A machine learning model screening method

Abstract

The present application relates to a kind of machine learning model screening method, comprising: obtaining the original data set in the preset time period in thermoelectric production process and the model set to be evaluated, the model set to be evaluated includes at least one machine learning model to be evaluated;For the original data set is preprocessed, obtains final data set;Final feature variable in final data set is input into any machine learning model to be evaluated in the model set to be evaluated, and the corresponding target variable predicted value is obtained;Based on final target variable actual value and target variable predicted value, the error score and similarity score of the machine learning model to be evaluated are obtained;Based on the error score and similarity score of the machine learning model to be evaluated, the final score of the machine learning model to be evaluated is obtained;Based on the final score of each machine learning model to be evaluated, the final machine learning model is screened out.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent power generation, and particularly relates to a machine learning model screening method. BACKGROUND

[0002] PID control is a common automatic adjustment method in the control of thermal power production process, but PID control is prone to cause problems such as overshoot and oscillation in the system. For the control of equipment in the thermal power production process, there is a machine learning model that is faster and more accurate than the traditional PID control. This model is a model established based on the corresponding relationship between the change amounts of input and output variables, and is mainly used when the equipment needs to be adjusted from one state to another target state. The model can directly output the operation amount that needs to be performed on the basis of the current state according to the difference between the input target state and the current state. This model is essentially a regression model, but the evaluation indicators R2, MAE and MSE of the conventional regression model cannot accurately evaluate the quality of this model, and the evaluation results often contradict the actual effect of the model. This is mainly due to the problems of time delay and fluctuation in the industrial production process data. In many cases, the measurement data in the unit operation process cannot truly reflect the relationship between the input and output of the system, so that the change amount obtained by subtracting the original data is not the true change amount. In addition, the change amount data is generally very small, and the data near 0 is mostly, which further reduces the signal-to-noise ratio of the change amount data.

[0003] In the thermal power production scene, due to the instability of the production conditions, it is necessary to continuously and effectively monitor the machine learning model to realize the timely updating and selection of the machine learning model. It is subjective to determine whether the machine learning model is available by manually observing the goodness of fit of the machine learning model and other conventional evaluation indicators. Not only is it time-consuming and laborious, but also the evaluation of the quality of the machine learning model is not intuitive and comprehensive. Therefore, the most suitable machine learning model for thermal power production cannot be screened out. SUMMARY

[0004] (I) Technical problems to be solved

[0005] In view of the above-mentioned shortcomings and deficiencies of the prior art, the present application provides a machine learning model screening method, which solves the technical problems of low efficiency caused by relying on manual observation and the inability to screen out the most suitable machine learning model for thermal power production due to the inability of the existing evaluation indicators to effectively evaluate the model.

[0006] (II) Technical solutions

[0007] In order to achieve the above-mentioned purposes, the main technical solutions adopted by the present application include:

[0008] The embodiment of the present application provides a machine learning model screening method, the machine learning model is for thermoelectric production process control, and the method comprises the following steps:

[0009] S1, obtaining a to-be-tested model set and an original data set in a preset time period in a thermoelectric production process, wherein the to-be-tested model set comprises at least one to-be-tested machine learning model;

[0010] The original data set comprises a plurality of original data collected in the thermoelectric production process at a preset frequency in the preset time period; wherein each original data comprises a timestamp, an original feature variable and an original target variable corresponding to the original data;

[0011] S2, preprocessing the original data set to obtain a final data set; the final data set comprises a plurality of pairs of final feature variables and final target variable actual values;

[0012] S3, inputting the final feature variable in the final data set into any to-be-tested machine learning model in the to-be-tested model set to obtain a corresponding target variable prediction value;

[0013] S4, obtaining an error score and a similarity score of the to-be-tested machine learning model based on the final target variable actual value and the target variable prediction value;

[0014] S5, obtaining a final score of the to-be-tested machine learning model based on the error score and the similarity score of the to-be-tested machine learning model;

[0015] S6, screening a final machine learning model based on the final score of each to-be-tested machine learning model.

