A prediction modeling system and method based on working condition division

Abstract

The present invention relates to a prediction modeling system and method based on working condition division, the system comprising: an equipment tree module, a data screening and processing module, a working condition slicing database module, a modeling module, a prediction module and a decision analysis module; the equipment tree module is connected to the data screening and processing module, the data screening and processing module is connected to the working condition slicing database module, the working condition slicing database module is connected to the modeling module, the modeling module is connected to the prediction module, and the prediction module is connected to the decision analysis module.

Description

Technical Field

[0001] The present invention belongs to the technical field of power plant equipment operation optimization, and specifically relates to a prediction modeling system and method based on operating condition division Background Art

[0002] Effectively predicting the relevant index parameters and operating status data of power plant equipment can fully understand the future trends and directions of the equipment, which is of great significance for adjusting equipment operation in advance. There are mainly two scenarios in the production and operation of power plants that urgently need to be predicted:

[0003] The first is that the acquisition of real-time values ​​of some indicators is delayed due to chemical experiments, value delays, etc., and real-time acquisition of indicator values ​​is conducive to real-time understanding of equipment status and adjustment of equipment operation;

[0004] The second is: knowing the indicator values ​​or trends in advance is conducive to grasping the development trend of the equipment in advance and formulating equipment operation or maintenance strategies and plans in advance.

[0005] The traditional prediction method in the existing technology is that professionals give prediction formulas for indicators based on experience and analysis. The disadvantage of this method is that the operating experience of professionals varies, the formulas given are not exactly the same, and there is no universality. With the development of big data technology, indicator prediction and modeling technology are becoming more mature, and more and more research is focused on data modeling of power plant equipment. However, when conducting system modeling, different equipment operating conditions and different rules between data are not taken into account, which may cause the accuracy of the data model to be unable to meet actual production needs.

[0006] In addition, the current status of power plant information construction is: lack of centralized analysis and modeling tools, unable to effectively divide production and operation data, and model according to actual application, resulting in the inability of power plant professional and technical personnel to centrally manage data, and data analysis and application are difficult. This is the shortcoming of the existing technology.

[0007] In view of this, the present invention provides a predictive modeling system and method based on operating condition division to solve the above-mentioned technical defects existing in the prior art. Summary of the invention

[0008] The purpose of the present invention is to provide a predictive modeling system and method based on working condition division to solve the above-mentioned technical problems in view of the defects of the above-mentioned prior art.

[0009] To achieve the above object, the present invention provides the following technical solutions:

[0010] A predictive modeling system based on working condition division, comprising:

[0011] Equipment tree module, data screening and processing module, working condition slicing database module, modeling module, prediction module and decision analysis module;

[0012] The device tree module is connected to the data screening and processing module, the data screening and processing module is connected to the working condition slicing database module, the working condition slicing database module is connected to the modeling module, the modeling module is connected to the prediction module, and the prediction module is connected to the decision analysis module;

[0013] The device tree module includes a device object unit, a device member unit and a modeling time unit. The device object unit is used to store the device name, the device member unit is used to store all key indicator data information of the device, and the modeling time unit is used to store the modeling time period data information;

[0014] The data screening and processing module includes an indicator data unit, a screening unit, a stability detection unit and a standard unit. The indicator data unit is used to obtain data of device tree research indicators. The screening unit is used to screen the obtained research indicator data to remove abnormal data values. The stability detection unit is used to detect the screened research indicator data to remove unstable data. The standard unit is used to normalize the research indicator data after stability detection.

[0015] The working condition slicing database module includes a working condition division unit and a working condition storage unit. The working condition division unit is used to receive data from the data screening and processing module unit, and then divide the working condition according to the working condition type. The working condition storage unit is used to store the divided working condition data;

[0016] The modeling module includes a model unit and a model storage unit. The model unit is used to receive data from the working condition slicing database module and perform predictive modeling on the data after the working conditions are divided. The model storage unit is used to store the model results of each modeling according to the working conditions.

[0017] The prediction module includes a working condition matching unit and an operation prediction unit. The working condition matching unit matches the stored working condition with the real-time data to obtain the working condition type. The operation prediction unit obtains the model stored in the modeling module according to the working condition type and runs the prediction model to obtain the prediction result.

[0018] The decision analysis module includes a diagnosis unit and an energy-saving optimization unit. The diagnosis unit is used to process the prediction index and assist in decision-making and analysis of the equipment. The energy-saving optimization unit optimizes the prediction index to bring the prediction index within a safe range.

[0019] Preferably, the equipment object unit is used to store equipment names, and the equipment names include: equipment name information of a boiler system, equipment name information of a steam turbine system, equipment name information of a desulfurization system, and equipment name information of a denitration system;

[0020] The device member unit is used to store all key indicator information of the device, and the key indicator information includes indicator name data information and indicator code data information;

[0021] The modeling time unit is used to store the modeling start time and end time, and the time period from the modeling start time to the end time represents the duration of the historical data used for modeling.

[0022] Preferably, the indicator data unit is used to obtain research indicator data of the device and obtain device tree data corresponding to the indicator data;

[0023] The screening unit is used to obtain the upper and lower limit values ​​of the indicator data, detect and screen the abnormal data that exceeds the upper and lower limit values ​​of the indicator, and clear the screened abnormal data;

[0024] The stability detection unit is used to calculate the change value of the indicator value within the continuous time window, and then determine whether the device is stable according to whether the change value exceeds the threshold, and remove unstable data;

[0025] The normalization unit is used to perform a normalization operation on the data.

