Data dynamic modeling analysis system for combustion atmosphere of biomass oxy-combustion boiler

Abstract

The application discloses a biomass oxygen-enriched combustion boiler combustion atmosphere data dynamic modeling analysis system and relates to the technical field of dynamic modeling analysis.The system comprises a working condition self-adaptive feature extraction module, a time sequence keeping model training module, an online self-adaptive correction module and an incremental model updating module.Through the working condition self-adaptive feature extraction module, the time sequence keeping model training module, the online self-adaptive correction module and the incremental model updating module, an adaptive feature vector is generated through feature extraction, a basic model is trained, online correction and incremental updating are determined and executed in combination with a feature vector check threshold, and finally, iterative upgrading is completed through incremental learning fine-tuning of the model, thereby improving the biomass oxygen-enriched combustion boiler combustion atmosphere data dynamic modeling analysis reliability and solving the problem of low biomass oxygen-enriched combustion boiler combustion atmosphere data dynamic modeling analysis reliability in the prior art.

Description

Technical Field

[0001] This invention relates to the field of dynamic modeling and analysis technology, and in particular to a dynamic modeling and analysis system for combustion atmosphere data of biomass oxygen-enriched combustion boilers. Background Technology

[0002] First, in the feature extraction stage, a fixed time window length and extraction rules are used throughout the process. For example, regardless of whether the boiler is under rated load, variable load, or fuel fluctuation conditions, the same window length and feature extraction logic are used. The adaptation strategy is not adjusted according to the dynamic changes of the combustion atmosphere under different operating conditions. This fixed feature extraction method directly results in the extracted features being unable to adapt to the differentiated needs of dynamic atmosphere changes under different operating conditions. Consequently, the model has extremely poor adaptability to operating conditions, cannot maintain stable modeling accuracy across the entire operating condition range, and cannot design a dynamically adapted feature extraction scheme based on the changing patterns of operating conditions. This leads to a serious disconnect between feature extraction and actual operating conditions.

[0003] Secondly, in the dataset partitioning stage, the time-series data related to the combustion atmosphere were randomly divided into training, validation, and test sets instead of being partitioned according to the time sequence of data generation. This unreasonable partitioning method allowed the model to learn the atmosphere data information of subsequent time dimensions in advance during the training process. As a result, the accuracy detection results on the test set seemed to be excellent, but when actually deployed in the industrial field, facing the real time-series combustion process, the model could not effectively capture the real-time changing trend of atmosphere parameters, and the accuracy dropped significantly. This further weakened the importance of the integrity of time-series information, resulting in a lack of engineering rationality in the dataset partitioning method.

[0004] Then, in the dynamic correction stage of the model, the constructed dynamic model was not effectively connected with the boiler DCS control system. It could not collect real-time data such as atmosphere parameters and process parameters during the combustion process, nor could it calculate the deviation between the model prediction value and the actual measured value on site in real time. When sudden operating conditions such as fuel fluctuations (such as sudden changes in biomass moisture content and composition) and fan pressure changes occur during boiler operation, the model cannot update parameters and correct the model in a timely manner according to the actual situation on site, resulting in a continuous increase in prediction deviation. It cannot provide accurate data support for combustion control and there is a disconnect between the model and the on-site system. In the model iteration optimization stage, there are problems such as excessively long offline iteration cycles and poor adaptability.

[0005] Finally, due to the failure of feature extraction to adapt to dynamic adjustments of operating conditions, the violation of time-series data segmentation by time patterns, the lack of online real-time update capability of the model, and the lag of offline iteration cycle behind changes in on-site operating conditions and equipment, the dynamic modeling and analysis of combustion atmosphere data of biomass oxygen-enriched combustion boilers ultimately suffers from low reliability. Summary of the Invention

[0006] To address the low reliability of dynamic modeling and analysis of combustion atmosphere data in existing biomass oxy-fuel combustion boilers, this invention provides a dynamic modeling and analysis system for combustion atmosphere data in biomass oxy-fuel combustion boilers. The technical solution is as follows: On the one hand, a dynamic modeling and analysis system for combustion atmosphere data of biomass oxygen-enriched combustion boilers is provided. This system includes: an adaptive feature extraction module, a time-series model training module, an online adaptive correction module, and an incremental model update module. Specifically, the adaptive feature extraction module collects and processes time-series data in real time to generate a standard time-series dataset. Based on a combination of the boiler's real-time load rate and load change rate, it dynamically determines the current operating condition category. According to different operating condition categories, it calls predefined feature extraction strategies and performs feature calculations on the standard time-series dataset based on the selected strategy, generating a feature vector dynamically adapted to the current operating condition, which serves as the model input. The time-series model training module receives data from the adaptive feature extraction module. The system adapts to the feature vectors extracted by the feature extraction module and optimizes the initial dynamic prediction model to obtain the basic model. The online adaptive correction module receives the current time-series process data in real time based on the basic model. After processing by the working condition adaptive feature extraction module, it generates real-time feature vectors. Based on the real-time feature vectors, it determines whether to perform online adaptive correction optimization. If so, incremental model update is performed after correction optimization; otherwise, incremental model update is performed directly. The incremental model update module collects new time-series process data from the previous model update to form an incremental dataset. Based on the incremental dataset and using incremental learning, the basic model is structurally fine-tuned to generate an updated model version, which replaces the original basic model, completing the iterative upgrade of the model.

[0007] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages: 1. By combining real-time boiler load rate and load change rate, the system accurately classifies operating conditions into three categories: steady-state, gradual, and transient. Predefined feature extraction strategies are matched to different operating conditions, and the time-series window length and sliding step size are dynamically adjusted to complete time-frequency feature calculation and high-dimensional feature vector construction. This ensures that the feature vectors are highly compatible with the dynamic changes in atmosphere data under different operating conditions of biomass oxygen-enriched combustion, guaranteeing the fit of the model input operating conditions from the feature extraction level. Simultaneously, the system dynamically corrects the feature vector verification threshold based on the input layer neuron dimension. The missing value correction coefficient and range correction coefficient are used to optimize the basic missing value verification threshold and the feature value process logic range verification threshold, respectively. This allows the feature vector verification threshold to dynamically adjust with the feature vector dimension, rigorously eliminating invalid feature vectors with excessive missing values ​​or those that violate the biomass oxygen-enriched combustion process logic. This effectively avoids model input distortion caused by missing features or abnormal feature values, providing high-quality and highly effective feature vector input for model training and online calibration from the data verification level, thus solidifying the data source foundation for the entire dynamic modeling and analysis.

[0008] 2. By segmenting the feature vector time sequence along the time axis, the system divides the training, validation, and testing time intervals, strictly maintaining the causal temporal relationship of the data. This completely avoids the "pseudo-high precision" problem caused by random partitioning and ensures the temporal rationality of the basic model training. Based on this, the system adjusts the number of training rounds in a correlation with the training set cutoff timestamp, matches the baseline number of training rounds to the actual time span of the training set, and introduces a data volume correction coefficient to calculate the actual number of training rounds. This allows the number of model training iterations to be dynamically optimized with the training set time span and the effective sample size, avoiding overtraining when the sample size is insufficient or undertraining when the sample size is sufficient. Simultaneously, the system dynamically adjusts the feature vector verification threshold driven by the neuron dimension, ensuring the validity of the training set feature vectors throughout the model training phase. The synergistic effect of the two allows the optimized training of the initial dynamic prediction model to match the temporal characteristics and sample size of the training data, and achieves accurate training based on high-standard feature vector verification, effectively improving the training accuracy of the basic model. At the same time, through the evaluation of the generalization ability of training and verification time interval data, the basic model can adapt to the changes in atmosphere data under different load conditions of biomass oxygen-enriched combustion boilers, greatly improving the generalization ability of the basic model under all operating conditions.

