficiency and subsequent economic benefits of coal-fired power plants.

[0140] Based on the same inventive concept, the present invention also proposes an economizer heating surface ash accumulation prediction system, comprising:

[0141] The acquisition module is used to acquire the real-time operation data of the economizer and analyze the real-time operation data, and obtain a curve of the clean factor data set that can characterize the cross-flow heat transfer efficiency of the economizer heating surface over time based on the analysis results.

[0142] The model building module is used to intercept the ash accumulation section y1, soot blowing section y2 and ash accumulation section y3 of the economizer heating surface from the curve of the cleaning factor data set changing with time; based on the soot blowing starting point, soot blowing end point and the fitting curve of the cleaning factor data set, the heat loss model of the economizer heating surface is constructed with the minimum heat loss area as the objective function; the soot blowing starting point is the intersection of the ash accumulation section y1 and the soot blowing section y2, and the soot blowing end point is the intersection of the soot blowing section y2 and the ash accumulation section y3.

[0143] The calculation module is used to solve the heat loss model of the economizer heating surface through an improved subtraction average algorithm, and obtain the soot blowing starting point and soot blowing end point when the heat loss area of ​​the economizer heating surface is the smallest; specifically, it includes introducing a piecewise chaotic mapping function, initializing the positions of the soot blowing starting point and the soot blowing end point, solving the heat loss area as a fitness function, evaluating the fitness corresponding to each soot blowing starting point and the soot blowing end point after initialization, and performing a global search by introducing a golden section search. At the same time, a roulette strategy is introduced in each round of iteration. Different probabilities are assigned according to the fitness corresponding to each soot blowing starting point and the soot blowing end point, and the soot blowing starting point and the soot blowing end point corresponding to the highest probability are screened out, so as to obtain the soot blowing starting point and the soot blowing end point corresponding to the minimum heat loss area.

[0144] The prediction module is used to obtain the target ash accumulation section of interval prediction in the curve of the cleaning factor data set changing with time according to the soot blowing starting point and soot blowing end point corresponding to the minimum heat loss area; and predict the ash accumulation interval of the economizer heating surface according to the target ash accumulation section of interval prediction.

[0145] The present invention also proposes a computer device for predicting ash accumulation on the heating surface of an economizer, comprising: a memory, a processor, and a computer program stored in the memory. When the processor executes the computer program, the steps of the method for predicting ash accumulation on the heating surface of an economizer are implemented.

[0146] The present invention also proposes a readable storage medium, which stores a computer program. The computer program includes program instructions. When the program instructions are executed by a processor, they are used to execute the steps of the ash accumulation prediction method for the economizer heating surface.

[0147] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for predicting ash accumulation on a heating surface of an economizer, characterized in that: The following steps are involved: Acquire the real-time operation data of the economizer, analyze the real-time operation data, and obtain a curve of the clean factor data set that can characterize the cross-flow heat transfer efficiency of the economizer heating surface over time according to the analysis results; The soot accumulation section y1, soot blowing section y2 and soot accumulation section y3 of the economizer heating surface are intercepted from the curve of the clean factor data set changing with time; based on the soot blowing starting point, soot blowing end point and the fitting curve of the clean factor data set, the heat loss model of the economizer heating surface is constructed with the minimum heat loss area as the objective function; the soot blowing starting point is the intersection of the soot accumulation section y1 and the soot blowing section y2, and the soot blowing end point is the intersection of the soot blowing section y2 and the soot accumulation section y3; The heat loss model of the economizer heating surface is solved by the improved subtraction average algorithm, and the soot blowing starting point and end point when the heat loss area of ​​the economizer heating surface is the smallest are obtained. Specifically, it includes introducing a piecewise chaotic mapping function, initializing the positions of the soot blowing starting point and the soot blowing end point, solving the heat loss area as the fitness function, evaluating the fitness of each soot blowing starting point and the soot blowing end point after initialization, and performing a global search by introducing the golden section search. At the same time, the roulette strategy is introduced in each round of iteration. Different probabilities are assigned to each soot blowing starting point and the soot blowing end point according to their corresponding fitness, and the soot blowing starting point and the soot blowing end point corresponding to the highest probability are selected, so as to obtain the soot blowing starting point and the soot blowing end point corresponding to the minimum heat loss area; According to the soot blowing starting point and end point corresponding to the minimum heat loss area, the target ash accumulation section of interval prediction is intercepted in the curve of the cleaning factor data set changing with time; according to the target ash accumulation section of interval prediction, the ash accumulation interval of the economizer heating surface is predicted.

2. The method for predicting ash deposition on the heating surface of an economizer according to claim 1, characterized in that: The method of intercepting the ash accumulation section y1, the soot blowing section y2 and the ash accumulation section y3 of the economizer heating surface from the curve of the clean factor data set changing with time specifically includes: smoothing the clean factor data set by a polynomial fitting method, and the fitting curve of the clean factor data set is expressed as: y=a0+a1x+a2x 2 +a3x 3 +a4x 4 The fitting curves of the ash accumulation section y1, the soot blowing section y2, and the ash accumulation section y3 are respectively expressed as: y1=0.78499-1.05678×10 -5 x 1 +8.63269×10 -10 x 2 -4.37152×10 -14 x 3 +6.99596×10 -19 x 4 y2=1416.15676-0.16281x 1 +7.02413×10 -6 x 2 -1.34737×10 -10 x 3 +9.69771×10 -16 x 4 y3=161.41513-0.01616x 1 +6.09207×10 -7 x 2 -1.0198×10 -11 x 3 +6.39346×10 -17 x 4 Wherein, a0, a1, a2, a3, and a4 represent coefficients of different terms of the polynomial; x and y represent time and the corresponding cleaning factor, respectively.