[0026] Preferably, the diagnosis unit is further used to determine whether the predicted indicator data exceeds the limit and issue an alarm based on the indicator threshold.

[0027] The present invention also provides a prediction modeling method based on working condition division, comprising the following steps:

[0028] Step S1: a step of obtaining a device tree of a research object;

[0029] Step S2: Reading the data of the research object and screening and processing the data, which specifically includes:

[0030] Step S21: receiving the device name, indicator information and modeling time of the device tree, and reading the modeling data corresponding to the research indicator according to the research indicator;

[0031] Step S22: obtaining the upper and lower limit values ​​of the indicator data, detecting and filtering the abnormal data exceeding the upper and lower limit values ​​of the indicator, and removing the filtered abnormal data;

[0032] Step S23: Calculate the change value of the indicator value in the continuous time window, and then determine whether the device is stable according to whether the change value exceeds the threshold, and remove unstable data;

[0033] Step S24: finally normalizing the data;

[0034] Step S3: dividing the processed data into working conditions and establishing a working condition slice database, which specifically includes:

[0035] Receive the data after data screening and processing, and divide the received data into working conditions;

[0036] Step S4: a step of establishing a prediction model for each operating condition data, which includes:

[0037] Receive data modeling data from the working condition slice database, establish prediction models for different working condition data, and store the model results by working condition;

[0038] The prediction models include linear regression model, autoregressive moving average model (ARMA), multilayer perceptron (MLP), extreme learning machine (ELM), deep learning long short-term memory model (LSTM) and radial basis function neural network model (RBF);

[0039] The prediction model is selected in two ways: free selection and automatic selection of the best method. The automatic selection of the best method is carried out through the following steps:

[0040] The root mean square error (RMSE) method is used to select the optimal model, and the calculation is shown in formula (1):

[0041]

[0042] where y i is the true value of a sample, y' i is the predicted value of the sample under the kth prediction model, and n is the number of samples;

[0043] The root mean square error of each prediction model is calculated respectively, and the model with the smallest error is automatically selected as the prediction model, thereby realizing automatic selection of the optimal model.

[0044] Preferably, in step S4, the default configuration uses the LSTM algorithm as the prediction model;

[0045] LSTM is an improved algorithm of the recurrent neural network algorithm. It can solve the gradient explosion and gradient vanishing problems in recurrent neural networks. It performs well in time series data modeling and is now widely used. The main components of LSTM are cells, and the cell state is updated or deleted by setting the gate structure.

[0046] Each LSTM neural network contains three gates to control the cell state, namely the forget gate, input gate and output gate. The input gate, output gate and forget gate are represented by i, f and o respectively, the neural network weight matrix is ​​represented by w, and the bias term is represented by b. The most critical gate is the forget gate. By setting the activation function, its output value determines the information that needs to be discarded for the cell state.

[0047] Assume that at time t, the LSTM hidden layer input vector is x t , the output vector is y t , the cell state is represented by state(t).

[0048] The value of the forget gate at time t is shown in formula (2):

[0049] f t =g(w f *(x t ,y t-1 )+b f ) (2)

[0050] Where g indicates that the activation function is the sigmoid function, w f 、b f Represent the weight matrix and bias of the forget gate respectively.

[0051] The input gate determines the cell information update, and the input gate value i t And the candidate state value of the input cell hs t The calculation is as follows:

[0052] i t =g(w i *(x t ,y t-1 )+b i ) (3)

[0053] hs t =tanh(w n *(x t ,y t-1 )+b n ) (4)

[0054] Among them, w i 、b i are the weight matrix and bias of the input gate, w n 、b n is the weight matrix and bias of the current state of the cell, and tanh indicates that the activation function is a hyperbolic tangent function.

[0055] By candidate state value hs t , calculate and update the cell state value at time t.

[0056] state(t)=i t *hs t +f t *state(t-1) (5)

[0057] Finally, the value of the output gate determines the output of the information, and the final output is y t By output gate o t Determined together with the cell state state.

[0058] o t =g(w out *(x t ,y t-1 )+b out ) (6)

[0059] y t =o t *tanh(state(t)) (7)

[0060] where w out 、b out Represent the weight matrix and bias of the output gate respectively.

[0061] The essence of the LSTM prediction model training process is the training of network model parameters. Preferably, the present invention adopts the back propagation algorithm (BPTT) when training the LSTM prediction model.

[0062] Step S5: a step of performing matching in real time and predicting indicators, which specifically includes:

[0063] Read the real-time data of the equipment, obtain the working condition of the real-time data matching the stored model results, obtain the working condition type, obtain the corresponding model according to the working condition type, run the model, and obtain the prediction result;

[0064] Step S6: Processing the prediction results to assist decision analysis.

[0065] Preferably, in step S1, obtaining the information of the device tree includes:

[0066] The name information of subordinate equipment of the boiler system, the name information of subordinate equipment of the steam turbine system, the name information of subordinate equipment of the desulfurization system, and the name information of subordinate equipment of the denitrification system are obtained, and all key indicator information of the equipment is obtained, and the key indicator information includes indicator name data information and indicator code data information; the modeling start time and end time are stored, and the time period from the modeling start time to the end time represents the duration of the historical data used for modeling.

[0067] Preferably, in step S3:

[0068] According to the configured working condition division method, the type of working condition division is determined. For data with clear working condition boundaries, each working condition is divided according to the working condition boundaries and the indicator range of the indicators. For research objects without clear working condition boundaries, the mathematical model clustering algorithm is used to divide the data into working conditions. Different operating modes or combination modes are used as the basis for working condition division. The results after working condition division are classified and stored according to working conditions.

