Coal-fired power plant boiler heating surface ash prediction method

Abstract

The invention discloses a coal-fired power plant boiler heating surface soot blowing prediction method, and relates to the field of coal-fired power plant boiler heating surface soot blowing prediction. Comprising the following steps: acquiring a coal-fired power plant boiler heating surface all-day cleaning factor data set, and intercepting ash deposition section cleaning factor time sequence data as original cleaning factor time sequence data; injecting white noise into the original cleaning factor time sequence data by adopting improved self-adaptive noise complete set empirical mode decomposition, constructing a to-be-decomposed sequence, and decomposing the to-be-decomposed sequence to obtain an intrinsic mode function component and a residual component; predicting each component by using a bidirectional recurrent neural network prediction model to obtain a predicted value of each component; and superposing the predicted values of the components to obtain a final predicted value for representing the ash condition of the heating surface of the boiler. The improved self-adaptive noise complete set empirical mode decomposition is combined with the bidirectional recurrent neural network, the soot condition of the boiler heating surface can be accurately predicted, and guidance is provided for soot blowing operation of a coal-fired power plant.

Description

Technical Field

[0001] The present invention relates to the field of prediction of soot blowing on heating surfaces of coal-fired power plant boilers, and in particular to a prediction method for soot blowing on heating surfaces of coal-fired power plant boilers. Background Art

[0002] At present, in coal-fired power plants, the efficient operation of the boiler is the key to ensuring the smoothness of the entire power generation process. The stability of the boiler's performance is directly related to the cleanliness of its heating surface. During the operation of the power plant, the heat generated by the combustion of coal is transferred to the heating surface, heating the working fluid to generate steam, which in turn drives the generator set. However, the ash produced during the combustion of coal may melt under high temperature conditions, forming a molten state, and flow to the heating surface with the flue gas, causing ash accumulation. Ash accumulation not only increases thermal resistance, resulting in the need to consume more raw coal on the working fluid side to maintain the predetermined temperature, but may also cause a series of problems, such as reduced heat exchange efficiency, accelerated metal corrosion, and even unplanned shutdown of the entire unit, which seriously affects the safety and economic benefits of the power plant.

[0003] In the prior art, the cleanliness of the boiler heating surface is predicted by analyzing the incremental distribution of ash accumulation in different time periods and combining it with the initial cleanliness state. Another strategy is to parse the historical contamination rate data into a fitting curve and its deviation from the original data, and then build a prediction model in combination with the latest contamination rate data.

[0004] However, these methods have limitations. Simply superimposing the incremental distributions at different time points may ignore nonlinear effects and complex patterns in time series, limiting the prediction accuracy. Decomposing historical data into fitting curves and deviations, although able to capture trends and fluctuations, may not be accurate enough when dealing with high-frequency noise and non-stationary characteristics. In particular, when the system is subject to external disturbances, the adaptability and generalization ability of the model may decrease, resulting in limitations in actual engineering applications.

[0005] Therefore, further research and development of more accurate sootblowing prediction models are needed. Summary of the invention

[0006] Based on this, it is necessary to provide a soot blowing prediction method for the heating surface of a coal-fired power plant boiler to address the above technical issues.

[0007] The present invention adopts the following technical solutions:

[0008] The present invention provides a method for predicting soot blowing of a heating surface of a coal-fired power station boiler, comprising:

[0009] Obtain the time series data of the cleaning factor of the ash accumulation section of the coal-fired power plant boiler as the original time series data of the cleaning factor;

[0010] An improved adaptive noise complete ensemble empirical mode decomposition algorithm is used to inject white noise into the original clean factor time series data to construct the time series data to be decomposed, and the time series data to be decomposed is decomposed to obtain a series of intrinsic mode function components and a residual component; wherein the improved adaptive noise complete ensemble empirical mode decomposition algorithm is obtained by an adaptive white noise injection strategy based on the ensemble empirical mode decomposition algorithm;

[0011] A bidirectional recurrent neural network prediction model is used to predict a series of intrinsic mode function components and a residual component to obtain predicted values ​​of a series of intrinsic mode function components and a residual component; wherein the bidirectional recurrent neural network prediction model is obtained by training a bidirectional recurrent neural network, and the bidirectional recurrent neural network is a recursive neural network with information flow in two directions constructed based on the Elman recurrent neural network;

[0012] The predicted values ​​of a series of intrinsic mode function components and the predicted value of a residual component are superimposed to obtain the final predicted value representing the ash accumulation condition of the boiler heating surface.