[0016] Preferably, the S2 specifically comprises:

[0017] S21, screening a stationary section data set from the original data set;

[0018] Wherein, any two stationary section data sets have no intersection, and any stationary section data set L i satisfies L i =[l 1i ,l 2i ,...,l ni ];

[0019] L i is the i th stationary section data set screened from the original data set;

[0020] l ni is the n th original data in the i th stationary section data set;

[0021] (lmi-x -l 1i-x ) / l1i-x ≤3%, m≤n;

[0022] (l mi-y -l 1i-y ) / l1i-y ≤3%, m≤n;

[0023] l mi-x is the original feature variable in the mth original data in the ith stationary section data set;

[0024] l mi-y is the original target variable in the mth original data in the ith stationary section data set;

[0025] The stationary section data set L i The first time corresponding to the stationary section data set L mi The time of the timestamp corresponding to the stationary section data set L 1i The time of the timestamp corresponding to the stationary section data set L

[0026] S22, for all stationary section data sets, obtaining the average value of the original feature variable and the average value of the original target variable in each stationary section data set, and the starting time and the ending time of the stationary section data set;

[0027] S23, based on the average value of the original feature variable and the average value of the original target variable in each stationary section data set, and the starting time and the ending time of the stationary section data set, obtaining a first data set;

[0028] The first data set includes information data corresponding to each stationary section data set one by one;

[0029] The information data includes: the average value of the original feature variable and the average value of the original target variable in the stationary section data corresponding to the information data, and the starting time and the ending time of the stationary section data set;

[0030] S24, sorting all information data in the first data set according to the starting time in the information data from small to large, to obtain a second data set;

[0031] S25, using a pre-set processing strategy to process the second data set to obtain a final data set.

[0032] Preferably, the S25 specifically includes:

[0033] S251, for any information data in the second data set, obtain the relative information between each information data after the information data and the information data based on the average value of the original characteristic variable in the information data, and group all the relative information in the second data set to form a third data set;

[0034] Wherein, the relative information between any information data after the information data and the information data in the second data set includes: first difference, second difference, first end time, first start time, second end time, first time difference;

[0035] The first difference is the difference between the average value of the original characteristic variable in any information data after the information data in the second data set and the average value of the original characteristic variable in the information data;

[0036] The second difference is the difference between the average value of the original target variable in any information data after the information data in the second data set and the average value of the original target variable in the information data;

[0037] The first end time is the end time of the stationary segment data set corresponding to the information data;

[0038] The first start time is the start time of the stationary segment data set corresponding to any information data after the information data in the second data set;

[0039] The second end time is the end time of the stationary segment data set corresponding to any information data after the information data in the second data set;

[0040] The first time difference is the difference between the first start time and the first end time;

[0041] S252, according to the relative information in the third data set, filter the third data set according to the pre-set filtering rule, and obtain a fourth data set;

[0042] S253, for any relative information in the fourth data set, the first difference is valued according to the pre-set value mode, and the new first difference corresponding to the relative information is obtained;

[0043] S254, based on the new first difference corresponding to any relative information in the fourth data set, the fourth data set is grouped to obtain at least one set of relative information;

[0044] Wherein, the new first difference corresponding to all the relative information in each set of relative information is equal;

[0045] S255. For each set of relative information, remove the relative information with a first time difference greater than or equal to 48 hours to obtain the set of relative information after removal.

[0046] S256. Sort the set of relative information after the removal according to a pre-set sorting method to obtain the final set of relative information.

[0047] S257. Extract the first 10 pieces of relative information from each final relative information set to form the final dataset;

[0048] In this dataset, the first difference in each relative information is used as the final feature variable; the second difference in each relative information is used as the final target variable's actual value.

[0049] Preferably,

[0050] The pre-defined filtering rules are as follows:

[0051] If the original feature variable and the original target variable are negatively correlated, delete the first relative information in the third dataset;

[0052] The first relative information is relative information where the product of the first difference and the second difference is greater than 0;

[0053] If the original feature variable and the original target variable are positively correlated, delete the second relative information in the third dataset;

[0054] The second relative information is the relative information that the product of the first difference and the second difference is less than 0.