[0069] In step S3, a mathematical model clustering algorithm is used to divide the data into working conditions; in this step, one of a k-means (kmeans) clustering algorithm, a hierarchical clustering algorithm, and a Gaussian mixture model (GMM) clustering method can be used to divide the working conditions;

[0070] Preferably, in step S3, when performing mathematical model clustering, a Gaussian mixture clustering algorithm is configured by default;

[0071] Gaussian mixture model clustering (GMM) is based on the Gaussian mixture model. It constructs the most appropriate mixed Gaussian distribution model by fitting the sample data set. It is the fastest learning probability model. The model outputs the probability value of each sample belonging to the model. It carries more information than the clustering method that simply assigns it to a certain category, and the clustering results obtained based on the probability value are more accurate.

[0072] Assuming that the sample obeys a mixed Gaussian distribution and there are n distribution models, for a certain sample x, the Gaussian mixture model can be expressed by formula (8):

[0073]

[0074] where θ i is the coefficient of the i-th model, and the sample is a weighted combination of n models, then θ is the weight vector of the model, p(x|u i ,δ i ) is the probability density function of the ith model.

[0075] where μ i is the mean vector of the ith model.

[0076] where δ i is the covariance matrix of the ith model.

[0077] The key to model training is to estimate the model parameter weights, mean and covariance parameters, which are θ, μ, δ respectively, and calculate the probability density of the model, as shown in formula (9).

[0078]

[0079] The steps of the Gaussian mixture model clustering algorithm are as follows:

[0080] Step T1: Construct n mixed component models and initialize the model parameters θ, μ, δ

[0081] Construct n Gaussian distribution mixture component models and initialize the parameters of the n component models respectively. For the i-th model, the parameter is θ i , μ i , δ i .

[0082] Step T2: Calculate the posterior probability

[0083] For the qth sample x q , the probability under the i-th model is the posterior probability, which can be represented as: ph qi =(y q =i|x q ).

[0084] The specific probability calculation is shown in formula (10):

[0085]

[0086] Further calculations yield:

[0087]

[0088] Step T3: Maximum likelihood solution to obtain model parameters

[0089] Solve the maximum likelihood function and the maximum likelihood value to obtain the values ​​of each parameter. For detailed calculation, see formulas (12)-(14). And update the parameters of the n component models, and update the parameters of the i-th model: θ i , μ i , δ i .

[0090] The mean μ of the i-th model i :

[0091]

[0092] The variance δ of the i-th model i :

[0093]

[0094] The weight θ of the i-th model i :

[0095]

[0096] Step T4: Repeat steps T2 and T3 until the algorithm converges

[0097] Repeat steps T2 and T3 until the algorithm converges and ends the algorithm learning, and save the final parameters of each component model.

[0098] The probability value of the sample under each component model is calculated, and then the sample is divided into clusters. The model to which the sample belongs depends on the probability value, and the model with the larger probability value belongs to the model.

[0099] Typically, when you run data clustering on a sample, you set the number of clusters.

[0100] Preferably, the present invention adopts the Bayesian Information Criterion (BIC) to determine the number of clusters, and automatically selects the number of clusters corresponding to the minimum BIC value as the optimal number of clusters. The BIC value calculation is detailed in formula (15).

[0101] BIC=mln(n)-2ln(L) (15)

[0102] Where m is the number of parameters in the model, n is the number of samples, and L is the likelihood function.

[0103] Preferably, in step S6:

[0104] According to the prediction results, the prediction indicators are applied to determine whether the prediction indicators are over-limit or optimal, and to assist in guiding the safe and economical operation of equipment production; for indicators that are over-limit, adjust the equipment operation indicators in advance to make the indicators reach a safe range; for indicators that are not optimal, adjust the controllable operation indicators in advance to make the indicators close to the optimized values.

[0105] The beneficial effect of the present invention is that the technical solution is based on the operation of the equipment, directly divides the working condition of the data with working condition boundaries, performs cluster modeling on the data without working condition boundaries, and divides the working condition according to the operation mode, and then performs predictive modeling on the divided working condition data sets. When the system is running in real time, it matches the working condition and the model, and gives the prediction results in real time to assist decision analysis. Dividing the working condition from the data perspective and modeling different data sets improves the accuracy of the prediction, and provides different working condition division methods to achieve different working condition divisions for different operating data, which is more in line with actual application requirements. The present invention proposes to classify the target data by working condition, and establish a data model based on the actual operation of the equipment, which is more reasonable, accurate, and more in line with actual production operations. In addition:

[0106] The present invention provides a tool for establishing a modeling of different operating conditions, which is suitable for various equipment in power plant operation and automatically analyzes and models the data based on actual operation data without requiring strong professional knowledge and experience.

[0107] The operating condition division system and method established by the present invention directly divides the operating condition of data with operating condition boundaries, performs cluster modeling on data without operating condition boundaries, and divides the operating condition according to the operation mode based on equipment operation, which conforms to actual production rules.

[0108] The operating condition division modeling tool established by the present invention analyzes and models the data, provides a data analysis tool for power plant professionals, and effectively manages and analyzes the large amount of data generated by production operations.

[0109] The present invention is beneficial to ensuring the safe and reliable operation of production equipment. Effective index prediction is beneficial to grasping the development trend of equipment, assisting equipment operators to adjust equipment operation, and ensuring the safe and reliable operation of equipment.

[0110] The present invention is an important means to improve the economic benefits of production equipment operation. Effective index prediction is conducive to knowing the value and trend of energy-saving index in advance, which has practical guiding significance for auxiliary energy-saving optimization.