[0009] 3. The system determines whether to perform online correction optimization by analyzing the prediction deviation of real-time feature vectors based on the base model. It dynamically adapts the model prediction deviation judgment threshold and the deviation persistence judgment duration threshold through the positive correlation between model prediction deviation and feature vector verification threshold. Simultaneously, it introduces a boiler operating condition correction coefficient to calculate the actual deviation persistence judgment duration threshold. This allows the deviation judgment parameters to dynamically adjust according to the stringency of feature vector verification and the actual boiler operating conditions. This avoids false triggering of correction due to instantaneous data fluctuations and accurately judges small model drifts under high-precision feature vector verification and significant model drifts during rapid changes in operating conditions, achieving precise control over the triggering timing of online adaptive correction optimization. For scenarios requiring correction, the system performs incremental updates after completing the correction optimization; otherwise, it directly performs incremental updates, balancing the accuracy and real-time performance of online model correction. Meanwhile, the incremental model update module collects new time-series process data to form an incremental dataset. Through incremental learning, the basic model is structurally fine-tuned. Combined with incremental updates after online correction, the model is iteratively upgraded, enabling it to adapt in real time to changes in atmosphere data caused by fuel fluctuations, equipment status changes, and load adjustments in biomass oxy-fuel combustion boilers. This effectively suppresses model drift and continuously ensures the accuracy of the dynamic modeling and analysis system's prediction of boiler combustion atmosphere data. It achieves dynamic optimization of the model throughout the entire process from online correction to incremental iteration, ensuring that the entire modeling and analysis system always remains adapted to the actual operating state of the biomass oxy-fuel combustion boiler, providing reliable model support for the precise control of boiler combustion atmosphere. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 A schematic diagram of the structure of the dynamic modeling and analysis system for combustion atmosphere data of a biomass oxygen-enriched combustion boiler provided in an embodiment of this application; Figure 2 The flowchart illustrates the dynamic correction process of the feature vector verification threshold driven by the execution neuron dimension of the dynamic modeling and analysis system for combustion atmosphere data of a biomass oxygen-enriched combustion boiler provided in this application embodiment. Detailed Implementation

[0012] The technical solution provided in this application will now be described in conjunction with the accompanying drawings.

[0013] To facilitate understanding of the embodiments of this application, the following points will be explained first: First, in this application, "at least one" means one or more, and "more than one" means two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A alone, A and B simultaneously, or B alone, where A and B can be singular or plural. The character " / " generally indicates an "or" relationship between the preceding and following related objects, but it does not exclude the possibility of indicating an "and" relationship; the specific meaning can be understood in context. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one of a, b, or c can mean: a, b, c; a and b; a and c; b and c; or a and b and c. Here, a, b, and c can be single or multiple.

[0014] Second, the use of prefixes such as "first" and "second" in this application is solely for the purpose of distinguishing and describing different things belonging to the same category, and does not constrain the order, size, or quantity of things. For example, "first message" and "second message" are simply different messages, and there is no chronological, size, or priority relationship between them.

[0015] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0016] like Figure 1 The diagram shown is a structural schematic of the dynamic modeling and analysis system for combustion atmosphere data of a biomass oxy-fuel combustion boiler provided in this application embodiment. The dynamic modeling and analysis system for combustion atmosphere data of a biomass oxy-fuel combustion boiler provided in this application embodiment includes: an adaptive feature extraction module for operating conditions, a time-series-maintaining model training module, an online adaptive correction module, and an incremental model update module.

[0017] As the first module of this system, the adaptive feature extraction module is used to collect and process time-series process data in real time to generate a standard time-series dataset. Based on the combination of boiler real-time load rate and load change rate, it dynamically determines the current operating condition category. According to different operating condition categories, it calls the predefined feature extraction strategy, performs feature calculation on the standard time-series dataset based on the selected feature extraction strategy, and generates a feature vector that is dynamically adapted to the current operating condition as the model input.

[0018] It should be understood that time-series process data includes, but is not limited to, fuel supply parameters, air distribution parameters, oxygen enrichment concentration parameters, boiler load parameters, and combustion atmosphere parameters measured by a flue gas analyzer; the time-series process data is preprocessed by timestamp alignment, outlier handling, and filtering smoothing to generate a standard time-series dataset; the operating condition categories include steady-state operating conditions, gradual operating conditions, and transient operating conditions.

[0019] It should be noted that the specific steps for dynamically determining the current operating condition category are as follows: Step 101: Obtain the load measurement value of the boiler in real time, and obtain the real-time load rate and the real-time load change rate. The real-time load rate is the percentage value of the current load measurement value to the rated load of the boiler. The real-time load change rate is the change in load rate per unit time. It is obtained by differential calculation of the load rate of the continuous time series and is processed by low-pass filtering to eliminate the interference of measurement noise. Step 102: If the real-time load rate is less than the load rate threshold and the real-time load change rate is less than the change rate threshold, then it is determined to be a steady-state condition. If the real-time load rate is less than the load rate threshold and the real-time load change rate is greater than the change rate threshold, it is determined to be a gradual change condition. If the real-time load rate is greater than the load rate threshold and the real-time load change rate is less than the change rate threshold, it is determined to be a gradual change condition. If the real-time load rate is greater than the load rate threshold and the real-time load change rate is greater than the change rate threshold, it is determined to be a transient operating condition.

[0020] In this embodiment, step 101 is the basic data acquisition and processing stage for boiler operating condition identification. First, the boiler load measurement value is acquired in real time to ensure the timeliness and accuracy of data acquisition, providing real operational data support for subsequent operating condition determination. Based on the acquired current load measurement value and the boiler's rated load, a percentage conversion is performed to calculate the real-time load rate. This parameter can intuitively reflect the proportion of the boiler's current actual load relative to the design rated load, achieving a quantitative representation of the boiler's load-bearing status. Simultaneously, using the change in load rate per unit time as the calculation basis, differential calculation is performed on the load rate under continuous time series to obtain initial load change rate data, which is then processed through low-pass filtering. Noise cancellation of the initial load change rate effectively filters out measurement noise generated during load measurement due to factors such as sensor accuracy and on-site electromagnetic interference, avoiding distortion of the load change rate calculation results caused by noise interference, and ensuring the accuracy and stability of real-time load change rate data. This parameter can accurately reflect the dynamic change trend of boiler load rate, providing reliable dynamic data indicators for dynamic judgment of operating conditions. Through the above data processing method, it is possible to simultaneously obtain the real-time load rate reflecting the static load bearing state of the boiler and the real-time load change rate reflecting the dynamic change characteristics of the load, forming a basic data system for judging operating conditions that combines static and dynamic data, laying a data foundation for subsequent accurate classification and judgment of operating conditions.Step 102 is the boiler operating condition classification and determination step based on the processed data described above. A load rate threshold and a change rate threshold are pre-set as quantitative benchmarks for operating condition determination, building upon step 101. The system precisely processes real-time load rate and real-time load change rate, using a combination of dual parameters to accurately classify boiler operating conditions: If the real-time load rate is less than the load rate threshold and the real-time load change rate is less than the change rate threshold, it indicates that the boiler is currently under low load and the load is not fluctuating significantly, thus classifying it as a steady-state condition. This judgment can accurately identify the characteristics of a stable boiler load and smooth operation, providing a basis for subsequent matching of stable operation control strategies. If the real-time load rate is less than the load rate threshold and the real-time load change rate is greater than the change rate threshold, it indicates that although the boiler is under low load, the load shows a clear dynamic change trend, thus classifying it as a gradual change condition. If the real-time load rate is greater than the load rate threshold and the real-time load change rate is less than the change rate threshold, it indicates that although the boiler is under high load, the load is not fluctuating significantly and the change trend is gentle, also classified as a gradual change condition. The judgment of these two scenarios can accurately identify boiler loads that are dynamically changing but have not reached a drastic fluctuation. The system effectively distinguishes between steady-state and transient operating conditions, providing precise operating condition data for subsequent gradual control strategies. If the real-time load rate is greater than the load rate threshold and the real-time load change rate is greater than the change rate threshold, it indicates that the boiler is currently under high load and the load is showing a drastic dynamic change trend, thus being identified as a transient operating condition. This judgment result can accurately identify the extreme operating condition characteristics of high and drastically fluctuating boiler load, providing a key operating condition judgment basis for timely activation of emergency control and ensuring the safe and stable operation of the boiler. Through the above-mentioned four-quadrant combination judgment method of dual parameters, comprehensive coverage and accurate differentiation of the three types of boiler operating conditions—steady-state, gradual, and transient—are achieved, effectively avoiding the operating condition identification deviation caused by single-parameter judgment, improving the accuracy and comprehensiveness of boiler operating condition judgment, and providing precise operating condition identification support for subsequent matching of differentiated operation control strategies for different operating conditions, ensuring that the boiler can achieve efficient, stable, and safe operation under different operating conditions.