[0013] The acquisition of the original cleaning factor time series data specifically includes the following steps:

[0014] The real-time heat transfer coefficient K of the boiler heating surface is obtained through the boiler distributed control system DCS throughout the day f ;

[0015] According to the real-time heat transfer coefficient K f and the theoretical heat transfer coefficient K of the boiler heating surface 0 The ratio of , obtains the real-time cleaning factor time series data for the whole day;

[0016] The real-time cleaning factor time series data of the dust accumulation section is intercepted from the real-time cleaning factor time series data of the whole day, and the real-time cleaning factor time series data of the dust accumulation section is used as the original cleaning factor time series data.

[0017] The improved adaptive noise complete ensemble empirical mode decomposition algorithm is obtained by an adaptive white noise injection strategy based on the ensemble empirical mode algorithm, and the specific steps include:

[0018] Inject white noise into the original clean factor time series data according to the preset noise standard deviation;

[0019] Performing ensemble empirical mode decomposition on the data injected with white noise to obtain a first group of ensemble empirical mode components, and using the first group of empirical mode components as a first group of adaptive white noise in improved adaptive complete ensemble empirical mode decomposition;

[0020] Injecting the first group of adaptive white noise into the original clean factor time series data and performing ensemble empirical mode decomposition and local averaging processing to obtain the first group of intrinsic mode function components and the first group of residual components;

[0021] Continue to perform ensemble empirical mode decomposition on the first group of residual components to obtain a second group of ensemble empirical mode components, and use the second group of empirical mode components as the second group of adaptive white noise in the improved adaptive complete ensemble empirical mode decomposition;

[0022] Injecting a second group of adaptive white noise into the first group of residual components, and again performing ensemble empirical mode decomposition and local averaging to obtain a second group of intrinsic mode function components and a second group of residual components;

[0023] Adaptive white noise is repeatedly injected into the previous set of residual components, and the decomposition is completed through ensemble empirical mode decomposition and local averaging until the residual components show a monotonic trend, so as to form an improved adaptive noise complete ensemble empirical mode decomposition algorithm.

[0024] The preset noise standard deviation is 0.2, the number of times the white noise is injected is 50, and the signal-to-noise ratio of the original cleaning factor time series data and the added white noise is 1.

[0025] The specific steps of obtaining the intrinsic mode function components and the residual components include:

[0026] Construct k times of time series data to be decomposed:

[0027] cf(i)=x(t)+β 0 E i (w(i))

[0028] In the formula, cf(i) is the time series data to be decomposed, i = 1, 2, 3, ..., k, x(t) is the original cleaning factor time series data, β 0 is the white noise weight coefficient, β 0 =ε 0 std(cf(i)), ε 0 is the noise amplitude, std(cf(i)) is the noise standard deviation of the time series data to be decomposed, E i (.) represents the i-th order modal component generated by the ensemble empirical mode decomposition, w(i) is the white noise injected during the i-th ensemble empirical mode decomposition, E i (w(i)) is the empirical mode component generated by the ensemble empirical mode decomposition after the i-th noise injection;

[0029] According to the k-times time series data to be decomposed, the residual components are determined:

[0030]

[0031] In the formula, RES i is the residual component obtained after the i-th decomposition, M(.) is the local mean operation, β i-1 is the white noise weight coefficient, β i-1 =ε 0 std(RES i-1 ), ε 0 is the noise amplitude, std(RES i-1 ) is the noise standard deviation of the i-1th residual component, E i (.) represents the i-th order modal component generated by the ensemble empirical mode decomposition, w(i) is the white noise injected during the i-th ensemble empirical mode decomposition, E i (w(i)) is the empirical mode component generated by the ensemble empirical mode decomposition after the i-th noise injection;

[0032] According to the residual components, the intrinsic mode function components are determined:

[0033] IMF i =RES i-1 -RES i

[0034] In the formula, IMF i The intrinsic mode function components obtained by the i-th decomposition, RES i-1 and RES i are the residual components obtained from the i-1th and ith decompositions respectively;

[0035] Until the decomposition is completed, a series of intrinsic mode function components and residual components obtained from the last decomposition are obtained.