[0111] In addition, the invention has a reliable design principle, a simple structure and a very broad application prospect.

[0112] It can be seen that compared with the prior art, the present invention has outstanding substantive features and significant progress, and the beneficial effects of its implementation are also obvious. BRIEF DESCRIPTION OF THE DRAWINGS

[0113] Figure 1 It is a principle block diagram of a prediction modeling system based on working condition division provided by the present invention.

[0114] Figure 2 It is a flow chart of a prediction modeling method based on working condition division provided by the present invention.

[0115] Among them, 1-equipment tree module, 2-data screening and processing module, 3-working condition slicing database module, 4-modeling module, 5-prediction module, 6-decision analysis module, 11-equipment object unit, 12-equipment member unit, 13-modeling time unit, 21-indicator data unit, 22-screening unit, 23-stability detection unit, 24-specification unit, 31-working condition division unit, 32-working condition storage unit, 41-model unit, 42-model storage unit, 51-working condition matching unit, 52-operation prediction unit, 61-diagnosis unit, 62-energy saving optimization unit. DETAILED DESCRIPTION

[0116] The present invention will be described in detail below with reference to the accompanying drawings and through specific embodiments. The following embodiments are explanations of the present invention, but the present invention is not limited to the following implementation modes.

[0117] Embodiment 1:

[0118] like Figure 1 As shown, this embodiment provides a prediction modeling system based on working condition division, including:

[0119] Equipment tree module 1, data screening and processing module 2, working condition slicing database module 3, modeling module 4, prediction module 5 and decision analysis module 6;

[0120] The device tree module 1 is connected to the data screening and processing module 2, the data screening and processing module 2 is connected to the working condition slicing database module 3, the working condition slicing database module 3 is connected to the modeling module 4, the modeling module 4 is connected to the prediction module 5, and the prediction module 5 is connected to the decision analysis module 6;

[0121] The equipment tree module 1 includes an equipment object unit 11, an equipment member unit 12 and a modeling time unit 13. The equipment object unit 11 is used to store equipment names, the equipment member unit 12 is used to store all key indicator data information of the equipment, and the modeling time unit 13 is used to store modeling time period data information; the equipment object unit is used to store equipment names, and the equipment names include: equipment name information of the boiler system, equipment name information of the steam turbine system, equipment name information of the desulfurization system, and equipment name information of the denitration system;

[0122] The device member unit is used to store all key indicator information of the device, and the key indicator information includes indicator name data information and indicator code data information;

[0123] The modeling time unit is used to store the modeling start time and end time, and the time period from the modeling start time to the end time represents the duration of the historical data used for modeling.

[0124] The data screening and processing module 2 includes an indicator data unit 21, a screening unit 22, a stability detection unit 23 and a standard unit 24. The indicator data unit 21 is used to obtain the data of the equipment tree research indicator, the screening unit 22 is used to screen the obtained research indicator data to remove abnormal data values, the stability detection unit 23 is used to detect the screened research indicator data to remove unstable data, and the standard unit 24 is used to normalize the research indicator data after stability detection; the indicator data unit is used to obtain the research indicator data of the equipment and obtain the equipment tree data of the corresponding indicator data;

[0125] The screening unit is used to obtain the upper and lower limit values ​​of the indicator data, detect and screen the abnormal data that exceeds the upper and lower limit values ​​of the indicator, and clear the screened abnormal data;

[0126] The stability detection unit is used to calculate the change value of the indicator value within the continuous time window, and then determine whether the device is stable according to whether the change value exceeds the threshold, and remove unstable data;

[0127] The normalization unit is used to perform a normalization operation on the data.

[0128] The working condition slicing database module 3 includes a working condition division unit 31 and a working condition storage unit 32. The working condition division unit 31 is used to receive data from the data screening and processing module unit, and then divide the working condition according to the working condition type. The working condition storage unit 32 is used to store the divided working condition data;

[0129] The modeling module 4 includes a model unit 41 and a model storage unit 42. The model unit 41 is used to receive data from the working condition slicing database module and perform predictive modeling on the data after the working conditions are divided. The model storage unit 42 is used to store the model results of each modeling according to the working conditions.

[0130] The prediction module 5 includes a working condition matching unit 51 and an operation prediction unit 52. The working condition matching unit 51 matches the stored working condition with the real-time data to obtain the working condition type. The operation prediction unit 52 obtains the model stored in the modeling module according to the working condition type, and runs the prediction model to obtain the prediction result.

[0131] The decision analysis module 6 includes a diagnostic unit 61 and an energy-saving optimization unit 62. The diagnostic unit 61 is used to process the prediction index and assist in decision-making and analysis of the equipment. The energy-saving optimization unit 62 optimizes the prediction index so that the prediction index reaches a safe range. The diagnostic unit is also used to determine whether the predicted index data exceeds the limit and alarm according to the index threshold.

[0132] Embodiment 2:

[0133] This embodiment takes the prediction of the pH value of the absorption tower of the desulfurization system of a thermal power plant as the research object. The pH value of the absorption tower (abbreviated as PH) affects the dissolution of limestone in the slurry and affects the efficiency of the entire desulfurization system. However, the acquisition of the pH value has a lag. The detailed elaboration of this embodiment further illustrates the implementation process of the present invention.

[0134] The system content for predicting the pH value of the absorption tower of the desulfurization system according to the embodiment of the present invention includes: an equipment tree module, a data screening and processing module, a working condition slicing database module, a modeling module, a prediction module and a decision analysis module.