3. The method for predicting ash deposition on the heating surface of an economizer according to claim 1, characterized in that: The method of predicting the ash accumulation interval of the economizer heating surface according to the target ash accumulation interval of the interval prediction includes the following steps: The target dust accumulation section is decomposed by using the ensemble empirical mode decomposition method to obtain high-frequency components, low-frequency components and trend items. The high-frequency component, low-frequency component, trend term and target ash accumulation section after wavelet threshold denoising are used as the input of a bidirectional gated recurrent neural network based on quantile regression to obtain the ash accumulation range of the economizer heating surface.

4. A method for predicting ash deposition on a heating surface of an economizer according to claim 3, characterized in that: The target dust accumulation section is decomposed by using the ensemble empirical mode decomposition method, wherein the ratio of the decomposed additional white noise standard deviation to the signal standard deviation is 0.2, and the number of times the signal is averaged is 100.

5. The method for predicting ash deposition on the heating surface of an economizer according to claim 3, characterized in that: The high-frequency component, the low-frequency component, the trend term and the target ash accumulation section after wavelet threshold denoising are used as the input of the bidirectional gated recurrent neural network based on quantile regression to obtain the ash accumulation interval of the economizer heating surface, which specifically includes the following steps: The input data is divided into multiple specific quantiles; the quantile absolute deviation loss is set as the loss function, and its loss function is expressed as: Where τ is the target quantile, y k is the actual observed value, is the predicted value, is the indicator function; Substitute the data divided into specific quantiles into the bidirectional gated recurrent neural network for prediction, and obtain the ash accumulation range of the economizer heating surface, including: A BiGRU model based on the quantile regression loss function is established, and the conditional quantization value of its output response variable Y at the quantile point τ is: Where J is the number of units in the hidden state, f is the activation function of the output layer, and h j (τ) is the output of the BiGRU hidden state, w j (τ) and b j (τ) are the weight and bias of the output layer respectively; the model parameter l is updated according to the gradient descent algorithm Adam to obtain the BiGRU model with different weights and biases; after a series of forward propagation and backward learning, the predicted value of the response variable at each magnitude at time t+h is obtained: Estimation of T t+h|t And confidence intervals for discrete conditional quartiles: in, is the prediction interval at the β significance level, and τ are the upper and lower bounds of the prediction interval respectively. When , the confidence level of the interval is 1-β.

6. The method for predicting ash deposition on the heating surface of an economizer according to claim 1, characterized in that: The distributed control system DCS is used to obtain the real-time operation data of the economizer, with a sampling period of once every 5 seconds, for a total of 24 hours.

7. A system for predicting ash accumulation on a heating surface of an economizer, comprising: An acquisition module is used to acquire the real-time operation data of the economizer, analyze the real-time operation data, and obtain a curve of the clean factor data set that can characterize the cross-flow heat transfer efficiency of the economizer heating surface over time according to the analysis result; A model building module is used to intercept the ash accumulation section y1, the soot blowing section y2 and the ash accumulation section y3 of the economizer heating surface from the curve of the clean factor data set changing with time; based on the soot blowing starting point, the soot blowing end point and the fitting curve of the clean factor data set, the heat loss model of the economizer heating surface is constructed with the minimum heat loss area as the objective function; the soot blowing starting point is the intersection of the soot accumulation section y1 and the soot blowing section y2, and the soot blowing end point is the intersection of the soot blowing section y2 and the ash accumulation section y3; A calculation module is used to solve the heat loss model of the economizer heating surface by using an improved subtraction average algorithm to obtain the soot blowing starting point and soot blowing end point when the heat loss area of ​​the economizer heating surface is the smallest; Specifically, it includes introducing a piecewise chaotic mapping function, initializing the positions of the soot blowing starting point and the soot blowing end point, solving the heat loss area as the fitness function, evaluating the fitness of each soot blowing starting point and the soot blowing end point after initialization, and performing a global search by introducing the golden section search. At the same time, the roulette strategy is introduced in each round of iteration. Different probabilities are assigned to each soot blowing starting point and the soot blowing end point according to their corresponding fitness, and the soot blowing starting point and the soot blowing end point corresponding to the highest probability are selected, so as to obtain the soot blowing starting point and the soot blowing end point corresponding to the minimum heat loss area; The prediction module is used to obtain the target ash accumulation section of interval prediction in the curve of the cleaning factor data set changing with time according to the soot blowing starting point and soot blowing end point corresponding to the minimum heat loss area; and predict the ash accumulation interval of the economizer heating surface according to the target ash accumulation section of interval prediction.

8. A computer device for predicting ash accumulation on the heating surface of an economizer, characterized in that: include: A memory, a processor and a computer program stored in the memory, wherein the processor implements the steps of the method for predicting ash accumulation on the heating surface of an economizer as described in any one of claims 1 to 6 when executing the computer program.

9. A readable storage medium, characterized in that: The readable storage medium stores a computer program, which includes program instructions. When the program instructions are executed by a processor, they are used to execute the steps of the method for predicting ash deposition on the heating surface of an economizer according to any one of claims 1 to 6.

Images