[0036] The bidirectional recurrent neural network prediction model is used to predict a series of intrinsic mode function components and a residual component, specifically including:

[0037] Part of the data of the intrinsic mode function component and the residual component are respectively intercepted as training sets, and the bidirectional recurrent neural network is trained using the training sets to obtain a bidirectional recurrent neural network prediction model;

[0038] The bidirectional recurrent neural network prediction model includes an input layer, a hidden layer, a receiving layer and an output layer;

[0039] Input any intrinsic mode function component or residual component into the input layer;

[0040] The modal function component or residual component received by the input layer is transferred to the hidden layer, the hidden layer is connected to the feedback link of the receiving layer, the hidden layer learns the nonlinear characteristics and complex patterns of the modal function component or residual component through a nonlinear activation function and transfers them to the receiving layer for storage, and the nonlinear characteristics and complex patterns are used as hidden layer outputs;

[0041] The output layer receives the hidden layer output and outputs the predicted value of the modal function component or the residual component through a linear activation function.

[0042] The bidirectional recurrent neural network includes a forward Elman recurrent neural network and a reverse Elman recurrent neural network; the forward Elman recurrent neural network starts from the first element of the sequence and processes step by step until the last element of the sequence, and the reverse Elman recurrent neural network starts from the end of the sequence and traces back to the starting position; the output of the bidirectional recurrent neural network is the fusion of the outputs of the Elman recurrent neural networks in two directions.

[0043] The number of hidden layer neurons of the bidirectional recurrent neural network is 5, the number of iterations is 5000, and the training target error is 0.0005.

[0044] The fusion method includes: splicing or weighted averaging.

[0045] The step of superimposing a series of predicted values ​​of intrinsic modal components and a predicted value of a residual component specifically includes:

[0046] The total number of a series of intrinsic mode function components and residual components decomposed is k, and the predicted value of any component is Y i , where i = 1, 2, ..., k, and the set of predicted values ​​of all components is {Y 1 ,Y 2 ,…,Y k};

[0047] The predicted values ​​of all components are superimposed as follows:

[0048] At least one of the above technical solutions adopted by the present invention can achieve the following beneficial effects:

[0049] The present invention introduces a cleanliness factor as an indicator to characterize the health status of the boiler heating surface, adopts an improved complete ensemble empirical mode decomposition algorithm with adaptive noise (ICEEMDAN), and significantly reduces the error in the signal reconstruction process through a fine-tuned adaptive noise injection strategy, and cleverly reduces the computational resource consumption of the algorithm, making the entire decomposition process more efficient and accurate.

[0050] A bidirectional recurrent neural network (Bidirectional Elman Neural Network, BIElman) is used to learn the decomposed intrinsic mode function components and residual components. This network architecture takes into account past information while forward propagating. It can not only capture the long-term dependencies of time series, but also enhance the understanding of sequence patterns through bidirectional information flow.

[0051] The present invention adopts a method combining an improved adaptive complete empirical mode decomposition algorithm and a bidirectional recurrent neural network (BIElman-ICEEMDAN), which can not only accurately identify and adapt to the nonlinear characteristics and complex frequency characteristics of the cleaning factor time series data, effectively process the high and low frequency noise and non-stationary characteristics in the data, and improve the data quality, but also accurately and comprehensively capture the potential information of the cleaning factor time series data, and improve the generalization ability and adaptability of the model when subjected to external interference, thereby achieving high-precision prediction of the ash condition of the boiler heating area, providing optimization guidance for the soot blowing operation of the thermal power plant, thereby improving the soot blowing efficiency, reducing energy waste, extending the life of the equipment, and ultimately improving the overall economic benefits and safety of the coal-fired power plant. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0053] Figure 1 A schematic diagram of the overall process of ash accumulation prediction of a soot blowing prediction method for a heating surface of a coal-fired power station boiler provided by the present invention;