[0021] It should be further explained that the specific process for invoking the predefined feature extraction strategy according to different working condition categories is as follows: Step 201: Select the received feature extraction strategy corresponding to the current working condition category from the predefined feature extraction strategy, and determine its temporal window length and sliding step size; Step 202: Using the current sampling time as a reference, extract a data sequence with a length equal to the length of the time series window from the standard time series dataset by backtracking along the time axis to form the current analysis window; Step 203: During continuous operation, repeat the interception operation from step 201 to step 202 at each sliding step interval to form a series of analysis window sequences that are continuous in time. Step 204: Perform time-frequency feature calculations in parallel for each process variable data sequence contained in each analysis window.

[0022] Based on different operating conditions, predefined feature extraction strategies are invoked, including: Step 205: For a single analysis window, arrange the multiple feature values ​​calculated according to the time-frequency characteristics of each process variable in a predetermined order to form the feature sub-vector of the variable. The multiple feature values ​​include, but are not limited to, the mean and standard deviation. Step 206: Concatenate the feature vectors of all process variables according to the predefined variable order to form a high-dimensional one-dimensional feature vector; Step 207: Each feature vector generated in a predefined time order is bound to the center time of its corresponding analysis window to form a feature vector with a timestamp. Step 208: Continuously output feature vectors with timestamps to form a time sequence of feature vectors after working condition adaptive feature extraction; Step 209: The time sequence of feature vectors is fed to the time-preserving model training module or the online adaptive correction module as input to the dynamic prediction model.

[0023] In this embodiment, step 201 serves as the strategy matching step for adaptive feature extraction under operating conditions. First, it precisely selects and receives feature extraction strategies from a predefined feature extraction strategy library that correspond one-to-one with the current boiler operating condition category. Based on this specific strategy, it clarifies the corresponding time-series window length and sliding step size. By matching customized core feature extraction parameters for different operating conditions, it avoids the problem of insufficient feature extraction accuracy caused by adapting a single parameter to all operating conditions. This achieves adaptive matching between the feature extraction strategy and the boiler operating conditions, ensuring that subsequent time-series data extraction can closely match the time-series change characteristics of process variables under the current operating conditions, laying a suitable parameter foundation for accurate feature extraction. Step 202 Based on the determined time window length, using the current sampling time as the time base, a complete data sequence with a length exactly equal to the time window length is extracted from the standardized boiler process variable standard time series dataset, tracing back along the time axis. This constitutes the current analysis window for a single point in time. Precise extraction with a fixed window length ensures that the time series data within each analysis window has a uniform time dimension scale, avoiding deviations in subsequent feature calculations due to inconsistent data lengths. Simultaneously, the use of standardized datasets eliminates interference from the original data's dimensions, noise, etc., further improving the effectiveness of the data within the analysis window. Step 203: During continuous boiler operation, steps 201 to 202 are continuously repeated according to the sliding step size determined in step 201. The time-series data extraction operation forms a series of continuous and uninterrupted analysis window sequences in the time dimension. Through the regular control of the sliding step size, the temporal continuity of feature extraction is ensured, which can completely capture the time-series changes of boiler process variables. At the same time, the time resolution of data extraction can be flexibly adjusted according to the operating conditions. While ensuring the comprehensiveness of feature extraction, the amount of data processing is effectively controlled, and the overall efficiency of feature extraction is improved. Step 204: For each generated analysis window, time-frequency feature calculation is performed synchronously and in parallel on all boiler process variable data sequences contained in the window. The parallel calculation method significantly shortens the time consumption of multi-variable feature extraction, which is adapted to the timeliness requirements of online real-time monitoring of boilers. The calculation of time-frequency features breaks through the limitations of single time domain or frequency domain features. It can simultaneously capture the change law of process variables in the time dimension and the feature information in the frequency dimension, fully explore the hidden operating conditions in the data, and improve the feature representation ability.Meanwhile, this technical solution also calls predefined feature extraction strategies based on different operating conditions, and designs subsequent feature vector construction and output steps: Step 205 For a single analysis window, the mean, standard deviation, and other feature values ​​obtained by time-frequency feature calculation for each process variable are arranged in a predefined fixed order to form a feature sub-vector specific to that process variable. The orderly arrangement of feature values ​​ensures the structural consistency of the feature sub-vector and avoids errors in subsequent feature splicing due to disordered order. The combination of basic features such as mean and standard deviation with other time-frequency features can characterize the process variable from multiple dimensions such as central tendency, dispersion, and frequency distribution, giving the feature sub-vector a comprehensive representation ability for a single variable; Step 206 After the feature sub-vectors of each process variable are constructed, they are seamlessly spliced ​​together according to the predefined order of boiler process variables to form a high-dimensional one-dimensional feature vector. The unified variable splicing order ensures the standardization and consistency of the high-dimensional feature vector structure. The construction of the high-dimensional feature vector realizes the fusion of the features of all process variables of the boiler, which can comprehensively represent the overall operating status of the boiler under the current operating conditions from the system level, avoiding the one-sidedness of single variable feature representation. Step 207: Each high-dimensional feature vector generated according to the predefined time order is uniquely bound to the center time of its corresponding analysis window to form a feature vector with a timestamp. The binding of the timestamp gives the feature vector a clear time attribute, which can accurately correspond to the specific operating time of the boiler, realize the accurate correlation between feature information and the actual operating time of the boiler, and provide accurate time dimension support for the subsequent training and online correction of the time series model. Step 208: Based on the continuous operation of the boiler, the aforementioned time-stamped feature vectors are continuously and orderly output, ultimately forming a feature vector time series sequence after adaptive feature extraction. This sequence retains the integrity of the feature information and continues the temporal continuity of the original time series data, achieving a precise transformation from the original process variable time series data to feature-level time series data. Step 209 directly inputs the aforementioned feature vector time series sequence to the time series-keeping model training module or the online adaptive correction module as the core input data of the boiler dynamic prediction model. Because the feature vector time series sequence has undergone adaptive feature extraction and time stamp binding, it has both the feature representation capability that fits the current operating condition and retains the complete temporal characteristics. It can provide high-quality and highly adaptable input data for the training and online correction of the dynamic prediction model, effectively improving the model's prediction accuracy and adaptive correction capability for the boiler's operating state under different operating conditions. This allows the dynamic prediction model to accurately match the actual operating conditions of the boiler, ensuring the reliability and effectiveness of the model's output results.

[0024] As the second module of this system, the time-preserving model training module receives feature vectors from the working condition adaptive feature extraction module and optimizes the initial dynamic prediction model to obtain the base model.

[0025] It needs to be explained that the specific process for obtaining the basic model is as follows: The feature vector time series is divided into training time interval, validation time interval and test time interval on the time axis to ensure that the timestamps of the training data are all earlier than the validation data, and the timestamps of the validation data are all earlier than the test data, thereby maintaining the causal time sequence relationship of the data. The generalization ability of the trained dynamic prediction model is evaluated by using data from the training and validation time intervals. Based on the evaluation results, the number of training rounds is adaptively adjusted according to the training set cutoff timestamp to dynamically optimize the number of model training iterations and dynamically correct the feature vector verification threshold driven by the neuron dimension, ensuring the effectiveness of feature input and training accuracy of the model under different operating conditions.

