Optimization methods for data processing and model training of time series online decision models

Abstract

The present invention discloses an optimization method for data processing and model training of a time series online decision model. The present invention obtains time series data such as the water level, superheated steam temperature and pressure, flue gas oxygen content, and furnace negative pressure of a steam boiler through multi-source sensors; during data processing, the precision of the floating-point type in each time series data is first unified, and then proportionally amplified into integer data, and the interpolation and time series alignment of the multi-source time series data are completed according to the set time interval; the aggregated time series data uses a compression algorithm, with the timestamp as the primary key, and is compressed and stored in a database; when making a decision, the compressed data in the database is read, and after decompression, a feature vector is constructed and input into the model, and the online decision model calculates and outputs the decision information; the operation and maintenance personnel promptly adjust the parameters of the boiler's external input equipment according to the model decision results, and whether the equipment needs to be shut down for maintenance. This method simplifies the complexity of the decision model and further improves the decision efficiency and storage efficiency.

Description

Technical Field

[0001] The present invention relates to the field of data processing and analysis such as time series processing technology, and in particular to an optimization method for data processing and model training of a time series online decision model. Background Art

[0002] Steam boilers are crucial thermal power plants for thermal power plants. Their mission is to supply stable, high-quality steam to meet load demands. Therefore, every key parameter in the boiler's production process must be strictly controlled. Key steam boiler parameters include water level (WL), superheated steam temperature (SST) and pressure (SSP), flue gas oxygen content (OC), and furnace negative pressure (CP). These parameters are collected and monitored by multiple sensors and then fed into an online decision-making model within the operation and maintenance system for calculation. Based on the model's decision-making results, operation and maintenance personnel can promptly adjust the parameters of the boiler's external input devices or decide whether to shut down the equipment for maintenance, thereby ensuring the long-term stable operation of the boiler system.

[0003] The time series data collection of steam boilers requires multi-source sensors. The original time series data obtained by sampling from different sensors needs to be merged and aggregated with other time series to generate time series with specific intervals before it can be used as the input of the time series model. This process is also called time series alignment.

[0004] However, all sensor clock sources are subject to clock drift. Even if the timestamps of each sensor are initially aligned, after a period of operation, the previously aligned results will deviate. Furthermore, the sampling times and frequencies of multiple sensors vary, making complete generation impossible. This makes it impossible to directly merge time series from multiple sources.

[0005] For example, the sampling frequency for superheated steam temperature and pressure in a steam boiler is both 10Hz. However, the superheated steam pressure sensor in a steam boiler samples data tens of milliseconds slower than the temperature sensor, making it impossible to obtain the sampled value of the boiler superheated steam pressure at a specific temperature. Furthermore, the sampling period for the water level gauge in a steam boiler is 500ms, while the sampling period for the furnace negative pressure sensor is 360ms. When the model sequence interval is 1000ms, the furnace negative pressure sampling data cannot be directly used as the model sequence input, nor can it be directly merged with the water level sampling data. This phenomenon makes it easy for default values to appear in the sampling records at the same time. Once the default values are input into the time series analysis model, the model must perform data preprocessing operations such as interpolation algorithms to align the time between multiple sources, or add a step to handle the default values in the decision model. This not only increases the complexity of the decision model but also significantly reduces the decision speed of the online model.

[0006] In addition, the measurement values of most multi-source sensors are floating-point numbers, and the processing efficiency of floating-point numbers in the online decision model of the decision tree is very low. Moreover, the compression rate of floating-point data is lower than that of integer data when compressed and stored. Therefore, directly storing a large amount of floating-point time series data reduces storage efficiency. Through experiments, it was found that for online decision models, proportional amplification of input data has no effect on the model decision results. Therefore, the present invention proportionally amplifies all floating-point data into integer data. Integer data not only facilitates the interpolation operations required for time series alignment, but also significantly improves the compression rate of time series data during compressed storage, and also improves the online decision calculation speed of the model.

[0007] Currently, the existing online decision-making model for steam boiler time series models uses a data acquisition module to obtain key parameters and directly store them in a database. The model then reads the data from the database and performs preprocessing in the decision-making model. This significantly increases the length of the decision-making process and reduces the efficiency of the online decision-making model. Summary of the Invention

[0008] The purpose of the present invention is to provide a method for optimizing data processing and model training of a time series online decision model in response to the deficiencies of the existing technology.

