Background
With the increase of population and the acceleration of urbanization process, the yield of Municipal Solid Waste (MSW) is increased year by year, and the MSW clearing volume in China reaches about 2.42 hundred million tons as late as 2019[1]. In recent 20 years, the MSW incineration (MSWI) power generation technology has rapidly developed in China due to the advantages of reduction, reclamation and harmlessness, and becomes one of the main modes of MSW disposal in China. However, pollution emission control remains a major problem for MSWI processes, such as nitrogen oxides (NOx, mainly including NO and NO)2). The MSWI has a very complex operation process, and if the MSWI is improperly controlled, the emission of nitrogen oxides exceeds the standard, so that secondary pollution is brought to the environment. At present, in MSWI plants, Selective non-catalytic reduction (SNCR) denitration technology is adopted to further reduce NOx emission, and urea diluent is sprayed into flue gas in a hearth, so that NOx in the flue gas and urea are subjected to chemical reaction to convert NOx into N2The denitration efficiency depends on the amount of sprayed urea diluent, and if the amount of sprayed urea diluent is small, the SNCR denitration efficiency is affected, so that the concentration of NOx is increased, and the pollutant emission exceeds the standard; if too much urea diluent is injected, excess urea can cause ammonia slip and unreacted NH3Discharged out of the boiler, not only pollutes the atmosphere, but also can be mixed with SO in the flue gas3And the acidic gases form ammonium salts, corroding and plugging downstream equipment. Obviously, timely and accurate urea diluent spraying can realize low emission of NOx, low ammonia leakage and low by-product.
The SNCR control system adjusts the spraying amount of the urea diluent according to the emission of NOx in tail gas, the MSWI measures the concentration of the NOx by adopting a Continuous flue gas emission monitoring system (CEMS), but the measurement result of the CEMS has larger hysteresis due to the characteristics of an SNCR reaction mechanism and longer transmission distance of flue gas, so that the denitration efficiency of the SNCR control system is reduced. Therefore, accurate prediction of the NOx concentration is the key to ensure reliable control of the SNCR system and to improve the denitration efficiency.
Currently, some researchers have focused on using soft-measurement methods to achieve prediction of NOx concentration. Li and the like[2]In order to solve time-varying characteristics and nonlinearity in NOx prediction, a soft measurement method of local weighted regression based on moving window partial least squares is provided to realize the prediction of NOx; due to the good nonlinear mapping capability of LSSVM, an adaptive LSSVM is also applied to the real-time prediction of NOx[3][4]. Although these methods achieve good predictive results, they rely on a single model. The MSWI process has the characteristics of high nonlinearity, dynamics, system uncertainty and the like, is influenced by time-varying characteristics, and industrial time series data follow different distributions under different operating conditions, so that a single global model is difficult to accurately describe local characteristics of a complex process.
The modularized neural network starts from a simulated human brain information processing mechanism, decomposes a complex task into a plurality of simpler subtasks through a divide-and-conquer strategy, establishes a corresponding sub-network model based on each subtask, and improves the precision of the model. To improve the accuracy of model prediction in time-varying processes, Hoori et al[5]Clustering data using K-d tree algorithm, constructing multi-column RBF neural network for power load prediction, and similarly, Wang et al[6]And a density-based clustering algorithm is adopted to carry out task decomposition on the data, and an LSTM sub-network prediction model is established, so that the power load prediction precision is improved. Compared with a single model, the modularized neural network reduces the complexity of tasks through task decomposition, establishes a local network model based on subtasks and improves the prediction precision of the model.
In summary, the MSWI process-oriented NOx prediction method based on the modular neural network is provided, and comprises four parts, namely firstly, carrying out normalization preprocessing on data to eliminate the influence of dimensions among different variables; then, dividing time sequence data by adopting an exponential smoothing prediction-based method, and dividing original data into subsets with different working condition distributions; then, aiming at each subset, establishing a corresponding sub-network model by using an RBF neural network to realize the prediction of NOx concentration; finally, in order to improve the prediction precision of the test sample, a sample matching method based on Euclidean distance is adopted to determine the type of the test sample, the prediction of the test sample is completed, and the effectiveness of the method is verified based on the actual industrial data of certain MSWI factory in Beijing.
Disclosure of Invention
The study object of this document is the MSWI factory in Beijing. The process flow of the MSWI plant is shown in figure 1.
The process comprises 5 subsystems, namely a solid waste storage and transportation system, a solid waste incineration system, a waste heat boiler system, a steam power generation system and a flue gas treatment system. Firstly, collecting solid waste by a special transport vehicle, and then transporting the collected solid waste to a solid waste pool for composting and fermentation; then, solid waste is put into a hopper by a manually operated grab bucket, enters a hearth through a feeder, and passes through a drying grate, a combustion grate 1, a combustion grate 2 and a burn-out grate to realize full combustion; then the waste heat boiler system and the steam power generation system use the heat generated in the process for steam power generation; finally, toxic substances and particles in the flue gas are purified in the flue gas treatment system.
