Multi-step K-means clustering-based nuclear power unit operation condition division and optimization method

Abstract

The invention belongs to the technical field of intelligent operation of nuclear power units, and particularly relates to a nuclear power unit operation condition division and optimization method based on multi-step K-means clustering. Comprising the following steps of 1, mining an optimal working condition in historical operation data, and outputting key parameters in a system under the working condition; 2, the similarity of boundary disturbance variables is judged; step 3, obtaining unit operation historical data, and removing abnormal values; 4, judging a steady-state working condition; 5, carrying out clustering analysis on the cleaned historical data; step 6, judging the clustering result of each step; step 7, obtaining values of boundary condition thermal power and seawater temperature of the current unit operation condition, forming a two-dimensional vector, and standardizing the vector; and 8, judging the working condition category formed by the clustering algorithm to which the current operation working condition belongs. The method has the beneficial effects that a reliable data training set is formed, and the accuracy of a clustering algorithm and working condition optimization is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent operation of nuclear power units, and specifically relates to a method for dividing and optimizing operating conditions of nuclear power units based on multi-step K-means clustering. Background Art

[0002] In the operation of nuclear power units, in order to improve the operating efficiency of the generating units, it is necessary to optimize the operating conditions of the units to find the best state that the units can achieve under historical operating conditions. In recent years, with the development of big data and artificial intelligence technologies, data-driven operating condition optimization methods have gradually become a research hotspot. In the existing technology, operating condition optimization is usually based on a single boundary condition or a simple mathematical model, or rule-based empirical optimization. The processing and cleaning of historical data is not sufficient, and some outliers still exist in the historical benchmark library, affecting the optimization accuracy. Summary of the invention

[0003] The purpose of the present invention is to provide a method for dividing and optimizing the operating conditions of a nuclear power unit based on multi-step K-means clustering. Through in-depth analysis of historical operating data and application of a multi-step clustering algorithm, the parameter that has the greatest impact on the unit's power generation is selected as a boundary condition, and the historical optimal operating conditions that are similar to the boundary conditions of the current operating conditions are mined. The deviation between the current operating conditions and the historical optimal conditions is analyzed, so as to achieve the real-time and dynamic performance of the condition optimization, and improve the unit's power generation efficiency, power generation capacity and the economic benefits of the power plant.

[0004] The technical solution of the present invention is as follows: A method for dividing and optimizing the operating conditions of a nuclear power unit based on multi-step K-means clustering, comprising the following steps:

[0005] Step 1: Under the condition that the control boundary disturbance variables are similar, with the maximum output power of the generator as the main goal, the optimal operating conditions in the historical operation data are mined, and the key parameters of the system under this condition are output;

[0006] Step 2: Determine the similarity of boundary disturbance variables;

[0007] Step 3: Obtain historical operation data of the unit and remove abnormal values;

[0008] Step 4: Use polynomial filtering algorithm to identify steady-state conditions;

[0009] Step 5: Use the multi-step Kmeans clustering algorithm to perform cluster analysis on the above cleaned historical data;

[0010] Step 6: Use the silhouette coefficient to judge the clustering results of each step;

[0011] Step 7: Obtain the boundary condition thermal power and seawater temperature of the current unit operating condition, form a two-dimensional vector, and standardize the vector;

[0012] Step 8: Determine the operating condition category formed by the above clustering algorithm to which the current operating condition belongs.

[0013] In step 2, according to the unit operation mechanism, two parameters, namely, the hot-end parameter reactor thermal power and the cold-end parameter seawater temperature, are selected as boundary disturbance variables, and the similarity of the boundary disturbance variables is determined by Euclidean distance.

[0014] In step 3, the 3σ principle is used to eliminate abnormal values, where σ is the standard deviation of the measured values ​​of a certain parameter over a period of time.

[0015] The polynomial filtering algorithm in step 4 is to express a certain measured variable in the form of a polynomial with time as the independent variable, as follows:

[0016] x(t)=p0+p1t+p2t 2 +p3t 3 +.......p m t m

[0017] Among them, x(t) is the function of time, m is the order of the model, p0, p1, ....p m is the parameter to be solved, let θ=[p0,p1,....p m ] T is the parameter vector of the model, r(t)=[t0,t1,....t m ] T is a regression variable. When the measured signal is known, the least square method can be used to obtain the optimal estimate of the parameter, as follows:

[0018]

[0019] In the above formula,

[0020] After polynomial filtering of the data in the window, the value of P1 is used to judge the steady-state effective working condition, and the judgment threshold is determined by the 3δ rule.