[0055] Preferably,

[0056] The preset value selection method is as follows:

[0057] When the first valid value obtained in advance is less than 1, one valid digit is retained in the first difference to obtain a new first difference;

[0058] When the first valid value obtained in advance is greater than or equal to 1 and less than 10, the first difference is retained to one decimal place to obtain a new first difference;

[0059] When the first valid value obtained in advance is greater than or equal to 10 and less than 100, the first difference is rounded to obtain a new first difference.

[0060] When the first valid value obtained in advance is greater than or equal to 100, the new first difference value is obtained by formula (1);

[0061] The formula (1) is:

[0062]

[0063] wherein, B is a new first difference value;

[0064] a is the bit number of the first effective value obtained in advance;

[0065] [] is a rounding symbol;

[0066] The first effective value is a value obtained by taking the mean of the absolute values of the original feature variables in the original data set and retaining 1 significant digit.

[0067] Preferably,

[0068] The pre-set sorting manner is: first, sorting in ascending order of the first time difference; when the first time difference is the same, sorting the relative information with the same first time difference in descending order of the second end time.

[0069] Preferably, the S4 comprises:

[0070] Based on the final target variable actual value and the final target variable predicted value, the error score A1 of the to-be-tested machine learning model is obtained by using formula (2), and the similarity score A2 of the to-be-tested machine learning model is obtained by using formula (3);

[0071] The formula (2) is:

[0072]

[0073] P is the proportion of the relative information with an err value less than or equal to 30% in the final data set;

[0074]

[0075] wherein, deltay is the final target variable actual value; pred·deltay is the final target variable predicted value;

[0076] The formula (3) is:

[0077] A2 = 1 - |1 - k|;

[0078] Wherein, k is a regression model fitting coefficient;

[0079] The regression model is established in advance by using a machine learning one-order polynomial algorithm, and the one-order polynomial thereof is deltay = k * pred·deltay.

[0080] Preferably, the S5 comprises:

[0081] Based on the error score A1 and the similarity score A2 of the to-be-tested machine learning model, the final score score of the to-be-tested machine learning model is obtained by using formula (4);

[0082] The formula (4) is:

[0083] score=(A1+A2) / 2.

[0084] Preferably, the S6 specifically comprises:

[0085] Based on the final score score of each to-be-tested machine learning model, the to-be-tested machine learning model with the highest final score score is screened out as the final machine learning model.

[0086] Preferably,

[0087] The preset time period is 20 days before the current time;

[0088] The preset time range is 5min-360min.

[0089] (Three) beneficial effects

[0090] The beneficial effects of the present application are: a machine learning model screening method, which is preprocessed by an original data set to obtain a final data set, the original data set including multiple original data collected at a preset frequency during the thermal power production process within the preset time period; wherein each original data includes a timestamp, an original feature variable and an original target variable corresponding to the original data; the final data set includes multiple pairs of final feature variables and final target variable actual values, further, the error and similarity between the final target variable actual value and the final target variable predicted value are combined to evaluate the to-be-tested machine learning model, and the final machine learning model is screened out through the final score of the to-be-tested machine learning model. In terms of prior art, it can evaluate the to-be-tested machine learning model, and the final machine learning model screened out further has better performance. BRIEF DESCRIPTION OF DRAWINGS

[0091] Figure 1 A machine learning model screening method flow chart;

[0092] Figure 2 A final score score of 0.96 to-be-tested machine learning model prediction error distribution graph;

[0093] Figure 3 A final score score of 0.3 to-be-tested machine learning model prediction error distribution graph. DETAILED DESCRIPTION

[0094] In order to better explain the present application, so as to be understood, the following will be combined with the drawings, through specific embodiments, the present application is described in detail.