[0135] The device tree module includes a device object unit, a device member unit and a modeling time unit. The device object unit is used to store the name of the desulfurization system. Specifically, the desulfurization system contains fewer sub-devices and the desulfurization system is treated as a device. The device member unit is used to store key indicator information of the desulfurization system, and the modeling time unit is used to store the modeling time period.

[0136] The data screening and processing module includes an indicator data unit, a screening unit, a stability detection unit and a standard unit. The indicator data unit is used to read the research indicator data of the desulfurization system. The research indicators include: unit load, slurry supply, pH value, inlet SO2 quality and slurry circulation pump current. The screening unit is used to screen and remove abnormal values. Specifically, the normal range of unit load, slurry supply and pH value is set, and abnormal data is filtered and deleted. The stability detection unit is used to remove unstable data. Specifically, the pH data stability detection threshold is set and unstable data is removed. The standard unit is used to normalize the processed desulfurization data.

[0137] The operating condition slicing database module is used to divide the processed desulfurization data into operating conditions and store them in the database in operating condition slices, including an operating condition division unit and an operating condition storage unit. The operating condition division unit is used to divide the operating conditions according to the operation mode of the desulfurization system, and the operating condition storage unit is used to store the desulfurization modeling data after the operating condition division.

[0138] The modeling module includes a model unit and a model storage unit. The model unit is used to receive desulfurization modeling data according to different working conditions and perform predictive modeling on it according to the working conditions. The model storage unit is used to store the models of each desulfurization system according to different working conditions.

[0139] The prediction module includes an operating condition matching unit and an operation prediction unit. The operating condition matching unit matches the operating condition stored in the desulfurization system model with the desulfurization real-time data to obtain the operating condition type. The operation prediction unit runs the model corresponding to the operating condition according to the operating condition type to obtain the predicted value of PH.

[0140] The decision analysis module determines whether the predicted value of pH exceeds the threshold value according to the threshold value of pH value, and thereby determines whether the operation of the desulfurization system equipment is energy-saving.

[0141] Embodiment 3:

[0142] like Figure 2 As shown, the present embodiment provides a prediction modeling method based on working condition division, comprising the following steps:

[0143] Step S1: a step of obtaining a device tree of a research object; the information of the device tree obtained includes:

[0144] The name information of subordinate equipment of the boiler system, the name information of subordinate equipment of the steam turbine system, the name information of subordinate equipment of the desulfurization system, and the name information of subordinate equipment of the denitrification system are obtained, and all key indicator information of the equipment is obtained, and the key indicator information includes indicator name data information and indicator code data information; the modeling start time and end time are stored, and the time period from the modeling start time to the end time represents the duration of the historical data used for modeling.

[0145] Step S2: Reading the data of the research object and screening and processing the data, which specifically includes:

[0146] Step S21: receiving the device name, indicator information and modeling time of the device tree, and reading the modeling data corresponding to the research indicator according to the research indicator;

[0147] Step S22: obtaining the upper and lower limit values ​​of the indicator data, detecting and filtering the abnormal data exceeding the upper and lower limit values ​​of the indicator, and removing the filtered abnormal data;

[0148] Step S23: Calculate the change value of the indicator value in the continuous time window, and then determine whether the device is stable according to whether the change value exceeds the threshold, and remove unstable data;

[0149] Step S24: finally normalizing the data;

[0150] Step S3: dividing the processed data into working conditions and establishing a working condition slice database, which specifically includes:

[0151] Receive the data after data screening and processing, and divide the received data into working conditions; determine the type of working condition division according to the configured working condition division method, and divide the data into various working conditions according to the working condition boundaries and the indicator range of the indicators for data with clear working condition boundaries; for research objects without clear working condition boundaries, use mathematical model clustering algorithms to divide the data into working conditions; divide the working conditions based on different operating modes or combinations; and classify the results after working condition division into different working conditions and store them.

[0152] Step S4: a step of establishing a prediction model for each operating condition data, which includes:

[0153] Receive data modeling data from the working condition slice database, establish prediction models for different working condition data, and store the model results by working condition;

[0154] The prediction models include linear regression model, autoregressive moving average model (ARMA), multilayer perceptron (MLP), extreme learning machine (ELM), deep learning long short-term memory model (LSTM) and radial basis function neural network model (RBF);

[0155] The prediction model is selected in two ways: free selection and automatic selection of the best method. The automatic selection of the best method is carried out through the following steps:

[0156] The root mean square error (RMSE) method is used to select the optimal model, and the calculation is shown in formula (1):

[0157]

[0158] where y i is the true value of a sample, y' i is the predicted value of the sample under the kth prediction model, and n is the number of samples;

[0159] The root mean square error of each prediction model is calculated respectively, and the model with the smallest error is automatically selected as the prediction model, thereby realizing automatic selection of the optimal model.

[0160] In step S4, the default configuration uses the LSTM algorithm as the prediction model;

[0161] LSTM is an improved algorithm of the recurrent neural network algorithm. It can solve the gradient explosion and gradient vanishing problems in recurrent neural networks. It performs well in time series data modeling and is now widely used. The main components of LSTM are cells, and the cell state is updated or deleted by setting the gate structure.

[0162] Each LSTM neural network contains three gates to control the cell state, namely the forget gate, input gate and output gate. The input gate, output gate and forget gate are represented by i, f and o respectively, the neural network weight matrix is ​​represented by w, and the bias term is represented by b. The most critical gate is the forget gate. By setting the activation function, its output value determines the information that needs to be discarded for the cell state.

[0163] Assume that at time t, the LSTM hidden layer input vector is x t , the output vector is y t , the cell state is represented by state(t).