[0054] Figure 2 A degradation curve of the cleaning factor of an economizer of a sootblowing prediction method for a heating surface of a coal-fired power station boiler provided by the present invention;

[0055] Figure 3 A degradation curve of the cleanliness factor of the economizer ash accumulation section of a soot blowing prediction method for a heating surface of a coal-fired power station boiler provided by the present invention;

[0056] Figure 4 The clean factor data set of the sootblowing prediction method for the heating surface of a coal-fired power station boiler provided by the present invention is the result of ICEEMDAN decomposition.

[0057] Figure 5 A bidirectional Elman neural network structure diagram of a sootblowing prediction method for a heating surface of a coal-fired power plant boiler provided by the present invention;

[0058] Figure 6 A comparison chart of prediction results of a sootblowing prediction method for a heating surface of a coal-fired power station boiler provided by the present invention;

[0059] Figure 7 A comparison and enlarged diagram of the prediction results of a sootblowing prediction method for a heating surface of a coal-fired power station boiler provided by the present invention;

[0060] Figure 8 A MAPE comparison diagram of a sootblowing prediction method for a heating surface of a coal-fired power plant boiler provided by the present invention;

[0061] Fig. 9 A schematic diagram of RMSE comparison of a sootblowing prediction method for a heating surface of a coal-fired power plant boiler provided by the present invention. DETAILED DESCRIPTION

[0062] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the specific embodiments of the present invention and the corresponding drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0063] At present, in actual engineering, the sootblowing prediction method of the heating surface of the boiler in coal-fired power plants is not accurate enough when dealing with high-frequency noise and non-stationary characteristics. Especially when the system is subject to external disturbances, the adaptability and generalization ability of the model may decrease.

[0064] The technical solutions provided by various embodiments of the present invention are described in detail below in conjunction with the accompanying drawings.

[0065] This example takes the boiler of Unit 2 of a coal-fired thermal power plant in Guizhou as the research object. The data set comes from the economizer of the coal-fired boiler, and the cleaning factor data of the distributed control system DCS detection system of the unit throughout the day is taken as the research object. Figure 1 The present invention is a flow chart of a sootblowing prediction method for a heating surface of a coal-fired power plant boiler, which specifically includes the following steps:

[0066] S101: Obtain a full-day cleaning factor data set through a DCS system of a coal-fired power plant, and extract the cleaning factor time series data of the dust accumulation section as the original cleaning factor time series data.

[0067] The real-time heat transfer coefficient K of the boiler heating surface is obtained through the boiler distributed control system DCS throughout the day f , according to the real-time heat transfer coefficient K f and the theoretical heat transfer coefficient K of the boiler heating surface 0 The real-time cleaning factor time series data of the whole day period is obtained by the ratio of , and the real-time cleaning factor time series data of the dust accumulation period is intercepted from the real-time cleaning factor time series data of the whole day period as the original cleaning factor time series data.

[0068] Using the clean factor as an indicator of the health of the heating surface can well reflect the real-time ash accumulation status. In fact, the daily variation curve of the clean factor is highly nonlinear. One of the reasons is the random noise generated by the environment when the economizer is operating normally. This noise can be eliminated by a reasonable denoising algorithm. The clean factor curve of the economizer before denoising and its corresponding load are shown in Figure 1. Figure 1 As shown;

[0069] It should be noted that S1 is not an effective soot blowing point, but S2 is. The S1 region is due to the surge in boiler load during this period, resulting in an effect similar to soot blowing. The soot accumulation section used in the present invention to verify the proposed model is the soot accumulation section between S2 and S3. Figure 2 As shown, the cleaning factor has a normal variation trend in this area.

[0070] S102: Decomposing the acquired original cleaning factor time series data by using the ICEEMDAN algorithm to obtain intrinsic mode function components (IMF components) and residual components.