[0095] In order to better understand the above technical solutions, the following will be described in more detail with reference to the drawings. Although the exemplary embodiments of the present application are shown in the drawings, it should be understood that the present application can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided in order to enable a clearer, more thorough understanding of the present application and to convey the scope of the present application to those skilled in the art.

[0096] Referring to Figure 1 The present embodiment provides a machine learning model screening method, which is oriented to thermal power production process control, and the method comprises:

[0097] S1, obtaining a set of to-be-evaluated models and an original data set in a preset time period in the thermal power production process, wherein the set of to-be-evaluated models comprises at least one to-be-evaluated machine learning model.

[0098] The original data set comprises a plurality of original data collected at a preset frequency in the thermal power production process in the preset time period; wherein each original data comprises a timestamp, an original feature variable and an original target variable corresponding to the original data.

[0099] Wherein, all to-be-evaluated machine learning models in the set of to-be-evaluated models are used to predict the predicted value of the original target variable corresponding to the original feature variable according to the original feature variable, that is, the input of any to-be-evaluated machine learning model in the set of to-be-evaluated models is the original target variable, and the output is the predicted value of the original target variable.

[0100] Specifically, the preset time period in the present embodiment is 20 days, that is, the original data 20 days before the current time in the thermal power production process is obtained to form the original data set; wherein each original data comprises a timestamp, an original feature variable and an original target variable corresponding to the original data, the original feature variable in the present embodiment is coal supply, and the original target variable is coal feeder frequency, as shown in part of the original data in the original data set in Table 1:

[0101] Table 1

[0102] Serial number time (timestamp) coal supply coal feeder frequency 0 1641881890000 11.292 26.0 1 1641881900000 11.328 26.0 2 1641881910000 11.292 26.0 3 1641881920000 11.339 26.0 4 1641881930000 11.446 26.0

[0103] S2, preprocessing the original data set to obtain a final data set; the final data set comprises a plurality of pairs of final feature variables and final target variable actual values.

[0104] S3, inputting the final feature variable in the final data set into any machine learning model to be tested in the set of machine learning models to be tested, to obtain a corresponding target variable prediction value.

[0105] S4, based on the final target variable actual value and the target variable prediction value, obtaining an error score and a similarity score of the machine learning model to be tested.

[0106] S5, based on the error score and the similarity score of the machine learning model to be tested, obtaining a final score of the machine learning model to be tested.

[0107] S6, based on the final score of each machine learning model to be tested, screening a final machine learning model.

[0108] Specifically, in actual application, in a thermoelectric production process, if the current coal supply amount (original feature variable) is 10 tons per hour, and if the target value of the coal supply amount is 12 tons per hour, then the value of 2 (target value of the coal supply amount - current coal supply amount) is input to the final machine learning model screened, and the final machine learning model outputs an adjustment value of the coal feeder frequency. At this time, the user can adjust the current coal feeder frequency according to the adjustment value of the coal feeder frequency output by the final machine learning model.

[0109] In actual application of the embodiment, the S2 specifically includes:

[0110] S21, screening a stationary section data set for the original data set.

[0111] Wherein, there is no intersection between any two stationary section data sets, and any stationary section data set L i satisfies L i = [l 1i , l 2i ,..., l ni ].

[0112] L i is the i-th stationary section data set screened from the original data set.

[0113] l ni is the n-th original data in the i-th stationary section data set.

[0114] (l mi-x -l 1i-x ) / l 1i-x ≤ 3%, m ≤ n.

[0115] (l mi-y -l 1i-y ) / l 1i-y ≤ 3%, m ≤ n.

[0116] l mi-x is the original feature variable in the mth original data in the ith stationary segment data set; l mi-y is the original target variable in the mth original data in the ith stationary segment data set.

[0117] The stationary segment data set L i The first time corresponding to the time stamp satisfies a preset time range; the first time is the time of the first original data in the stationary segment data set L mi The time of the time stamp corresponding to the time of the last original data in the stationary segment data set L 1i The time of the time stamp corresponding to the time of the last original data in the stationary segment data set L

[0118] S22, for all stationary segment data sets, obtaining the average value of the original feature variable and the average value of the original target variable in each stationary segment data set, and the starting time and the ending time of the stationary segment data set.