[0021]

[0022] Where δ is the standard deviation of the measurement error, H is the length of the sliding window of the data, and when |p1| < λ, it is considered that the unit load is in a stable state within the observation window, otherwise it is unstable.

[0023] In step 5, thermal power and seawater temperature are used as clustering features, and only one feature is used to perform clustering training on historical data each time. This is multi-step clustering, and multiple groups of working condition categories are formed after clustering.

[0024] The calculation formula for a single sample of the silhouette coefficient in step 6 is as follows:

[0025]

[0026] Among them, a(i) represents the similarity within the cluster, which means the average distance between sample i and other samples in the same cluster; b(i) is the dissimilarity between clusters, which means the average distance between the sample and all samples in other clusters. The larger the silhouette coefficient, the better the clustering effect.

[0027] The expression for normalization to Min-Max in step 7 is as follows:

[0028]

[0029] x is the current measured value of a parameter, x max The maximum value of this parameter in the historical data, x min The minimum value in the history of this parameter.

[0030] In step 8, based on the center point of the above-mentioned category, a group of historical operating data with the largest power generation is searched for as the optimal operating condition for the current unit operation.

[0031] The beneficial effects of the present invention are: 1. The historical data is deeply cleaned, and the data due to measurement point drift and non-steady-state conditions are deleted to form a reliable data training set, which is beneficial to improving the accuracy of clustering algorithms and working condition optimization. 2. Multi-step Kmeans clustering is adopted, and only one feature is selected for each clustering, which avoids data normalization before clustering begins, and also avoids the unreliability of distance in multi-feature clustering in Euclidean space. 3. The working condition optimization is based on the category center point for diffusion optimization, avoiding the influence of individual outliers on the optimization results. 4. The current working condition and the historical database are benchmarked in real time to achieve the real-time and dynamic nature of working condition optimization. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 It is a schematic diagram of the multi-step Kmeans clustering algorithm;

[0033] Figure 2 Schematic diagram of the model training and working condition optimization process. DETAILED DESCRIPTION

[0034] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0035] A method for dividing and optimizing the operating conditions of nuclear power units based on multi-step K-means clustering includes the following steps:

[0036] Step 1: Under the condition that the control boundary disturbance variables are similar, with the maximum generator output power as the main goal, find a set of operating parameters with the maximum generator output power from the category of similar operating conditions as the historical best operating condition, and output the key parameters of the system under this condition. The key parameters include reactor thermal power, unit back pressure, main steam pressure, main steam temperature, final feed water pressure, and final feed water temperature.

[0037] Step 2: According to the unit operation mechanism (the reactor provides heat for the steam turbine generator set, and the seawater takes away the remaining heat after the steam turbine generator set works, and these two parameters have the greatest impact on the power generation), the hot end parameter reactor thermal power and the cold end parameter seawater temperature are selected as boundary disturbance variables, and the Euclidean distance is used to judge the similarity of boundary disturbance variables. The operating parameters corresponding to similar boundary disturbance variables will be divided into the same category of operating condition groups.

[0038] Step 3: Obtain the historical operation data of the unit (operation data collected by sensors at various locations in the thermal system of the power plant, including physical parameters such as temperature, pressure, and flow) and use the 3σ principle to eliminate abnormal values, where σ is the standard deviation of the measured value of a certain parameter over a period of time.

[0039] Step 4: There are large fluctuations in the operating history data, which is in a non-steady state. The operating condition optimization needs to be performed when the unit is operating in a steady state. The present invention uses a polynomial filtering algorithm to determine the steady-state operating condition;

[0040] Polynomial filtering is a low-pass filtering algorithm based on a polynomial regression model, which can effectively remove the high-frequency part of the measured signal noise and improve the detection accuracy.

[0041] Specifically, a certain measured variable is expressed as a polynomial with time as the independent variable, as follows:

[0042] x(t)=p0+p1t+p2t 2 +p3t 3 +.......p m t m

[0043] Among them, x(t) is the function of time, m is the order of the model, p0, p1, ....p m is the parameter to be solved, let θ=[p0,p1,....p m ] T is the parameter vector of the model, r(t)=[t0,t1,....t m ]T is the regression variable. When the measured signal is known, the least square method can be used to obtain the optimal estimate of the parameters, as follows:

[0044]

[0045] In the above formula,

[0046] After polynomial filtering of the data in the window, the value of P1 can be used to judge the steady-state effective working condition, and the judgment threshold is determined by the 3δ rule.