[0026] In this embodiment, the time-series sequence of feature vectors obtained after adaptive feature extraction is first subjected to ordered segmentation along the time dimension. It is divided into independent training, validation, and testing time intervals according to the chronological order of the timeline. During the segmentation process, the time sequence rule is strictly followed: all training data timestamps are earlier than the validation data timestamps, and all validation data timestamps are earlier than the testing data timestamps. This completely avoids the possibility of future data interfering with model training, validation, and testing prematurely, strictly maintaining the causal temporal relationship between feature vectors and boiler operating states. This ensures that the process of the model learning patterns, validating effects, and testing performance based on historical feature data perfectly matches the actual operating sequence logic of the boiler. From the data source, this guarantees the authenticity and effectiveness of subsequent model training and evaluation results, avoiding the model learning incorrect operating condition correlation patterns due to chaotic data temporal sequence, and improving the accuracy of the model's generalization ability evaluation. Subsequently, based on the feature vector time-series data within the divided training and validation time intervals, a comprehensive evaluation of the generalization ability of the dynamically predicted model that has completed basic training is conducted. This evaluation process is not a single-dimensional performance test, but rather uses the evaluation results as the core basis for dynamic model optimization, specifically performing two key model optimization operations: First, adaptive adjustment of the number of training epochs in relation to the training set deadline timestamp. Based on the evaluation results, the number of iterations in the training epochs is correlated and matched with the training set deadline timestamps. The number of iterations is dynamically optimized according to the model learning effect under different time stages and working conditions. This avoids both underfitting due to insufficient iterations and overfitting due to excessive iterations, allowing the model to achieve optimal performance. The number of training iterations can match the temporal and operational characteristics of the feature data, improving the learning accuracy of the model at different time stages and under different operating conditions. Secondly, the feature vector verification threshold is dynamically corrected by the neuron dimension. With the feature learning feedback of the model's neuron dimension as the core driver, the verification threshold of the feature vector is adjusted in real time and dynamically. This can accurately identify the effective representation range of the feature vector under different operating conditions, promptly eliminate invalid feature information caused by operating condition fluctuations and data noise, and retain the effective feature input that can truly reflect the boiler's operating status. From the feature verification level, this ensures the effectiveness and accuracy of the model's feature input under different operating conditions such as steady state, gradual change, and transient change, and avoids invalid features interfering with the model's learning and prediction process.By using the aforementioned time-axis ordered dataset partitioning method, combined with two-dimensional model dynamic optimization based on generalization ability evaluation results, the entire process of adaptation and optimization from data foundation to model training was achieved. This ensured the temporal rationality of the model training data and allowed the number of model iterations and feature verification thresholds to be dynamically adjusted according to operating conditions and data features. This effectively improved the training accuracy, feature recognition ability, and generalization adaptation ability of the dynamic prediction model under different boiler operating conditions, laying a solid foundation for the reliable application of the model in subsequent online adaptive correction and actual operating status prediction.

[0027] It should be further explained that the specific process for adaptively adjusting the number of training rounds based on the training set deadline timestamp is as follows: The predefined association mapping relationship between the training set end timestamp and the benchmark training rounds is divided into multiple time intervals according to the training set time span. Each time interval corresponds to a benchmark training round. The training set time span is the time difference between the start timestamp and the end timestamp of the training set. The larger the time span, the more benchmark training rounds are matched. The actual cutoff timestamp of the training set in the time-preserving model training module is obtained in real time, the actual time span of the training set is calculated, and the corresponding baseline training round number is obtained by matching the correlation mapping relationship. A training set data volume correction coefficient is introduced. The data volume correction coefficient is determined based on the ratio of the effective feature vector sample size of the training set corresponding to the actual cutoff timestamp to the preset sample size threshold. When the sample size exceeds the threshold, the correction coefficient is greater than 1, and when the sample size is less than the threshold, the correction coefficient is less than 1. Based on the baseline number of training rounds, the actual number of training rounds of the model is calculated by combining the data volume correction coefficient. The actual number of training rounds is the result of combining the baseline number of training rounds and the data volume correction coefficient. The actual number of training rounds is constrained between a preset minimum training rounds threshold and a maximum training rounds threshold. Finally, the initial dynamic prediction model is optimized and trained using the actual number of training rounds.

[0028] In this embodiment, a predefined mapping relationship between the training set cutoff timestamp and the baseline training rounds is first defined. The training set is divided into multiple independent time intervals according to its time span, and a unique baseline training round is matched for each time interval. The training set time span is the actual time difference between the start and end timestamps of the training set, and follows the core mapping rule that the larger the time span, the more baseline training rounds are matched. This predefined mapping relationship achieves the initial adaptation between the training rounds and the time dimension features of the training set, allowing the baseline training rounds to be matched in a stepwise manner according to the boiler operating time range covered by the training set. This avoids underfitting or overfitting problems caused by a single training round matching different time span training sets, and lays a baseline foundation that fits the time features for the accurate calculation of the subsequent training rounds. Subsequently, the actual cutoff timestamp of the training set in the time-preserving model training module is obtained in real time. Based on this actual cutoff timestamp and the fixed start timestamp of the training set, the actual time span of the training set is accurately calculated. Then, the predefined association mapping relationship is directly called to match the corresponding baseline training epochs for this actual time span. By matching the real-time time span calculation with the baseline epochs, the baseline training epochs can be made to fit the actual data time characteristics of the model training, ensuring the real-time performance and adaptability of the baseline epochs. On this basis, a training set data volume correction coefficient is introduced. This correction coefficient is accurately determined based on the ratio of the effective feature vector sample size of the training set corresponding to the actual cutoff timestamp to a preset sample size threshold. A quantization rule is set that the correction coefficient is greater than 1 when the sample size exceeds the threshold and less than 1 when the sample size is less than the threshold. By introducing the data volume correction coefficient, a secondary adaptation of the training epochs to the dimension of the training set sample size is achieved. The baseline training epochs can be adjusted in a targeted manner according to the amount of effective samples, making up for the single-dimensional defect of matching the epochs based solely on the time span. This allows the calculation of the training epochs to take into account both the time span and sample size of the training set. Next, based on the baseline training rounds obtained through matching, the actual training rounds are calculated by combining the training set data volume correction coefficient determined above. During the calculation process, strict numerical constraints are set for the actual training rounds, limiting them to between the preset minimum training round threshold and the maximum training round threshold. This numerical constraint can effectively avoid the actual training rounds being too high or too low due to abnormal fluctuations in the sample size, and prevent problems such as overfitting, underfitting, or low training efficiency caused by extreme rounds, ensuring that the actual training rounds are always within a reasonable and effective training range.Finally, the initial dynamic prediction model was optimized by using the actual number of training rounds after time span matching, sample size correction, and numerical constraints. This ensured that the number of training iterations matched both the time and quantity dimensions of the training set, while remaining within a reasonable range. This guaranteed that the model could learn sufficiently and appropriately from training sets with different time spans and sample sizes, effectively improving the model's training accuracy and generalization ability. It also avoided ineffective over-iteration, improving the model's training efficiency. Furthermore, the number of training rounds provided precise support for adapting the model to the feature data characteristics of different boiler operating conditions, ensuring that the optimized dynamic prediction model could more accurately capture the temporal patterns of boiler operation and improve the accuracy of predicting boiler operating status.

[0029] It should be understood that, such as Figure 2 The diagram shows a flowchart of the dynamic correction process for feature vector verification thresholds driven by the neuron dimension in the dynamic modeling and analysis system for combustion atmosphere data of a biomass oxygen-enriched combustion boiler provided in this application embodiment. The specific process is as follows: Starting with "Feature Vector Verification Threshold Association Correction," the "Predefined Association Correction Rules" step is executed first, clarifying three core rules: the more neurons there are, the higher the corresponding feature vector dimension; the more neurons there are, the lower the corresponding feature value missing rate verification threshold; and the more neurons there are, the narrower the fluctuation range of the feature vector verification threshold. Simultaneously, this step acquires the actual number of neurons in the input layer of the initial dynamic prediction model in the time-series-preserving model training module in real time, and matches the corresponding threshold correction coefficient according to the predefined association correction rules. Next, the process enters a judgment phase. If the correction coefficient is a missing rate correction coefficient, it is used to combine the preset feature value basic missing rate verification threshold to obtain the actual missing rate verification threshold. If the correction coefficient is a range correction coefficient, its complementary coefficient is combined with the basic upper limit threshold to obtain the actual lower limit threshold, and the result of the superposition of the coefficient and 1 is combined with the basic lower limit threshold to obtain the actual lower limit threshold. Finally, the actual missing rate verification threshold is compared with the actual upper limit.