[0164] The value of the forget gate at time t is shown in formula (2):

[0165] f t =g(w f *(x t ,y t-1 )+b f ) (2)

[0166] Where g indicates that the activation function is the sigmoid function, w f 、b f Represent the weight matrix and bias of the forget gate respectively.

[0167] The input gate determines the cell information update, and the input gate value i t And the candidate state value of the input cell hs t The calculation is as follows:

[0168] i t =g(w i *(x t ,y t-1 )+b i ) (3)

[0169] hs t =tanh(w n *(x t ,y t-1 )+b n ) (4)

[0170] Among them, w i 、b i are the weight matrix and bias of the input gate, w n 、b n is the weight matrix and bias of the current state of the cell, and tanh indicates that the activation function is a hyperbolic tangent function.

[0171] By candidate state value hs t , calculate and update the cell state value at time t.

[0172] state(t)=i t *hs t +f t *state(t-1) (5)

[0173] Finally, the value of the output gate determines the output of the information, and the final output is y t By output gate o t Determined together with the cell state state.

[0174] o t =g(w out *(x t ,y t-1 )+b out ) (6)

[0175] y t =o t *tanh(state(t)) (7)

[0176] where w out 、b out Represent the weight matrix and bias of the output gate respectively.

[0177] The essence of the LSTM prediction model training process is the training of network model parameters. Preferably, the present invention adopts the back propagation algorithm (BPTT) when training the LSTM prediction model.

[0178] Step S5: a step of performing matching in real time and predicting indicators, which specifically includes:

[0179] Read the real-time data of the equipment, obtain the working condition of the real-time data matching the stored model results, obtain the working condition type, obtain the corresponding model according to the working condition type, run the model, and obtain the prediction result;

[0180] Step S6: Processing the prediction results to assist decision-making analysis;

[0181] According to the prediction results, the prediction indicators are applied to determine whether the prediction indicators are over-limit or optimal, and to assist in guiding the safe and economical operation of equipment production; for indicators that are over-limit, adjust the equipment operation indicators in advance to make the indicators reach a safe range; for indicators that are not optimal, adjust the controllable operation indicators in advance to make the indicators close to the optimized values.

[0182] Embodiment 4:

[0183] The method steps for predicting the pH value of the absorption tower of the desulfurization system in the embodiment of the present invention are as follows:

[0184] Step S1: Obtain the equipment tree of the desulfurization system. This step specifically includes:

[0185] Use the equipment tree module to obtain the name of the desulfurization system, key indicator information and modeling time. Specifically, use the equipment object unit to obtain the name of the equipment object desulfurization system. The desulfurization system contains fewer sub-equipment, and the desulfurization system is treated as a device. Use the equipment member unit to obtain the equipment object of the desulfurization system, including the key indicators and indicator codes of the desulfurization system. Specific key indicators include: unit load, slurry circulation pump current, inlet flue gas SO2 conversion concentration, absorption tower liquid level, pH value, absorption tower inlet flue gas conversion flow, absorption tower outlet pressure, etc. Use the modeling time unit to obtain the start and end time of the desulfurization system modeling.

[0186] Step S2: reading the data of the desulfurization system and screening and processing the data, which specifically includes:

[0187] The indicator data unit is used to receive the equipment tree module information, and research indicators are selected from the key indicators of the desulfurization system: unit load, slurry supply, pH value, inlet SO2 quality and slurry circulation pump current. The data within the modeling time period of the above indicators is obtained according to the received modeling time as the modeling analysis data of the desulfurization system.

[0188] The screening and detection unit is used to screen the indicators of the modeling and analysis data, and the range of the indicator unit load, slurry supply and pH value is set to screen the indicators and delete the data that exceeds the indicator range.

[0189] Using the screening detection unit, the data is tested for stability according to the threshold value of the pH data stability detection, and the stability data is retained.

[0190] The received desulfurization data is normalized by the standardization unit and transmitted to the working condition slice database module.

[0191] Step S3: dividing the processed desulfurization data into different working conditions and establishing a working condition slice database, which specifically includes:

[0192] The working condition division unit is used to divide the working conditions of the processed desulfurization data. Specifically, the slurry circulation pump of the desulfurization system is an important equipment of the desulfurization system. Its main function is to provide slurry for desulfurization. The opening of different pumps of the slurry circulation pump has a significant impact on the slurry volume. The basis for the working condition division is to use different operation modes or combinations as the clear basis for the working condition division. Specifically, the desulfurization system of the power plant has 6 slurry circulation pumps, and the circulation pumps are represented by the numbers AF. The working conditions are divided according to the opening and operation of the pumps. For example, the ACD pump working condition represents the opening of the A, C and D pumps. Whether the pump is turned on and running is determined according to the pump current value. 0 represents that the pump is not turned on, and 1 represents that the pump is turned on. The ACD pump working condition is represented by the working condition code "101100".

[0193] The working condition storage unit is used to store the data after the working conditions of the desulfurization system are divided. Specifically, it includes the working condition code of each working condition and the corresponding PH modeling index data after the working conditions are divided. The modeling indicators include: inlet SO2 quality, slurry supply value and PH value.

[0194] In step S3, a mathematical model clustering algorithm is used to divide the data into working conditions; in this step, one of a k-means clustering algorithm, a hierarchical clustering algorithm, and a Gaussian mixture model (GMM) clustering method can be used to divide the working conditions;

[0195] In step S3, when performing mathematical model clustering, the Gaussian mixture clustering algorithm is configured by default;

[0196] Gaussian mixture model clustering (GMM) is based on the Gaussian mixture model. It constructs the most appropriate mixed Gaussian distribution model by fitting the sample data set. It is the fastest learning probability model. The model outputs the probability value of each sample belonging to the model. It carries more information than the clustering method that simply assigns it to a certain category, and the clustering results obtained based on the probability value are more accurate.