The mechanism of NOx formation and elimination is shown in FIG. 2. In the actual MSWI process, there are two main sources of NOx, respectively: n in air2And nitrogen-containing organic matters in the solid waste. When the temperature is about 1800K, N in the air2Is oxidized into NOx at high temperature, and moreover, nitrogen-containing organic matters (mainly comprising CHi) in the solid waste are extremely easy to react with N2NOx is generated. In order to eliminate NOx produced during combustion, MSWI plants use SNCR processes for denitration. The SNCR control system sprays dilute solution of urea into the hearth, the urea is decomposed into NH3 at high temperature, and NOx is converted into N through chemical reaction2So as to achieve the aim of denitration. In the denitration process, the injection amount of the urea diluent is a key influence on the denitration efficiency, if the injection amount is too small, the emission amount of NOx is increased, and if the injection amount is too large, excessive urea can cause ammonia to escape and corrode a flue gas sampling device, so that the NOx concentration needs to be predicted to ensure the denitration efficiency of SNCR, and the NOx emission is reduced.
Due to the influence of a plurality of process variables such as the composition of garbage entering a furnace, the temperature, the air volume and the like, the MSWI process has strong time-varying characteristics and uncertainty, and NOx has different distribution characteristics under different working conditions, aiming at the characteristics, a modular neural network method is constructed to predict the NOx concentration under different working conditions, and a modeling strategy is shown in figure 3.
In FIG. 3, xori_1,xori_2,...,xori_NRespectively representing original process variables collected from a hearth under the influence of complex environmental changes in the hearth, wherein abnormal values are easy to appear in data collected by a sensor, and different variable dimensions are different, so that abnormal value elimination and normalization processing operation needs to be carried out on the original data, and after data preprocessing, the obtained data is represented as x1,x2,...,xNThen, the time series data is divided by adopting a method based on exponential smooth prediction, and the data is divided into K +1 sections which are respectively expressed as X1,X2,...,XK+1(ii) a Then, aiming at each subset, establishing a corresponding sub-network by adopting an RBF neural network, and carrying out NOx prediction on data with different working condition distributions; finally, carrying out optimal working condition matching on the test sample on the test set according to the distance between the sample point and the center of each time sequence subset, and realizing the prediction of NOx on the test set;
the functions of the modules are respectively as follows:
(1) a data preprocessing module: and carrying out abnormal value elimination and normalization processing on the original process variable.
(2) A time sequence data segmentation module: dividing time series data by adopting an exponential smooth prediction-based method, and dividing the time series data into subsets with different distribution characteristics;
(3) a sub-network construction module: a sub-network model based on subset driving is established by adopting RBF, so that the accuracy of NOx prediction is improved;
(4) a test sample matching module: and determining the membership class of the sample by calculating the shortest Euclidean distance between the test sample and the central point of each subset, and selecting a corresponding sub-network model for prediction.
NOx values are collected by the CEMS and are influenced by the complex environment in the hearth, and the collected NO is caused by faults of the CEMS due to various reasons in the use processAbnormal values of the x value are generated, the abnormal values bring great interference to the analysis of the data, and the abnormal values need to be detected and eliminated. Using Rajda's criterion[7]Namely, the method which is three times higher than the standard deviation of the data eliminates the abnormal value in the original data, as shown in formula (1):
wherein, mu
NOxAnd σ
NOxThe abnormal value satisfying the formula (1) is the mean and variance of NOx, respectively
And (5) removing from the data set. In order to eliminate dimensional differences between variables and improve the accuracy of the model, the Z-score method is used to normalize the original process variables as shown in equation (2):
wherein x ismData, x, representing the normalized mth process variable affecting NOx emissionsori_mRepresents the mth process variable, μ, in the collected data setmAnd σmRespectively, mean and standard deviation of the mth process variable. The preprocessed data set is denoted X, X ═ X1,x2,...,xN]=[x1,x2,...,xM]T,X∈RN×MWhere N represents the sample size and M represents the number of process variables.