[0047]

[0048] Where δ is the standard deviation of the measurement error, H is the length of the sliding window of the data, and when |p1| < λ, it is considered that the unit load is in a stable state within the observation window, otherwise it is unstable.

[0049] After the discrimination is completed, stable running data is output for training the clustering algorithm.

[0050] Step 5: Use the multi-step Kmeans clustering algorithm to perform cluster analysis on the above-mentioned cleaned historical data (data after steps 3 and 4), with thermal power and seawater temperature as clustering features. Only one feature is used to cluster the historical data each time. This is multi-step clustering, and multiple groups of working condition categories are formed after clustering.

[0051] Step 6: Use the silhouette coefficient to judge the clustering results of each step to achieve the best clustering effect.

[0052] Silhouette coefficient is a way to evaluate the quality of clustering. It combines the two factors of cohesion and separation. It can be used to evaluate the impact of different algorithms or different operation modes of algorithms on clustering results based on the same original data. The calculation formula for a single sample is as follows:

[0053]

[0054] Among them, a(i) represents the similarity within the cluster, which means the average distance between sample i and other samples in the same cluster; b(i) represents the dissimilarity between clusters, which means the average distance between the sample and all samples in other clusters. The larger the silhouette coefficient, the better the clustering effect.

[0055] Step 7: Obtain the boundary conditions of the current unit operating conditions, thermal power and seawater temperature, form a two-dimensional vector, and standardize the vector so that each element of the vector is between 0 and 1, in order to reduce the impact of different numerical values ​​on the Euclidean distance calculation. Min-Max is a commonly used vector standardization calculation method. This invention uses Min-Max for vector standardization.

[0056] The expression of Min-Max is as follows:

[0057]

[0058] x is the current measured value of a parameter, x max The maximum value of this parameter in the historical data, x min The minimum value in the history of this parameter.

[0059] Step 8: Determine the operating condition category formed by the above clustering algorithm to which the current operating condition belongs.

[0060] Based on the center point of the above categories, a group of historical operating data with the largest power generation is searched as the best operating condition for the current unit. The standardized vector obtained in step 7 is compared with the category center formed in step 5 one by one, and the category with the closest distance is the category to which the current operating condition belongs.

[0061] Example

[0062] A nuclear power unit operating condition division and optimization method based on multi-step K-means clustering is proposed. The historical operating data of the measuring points of the operating nuclear power units are obtained through the power plant time series database, including various physical parameters of the unit operation, such as steam generator outlet pressure, main steam valve front pressure, feed water flow, high pressure heater outlet feed water mother pipe temperature, seawater temperature, thermal power, etc.

[0063] For the above-obtained measurement point operation data, the 3σ principle is used to eliminate abnormal values. Specifically, for each measurement point, the average value x^ and standard deviation σ of the historical data of this measurement point are calculated, and the data outside x^±3σ are eliminated.

[0064] Taking electric power as the target, the data under 90% of the designed electric power in the cleaned data in

[0003] are deleted again.

[0065] Taking electric power as the characteristic basis for judging steady-state conditions, the model order is set to 2, the sliding window length is set to 30, and the polynomial filtering algorithm is calculated to determine the steady-state condition of the unit operation and delete the historical data of non-steady-state conditions.

[0066] Taking thermal power as the clustering feature, the first step of Kmeans clustering is performed. The number of clusters is set to 2-6. The clustering algorithm is trained for each number of clusters, and the corresponding silhouette coefficient is calculated. The number of clusters when the silhouette coefficient is the largest is taken as the clustering result of the first step. At the same time, the encoding and centroid data of each category are output.

[0067] Taking sea water temperature as the clustering feature, the first step of Kmeans clustering is carried out. For each category formed by

[0006] above, the number of clusters is set to 2-6, the clustering algorithm is trained, and the corresponding silhouette coefficient is calculated. The number of clusters when the silhouette coefficient is the largest is taken as the clustering result under the corresponding category in

[0006] . At the same time, the code and centroid data of each category are output.

[0068] After completing all the clustering training in the two steps, a historical data benchmark library is formed and presented in the form of categories. Each category contains multiple groups of operating conditions and has corresponding encoding and centroid data. Each centroid data is a two-dimensional vector composed of thermal power and seawater temperature.

[0069] Obtain the current unit operation data, extract the measured values ​​of thermal power and seawater temperature, and perform normalization. At the same time, normalize each centroid data in the historical data benchmark library.