[0009] The object of the present invention is achieved through the following technical solution: a method for optimizing data processing and model training of a time series online decision model, the method comprising the following steps:

[0010] S1. Obtain time series data from multiple sensors in the steam boiler system, including water level WL, superheated steam temperature SST, pressure SSP, flue gas oxygen content OC, and furnace negative pressure CP. The time series data values are all floating point numbers, and each sampled data includes its corresponding integer timestamp TS.

[0011] S2. Steam boiler multi-source sensor data processing, including:

[0012] S2-1. Unify the floating point precision: retain N decimal places for the obtained floating point number;

[0013] S2-2, floating point integer conversion; multiply the floating point data after precision unification in S2-1 by a factor of M = 10 N ;

[0014] S2-3, Multi-source time series data alignment: Set the custom time interval T and start time ST, for a certain time CT n =ST+nT, if there is a corresponding sampling timestamp TS in the time series data n =CT n , then directly select the integer TS nThe sampling value at the moment is used as the summary value of each parameter; if it does not exist, the system preset interpolation algorithm is selected to calculate TS using different interpolation algorithms for the integer multi-source time series data. n The default value at the moment, and the result of the interpolation algorithm is used as the value at TS n Aggregate values of parameters with default values at a given moment;

[0015] S3, time series data compression storage: After the multi-source time series data is aligned, the time series data is compressed into integers and stored in the database with the timestamp as the primary key;

[0016] S4. Construct feature vector: Using timestamp as primary key, read the time series data in the database and decompress it to construct boiler feature vector B. v ;

[0017] S5. The boiler feature vector B v The decision tree model is input into the time series online decision model to train the steam boiler operation and maintenance decision model; if the rate of change of accuracy is less than or equal to the preset change rate threshold in consecutive preset number of trainings, the training is temporarily stopped. At this time, the error rate, recall rate and false alarm rate of the model are calculated based on the confusion matrix. If the error rate, recall rate and false alarm rate are all less than the preset threshold, the steam boiler operation and maintenance decision model is saved as a model file, stored in the database and training is stopped; otherwise, the initialization parameters of the decision tree model are readjusted and training is repeated;

[0018] S6. Obtain the current production parameters of the steam boiler system and input them into the trained steam boiler operation and maintenance decision model; the threshold conditions of the steam boiler operation and maintenance decision model are customized according to the boiler model and working scenario; and the decision is output as a basis for guiding the operation and maintenance personnel to adjust the parameters of the boiler's external input equipment at the next moment or decide whether to shut down the equipment for maintenance.

[0019] Furthermore, the value of N in step S2-2 is set to the maximum value of the floating-point precision of all sensors or is customized according to the actual scenario.

[0020] Furthermore, the time interval T of step S2-3 is not less than 600 seconds.

[0021] Furthermore, the boiler feature vector B in S4 v =(WL ST , WL ST+T ,...,WL ST+nT , SST ST ,...,SST ST+nT ,...,SSP ST ,...,SSP ST+nT ,...,OC ST ,...,OCST+nT ,...,CP ST ,...,CP ST+nT ); among them, WL ST is the water level at the start of sampling ST, WL ST+T is the water level at ST after a time interval T, WL ST+nT is the water level at ST after n time intervals T, SST ST is the superheated steam temperature at the sampling start time ST, SST ST+nT is the superheated steam temperature after n time intervals T at ST, SSP ST is the superheated steam pressure at the sampling start time ST, SSP ST+nT is the superheated steam pressure after n time intervals T at ST, OC ST is the oxygen content of flue gas at the sampling start time ST, OC ST+nT is the amount of oxygen in the flue gas after n time intervals T at ST, CP ST is the furnace negative pressure at the sampling start time ST, CP ST+nT It is the furnace negative pressure after n time intervals T at ST moment.

[0022] Furthermore, in S5, the preset number of times is set to 100 times, the preset change rate threshold is set to 1%, and the preset threshold is set to 5%.