[0197] Assuming that the sample obeys a mixed Gaussian distribution and there are n distribution models, for a certain sample x, the Gaussian mixture model can be expressed by formula (8):

[0198]

[0199] where θ i is the coefficient of the i-th model, and the sample is a weighted combination of n models, then θ is the weight vector of the model, p(x|u i ,δ i ) is the probability density function of the ith model.

[0200] where μ i is the mean vector of the ith model.

[0201] where δ i is the covariance matrix of the ith model.

[0202] The key to model training is to estimate the model parameter weights, mean and covariance parameters, which are θ, μ, δ respectively, and calculate the probability density of the model, as shown in formula (9).

[0203]

[0204] The steps of the Gaussian mixture model clustering algorithm are as follows:

[0205] Step T1: Construct n mixed component models and initialize the model parameters θ, μ, δ

[0206] Construct n Gaussian distribution mixture component models and initialize the parameters of the n component models respectively. For the i-th model, the parameter is θ i , μ i , δ i .

[0207] Step T2: Calculate the posterior probability

[0208] For the qth sample x q , the probability under the i-th model is the posterior probability, which can be represented as: ph qi =(y q =i|x q ).

[0209] The specific probability calculation is shown in formula (10):

[0210]

[0211] Further calculations yield:

[0212]

[0213] Step T3: Maximum likelihood solution to obtain model parameters

[0214] Solve the maximum likelihood function and the maximum likelihood value to obtain the values ​​of each parameter. For detailed calculation, see formulas (12)-(14). And update the parameters of the n component models, and update the parameters of the i-th model: θ i , μ i , δ i .

[0215] The mean μ of the i-th model i :

[0216]

[0217] The variance δ of the i-th model i :

[0218]

[0219] The weight θ of the i-th model i :

[0220]

[0221] Step T4: Repeat steps T2 and T3 until the algorithm converges

[0222] Repeat steps T2 and T3 until the algorithm converges and ends the algorithm learning, and save the final parameters of each component model.

[0223] The probability value of the sample under each component model is calculated, and then the sample is divided into clusters. The model to which the sample belongs depends on the probability value, and the model with the larger probability value belongs to the model.

[0224] Typically, when you run data clustering on a sample, you set the number of clusters.

[0225] The present invention adopts the Bayesian Information Criterion (BIC) to determine the number of clusters, and selects the number of clusters corresponding to the smallest BIC value as the optimal number of clusters. The calculation of the BIC value is detailed in formula (15).

[0226] BIC=mln(n)-2ln(L) (15)

[0227] Where m is the number of parameters in the model, n is the number of samples, and L is the likelihood function.

[0228] Step S4: for the data of the divided working conditions, a step of establishing a pH value prediction model is performed. The step specifically includes:

[0229] The model unit is used to slice the data of different working conditions for predictive modeling, and the modeling algorithm LSTM is used to establish the predictive model. The specific model inputs are: inlet SO2 quality at time t, slurry supply value at time t, pH value at time t, and model output: pH value at time t+1.

[0230] The model storage unit is used to store the modeling results of the prediction model according to the working conditions, specifically including the working condition code and the modeling results of the prediction model corresponding to the working condition.

[0231] Step S5: Real-time matching and pH value prediction, which specifically includes:

[0232] The operating condition matching unit is used to read the real-time data of the desulfurization system. The specific real-time data includes indicators: inlet SO2 quality, slurry supply value, slurry circulation pump current, and the operating condition is matched according to the slurry circulation pump current value to determine the operating condition type.

[0233] Using the running prediction unit, first use the corresponding model results to predict the pH value based on the matched operating conditions. The specific model inputs are: the inlet SO2 mass at the previous moment, the slurry supply value at the previous moment, and the pH value at the previous moment. The model prediction output is: the pH value at the current moment, which is the predicted value of the pH.

[0234] Step S6: Setting the pH threshold to assist in the decision-making analysis of the desulfurization system, which specifically includes:

[0235] Using the energy-saving optimization unit, the normal upper and lower limits of the pH value for energy saving are set in the range of 4.5-6.4. If the predicted pH value deviates from the normal range, an alarm notification is issued to remind production and operation personnel to pay attention to the status of the desulfurization system in time and adjust the desulfurization operation until the pH reaches a safe energy-saving range.

[0236] The above disclosure is only a preferred embodiment of the present invention, but the present invention is not limited thereto. Any non-creative changes that can be thought of by a person skilled in the art, as well as several improvements and modifications made without departing from the principle of the present invention, should fall within the protection scope of the present invention.