[0030] It should be further explained that the specific steps for dynamically adjusting the feature vector verification threshold driven by the neuron dimension are as follows: A predefined correlation correction rule between the number of input layer neurons and the feature vector verification threshold is established. The number of input layer neurons corresponds one-to-one with the feature vector dimension. The more neurons there are, the higher the corresponding feature vector dimension, the lower the matching feature value missing rate verification threshold, and the narrower the fluctuation range of the feature vector verification threshold. The actual number of neurons in the input layer of the initial dynamic prediction model in the time-preserving model training module is obtained in real time, and the corresponding threshold correction coefficient is obtained by matching according to the association correction rule. The threshold correction coefficient includes the missing rate correction coefficient and the range correction coefficient. The preset feature value basic missing rate verification threshold is corrected based on the missing rate correction coefficient to obtain the actual missing rate verification threshold. The actual missing rate verification threshold is the result of combining the feature value basic missing rate verification threshold and the missing rate correction coefficient, and the actual missing rate verification threshold is not higher than the preset maximum missing rate threshold. The upper and lower limits of the preset feature value basic missing rate verification threshold are corrected based on the range correction coefficient to obtain the actual verification threshold. The actual upper limit threshold is the result of combining the complementary coefficient of the basic upper limit threshold and the range correction coefficient, and the actual lower limit threshold is the result of combining the basic lower limit threshold with the superposition result of 1 and the range correction coefficient. The actual missing rate verification threshold and the actual upper limit threshold are used as the final verification thresholds for the feature vectors to verify their validity.

[0031] In this embodiment, a predefined rule is first established to correlate the number of neurons in the input layer with the feature vector verification threshold. This rule establishes a precise one-to-one correspondence between the number of neurons in the input layer and the dimension of the feature vector. It follows the core correction logic that the more neurons there are, the higher the dimension of the corresponding feature vector, the lower the matching feature value missing rate verification threshold, and the narrower the fluctuation range of the feature vector verification threshold. This predefined rule achieves a deep fit between the verification threshold and the model input layer structure and the dimension of the feature vector. Setting a lower missing rate threshold and a narrower fluctuation range for high-dimensional feature vectors can impose stricter verification requirements on the integrity of high-dimensional features, avoiding feature representation distortion caused by feature missingness in high dimensions. On the other hand, a relatively lenient verification threshold is matched for low-dimensional feature vectors, which can reduce the false rejection of effective features while ensuring feature validity. From a rule perspective, this ensures the adaptability of the verification threshold to feature vectors of different dimensions, laying a structured rule foundation for subsequent accurate threshold correction. Subsequently, the actual number of neurons in the input layer of the initial dynamic prediction model in the time-preserving model training module is obtained in real time. Based on this actual number of neurons, the predefined association correction rules mentioned above are directly called to accurately match and obtain the corresponding threshold correction coefficients. These correction coefficients include two core types: missing rate correction coefficients and range correction coefficients. They correspond to the numerical correction of the feature value missing rate threshold and the range correction of the verification threshold fluctuation range, respectively. By obtaining the number of neurons in real time and matching the correction coefficients, the threshold correction can fit the current actual input layer structure and feature vector dimension of the model, ensuring the real-time performance and accuracy of the threshold correction and avoiding verification bias caused by adapting a fixed threshold to feature vectors of different dimensions.Based on this, the preset feature value base missing rate verification threshold is quantitatively corrected using the missing rate correction coefficient obtained from the matching, resulting in the actual missing rate verification threshold. This actual missing rate verification threshold is the result of a combination operation between the feature value base missing rate verification threshold and the missing rate correction coefficient. A hard constraint is set that the actual missing rate verification threshold does not exceed the preset maximum missing rate threshold. This constraint effectively prevents the missing rate threshold from being too high due to abnormal correction coefficients, preventing a large number of invalid vectors with missing features from passing the verification. From the perspective of feature integrity, this ensures that the feature vectors of the input model have sufficient effective information, avoiding model learning bias caused by too many missing features. Simultaneously, the upper and lower limits of the preset feature value base verification threshold are specifically corrected based on the range correction coefficient. The actual upper limit threshold is processed as a combination of the complementary coefficients of the base upper limit threshold and the range correction coefficient, and the actual lower limit threshold is processed as a combination of the base lower limit threshold and 1. The calculation rules for combining the superposition results of the range correction coefficient and the actual upper and lower thresholds of the feature vector verification are obtained. This correction method can dynamically adjust the fluctuation range of the verification threshold according to the feature vector dimension, making the verification range of high-dimensional feature vectors narrower and low-dimensional feature vectors more reasonable. It achieves accurate matching between the fluctuation range of the verification threshold and the feature dimension, avoiding invalid feature vectors from passing the verification due to the range being too wide, or valid feature vectors from being mistakenly removed due to the range being too narrow. Finally, the actual missing rate verification threshold and the actual upper limit threshold, obtained after correction, are used as the final verification thresholds for the feature vectors. A comprehensive validity verification is carried out on the feature vectors of the received input model. This final verification threshold has undergone deep adaptation and correction with the number of neurons in the model input layer and the dimension of the feature vectors, and strict numerical constraints have been set. It can perform differentiated and precise validity judgments for feature vectors of different dimensions, effectively eliminating invalid feature vectors with missing features or feature values ​​exceeding the reasonable range, and accurately retaining valid feature vectors that can truly and completely represent the boiler operating status. From the source of feature input, the validity, completeness and accuracy of the feature vectors input to the dynamic prediction model are guaranteed, and invalid features are avoided from interfering with the model training and prediction process. This greatly improves the model's efficiency in recognizing and utilizing feature inputs under different operating conditions, and further ensures the training accuracy of the dynamic prediction model and the prediction accuracy of the boiler operating status.

[0032] As the third module of this system, the online adaptive correction module is used to receive the current time-series process data in real time based on the basic model. After being processed by the working condition adaptive feature extraction module, a real-time feature vector is generated. Based on the real-time feature vector, it is determined whether to perform online adaptive correction optimization. If yes, incremental model update is performed after correction optimization; otherwise, incremental model update is performed directly.

[0033] It should be noted that the specific steps for determining whether to perform online adaptive correction optimization based on real-time feature vectors are as follows: By inputting the real-time feature vector into the basic model, the real-time predicted values ​​of the combustion atmosphere parameters are obtained. The real-time predicted value is compared with the actual value measured by the flue gas analyzer to calculate the real-time prediction deviation. When the real-time prediction deviation continuously exceeds the preset threshold, the deviation judgment parameter is dynamically adapted and adjusted. The specific steps for dynamic adaptation and adjustment of deviation judgment parameters are as follows: The predefined association mapping rule between the model prediction deviation judgment threshold and the deviation duration judgment threshold is defined. The model prediction deviation judgment threshold is the maximum allowable percentage deviation between the predicted value output by the basic model based on the real-time feature vector and the actual field-collected value. The deviation duration judgment threshold is the maximum allowable duration for which the model prediction deviation exceeds the model prediction deviation judgment threshold. The value of the model prediction deviation judgment threshold is positively correlated with the feature vector verification threshold. The more stringent the feature vector verification threshold, the smaller the value of the model prediction deviation judgment threshold, and the shorter the corresponding matched deviation duration judgment threshold. The actual value of the prediction deviation judgment threshold currently in effect in the online adaptive correction module is obtained in real time. The actual value is dynamically matched by the real-time correction result of the feature vector verification threshold. The feature vector verification threshold includes the feature value missing rate verification threshold and the feature value process logic range verification threshold. The lower the feature value missing rate verification threshold and the narrower the fluctuation range of the feature value process logic range verification threshold, the more stringent the feature vector verification threshold, and the smaller the actual value of the matched model prediction deviation judgment threshold.