[0023] Furthermore, the specific steps of step S6 are as follows:

[0024] S6-1, using the collected production parameters as input to the online decision model after pre-processing in step S2;

[0025] S6-2, based on the actual operating conditions of the boiler, determine the threshold conditions, including: water level and its allowable fluctuation range, upper limit of superheated steam temperature, upper and lower limits of superheated steam pressure, range of flue gas oxygen content, and upper and lower limits of furnace negative pressure;

[0026] S6-3, based on the input data, the decision information of the online decision model is obtained for the operation and maintenance personnel to make decisions.

[0027] Furthermore, the integer compression and decompression are performed using a ZigZag algorithm.

[0028] Furthermore, the current production parameters of the steam boiler system include water level WL, superheated steam temperature SST and pressure SSP, flue gas oxygen content OC, and furnace negative pressure CP.

[0029] Beneficial effects of the present invention:

[0030] This method proportionally scales floating-point data from multi-source time series data to integer data, interpolates and time-series aligns the integer data, and uses the aligned integer time series data as input for the decision model. Integer data improves model computational efficiency. Aligned time series data eliminates the need to consider default values, simplifying the complexity of the decision model and further improving decision efficiency. Furthermore, integer data has a higher storage compression ratio than floating-point data, thus improving storage efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 A flow chart of the optimization method for data processing and model training of the time series online decision model of the present invention;

[0032] Figure 2 The figure is a comparison diagram of the implementation effects of the method of the present invention and the traditional method. DETAILED DESCRIPTION

[0033] The specific embodiments of the present invention are further described in detail below with reference to the accompanying drawings.

[0034] like Figure 1 As shown, the present invention provides a method for optimizing data processing and model training of a time series online decision model, the steps of which are as follows:

[0035] S1. Obtain historical data from multiple sensors in the steam boiler. Obtain the water level WL, superheated steam temperature SST and pressure SSP, and flue gas oxygen content OC and CP of the steam boiler system. WL, SST, SSP, OC, and CP are all time series data values, all floating-point numbers. Each sampled data point includes its corresponding timestamp TS, which is an integer.

[0036] S2. Steam boiler multi-source sensor data processing.

[0037] S2-1. Unify floating-point precision. For floating-point numbers such as WL, SST, SSP, OC, and CP, retain N = 8 decimal places, and set the value of N to the maximum floating-point precision of all sensors.

[0038] S2-2, floating point integerization. Multiply the floating point data after precision unification in S2-1 by a factor of M = 10 N , convert all floating-point data into integer data in equal proportions.

[0039] S2-3, multi-source time series data alignment. Set the custom time interval T = 600s, the start time ST, for a certain time CT n =ST+nT, if the sampling timestamp TS exists in the S2-2 data n =CT n , then directly select TS in S2-2 nThe sampling value at the moment is used as the summary value; if it does not exist, the operation and maintenance personnel will customize the system preset interpolation method, such as linear interpolation, polynomial interpolation, Lagrange interpolation, etc., and use different interpolation algorithms for the integers such as WL, SST, SSP, OC, CP processed in S2-2 to calculate TS n The default value at the moment, and use this interpolation result as the default value at TS n The summary value at the moment.

[0040] S3. Compress and store time series data. The time series data aligned in steps S2-1, S2-2, and S2-3 is compressed using the ZigZag algorithm and stored in the database with the timestamp as the primary key.

[0041] S4. Construct the feature vector. Using the timestamp as the primary key, read the time series data in the database, decompress it using the ZigZag algorithm, and construct the boiler feature vector B according to the needs of the online model input. v =(WL ST , WL ST+T ,...,WL ST+nT , SST ST ,...,SST ST+nT ,...,SSP ST ,...,SSP ST+nT ,...,OC ST ,...,OC ST+nT ,...,CP ST ,...,CP ST+nT ). Among them, WL ST is the water level at the start of sampling ST, WL ST+T is the water level at ST after a time interval T, WL ST+nT is the water level at ST after n time intervals T, SST ST is the superheated steam temperature at the sampling start time ST, SST ST+nT is the superheated steam temperature after n time intervals T at ST, SSP ST is the superheated steam pressure at the sampling start time ST, SSP ST+nT is the superheated steam pressure after n time intervals T at ST, OC ST is the oxygen content of flue gas at the sampling start time ST, OC ST+nT is the amount of oxygen in the flue gas after n time intervals T at ST, CP ST is the furnace negative pressure at the sampling start time ST, CP ST+nT It is the furnace negative pressure after n time intervals T at ST moment.