[0070] The Euclidean distance between the current operating data and each centroid is calculated respectively, and the category code corresponding to the minimum Euclidean distance is taken as the category to which the current operating condition belongs.

[0071] Taking the seawater temperature in the centroid of the category marked in

[0010] as the center, search for the historical operating conditions when the power generation power is maximum within the range of ±10°C. If the search is successful, the corresponding unit operating condition is the optimal operating condition under the current state; if the search fails, the current operating condition is the optimal one, which is updated to the historical data benchmark library.

Claims

1. A method for dividing and optimizing the operating conditions of nuclear power units based on multi-step K-means clustering, characterized in that: The steps include: Step 1: Under the condition that the control boundary disturbance variables are similar, with the maximum output power of the generator as the main goal, the optimal operating conditions in the historical operation data are mined, and the key parameters of the system under this condition are output; Step 2: Determine the similarity of boundary disturbance variables; Step 3: Obtain historical operation data of the unit and remove abnormal values; Step 4: Determine the steady-state condition; Step 5: Perform cluster analysis on the above cleaned historical data; Step 6: Judge the clustering results of each step; Step 7: Obtain the boundary condition thermal power and seawater temperature of the current unit operating condition, form a two-dimensional vector, and standardize the vector; Step 8: Determine the operating condition category formed by the above clustering algorithm to which the current operating condition belongs.

2. A method for dividing and optimizing the operating conditions of a nuclear power unit based on multi-step K-means clustering as claimed in claim 1, characterized in that: In step 2, according to the unit operation mechanism, two parameters, namely, the hot-end parameter reactor thermal power and the cold-end parameter seawater temperature, are selected as boundary disturbance variables, and the similarity of the boundary disturbance variables is determined by Euclidean distance.

3. The method for dividing and optimizing the operating conditions of a nuclear power unit based on multi-step K-means clustering according to claim 1, characterized in that: In step 3, the 3σ principle is used to eliminate abnormal values, where σ is the standard deviation of the measured values ​​of a certain parameter over a period of time.

4. The method for dividing and optimizing the operating conditions of a nuclear power unit based on multi-step K-means clustering according to claim 1, characterized in that: The polynomial filtering algorithm in step 4 is to express a certain measured variable in the form of a polynomial with time as the independent variable, as follows: x(t)=p0+p1t+p2t 2 +p3t 3 +.......p m t m Among them, x(t) is the function of time, m is the order of the model, p0, p0,....p m is the parameter to be solved, let θ=[p0,p1,....p m ] T is the parameter vector of the model, r(t)=[t0,t1,....t m ] T is a regression variable. When the measured signal is known, the least square method can be used to obtain the optimal estimate of the parameter, as follows: In the above formula, After polynomial filtering of the data in the window, the value of P1 is used to judge the steady-state effective working condition, and the judgment threshold is determined by the 3δ rule. Where δ is the standard deviation of the measurement error, H is the length of the sliding window of the data, and when N < λ, it is considered that the unit load is in a stable state within the observation window, otherwise it is unstable.

5. The method for dividing and optimizing the operating conditions of a nuclear power unit based on multi-step K-means clustering according to claim 1, characterized in that: In step 5, thermal power and seawater temperature are used as clustering features, and only one feature is used to perform clustering training on historical data each time. This is multi-step clustering, and multiple groups of working condition categories are formed after clustering.

6. The method for dividing and optimizing the operating conditions of a nuclear power unit based on multi-step K-means clustering according to claim 1, characterized in that: The calculation formula for a single sample of the silhouette coefficient in step 6 is as follows: Among them, a(i) represents the similarity within the cluster, which means the average distance between sample i and other samples in the same cluster; b(i) is the dissimilarity between clusters, which means the average distance between the sample and all samples in other clusters. The larger the silhouette coefficient, the better the clustering effect.

7. The method for dividing and optimizing the operating conditions of a nuclear power unit based on multi-step K-means clustering according to claim 1, characterized in that: The expression for normalization to Min-Max in step 7 is as follows: x is the current measured value of a parameter, x max The maximum value of this parameter in the historical data, x min The minimum value in the history of this parameter.

8. The method for dividing and optimizing the operating conditions of a nuclear power unit based on multi-step K-means clustering according to claim 1, characterized in that: In step 8, based on the center point of the above-mentioned category, a group of historical operating data with the largest power generation is searched for as the optimal operating condition for the current unit operation.

Images