[0034] Dynamic adaptation and adjustment of deviation judgment parameters also include: Based on the association mapping rules, the baseline threshold for the duration of deviation determination is obtained by matching the actual value of the deviation determination threshold predicted by the model with the actual value of the deviation determination threshold. If the actual value of the model prediction deviation judgment threshold is less than or equal to the preset base threshold, the matching deviation duration judgment benchmark threshold is the short duration level. If the actual value of the model prediction deviation judgment threshold is greater than the preset base threshold, the matching deviation duration judgment benchmark threshold is the long duration level. A condition correction factor is introduced, which is determined according to the real-time operating condition category of the boiler (for example, the condition correction factor is 1.2 under rated stable load conditions and 0.8 under variable load and low load conditions). Based on the baseline threshold for the duration of deviation determination, the actual threshold for the duration of deviation determination is calculated by combining the condition correction factor. The actual threshold for the duration of deviation determination is the result of combining the baseline threshold for the duration of deviation determination and the condition correction factor, and the actual threshold for the duration of deviation determination is constrained between the preset shortest duration threshold and the longest duration threshold.

[0035] In this embodiment, during the dynamic operation monitoring and control of the boiler, the real-time feature vector, after being extracted and validated by the adaptive feature extraction, is first accurately input into the basic prediction model. The model's inference calculations yield real-time predicted values ​​of the combustion atmosphere parameters. These real-time predicted values ​​quickly reflect the theoretical state of the boiler's current combustion atmosphere, providing a forward-looking reference for subsequent combustion control and effectively avoiding the control lag problem caused by relying solely on on-site measurements. Subsequently, the real-time predicted values ​​are compared one by one with the actual values ​​of the combustion atmosphere parameters obtained from the flue gas analyzer's on-site measurements. The real-time prediction deviation between the two is accurately calculated. This deviation value is a core quantitative indicator for measuring the prediction accuracy of the basic model and determining whether the model is suitable for the current operating state. Real-time deviation calculation can promptly capture fluctuations in the model's prediction effect, providing accurate data support for subsequent control decisions. During continuous monitoring, the real-time prediction deviation is continuously tracked and judged. When the real-time prediction deviation continuously exceeds a preset deviation threshold, the dynamic adaptation control process for the deviation judgment parameters is immediately triggered. This triggering mechanism can quickly respond to a decrease in model prediction accuracy, preventing combustion control from becoming inaccurate due to continuously exceeding the deviation threshold, and ensuring the stability and safety of the boiler combustion process.

[0036] The dynamic adaptation and control process for deviation judgment parameters has complete rule presets and real-time adaptation logic. The specific operation and technical effects are as follows: First, a correlation mapping rule is predefined between the model prediction deviation judgment threshold and the deviation duration judgment threshold. The model prediction deviation judgment threshold is clearly defined as the maximum allowable percentage deviation between the predicted value output by the basic model based on the real-time feature vector and the actual field-collected value, used to define the reasonable range of prediction deviation. The deviation duration judgment threshold is the maximum allowable continuous duration for which the model prediction deviation exceeds the above judgment threshold, used to avoid situations where instantaneous deviations falsely trigger control. Simultaneously, the rules clearly define the model prediction deviation threshold. The value of the prediction deviation judgment threshold is positively correlated with the feature vector verification threshold. That is, the more stringent the feature vector verification threshold (the higher the requirements for the integrity and rationality of the feature vector), the smaller the value of the model prediction deviation judgment threshold, and the shorter the corresponding deviation judgment duration threshold. This rule design realizes the deep binding between the deviation judgment standard and the feature input quality. When the feature input quality is high (after rigorous verification), the requirements for model prediction accuracy are increased simultaneously. This not only ensures the prediction accuracy under high input quality, but also avoids the risk of continuous deviation after rigorous verification, thus improving the rationality and pertinence of deviation judgment.

[0037] Subsequently, the actual value of the model prediction deviation judgment threshold currently in effect in the online adaptive correction module is obtained in real time. This actual value is not a fixed value, but is dynamically matched with the real-time correction result of the feature vector verification threshold. The feature vector verification threshold includes two core indicators: the feature value missing rate verification threshold and the feature value process logic range verification threshold. Specifically, the lower the feature value missing rate verification threshold (the smaller the proportion of missing features allowed) and the narrower the fluctuation range of the feature value process logic range verification threshold (the stricter the reasonable range of feature values), the more stringent the feature vector verification threshold is, and the smaller the actual value of the corresponding model prediction deviation judgment threshold is. Through this real-time dynamic matching method, the deviation judgment standard can be adjusted in real time according to the changes in feature input quality, avoiding the misjudgment or missed judgment problem caused by a fixed judgment standard adapting to different input quality scenarios, and ensuring the real-time performance and adaptability of deviation judgment.

[0038] The dynamic adaptation and control process for deviation judgment parameters also includes a precise adaptation step for the deviation duration judgment threshold: First, based on the predefined association mapping rules mentioned above, and the actual value of the deviation judgment threshold predicted by the model, a baseline threshold for the duration of deviation judgment is precisely matched. To further improve the adaptation accuracy, a preset base threshold is set as the basis for level division. If the actual value of the deviation judgment threshold predicted by the model is less than or equal to the preset base threshold (i.e., the prediction accuracy requirement is high), the matched baseline threshold for the duration of deviation judgment is a short duration level, ensuring a rapid response to continuous deviation situations. If the actual value of the deviation judgment threshold predicted by the model is greater than the preset base threshold (i.e., the prediction accuracy requirement is relatively lenient), the matched baseline threshold for the duration of deviation judgment is a long duration level, avoiding accidental triggering of control by instantaneous deviations and balancing control sensitivity and stability.

[0039] Based on this, a condition correction coefficient is introduced to further optimize the benchmark threshold. This condition correction coefficient is precisely determined according to the real-time operating condition of the boiler. For example, under rated stable load conditions, the boiler operates smoothly, and the tolerance for deviation response can be appropriately increased, so the condition correction coefficient is set to 1.2. However, under variable load and low load conditions, the boiler operating condition fluctuates greatly, and a rapid response to deviation is required to avoid combustion instability, so the condition correction coefficient is set to 0.8. Through this differentiated coefficient setting, the deviation judgment duration is adapted to the real-time operating status of the boiler. Subsequently, based on the benchmark threshold for the duration of deviation judgment, the actual threshold for the duration of deviation judgment is obtained by combining the condition correction coefficient. At the same time, to avoid the actual threshold from exceeding a reasonable range, it is strictly constrained between the preset shortest duration threshold and the longest duration threshold. This constraint mechanism can effectively avoid the problem of the actual threshold being too long (missed continuous deviation judgment) or too short (false triggering of control) due to extreme fluctuations in operating conditions or abnormal coefficients, ensuring the rationality and effectiveness of the duration threshold.

[0040] The entire deviation judgment parameter dynamic adaptation and control process deeply binds the deviation judgment standard with the feature input quality and the real-time operating conditions of the boiler, realizing full-process dynamic adaptation of the deviation judgment parameters. It can accurately identify the continuous deviation scenarios that truly need to be controlled, and effectively avoid the problem of mis-control caused by factors such as instantaneous deviation, input quality fluctuations, and changes in operating conditions. This ensures the accuracy and reliability of the deviation judgment predicted by the basic model, and provides a precise trigger basis for the subsequent online adaptive correction of the model. In this way, it ensures that the boiler combustion atmosphere parameters are always within a reasonable range, improving the efficiency, stability and safety of boiler combustion.

[0041] As the fourth module of this system, the incremental model update module is used to collect new time series process data from the previous round of model update to form an incremental dataset. Based on the incremental dataset and using incremental learning, the basic model is structurally fine-tuned to generate an updated model version, which replaces the original basic model and completes the iterative upgrade of the model.