[0119] In this embodiment, the starting time of the stationary segment data set is the time of the time stamp corresponding to the first original data in the stationary segment data set, and the ending time of the stationary segment data set is the time of the time stamp corresponding to the last original data in the stationary segment data set.

[0120] S23, based on the average value of the original feature variable and the average value of the original target variable in each stationary segment data set, and the starting time and the ending time of the stationary segment data set, obtaining a first data set.

[0121] The first data set includes information data corresponding to each stationary segment data set one by one.

[0122] The information data includes: the average value of the original feature variable and the average value of the original target variable in the stationary segment data set corresponding to the information data, and the starting time and the ending time of the stationary segment data set.

[0123] S24, sorting all information data in the first data set according to the starting time in the information data from small to large, to obtain a second data set.

[0124] S25, using a pre-set processing strategy to process the second data set to obtain a final data set.

[0125] In this embodiment, the S25 specifically includes:

[0126] S251. For the average value of the original feature variables in any information data in the second dataset, obtain the relative information between each information data after the information data in the second dataset and the information data, and combine all the relative information in the second dataset into a third dataset.

[0127] The relative information between any information data following this information data in the second dataset and this information data includes: first difference, second difference, first end time, first start time, second end time, and first time difference.

[0128] The first difference is the difference obtained by subtracting the average of the original feature variables in any information data after the information data in the second dataset from the average of the original feature variables in the information data.

[0129] The second difference is the difference obtained by subtracting the average of the original target variable in any information data after the information data in the second dataset from the average of the original target variable in the information data.

[0130] The first end time is the end time of the stationary segment dataset corresponding to the information data.

[0131] The first start time is the start time of the stationary segment dataset corresponding to any information data after the information data in the second dataset.

[0132] The second end time is the end time of the stationary segment dataset corresponding to any information data after the information data in the second dataset.

[0133] The first time difference is the difference between the first start time and the first end time.

[0134] S252. Based on the relative information in the third dataset, filter the third dataset according to a pre-set filtering rule to obtain a fourth dataset; wherein, the pre-set filtering rule is: if the original feature variable and the original target variable are negatively correlated, delete the first relative information in the third dataset; the first relative information is relative information where the product of the first difference and the second difference is greater than 0; if the original feature variable and the original target variable are positively correlated, delete the second relative information in the third dataset; the second relative information is relative information where the product of the first difference and the second difference is less than 0.

[0135] S253. For any first difference in relative information in the fourth dataset, perform value processing according to a preset value taking method to obtain a new first difference corresponding to the relative information.

[0136] The preset value selection method is as follows:

[0137] When the first effective value obtained in advance is less than 1, the first difference value is kept to one significant figure to obtain a new first difference value.

[0138] When the first effective value obtained in advance is greater than or equal to 1 and less than 10, the first difference value is kept to one decimal place to obtain a new first difference value.

[0139] When the first effective value obtained in advance is greater than or equal to 10 and less than 100, the first difference value is rounded to obtain a new first difference value.

[0140] When the first effective value obtained in advance is greater than or equal to 100, the formula (1) is used to obtain a new first difference value.

[0141] The formula (1) is:

[0142]

[0143] Wherein, B is the new first difference value.

[0144] a is the number of digits of the first effective value obtained in advance; for example, if the first effective value is 111, then the value of a is equal to 3.

[0145] [] is a rounding symbol.

[0146] The first effective value is the value obtained by taking the mean of the absolute values of the original feature variables in the original data set and keeping one significant figure.

[0147] S254, grouping the fourth data set based on the new first difference value corresponding to any relative information in the fourth data set to obtain at least one relative information set.

[0148] Wherein, the new first difference values corresponding to all relative information in each relative information set are equal.

[0149] S255, for each relative information set, relative information with a first time difference greater than or equal to 48 hours is removed to obtain a removed relative information set.