The MSWI process has the characteristics of high nonlinearity, dynamics, system uncertainty and the like, is influenced by time-varying characteristics, and industrial time series data follow different distributions under different operating conditions, so that a single global model is difficult to accurately describe local characteristics of a complex process. The method for exponential smooth prediction is simple in calculation, high in solving speed andthe accuracy of the division is high[8]Therefore, the method of single-based exponential smooth prediction is adopted in the text to segment the NOx time series data. The single exponential smoothing method is shown in equation (3):
St=a·yt-1+(1-a)St-1 (3)
wherein S istAnd St-1Indicating the exponentially smoothed values at time t and t-1, respectively, yt-1Is the true value at time t-1, a represents the exponential smoothing factor;
by iterative computation, equation (4) can be written as:
in order to obtain good prediction performance, two indexes, namely a smoothing factor a and a smoothing initial value S, need to be considered0. The smoothing factor a reflects the difference between the smooth value and the real value, if the value of a is close to 1, the real value of the historical time has little influence on the smooth value of the t moment, and the accumulated error is increased therewith, otherwise, if the value of a is close to 0, the real value of the historical time has great influence on the smooth value of the t moment, and a good smoothing effect can be generated. Smoothed initial value S0Determined by equation (5):
the time sequence segmentation method based on the exponential smoothing prediction is summarized as follows:
step 1: initializing a segmentation point set SegNum, an error vector Err, a total residual error TSE and a smooth initial value S0The compression ratio p;
step 2: calculating a smoothed value S at time tt;
Step 3: calculating the absolute prediction error, Err, at time tt=|yt-St|;
Step 4: updating Err mean μErrAnd standard deviation σErr;
Step 5: determination of ErrtIf Err is in the distribution interval oftIn the interval [ mu ]Err-p*σErr,μErr+p*σErr]Then y is determinedtSegNum (t) y as a division pointtAnd reinitializing S according to equation (5)t(ii) a Otherwise, return to Step 2;
dividing the time sequence data to obtain K division points, and dividing the data set X into K +1 subsets according to the K division points, wherein the X subsets are respectively expressed as X1,X2,...,XK+1;
After K +1 subsets are obtained by time-sequential segmentation, the different subsets are processed "divide-by-divide" using a modular neural network. Because the RBF network has a simple structure and can approach any nonlinear function, a corresponding RBF neural network is established for each subset, and the RBF neural network consists of three parts, namely an input layer, a hidden layer and an output layer as shown in FIG. 4; the input layer transmits potential auxiliary variables affecting NOx into the network in a first subset X of segments
1For example, the input of the neural network is
These variables represent the first segment subset X, respectively
1The 1 st, 2 nd, as well as the M auxiliary variables under the nth training sample; the hidden layer comprises H hidden layer nodes, each node selects a Gaussian kernel function as a basis function, and the basis functions are respectively used
And (3) representing that the nonlinear mapping from the input space to the hidden layer space is completed, and the calculation of the Gaussian kernel function is shown as the formula (7):
wherein x
1,nRepresents a subset X
1The nth sample in (1), c
hAnd σ
hRespectively representing the center and width of the h-th hidden node,
and representing the output of the h hidden layer node, the output of the RBF neural network for predicting the NOx concentration is as follows:
wherein, w0Indicates a deviation, wh(H ═ 1, …, H) represents the weight between the hidden node and the output node;
parameter c of the networkh,σh,whUsing a second-order LM algorithm[9]Training adjustment is carried out until the desired precision is achieved, and the updating rule of the algorithm is as follows:
Δn+1=Δn-(Qn+μnI)-1gn (9)
where Δ is all the parameters to be trained of the network, including center c, width σ and weight w, i.e., Δ ═ c1,...,cH,σ1,...,σH,w0,w1,...,wH]Q is the Hessian matrix, μnIs step size, I is identity matrix, g represents gradient vector;
the Hessian matrix Q is obtained by summing Hessian submatrices:
wherein N is1Is a subset X1Number of samples j contained innIs a jacobian component, represented by equation (11):
the gradient vector g is obtained by summing the gradient sub-vectors:
wherein
Predicted value indicating NOx concentration
And expected value
The difference of (a):
the Root Mean Square Error (RMSE) is used as a model performance index, and is shown as a formula (17):
in order to select a proper sub-network to accurately evaluate a test sample, a membership class k of the test sample needs to be determined, Euclidean distance is selected as a basis for measuring the similarity of the sample, namely, the Euclidean distance between the test sample and the center of each sub-set is calculated, the class corresponding to the center of the sub-set with the shortest distance to the test sample is determined as the class of the current test sample, and the test result is tested by the corresponding sub-network;
the center point of the subset should satisfy the following condition: the sum of the distances to the point, except the center point, of all the other points is minimal, i.e. the
A point x satisfying the formulas (18) and (19)c1As subset X1Center point of (1), same principle, xc2,...,xc(K+1)Is a subset X2,...,XK+1A center point of (a);
determining the membership class of the test sample by selecting the central point closest to the test sample:
cluster(xtest)=min dist(xtest,xck),k=1,2,...,K+1 (20)
the k value satisfying the formula (20) was taken as a test sample xtestIs of a membership class of xtestBy the sub-network RBFkGiven, as shown in equation (21):
wherein, ck,σk,wkAs a subnetwork RBFkParameters obtained after training.