[0042] In this embodiment, the incremental model update module first performs a targeted acquisition operation of new time series data, collecting only the completely new time series data generated after the previous model update. This targeted acquisition method effectively avoids the introduction of redundant historical data, significantly reducing the computational cost and storage overhead of data processing, while ensuring the timeliness and relevance of the collected data. An incremental dataset is then constructed based on this data. Subsequently, the module starts the incremental learning process based on the incremental dataset. Unlike the traditional full retraining method, incremental learning only performs targeted structural fine-tuning on the network layers and parameters in the basic model that are related to the features of the new time series data, without retraining the entire model. This effectively preserves the historical data features and model information already learned by the basic model. This approach not only improves model performance but also enables the model to quickly learn new features and patterns in new time series data, avoiding catastrophic forgetting issues and significantly improving model update efficiency while reducing computational consumption. After targeted structural fine-tuning, the module generates an updated model version adapted to the new data features. Subsequently, the original base model is replaced with this updated version through a model replacement mechanism, completing the iterative upgrade of the model. Through this entire set of incremental update steps, the model can continuously adapt to the dynamic changes in time series data, achieving gradual optimization of model performance and continuous improvement of capabilities. At the same time, it significantly shortens the model iteration and upgrade cycle, reduces the maintenance cost throughout the model's lifecycle, and ensures that the model always possesses accurate and efficient prediction and analysis capabilities in time series analysis scenarios.

[0043] The various features and processes described above can be used independently of each other or can be combined in various ways. All possible combinations and sub-combinations are intended to fall within the scope of this disclosure. Furthermore, certain method or process blocks may be omitted in some embodiments. The methods and processes described herein are not limited to any particular order, and the blocks or states associated with them may be performed in other suitable orders. For example, the described blocks or states may be performed in an order different from the order specifically disclosed, or multiple blocks or states may be combined in a single block or state. Example blocks or states may be performed serially, in parallel, or in some other manner. Blocks or states may be added to or removed from the disclosed example embodiments. The exemplary systems and components described herein may be configured differently from those described. For example, elements may be added to, removed from, or rearranged compared to the disclosed example embodiments.

[0044] The various operations of the example methods described herein can be performed at least in part by an algorithm. This algorithm can be contained in program code or instructions stored in memory (e.g., the aforementioned non-transitory computer-readable storage medium). Such an algorithm may include a machine learning algorithm. In some embodiments, the machine learning algorithm may not be explicitly programmed into the computer to perform the function, but can learn from training data to create a predictive model that performs the function.

[0045] The various operations of the example methods described herein can be performed, at least in part, by one or more processors that are temporarily configured (e.g., by software) or permanently configured to perform the relevant operations. Whether temporarily or permanently configured, such processors can constitute the engine of a processor implementation that operates to perform one or more of the operations or functions described herein.

[0046] Similarly, the methods described herein can be implemented at least in part by a processor, where one or more specific processors are examples of hardware. For example, at least some operations of a method can be performed by one or more processors or an engine implemented by a processor. Furthermore, one or more processors can also be operated to support the performance of related operations in a “cloud computing” environment or as “Software as a Service” (SaaS). For example, at least some operations can be performed by a set of computers (as an example of a machine including processors), where these operations are accessible via a network (e.g., the Internet) and via one or more suitable interfaces (e.g., application programming interfaces (APIs)).

[0047] The performance of certain operations can be distributed across processors, residing not only within a single machine but also deployed across multiple machines. In some example embodiments, the processor or processor-implemented engine may reside in a single geographic location (e.g., within a home environment, office environment, or server cluster). In other example embodiments, the processor or processor-implemented engine may be distributed across multiple geographic locations.

[0048] In this specification, multiple instances may implement components, operations, or structures described as single instances. Although individual operations of one or more methods are shown and described as separate operations, one or more of the separate operations may be performed simultaneously and do not need to be performed in the order shown. Structures and functions presented as separate components in the example configuration may be implemented as composite structures or components. Similarly, structures and functions presented as single components may be implemented as separate components. These and other variations, modifications, additions, and improvements fall within the scope of this document.

[0049] While an overview of the subject matter has been described with reference to specific example embodiments, various modifications and changes can be made to these embodiments without departing from the broader scope of embodiments of this disclosure. Such embodiments of the subject matter are referred to herein, individually or collectively, by the term "invention," and are used for convenience only and are not intended to limit the scope of this application to any single disclosure or concept, should more than one disclosure or concept be disclosed in fact.

[0050] The embodiments described herein have been described in sufficient detail to enable those skilled in the art to practice the disclosed teachings. Other embodiments may be used and derived therefrom, such that structural and logical substitutions and changes may be made without departing from the scope of this disclosure. Therefore, the detailed description should not be construed as limiting, and the scope of the various embodiments is defined only by the appended claims and the full scope of their equivalents.

Claims

1. A dynamic modeling and analysis system for combustion atmosphere data in a biomass oxygen-enriched combustion boiler, characterized in that, It includes a working condition adaptive feature extraction module, a time-preserving model training module, an online adaptive correction module, and an incremental model update module: The adaptive feature extraction module is used to collect and process time-series data in real time to generate a standard time-series dataset. Based on a combination of the boiler's real-time load rate and load change rate, it dynamically determines the current operating condition category. According to different operating condition categories, it calls a predefined feature extraction strategy and performs feature calculation on the standard time-series dataset based on the selected feature extraction strategy to generate a feature vector that dynamically adapts to the current operating condition, which serves as the model input. The time-series data includes, but is not limited to, fuel supply parameters, air distribution parameters, oxygen enrichment concentration parameters, boiler load parameters, and combustion atmosphere parameters measured by a flue gas analyzer. The time-preserving model training module is used to receive feature vectors from the working condition adaptive feature extraction module and optimize the initial dynamic prediction model to obtain the basic model. The online adaptive correction module is used to receive the current time-series process data in real time based on the base model. After being processed by the working condition adaptive feature extraction module, a real-time feature vector is generated. Based on the real-time feature vector, it is determined whether to perform online adaptive correction optimization. If yes, incremental model update is performed after correction optimization; otherwise, incremental model update is performed directly. The incremental model update module is used to collect new time-series process data from the previous round of model update to form an incremental dataset. Based on the incremental dataset and using incremental learning, the basic model is structurally fine-tuned to generate an updated model version, which replaces the original basic model and completes the iterative upgrade of the model.

2. The dynamic modeling and analysis system for combustion atmosphere data of a biomass oxygen-enriched combustion boiler as described in claim 1, characterized in that, The time series data is preprocessed with timestamp alignment, outlier handling, and filtering smoothing to generate a standard time series dataset. The operating condition categories include steady-state operating conditions, gradual-change operating conditions, and transient operating conditions.

3. The dynamic modeling and analysis system for combustion atmosphere data of a biomass oxygen-enriched combustion boiler as described in claim 1, characterized in that, The specific steps for dynamically determining the current operating condition category are as follows: Step 101: Obtain the load measurement value of the boiler in real time, and obtain the real-time load rate and the real-time load change rate. The real-time load rate is the percentage value of the current load measurement value to the rated load of the boiler. The real-time load change rate is the change in load rate per unit time, which is obtained by differential calculation of the load rate of the continuous time series and then processed by low-pass filtering to eliminate the interference of measurement noise. Step 102: If the real-time load rate is less than the load rate threshold and the real-time load change rate is less than the change rate threshold, then it is determined to be a steady-state condition. If the real-time load rate is less than the load rate threshold and the real-time load change rate is greater than the change rate threshold, it is determined to be a gradual change condition. If the real-time load rate is greater than the load rate threshold and the real-time load change rate is less than the change rate threshold, it is determined to be a gradual change condition. If the real-time load rate is greater than the load rate threshold and the real-time load change rate is greater than the change rate threshold, it is determined to be a transient operating condition.

4. The dynamic modeling and analysis system for combustion atmosphere data of a biomass oxygen-enriched combustion boiler as described in claim 1, characterized in that, The specific process of invoking the predefined feature extraction strategy according to different working condition categories is as follows: Step 201: Select the received feature extraction strategy corresponding to the current working condition category from the predefined feature extraction strategy, and determine its temporal window length and sliding step size; Step 202: Based on the current sampling time, extract a data sequence with a length equal to the length of the time series window from the standard time series dataset along the time axis to form the current analysis window; Step 203: During continuous operation, repeat the interception operation from step 201 to step 202 at each sliding step interval to form a series of analysis window sequences that are continuous in time. Step 204: Perform time-frequency feature calculations in parallel for each process variable data sequence contained in each analysis window.