[0071] Directly using deep learning models for prediction may result in poor accuracy and stability in time series prediction because the model is difficult to simultaneously capture and adapt to all the complex frequency characteristics in time series data. By using the improved adaptive noise complete set empirical mode decomposition algorithm to decompose time series data, time series data can be decomposed into multiple subsequences that are easier to manage and analyze, thereby improving the performance of the prediction model.

[0072] The improved adaptive noise complete ensemble empirical mode decomposition algorithm ICEEMDAN is used to inject white noise into the original clean factor time series data and construct the time series to be decomposed. The calculation process of decomposing the time series to be decomposed and obtaining the IMF component and residual component specifically includes:

[0073] For the economizer original cleaning factor time series data obtained in this example, the optimal algorithm parameters need to be determined before decomposition. After parameter debugging tests, the optimal parameters of the ICEEMDAN algorithm are set, including: the noise standard deviation is 0.2, the number of noise additions is 50, and the signal-to-noise ratio is 1.

[0074] For the first set of ICEEMDAN decomposition, the first set of white noise is injected into the original clean factor time series data, and the formula cf(i) = x(t) + β 0 E i (w(i)) constructs the first set of decomposition sequences cf(1), takes the average value of cf(1), and obtains the first set of residual components RES 1 =M(cf(1)), through the formula IMF 1 =x(t)-RES 1 Get the first group of intrinsic mode function components IMF 1 .

[0075] Among them, cf(i) is the time series data to be decomposed, i = 1, 2, 3, ..., k, x(t) is the time series data of the cleaning factor, β 0 is the white noise weight coefficient, E i (.) represents the i-th order modal component generated by the ensemble empirical mode decomposition, w(i) is the white noise injected during the i-th ensemble empirical mode decomposition;

[0076] For the second group of ICEEMDAN decomposition, the first group of residual signals is used as the benchmark, and the second group of adaptive white noise is injected to construct the second group of time series data to be decomposed cf(2). i =M(RES i-1 +β i E i The second set of residual components RES is obtained by local mean decomposition of (w(i)) 2 , and finally through the formula IMF 2 =RES 1 -RES 2 The second group of intrinsic mode function components IMF is calculated 2 .

[0077] Repeat the above decomposition steps until the residual component obtained by decomposition shows a monotonic trend, and then stop the decomposition. For the time series data of the cleaning factor of the economizer dust accumulation section training set obtained in this example, after continuous iteration, the number of decomposition groups k = 10 is obtained, that is, a total of 10 groups of intrinsic mode function components IMF are obtained. 1 ~IMF 10 And the residual component RES obtained from the last decomposition 10 .

[0078] S103: Use the bidirectional recurrent neural network BIElamn prediction model to predict each component. The network structure diagram is as follows Figure 4 As shown, the input layer, hidden layer and output layer are static layers, and the receiving layer is a dynamic layer.

[0079] Each component is used as the input of the prediction model to obtain the predicted value of each component.

[0080] The predicted values ​​of all components are superimposed to obtain the final predicted value of the cleaning factor, which is used to characterize the ash accumulation condition of the boiler heating surface.

[0081] Assume that the length of the cleaning factor time series data input to the forward Elman network is n, u F =IMF i ={a i1 ,a i2 ,…,a in}, the clean factor time series data input to the reverse Elman network is u P =IMF' i ={a in ,a in-1 ,…,a i1}, the hidden layer state space expressions of the forward network and the reverse network are:

[0082]

[0083] In the formula, x F is the hidden layer activation state in the forward network, x P is the hidden layer activation state in the reverse network, f is the nonlinear activation function, w xuF and w xuP is the weight coefficient from the input layer to the hidden layer in the forward and reverse networks, is the weight coefficient from the receiving layer to the hidden layer in the forward and reverse networks, is the weight coefficient from the reverse network receiving layer to the forward network hidden layer, is the weight coefficient from the forward network receiving layer to the reverse network hidden layer, x cF (t-1) and x cP (t-1) are the activation states of the receiving layers in the forward network and the reverse network at the previous time step, b uF and b uP are the bias items of the hidden layers in the positive and negative directions respectively;