[0042] S5. The boiler feature vector B vInput into the decision tree model and train the boiler decision model. If the change rate of precision Precision is less than or equal to 1% in 100 consecutive trainings, the training is temporarily stopped. At this time, the error rate ErrorRate, recall rate Recall, and false alarm rate FPrate of the model are calculated based on the confusion matrix CM. If the error rate, recall rate, and false alarm rate are all less than 5%, the boiler performance model is saved as a model file, stored in the database, and training is stopped. Otherwise, the initialization parameters of the decision tree model are readjusted and retrained.

[0043] S6. Obtain the current production parameters of the steam boiler system, mainly including water level WL, superheated steam temperature SST and pressure SSP, flue gas oxygen content OC, and furnace negative pressure CP. After preprocessing, these parameters are input into the boiler's online decision model. The decision model threshold is customized based on the experience of operation and maintenance experts and combined with the boiler model and work scenario. The decision output by the model serves as a basis for guiding operation and maintenance personnel to adjust the parameters of the boiler's external input equipment at the next moment or decide whether to shut down the equipment for maintenance.

[0044] S6-1, the collected production parameters are pre-processed in step S2 and used as input of the online decision model.

[0045] S6-2, based on the actual operating conditions of the steam boiler, determine the threshold conditions of the online decision model, including: the low water level cannot be lower than 25mm from the lower visible edge of the water level gauge, the drum water level is maintained at 0mm during normal operation, the fluctuation range is allowed to be ±30mm, and the maximum does not exceed ±50mm; the upper limit of the superheated steam temperature is 540℃, and the upper and lower limits of the superheated steam pressure are 9.8MPa and 13.7MPa respectively; the flue gas oxygen content is controlled at 6% to 8%; the upper and lower limits of the furnace negative pressure are 19.6Pa and 49Pa respectively.

[0046] S6-3, obtain online model decision information based on input data for operation and maintenance personnel to make decisions.

[0047] To demonstrate the practical benefits of the method, we conducted a set of experiments based on real-world data from a production environment. By comparing the method with traditional methods, we found that the method produced results in significantly less time than the traditional method's sequential decision model when processing the same amount of data.

[0048] The experimental scenario is to collect time series data of water level, superheated steam temperature and pressure, flue gas oxygen content and furnace negative pressure parameters of a steam boiler monitoring system within 24 hours. Each set of test data includes a timestamp, which represents the time series of the data. The test data is distributed over 24 consecutive hours, with 10,000 sets per hour, and the timestamps of the data are evenly distributed within the hour.

[0049] The test method is: using the data processing method of the present invention and the traditional data processing method respectively to calculate the calculation time of the steam boiler online decision model for 1, 2...24 hours of data input and record them, and finally draw a chart.

[0050] The test results are as follows Figure 2 As shown. The results show that:

[0051] (1) When the data volume is the same, the calculation time of the traditional steam boiler time series online decision model is longer than the calculation time of the decision model optimized by the method of the present invention.

[0052] (2) As the amount of data increases, the time-consuming growth rate of the timing decision model optimized by the method of the present invention is smaller than that of the traditional decision model.

[0053] The above embodiments are used to illustrate the present invention rather than to limit the present invention. Any modifications and changes made to the present invention within the spirit of the present invention and the protection scope of the claims shall fall within the protection scope of the present invention.