Detailed Description
The text is based on the results of the MSWI factory, Beijing, 10 months, 17 days, 15, 2019: 19: 23-21: 36: a total of 11314 actual data sets were subjected to industrial experiments during 39 hours. After the preprocessing operation, 273 groups of abnormal values are eliminated altogether, and 11041 groups of normal data are reserved. Because MSWI process operating mode is complicated and is influenced by different staff's operation means and presents a plurality of distribution characteristics, in order to make the training process of model cover a plurality of operating modes as far as possible, select 9500 group of data as the training set in the experimentation, 1541 group of data is as the test set. On the training set, the NOx time series data is divided by using a method based on exponential smoothing prediction, and the obtained division result is shown in fig. 5.
In the time series data segmentation stage, two parameters, namely a smoothing factor a and a compression rate p, need to be artificially determined. These two parameters affect the prediction performance and the error distribution, respectively, and therefore, the two parameter values are 0.58 and 0.6, respectively, as determined by trial and error in the experimental process. From the dividing line in fig. 5, 2 dividing points can be obtained: 3658 and 6830. The two division points divide the NOx time sequence data into three subsets, the distribution situation in each subset is different as can be seen from the graph, and when the distribution characteristics are changed, the change can be captured by detecting an error interval based on an exponential smoothing prediction method, so that the division of the working conditions is realized.
And establishing corresponding RBF neural network prediction models aiming at the three subsets. The RBF neural network structure adopted in the method is 11-8-1, 11-8-1, 11-10-1, and the number of nodes of the hidden layer is obtained by a trial and error method. The training results for the subnetworks are shown in fig. 6-8, the training RMSE is 4.0957, and the fitting effect of the modular neural network on the training set is shown in fig. 9. Under different working conditions, the sub-network has good training performance.
And after the network training is finished, testing the network by adopting the test sample. In order to determine the class to which the sample to be tested belongs, the test sample needs to be matched with the training subset, and fig. 10 shows the matching result. The 916 groups of data in the test sample belong to the first class subset, the 617 groups of data belong to the third class subset, and the test results are respectively tested by using the sub-networks 1 and 3, as shown in fig. 11 and 12, the predicted values of the modular neural network can be better fitted to expected values, have higher prediction accuracy, the test RMSE is 6.0953,
a modular neural network based MSWI process NOx prediction method is presented herein. Different from the traditional modularized neural network method, the method based on exponential smoothing prediction is adopted in the text to segment data so as to complete task decomposition, and the data under different working conditions are divided by capturing the change of data distribution; in addition, the RBF sub-network model established based on the specific subset has high prediction precision, and the prediction performance of the modular neural network model is improved.
_____________________________________
[1] National statistical office of the people's republic of china, china statistical yearbook [ M ]. beijing: china statistics press 2020.
[2]Li Z,Lee Y S,Chen J H,et al.Developing variable moving window PLS models:Using case of NOx emission prediction of coal-fired power plants[J].Fuel,2021,296:120441.
[3]Lv Y,Yang T,Liu J.An adaptive least squares support vector machine model with a novel update for NOx emission prediction[J].Chemometrics and Intelligent Laboratory Systems,2015,145:103-113.
[4]Yang T T,Ma K F,Lv Y,et al.Real-time dynamic prediction model of NOx emission of coal-fired boilers under variable load conditions.Fuel,2020,274:117811.
[5]Hoori AO,Kazzaz A A,Khimani R,et al.Electric Load Forecasting Model Using a Multicolumn Deep Neural Networks[J].IEEE Transactions on Industrial Electronics,2020,67(8):6473-6482.
[6]Wang L X,Mao S W,Wilamowski B M,et al.Ensemble Learning for Load Forecasting[J].IEEE Transactions on Green Communications and Networking,2020,4(2):616-628.
[7]Tian Z D,Li S J,Wang Y H.A Multi-model Fusion Soft Sensor Modeling Method and Its Application in Rotary Kiln Calcination Zone Temperature Prediction[J].Transactions of the Institute of Measurement&Control,2015,38(1):110-124.
[8]Li G L,Cai Z H,Kang X J,et al.ESPSA:A prediction-based algorithm for streaming time series segmentation[J].Expert Systems with Applications,2014,41(14):6098-6105.
[9] Jialieger, Liwenjing, Jojunfei radial basis function neural network self-organizing design method based on neuron characteristics [ J ] control theory and application 2020,37(12):2618-2626.