[0084] The state space expressions of the forward network and the reverse network are:

[0085] x cF =f 2 (x F +bcF )

[0086] x cP =f 2 (x P +b cP )

[0087] In the formula, x cF and x cP are the activation states of the receiving layers of the forward network and the reverse network at the current time step, respectively, 2 is the linear activation function of the receiving layer, b cF and b cP They are the bias items of the forward network and the reverse network receiving layer respectively;

[0088] The final output of the bidirectional recurrent neural network BIElman is the fusion of the outputs of the Elman network in the positive and negative directions. Through the splicing of the state matrix, the calculation steps of the final prediction result are:

[0089] y F =f 3 (w yuF x F +b yF )

[0090] y P =f 3 (w yuP x P +b yP )

[0091] Y i =[y F ;y P ]

[0092] In the formula, y F and P are the outputs of the forward network and the reverse network, respectively, i is the activation state matrix of the output layer in both the forward and reverse directions, i.e., the predicted value of each component, f 3 is the output layer activation function, w yuF and w yuP x P are the weight coefficients from the hidden layer to the output layer of the forward network and the reverse network, respectively, yF and b yP They are the bias terms of the output layers of the forward network and the reverse network respectively;

[0093] The predicted values ​​of each component are superimposed to obtain the final predicted value that can characterize the dust condition of the boiler heating area.

[0094] The monitoring data of the unit in the ash accumulation section of 510 minutes is used as sample data, and the first 75% of the data is set as the training set, and the last 25% of the data is set as the test set. ICEEMDAN is used to process the high and low frequency noise and nonlinear characteristics in the data to improve the quality of the data. Then, a bidirectional recurrent neural network prediction model is constructed based on the training set data, and each component after ICEEMDAN decomposition is trained. The number of hidden layer neurons is set to 5, the number of iterations is 5000, and the training target error is 0.0005. When the preset training target error is reached, the training ends and the prediction model is obtained. The prediction model is used to predict the cleaning factor, and the final prediction value Y(t) is obtained by superimposing the predicted values ​​of each component.

[0095] Next, the method of the present invention was strictly evaluated by a series of evaluation indicators, including root mean square error RMSE and mean absolute percentage error MAPE, and the calculation formulas are:

[0096]

[0097] Where n is the number of samples to be predicted, Y(t) is the true value of the tth sample, is the predicted value of the tth sample, and abs is the absolute value sign.

[0098] Subsequently, in order to verify the beneficial effects produced by the present invention, the ICEEMDAN-BIElman model prediction method was selected for comparison with the BIElman model prediction method to verify the effectiveness of the ICEEMDAN method; the ICEEMDAN-BIElman model prediction method and the ICEEMDAN-SVR model prediction method were selected to verify the effectiveness of the BIElman model.

[0099] Finally, the prediction error of each model is calculated by the root mean square error and mean absolute percentage error calculation formula. Figure 4 , Figure 5 The visualization results show that compared with the single BIElman model prediction method, the prediction accuracy of the ICEEMDAN-BIElman model prediction method proposed in this paper is significantly improved, mainly because the ICEEMDAN preprocessing step effectively separates and processes different frequency components, thereby reducing the mutual interference of signals between frequency domains and improving the prediction accuracy. In addition, according to the results, it is found that the ICEEM DAN-BIElman model not only demonstrates the excellent ability of the BIElman network in capturing the intrinsic information of time series, but also reflects its advantage in efficient data processing, which ultimately makes ICEEMDAN-BIElman perform best in all tests, proving the applicability and advancement of the model in complex sequence prediction tasks.

[0100] In summary, the verification of this embodiment shows that the present invention predicts the ash condition of the boiler heating area through a relatively accurate multi-step prediction method, and through reasonable judgment of the prediction results, the soot blowing and cleaning work is carried out in a timely and accurate manner to avoid the problems of excessive soot blowing and insufficient soot blowing. It can provide guidance for the soot blowing operation of thermal power plants and provide a new method for solving the problems of energy waste and industrial safety accidents caused by unreasonable soot blowing time in coal-fired power plants at this stage.