[0150] S256, for the removed relative information set, sorting is performed according to a pre-set sorting manner to obtain a final relative information set.

[0151] The pre-set sorting manner is: first, sorting is performed according to the order of the first time difference from small to large; when the first time difference is the same, the relative information with the same first time difference is sorted according to the order of the second end time from large to small.

[0152] S257, extracting the first 10 relative information in each final relative information set to form a final data set.

[0153] wherein the first difference value in each relative information in the final dataset is taken as a final feature variable; and the second difference value in each relative information in the final dataset is taken as a final target variable actual value.

[0154] Specifically, the S4 in the embodiment includes:

[0155] Based on the final target variable actual value and the final target variable predicted value, an error score A1 of the to-be-tested machine learning model is obtained by using formula (2), and a similarity score A2 of the to-be-tested machine learning model is obtained by using formula (3).

[0156] The formula (2) is:

[0157]

[0158] p is a proportion of relative information with an err value less than or equal to 30% in the final dataset.

[0159]

[0160] wherein deltay is the final target variable actual value; and pred·deltay is the final target variable predicted value.

[0161] The formula (3) is:

[0162] A2 = 1 - |1 - k|; wherein k is a fitting coefficient of the regression model.

[0163] wherein k is a fitting coefficient of the regression model.

[0164] The regression model is established in advance by using a machine learning one-order polynomial algorithm, and the one-order polynomial is deltay = k*pred·deltay.

[0165] In the specific implementation of the embodiment, the S5 specifically includes:

[0166] Based on the error score A1 and the similarity score A2 of the to-be-tested machine learning model, a final score score of the to-be-tested machine learning model is obtained by using formula (4).

[0167] The formula (4) is:

[0168] score = (A1 + A2) / 2.

[0169] In the embodiment, the S6 specifically includes: based on the final score score of each to-be-tested machine learning model, screening the to-be-tested machine learning model with the highest final score score as the final machine learning model. For example, the higher the final score score of the to-be-tested machine learning model, the smaller the error distribution interval of the to-be-tested machine learning model, see Figure 2 and Figure 3 , Figure 2 is a prediction error distribution diagram of the to-be-tested machine learning model with the final score score of 0.96, Figure 3 is a prediction error distribution diagram of the to-be-tested machine learning model with the final score score of 0.3, Figure 2 and Figure 3 The horizontal axis is the prediction error, and the vertical axis is the probability of error occurrence. By comparison, it can be found that the to-be-tested machine learning model with a low final score score has a wide error distribution interval and greater error. Therefore, the final score score of the to-be-tested machine learning model conforms to the true effect of prediction.

[0170] The machine learning model screening method in the embodiment, since the final data set is obtained by preprocessing the original data set, the original data set includes multiple pieces of original data collected at a preset frequency during the thermal power production process in the preset time period; each piece of original data includes a time stamp, an original feature variable, and an original target variable corresponding to the piece of original data; the final data set includes multiple pairs of final feature variables and final target variable actual values. Further, the to-be-tested machine learning model is evaluated in combination with the error and similarity between the final target variable actual value and the final target variable predicted value, and the final machine learning model is screened out through the final score of the to-be-tested machine learning model. In terms of the prior art, the to-be-tested machine learning model can be evaluated, and the final machine learning model screened out further has better performance.

[0171] The embodiment further provides a machine learning model screening system, including: at least one processor; and at least one memory in communication connection with the processor, wherein the memory stores program instructions executable by the processor, and the processor calling the program instructions can execute the machine learning model screening method in the above embodiment.

[0172] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system or a computer program product. Therefore, the present application can adopt a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer usable program codes.

[0173] The present application is described in reference to the flowchart and / or block diagram of the method, apparatus (system) and computer program product according to an embodiment of present application. It is understood that each block of the flowchart and / or block diagram, and combinations of blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general purpose computer, special purpose computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions of the flowchart and / or block diagram blocks.