5. The dynamic modeling and analysis system for combustion atmosphere data of a biomass oxygen-enriched combustion boiler as described in claim 4, characterized in that, The step of invoking a predefined feature extraction strategy according to different working condition categories also includes: Step 205: For a single analysis window, arrange the multiple feature values ​​calculated according to the time-frequency characteristics of each process variable in a predetermined order to form the feature sub-vector of the process variable. The multiple feature values ​​include, but are not limited to, the mean and standard deviation. Step 206: Concatenate the feature vectors of all process variables according to the predefined variable order to form a high-dimensional one-dimensional feature vector; Step 207: Each feature vector generated in a predefined time order is bound to the center time of its corresponding analysis window to form a feature vector with a timestamp. Step 208: Continuously output feature vectors with timestamps to form a time sequence of feature vectors after working condition adaptive feature extraction; Step 209: The time sequence of the feature vector is fed to the time-preserving model training module or the online adaptive correction module as input to the dynamic prediction model.

6. The dynamic modeling and analysis system for combustion atmosphere data of a biomass oxygen-enriched combustion boiler as described in claim 5, characterized in that, The specific process for obtaining the basic model is as follows: The time sequence of the feature vector is divided into training time interval, validation time interval and test time interval on the time axis to ensure that the timestamps of the training data are all earlier than the validation data, and the timestamps of the validation data are all earlier than the test data, thereby maintaining the causal time sequence relationship of the data. The generalization ability of the trained dynamic prediction model is evaluated using data from the training and validation time intervals. Based on the evaluation results, the number of training rounds is adaptively adjusted according to the training set cutoff timestamp to dynamically optimize the number of model training iterations and dynamically correct the feature vector verification threshold driven by the neuron dimension, ensuring the effectiveness of feature input and training accuracy of the model under different operating conditions.

7. The dynamic modeling and analysis system for combustion atmosphere data of a biomass oxygen-enriched combustion boiler as described in claim 6, characterized in that, The specific process for adaptively adjusting the number of training rounds based on the training set deadline timestamp is as follows: The predefined association mapping relationship between the training set end timestamp and the benchmark training rounds is divided into multiple time intervals according to the training set time span. Each time interval corresponds to a benchmark training round. The training set time span is the time difference between the start timestamp and the end timestamp of the training set. The larger the time span, the more benchmark training rounds are matched. The actual cutoff timestamp of the training set in the time-preserving model training module is obtained in real time, the actual time span of the training set is calculated, and the corresponding baseline training round number is matched according to the association mapping relationship. A training set data volume correction coefficient is introduced. The data volume correction coefficient is determined based on the ratio of the effective feature vector sample size of the training set corresponding to the actual cutoff timestamp to a preset sample size threshold. When the sample size exceeds the threshold, the correction coefficient is greater than 1, and when the sample size is less than the threshold, the correction coefficient is less than 1. Based on the baseline number of training rounds, the actual number of training rounds of the model is calculated by combining the data volume correction coefficient. The actual number of training rounds is the result of combining the baseline number of training rounds and the data volume correction coefficient. The actual number of training rounds is constrained between a preset minimum training rounds threshold and a maximum training rounds threshold. Finally, the initial dynamic prediction model is optimized and trained using the actual number of training rounds.

8. The dynamic modeling and analysis system for combustion atmosphere data of a biomass oxygen-enriched combustion boiler as described in claim 6, characterized in that, The specific steps for dynamically correcting the feature vector verification threshold driven by the neuron dimension are as follows: A predefined correlation correction rule between the number of input layer neurons and the feature vector verification threshold is established. The number of input layer neurons corresponds one-to-one with the feature vector dimension. The more neurons there are, the higher the corresponding feature vector dimension, the lower the matching feature value missing rate verification threshold, and the narrower the fluctuation range of the feature vector verification threshold. The actual number of neurons in the input layer of the initial dynamic prediction model in the time-preserving model training module is obtained in real time, and the corresponding threshold correction coefficient is obtained by matching according to the association correction rule. The threshold correction coefficient includes the missing rate correction coefficient and the range correction coefficient. The preset feature value basic missing rate verification threshold is corrected based on the missing rate correction coefficient to obtain the actual missing rate verification threshold. The actual missing rate verification threshold is the result of combining the feature value basic missing rate verification threshold and the missing rate correction coefficient, and the actual missing rate verification threshold is not higher than the preset maximum missing rate threshold. The upper and lower limits of the preset feature value basic missing rate verification threshold are corrected based on the range correction coefficient to obtain the actual verification threshold. The actual upper limit threshold is the result of combining the complementary coefficient of the basic upper limit threshold and the range correction coefficient, and the actual lower limit threshold is the result of combining the basic lower limit threshold with the superposition result of 1 and the range correction coefficient. The actual missing rate verification threshold and the actual upper limit threshold are used as the final verification thresholds for the feature vectors to verify the validity of the received feature vectors.

9. The dynamic modeling and analysis system for combustion atmosphere data of a biomass oxygen-enriched combustion boiler as described in claim 1, characterized in that, The specific steps for determining whether to perform online adaptive correction optimization based on real-time feature vectors are as follows: The real-time feature vector is input into the basic model to obtain the real-time predicted values ​​of the combustion atmosphere parameters; The real-time predicted value is compared with the actual value measured by the flue gas analyzer to calculate the real-time prediction deviation. When the real-time prediction deviation continuously exceeds the preset threshold, dynamic adaptation and adjustment of the deviation judgment parameter is triggered. The specific steps for dynamically adapting and adjusting the deviation judgment parameters are as follows: A predefined association mapping rule is established between the model prediction deviation judgment threshold and the deviation duration judgment threshold. The model prediction deviation judgment threshold is the maximum allowable percentage deviation between the predicted value output by the basic model based on the real-time feature vector and the actual field-collected value. The deviation duration judgment threshold is the maximum allowable duration for which the model prediction deviation exceeds the model prediction deviation judgment threshold. The value of the model prediction deviation judgment threshold is positively correlated with the feature vector verification threshold. The more stringent the feature vector verification threshold, the smaller the value of the model prediction deviation judgment threshold, and the shorter the corresponding deviation duration judgment threshold. The actual value of the currently effective model prediction deviation judgment threshold in the online adaptive correction module is obtained in real time. The actual value is dynamically matched by the real-time correction result of the feature vector verification threshold. The feature vector verification threshold includes the feature value missing rate verification threshold and the feature value process logic range verification threshold. The lower the feature value missing rate verification threshold and the narrower the fluctuation range of the feature value process logic range verification threshold, the more stringent the feature vector verification threshold, and the smaller the actual value of the matched model prediction deviation judgment threshold.

10. The dynamic modeling and analysis system for combustion atmosphere data of a biomass oxygen-enriched combustion boiler as described in claim 9, characterized in that, The dynamic adaptation and adjustment of the deviation judgment parameters also includes: Based on the association mapping rule, the baseline threshold for the duration of deviation determination is obtained by matching the actual value of the model prediction deviation determination threshold with the actual value of the threshold. If the actual value of the model prediction deviation judgment threshold is less than or equal to the preset base threshold, the matching deviation duration judgment benchmark threshold is the short duration level. If the actual value of the model prediction deviation judgment threshold is greater than the preset base threshold, the matching deviation duration judgment benchmark threshold is the long duration level. An operating condition correction coefficient is introduced, which is determined according to the real-time operating condition category of the boiler. Based on the baseline threshold for the duration of deviation determination, the actual threshold for the duration of deviation determination is calculated by combining the operating condition correction coefficient. The actual threshold for the duration of deviation determination is the result of combining the baseline threshold for the duration of deviation determination and the operating condition correction coefficient, and the actual threshold for the duration of deviation determination is constrained between the preset shortest duration threshold and the longest duration threshold.

Images