[0101] The technical solution of the present invention is not limited to the above-mentioned specific embodiments. All technical variations made according to the technical solution of the present invention fall within the protection scope of the present invention.

Claims

1. A sootblowing prediction method for a heating surface of a coal-fired power plant boiler, characterized in that: The following steps are involved: Obtain the time series data of the cleaning factor of the ash accumulation section of the coal-fired power plant boiler as the original time series data of the cleaning factor; An improved adaptive noise complete ensemble empirical mode decomposition algorithm is used to inject white noise into the original clean factor time series data to construct the time series data to be decomposed, and the time series data to be decomposed is decomposed to obtain a series of intrinsic mode function components and a residual component; wherein the improved adaptive noise complete ensemble empirical mode decomposition algorithm is obtained by an adaptive white noise injection strategy based on the ensemble empirical mode decomposition algorithm; A bidirectional recurrent neural network prediction model is used to predict a series of intrinsic mode function components and a residual component to obtain predicted values ​​of a series of intrinsic mode function components and a residual component; wherein the bidirectional recurrent neural network prediction model is obtained by training a bidirectional recurrent neural network, and the bidirectional recurrent neural network is a recursive neural network with information flow in two directions constructed based on the Elman recurrent neural network; The predicted values ​​of a series of intrinsic mode function components and the predicted value of a residual component are superimposed to obtain the final predicted value representing the ash condition of the boiler heating surface.

2. A sootblowing prediction method for a heating surface of a coal-fired power station boiler according to claim 1, characterized in that: The acquisition of the original cleaning factor time series data specifically includes the following steps: The real-time heat transfer coefficient K of the boiler heating surface is obtained through the boiler distributed control system DCS throughout the day f ; According to the real-time heat transfer coefficient K f The ratio of the theoretical heat transfer coefficient K0 of the boiler heating surface is used to obtain the real-time cleaning factor time series data for the entire day; The real-time cleaning factor time series data of the dust accumulation section is intercepted from the real-time cleaning factor time series data of the whole day, and the real-time cleaning factor time series data of the dust accumulation section is used as the original cleaning factor time series data.

3. A sootblowing prediction method for a heating surface of a coal-fired power station boiler according to claim 1, characterized in that: The improved adaptive noise complete ensemble empirical mode decomposition algorithm is obtained by an adaptive white noise injection strategy based on the ensemble empirical mode algorithm, and the specific steps include: Inject white noise into the original clean factor time series data according to the preset noise standard deviation; Performing ensemble empirical mode decomposition on the data injected with white noise to obtain a first group of ensemble empirical mode components, and using the first group of empirical mode components as a first group of adaptive white noise in improved adaptive complete ensemble empirical mode decomposition; Injecting the first group of adaptive white noise into the original clean factor time series data and performing ensemble empirical mode decomposition and local averaging processing to obtain the first group of intrinsic mode function components and the first group of residual components; Continue to perform ensemble empirical mode decomposition on the first group of residual components to obtain a second group of ensemble empirical mode components, and use the second group of empirical mode components as the second group of adaptive white noise in the improved adaptive complete ensemble empirical mode decomposition; Injecting a second group of adaptive white noise into the first group of residual components, and again performing ensemble empirical mode decomposition and local averaging to obtain a second group of intrinsic mode function components and a second group of residual components; Adaptive white noise is repeatedly injected into the previous set of residual components, and the decomposition is completed through ensemble empirical mode decomposition and local averaging until the residual components show a monotonic trend, so as to form an improved adaptive noise complete ensemble empirical mode decomposition algorithm.

4. A sootblowing prediction method for a heating surface of a coal-fired power station boiler according to claim 3, characterized in that: The preset noise standard deviation is 0.2, the number of times the white noise is injected is 50, and the signal-to-noise ratio of the original cleaning factor time series data and the added white noise is 1.