Claims

1. A prediction modeling system based on working condition division, characterized in that: include: Equipment tree module, data screening and processing module, working condition slicing database module, modeling module, prediction module and decision analysis module; The device tree module is connected to the data screening and processing module, the data screening and processing module is connected to the sub-condition slicing database module, the sub-condition slicing database module is connected to the modeling module, the modeling module is connected to the prediction module, and the prediction module is connected to the decision analysis module; The device tree module includes a device object unit, a device member unit and a modeling time unit. The device object unit is used to store the device name, the device member unit is used to store all key indicator data information of the device, and the modeling time unit is used to store the modeling time period data information; The data screening and processing module includes an indicator data unit, a screening unit, a stability detection unit and a standard unit. The indicator data unit is used to obtain data of device tree research indicators. The screening unit is used to screen the obtained research indicator data to remove abnormal data values. The stability detection unit is used to detect the screened research indicator data to remove unstable data. The standard unit is used to normalize the research indicator data after stability detection. The working condition slicing database module includes a working condition division unit and a working condition storage unit. The working condition division unit is used to receive data from the data screening and processing module unit, and then divide the working condition according to the working condition type. The working condition storage unit is used to store the divided working condition data; The modeling module includes a model unit and a model storage unit. The model unit is used to receive data from the working condition slicing database module and perform predictive modeling on the data after the working conditions are divided. The model storage unit is used to store the model results of each modeling according to the working conditions. The prediction module includes a working condition matching unit and an operation prediction unit. The working condition matching unit matches the stored working condition with the real-time data to obtain the working condition type. The operation prediction unit obtains the model stored in the modeling module according to the working condition type and runs the prediction model to obtain the prediction result. The decision analysis module includes a diagnosis unit and an energy-saving optimization unit. The diagnosis unit is used to process the prediction index and assist in decision-making and analysis of the equipment. The energy-saving optimization unit optimizes the prediction index to make it reach a safe range. The equipment object unit is used to store equipment names, and the equipment names include: equipment name information of a boiler system, equipment name information of a steam turbine system, equipment name information of a desulfurization system, and equipment name information of a denitration system; The device member unit is used to store all key indicator information of the device, and the key indicator information includes indicator name data information and indicator code data information; The modeling time unit is used to store the modeling start time and end time, and the time period from the modeling start time to the end time represents the duration of the historical data used for modeling.

2. A prediction modeling system based on working condition division according to claim 1, characterized in that: The indicator data unit is used to obtain research indicator data of the device and obtain device tree data corresponding to the indicator data; The screening unit is used to obtain the upper and lower limit values ​​of the indicator data, detect and screen the abnormal data that exceeds the upper and lower limit values ​​of the indicator, and clear the screened abnormal data; The stability detection unit is used to calculate the change value of the indicator value within the continuous time window, and then determine whether the device is stable according to whether the change value exceeds the threshold, and remove unstable data; The normalization unit is used to perform a normalization operation on the data.

3. A prediction modeling system based on working condition division according to claim 2, characterized in that: The diagnosis unit is also used to determine whether the predicted indicator data exceeds the limit and issue an alarm based on the indicator threshold.

4. A modeling method of a predictive modeling system based on working condition division according to any one of claims 1 to 3, characterized in that: The following steps are involved: Step S1: a step of obtaining a device tree of a research object; Step S2: Reading the data of the research object and screening and processing the data, specifically including: Step S21: receiving the device name, indicator information and modeling time of the device tree, and reading the modeling data corresponding to the research indicator according to the research indicator; Step S22: obtaining the upper and lower limit values ​​of the indicator data, detecting and filtering the abnormal data exceeding the upper and lower limit values ​​of the indicator, and removing the filtered abnormal data; Step S23: Calculate the change value of the indicator value in the continuous time window, and then determine whether the device is stable according to whether the change value exceeds the threshold, and remove unstable data; Step S24: finally normalizing the data; Step S3: dividing the processed data into working conditions and establishing a working condition slice database, specifically including: Receive the data after data screening and processing, and divide the received data into working conditions; Step S4: The step of establishing a prediction model for each operating condition data includes: Receive data modeling data from the working condition slice database, establish prediction models for different working condition data, and store the model results by working condition; The prediction models include linear regression model, autoregressive moving average model, multi-layer perceptron, extreme learning machine, deep learning long short-term memory model and radial basis neural network model; The prediction model is selected in two ways: free selection and automatic selection of the best method. The automatic selection of the best method is carried out through the following steps: The root mean square error method is used to select the optimal model, and the calculation is shown in formula (1): (1) in is the true value of a sample, is the predicted value of the sample under the kth prediction model, and n is the number of samples; Calculate the root mean square error of each prediction model respectively, and automatically select the model with the smallest error as the prediction model, so as to automatically select the optimal model; Step S5: Real-time matching and index prediction steps, specifically including: Read the real-time data of the equipment, obtain the working condition of the real-time data matching the stored model results, obtain the working condition type, obtain the corresponding model according to the working condition type, run the model, and obtain the prediction result; Step S6: Processing the prediction results to assist decision-making analysis; In step S1, obtaining the information of the device tree includes: The name information of the subordinate equipment of the boiler system, the name information of the subordinate equipment of the steam turbine system, the name information of the subordinate equipment of the desulfurization system, and the name information of the subordinate equipment of the denitrification system are obtained, and all the key indicator information of the equipment is obtained, and the key indicator information includes the indicator name data information and the indicator code data information; the modeling start time and end time are stored, and the time period from the modeling start time to the end time represents the duration of the historical data used for modeling; In step S3: According to the configured working condition division method, the type of working condition division is determined. For data with clear working condition boundaries, each working condition is divided according to the working condition boundaries and the indicator range of the indicators. For research objects without clear working condition boundaries, the mathematical model clustering algorithm is used to divide the data into working conditions. Different operating modes or combination modes are used as the basis for working condition division. The results after working condition division are classified and stored according to working conditions. In step S6: According to the prediction results, the prediction indicators are applied to determine whether the prediction indicators are over-limit or optimal, and to assist in guiding the safe and economical operation of equipment production; for indicators that are over-limit, adjust the equipment operation indicators in advance to make the indicators reach a safe range; for indicators that are not optimal, adjust the controllable operation indicators in advance to make the indicators close to the optimized values.

Images