Claims

1. A method for optimizing data processing and model training for a time series online decision model, characterized in that: The method comprises the following steps: S1. Obtain time series data from multiple sensors in the steam boiler system, including water level WL, superheated steam temperature SST, pressure SSP, flue gas oxygen content OC, and furnace negative pressure CP. The time series data values are all floating point numbers, and each sampled data includes its corresponding integer timestamp TS. S2. Steam boiler multi-source sensor data processing, including: S2-1. Unify the floating point precision: retain N decimal places for the obtained floating point number; S2-2, floating point integer conversion; multiply the floating point data after precision unification in S2-1 by a factor of M = 10 N ; S2-3, Multi-source time series data alignment: Set the custom time interval T and start time ST, for a certain time CS n =ST+nT, if there is a corresponding sampling timestamp TS in the time series data n =CT n , then directly select the integer TS n The sampling value at the moment is used as the summary value of each parameter; if it does not exist, the system preset interpolation algorithm is selected to calculate TS using different interpolation algorithms for the integer multi-source time series data. n The default value at the moment, and the result of the interpolation algorithm is used as the value at TS n Aggregate values of parameters with default values at a given moment; S3, time series data compression storage: After the multi-source time series data is aligned, the time series data is compressed into integers and stored in the database with the timestamp as the primary key; S4. Construct feature vector: Using timestamp as primary key, read the time series data in the database and decompress it to construct boiler feature vector B. v ; S5. The boiler feature vector B v The decision tree model is input into the time series online decision model to train the steam boiler operation and maintenance decision model; if the rate of change of accuracy is less than or equal to the preset change rate threshold in consecutive preset number of trainings, the training is temporarily stopped. At this time, the error rate, recall rate and false alarm rate of the model are calculated based on the confusion matrix. If the error rate, recall rate and false alarm rate are all less than the preset threshold, the steam boiler operation and maintenance decision model is saved as a model file, stored in the database and training is stopped; otherwise, the initialization parameters of the decision tree model are readjusted and training is repeated; S6. Obtain the current production parameters of the steam boiler system and input them into the trained steam boiler operation and maintenance decision model; the threshold conditions of the steam boiler operation and maintenance decision model are customized according to the boiler model and working scenario; and the decision is output as a basis for guiding the operation and maintenance personnel to adjust the parameters of the boiler's external input equipment at the next moment or decide whether to shut down the equipment for maintenance.

2. The method for optimizing data processing and model training of a time series online decision model according to claim 1, characterized in that: The value of N in step S2-2 is set to the maximum value of the floating-point precision of all sensors or is customized according to the actual scenario.

3. The method for optimizing data processing and model training of a time series online decision model according to claim 1, characterized in that: The time interval T of step S2-3 is not less than 600 seconds.

4. The method for optimizing data processing and model training of a time series online decision model according to claim 1, characterized in that: The boiler feature vector B described in S4 v =(WL ST ,WL ST+T ,…,WL ST+nT ,SST ST ,…,SST ST+nT ,…,SSP ST ,…,SSP ST+nT ,…,OC ST ,…,OC ST+nT ,…,CP ST ,…,CP ST+nT ); Among them, WL ST is the water level at the start of sampling ST, WL ST+T is the water level at ST after a time interval T, WL ST+nT is the water level at ST after n time intervals T, SST ST is the superheated steam temperature at the sampling start time ST, SST ST+nT is the superheated steam temperature after n time intervals T at ST, SSP ST is the superheated steam pressure at the sampling start time ST, SSP ST+nT is the superheated steam pressure after n time intervals T at ST, OC ST is the oxygen content of flue gas at the sampling start time ST, OC ST+nT is the amount of oxygen in the flue gas after n time intervals T at ST, CP ST is the furnace negative pressure at the sampling start time ST, CP ST+nT It is the furnace negative pressure after n time intervals T at ST moment.

5. The method for optimizing data processing and model training of a time series online decision model according to claim 1, characterized in that: In S5, the preset number of times is set to 100 times, the preset change rate threshold is set to 1%, and the preset threshold is set to 5%.

6. The method for optimizing data processing and model training of a time series online decision model according to claim 1, characterized in that: The specific steps of step S6 are as follows: S6-1, using the collected production parameters as input to the online decision model after pre-processing in step S2; S6-2, based on the actual operating conditions of the boiler, determine the threshold conditions, including: water level and its allowable fluctuation range, upper limit of superheated steam temperature, upper and lower limits of superheated steam pressure, range of flue gas oxygen content, and upper and lower limits of furnace negative pressure; S6-3, based on the input data, the decision information of the online decision model is obtained for the operation and maintenance personnel to make decisions.

7. The method for optimizing data processing and model training of a time series online decision model according to claim 1, characterized in that: The integer compression and decompression are performed using the ZigZag algorithm.

8. The method for optimizing data processing and model training of a time series online decision model according to claim 1, characterized in that: The current production parameters of the steam boiler system include water level WL, superheated steam temperature SST and pressure SSP, flue gas oxygen content OC, and furnace negative pressure CP.

Images