5. A sootblowing prediction method for a heating surface of a coal-fired power station boiler according to claim 1, characterized in that: The specific steps of obtaining the intrinsic mode function components and the residual components include: Construct k times of time series data to be decomposed: cf(i)=x(t)+β0E i (w(i)) Where cf(i) is the time series data to be decomposed, i = 1, 2, 3, ..., k, x(t) is the original clean factor time series data, β0 is the white noise weight coefficient, β0 = ε0std(cf(i)), ε0 is the noise amplitude, std(cf(i)) is the noise standard deviation of the time series data to be decomposed, E i (.) represents the i-th order modal component generated by the ensemble empirical mode decomposition, w(i) is the white noise injected during the i-th ensemble empirical mode decomposition, E i (w(i)) is the empirical mode component generated by the ensemble empirical mode decomposition after the i-th noise injection; According to the k-times time series data to be decomposed, the residual components are determined: In the formula, RES i is the residual component obtained after the i-th decomposition, M(.) is the local mean operation, β i-1 is the white noise weight coefficient, β i-1 =ε0std(RES i-1 ), ε0 is the noise amplitude, std(RES i-1 ) is the noise standard deviation of the i-1th residual component, E i (.) represents the i-th order modal component generated by the ensemble empirical mode decomposition, w(i) is the white noise injected during the i-th ensemble empirical mode decomposition, E i (w(i)) is the empirical mode component generated by the ensemble empirical mode decomposition after the i-th noise injection; According to the residual components, the intrinsic mode function components are determined: IMF i =RES i-1 -RES i In the formula, IMF i The intrinsic mode function components obtained by the i-th decomposition, RES i-1 and RES i are the residual components obtained from the i-1th and ith decompositions respectively; Until the decomposition is completed, a series of intrinsic mode function components and residual components obtained from the last decomposition are obtained.

6. A sootblowing prediction method for a heating surface of a coal-fired power station boiler according to claim 1, characterized in that: The bidirectional recurrent neural network prediction model is used to predict a series of intrinsic mode function components and a residual component, specifically including: Part of the data of the intrinsic mode function component and the residual component are respectively intercepted as training sets, and the bidirectional recurrent neural network is trained using the training sets to obtain a bidirectional recurrent neural network prediction model; The bidirectional recurrent neural network prediction model includes an input layer, a hidden layer, a receiving layer and an output layer; Input any intrinsic mode function component or residual component into the input layer; The modal function component or residual component received by the input layer is transferred to the hidden layer, the hidden layer is connected to the feedback link of the receiving layer, the hidden layer learns the nonlinear characteristics and complex patterns of the modal function component or residual component through a nonlinear activation function and transfers them to the receiving layer for storage, and the nonlinear characteristics and complex patterns are used as hidden layer outputs; The output layer receives the hidden layer output and outputs the predicted value of the modal function component or the residual component through a linear activation function.

7. A sootblowing prediction method for a heating surface of a coal-fired power station boiler according to claim 1 or 6, characterized in that: The bidirectional recurrent neural network includes a forward Elman recurrent neural network and a reverse Elman recurrent neural network; the forward Elman recurrent neural network starts from the first element of the sequence and processes step by step until the last element of the sequence, and the reverse Elman recurrent neural network starts from the end of the sequence and traces back to the starting position; the output of the bidirectional recurrent neural network is the fusion of the outputs of the Elman recurrent neural networks in two directions.

8. A sootblowing prediction method for a heating surface of a coal-fired power station boiler according to claim 6, characterized in that: The number of hidden layer neurons of the bidirectional recurrent neural network is 5, the number of iterations is 5000, and the training target error is 0.0005.

9. A sootblowing prediction method for a heating surface of a coal-fired power station boiler according to claim 7, characterized in that: The fusion method includes: splicing or weighted averaging.

10. A sootblowing prediction method for a heating surface of a coal-fired power station boiler according to claim 1, characterized in that: The step of superimposing a series of predicted values ​​of intrinsic modal components and a predicted value of a residual component specifically includes: The total number of a series of intrinsic mode function components and residual components decomposed is k, and the predicted value of any component is Y i , where i = 1, 2, ..., k, and the set of predicted values ​​of all components is {Y1, Y2, ..., Y k }; The predicted values ​​of all components are superimposed as follows:

Images