[0174] It should be noted that any references made in the claims to an element or steps in the flowchart or block diagram utilizing a "means" for performing the recited function are not to be construed as limiting the claim to the best mode of performing the recited function. Rather, the means can be interpreted as simply a further recitation of a claim limitation. The word "comprising" does not exclude the presence of elements or steps not listed in a claim. The word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements. The application can be implemented by means of hardware comprising several distinct elements, and by means of a programmed computer. In the claims, the word "first", "second", "third", etc. does not imply any order. The terms "first", "second", "third", etc. are to be understood only as distinguishing between two or more elements or steps. The use of the terms "first" and "second" to describe a claim limitation does not imply any order or sequence.

[0175] Furthermore, it is noted that the described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments or examples. It will also be readily understood to those skilled in the art that the above description is merely illustrative of the application and should not be

[0176] While the preferred embodiments of the application have been described above, it will be recognized and understood that various modifications and changes can be made to the application by those skilled in the art that fall within the spirit and scope of the application. Accordingly, the scope of the application should not be limited by the preferred embodiments, but should be given the broadest possible interpretation accessible under the law.

[0177] Obviously, many modifications and variations of the present application are possible in light of the above teachings. It is, therefore, to be understood that within the scope of the appended claims and their equivalents, the application can be practiced otherwise than as specifically described.

Claims

1. A method for selecting machine learning models, wherein the machine learning models are designed for control of thermoelectric production processes, characterized in that, The method comprises: S1, acquiring a set of to-be-evaluated models and an original data set in a preset time period in a thermal power production process, wherein the set of to-be-evaluated models comprises at least one to-be-evaluated machine learning model; The original data set comprises a plurality of original data collected in the thermal power production process at a preset frequency in the preset time period; wherein each original data comprises a timestamp, an original feature variable and an original target variable corresponding to the original data; S2, preprocessing the original data set to obtain a final data set; the final data set comprises a plurality of pairs of final feature variables and final target variable actual values; S3, inputting the final feature variables in the final data set into any to-be-tested machine learning model in the set of to-be-evaluated models to obtain corresponding target variable predicted values; S4, based on the final target variable actual values and the target variable predicted values, obtaining an error score and a similarity score of the to-be-tested machine learning model; S5, based on the error score and the similarity score of the to-be-tested machine learning model, obtaining a final score of the to-be-tested machine learning model; S6, based on the final score of each to-be-tested machine learning model, screening a final machine learning model; The S4 comprises: Based on the final target variable actual values and the final target variable predicted values, the error score A1 of the to-be-tested machine learning model is obtained by using formula (2), and the similarity score A2 of the to-be-tested machine learning model is obtained by using formula (3); The formula (2) is: ; P is a proportion of the err value less than or equal to 30% in the final data set; ； Wherein, deltay is the final target variable actual value; pred·deltay is the final target variable predicted value; the formula (3) is: ； wherein k is a regression model fitting coefficient; the regression model is established in advance using a machine learning one-order polynomial algorithm, and the one-order polynomial is ; The S5 specifically comprises: based on the error score A1 and the similarity score A2 of the to-be-tested machine learning model, the final score score of the to-be-tested machine learning model is obtained by using formula (4); The formula (4) is: score = (A1+A2) / 2. 2.The machine learning model screening method of claim 1, wherein, The S2 specifically comprises: S21, screening a stationary section data set for the original data set; Wherein, there is no intersection between any two stationary segment data sets, and any stationary segment data set L i satisfies ; L i is the i-th stationary segment data set screened out from the original data set; is the nth original data in the ith stationary segment data set; ； ； is the original feature variable in the mth original data in the ith stationary segment data set; is the original target variable in the mth original data in the ith stationary segment data set; The stationary section data set L i The corresponding first time meets a preset time range; the first time is L mi The time of the corresponding timestamp minus L 1i The time of the corresponding timestamp S22, for all stationary section data sets, obtaining the average value of the original feature variable and the average value of the original target variable in each stationary section data set, and the start time and the end time of the stationary section data set; S23, based on the average value of the original feature variable and the average value of the original target variable in each stationary section data set, and the start time and the end time of the stationary section data set, a first data set is obtained; The first data set comprises information data corresponding to each stationary section data set one by one; The information data comprises: the average value of the original feature variable and the average value of the original target variable in the stationary section data corresponding to the information data, and the start time and the end time of the stationary section data set; S24, sorting all information data in the first data set according to the start time in the information data in ascending order to obtain a second data set; S25, processing the second data set by using a preset processing strategy to obtain a final data set. 3.The machine learning model screening method of claim 2, wherein, The S25 specifically comprises: S251, for the average value of the original feature variable in any information data based on the second data set, obtaining the relative information between each information data after the information data in the second data set and the information data, and grouping all the relative information in the second data set into a third data set; Wherein, the relative information between any information data after the information data in the second data set and the information data includes: first difference, second difference, first end time, first start time, second end time, first time difference; The first difference is the difference between the average value of the original feature variable in any information data after the information data in the second data set and the average value of the original feature variable in the information data; The second difference is the difference between the average value of the original target variable in any information data after the information data in the second data set and the average value of the original target variable in the information data; The first end time is the end time of the stable segment data set corresponding to the information data; The first start time is the start time of the stable segment data set corresponding to any information data after the information data in the second data set; The second end time is the end time of the stable segment data set corresponding to any information data after the information data in the second data set; The first time difference is the difference between the first start time and the first end time; S252, according to the relative information in the third data set, filtering the third data set according to the preset filtering rule to obtain a fourth data set; S253, for the first difference in any relative information in the fourth data set, taking value processing according to the preset value taking mode to obtain the new first difference corresponding to the relative information; S254, based on the new first difference corresponding to any relative information in the fourth data set, grouping the fourth data set to obtain at least one relative information set; Wherein, the new first difference corresponding to all relative information in each relative information set is equal; S255, for each relative information set, the relative information with the first time difference greater than or equal to 48 hours is removed to obtain the removed relative information set; S256, for the removed relative information set, sorting according to the preset sorting mode to obtain the final relative information set; S257, extracting the first 10 relative information in each final relative information set to form a final data set; Wherein, the first difference in each relative information in the final data set is taken as the final feature variable; the second difference in each relative information in the final data set is taken as the final target variable actual value.

4. The machine learning model screening method according to claim 3, wherein the preset filtering rule is: if the original feature variable and the original target variable are negatively correlated, delete the first relative information in the third data set; ​ The first relative information is relative information in which the product of the first difference value and the second difference value is greater than 0; If the original feature variable and the original target variable are positively correlated, the second relative information in the third data set is deleted; The second relative information is relative information in which the product of the first difference value and the second difference value is less than 0.

5. The machine learning model screening method according to claim 4, wherein the preset value mode is: When the first effective value obtained in advance is less than 1, the first difference value is kept to have one significant digit to obtain a new first difference value; When the first effective value obtained in advance is greater than or equal to 1 and less than 10, the first difference value is kept to have one decimal place to obtain a new first difference value; When the first effective value obtained in advance is greater than or equal to 10 and less than 100, the first difference value is rounded to obtain a new first difference value; When the first effective value obtained in advance is greater than or equal to 100, a formula (1) is used to obtain a new first difference value; The formula (1) is: Wherein, B is the new first difference value; ； a is the number of bits of the first effective value obtained in advance; [] is a rounding symbol; The first effective value is a value obtained by taking the absolute value of the original feature variable in the original data set, taking the mean value, and keeping one significant digit.

6. The machine learning model screening method according to claim 5, wherein the preset sorting mode is: first, sorting in ascending order of the first time difference; when the first time difference is the same, sorting the relative information with the same first time difference in descending order of the second end time. The S6 specifically includes: Based on the final score score of each machine learning model to be tested, the machine learning model to be tested with the highest final score score is screened as the final machine learning model.

7. The machine learning model screening method of claim 6, wherein, 8. The machine learning model screening method according to claim 7, wherein the preset time period is 20 days before the current time; The preset time range is 5min-360min. ​ ​ ​

Images