Biomass power generation combustion parameter deep learning method and system

Abstract

The invention relates to the field of power equipment data processing, in particular to a biomass power generation combustion parameter deep learning method and system, and aims to solve the problem of phase mismatch caused by sampling frequency difference and clock reference offset of multi-source heterogeneous time sequence data. A parallel multi-scale convolution and bidirectional long-short-term memory network hybrid model is constructed, transient fluctuation and long-period trend features are extracted, and combustion stage feature weights are dynamically distributed through a gating attention mechanism. The optimization control module generates a multi-target constraint condition, an operation instruction is output in combination with a fuzzy inference engine, and a digital twin platform simulates an extreme working condition to enhance model robustness. A closed-loop feedback mechanism dynamically adjusts model parameters through combustion efficiency monitoring data and simulation results, and a two-stage fault-tolerant strategy realizes sensor abnormity compensation and historical control strategy backtracking. The problem of asynchronous data stream feature misalignment is effectively solved, and the combustion efficiency prediction precision and the control decision reliability are improved.

Description

Technical Field

[0001] The present invention relates to the field of power equipment data processing, and in particular to a method and system for deep learning of biomass power generation combustion parameters. Background Art

[0002] In the scenario of biomass power generation in new energy power plants, dynamic optimization of combustion parameters is a key link to improve energy conversion efficiency and reduce pollutant emissions. The diversity of components and calorific value fluctuations of biomass fuels lead to significant nonlinear and strongly coupled characteristics of the combustion process, and traditional control strategies based on empirical formulas or linear regression are difficult to accurately model. Deep learning constructs a multi-layer convolutional neural network and a long-short-term memory network fusion model to extract features and perform correlation analysis on multi-source time series data such as combustion temperature, flue gas oxygen content, feed rate, furnace pressure, etc., and can independently mine the high-order nonlinear mapping relationship between input parameters and combustion efficiency and nitrogen oxide generation. By collecting sensor data in real time and inputting a well-trained model, the system can dynamically adjust operating variables such as blast volume and feed ratio to achieve adaptive control of combustion conditions, effectively balance the impact of fuel calorific value fluctuations on steam parameter stability, and continuously optimize the objective function through the gradient back propagation algorithm to maximize power generation efficiency while meeting environmental emission constraints.

[0003] In the scenario of biomass power generation in new energy power plants, deep learning of combustion parameters faces the problem of time synchronization and feature alignment of multi-source heterogeneous time series data. During the biomass combustion process, the real-time monitoring data generated by devices such as temperature sensors, gas analyzers, and pressure transmitters have sampling frequency differences and clock reference offsets, resulting in phase mismatches in the time series of parameters such as combustion temperature, flue gas composition, and furnace pressure at the millisecond scale. This type of time series misalignment will distort the dynamic coupling relationship between combustion operating parameters, especially in the scenario of sudden changes in fuel calorific value or rapid load adjustment. When asynchronous data flows through the convolutional neural network to extract spatial features, the pseudo-correlation signal caused by the device communication delay is misjudged as the real correlation of the combustion state, thereby reducing the prediction accuracy of the long short-term memory network for combustion efficiency and pollutant generation. Existing data preprocessing methods are difficult to completely eliminate the time domain aliasing effect between high-frequency acquisition equipment and low-frequency control signals, resulting in damage to the causal representation of input data during model training, which directly affects the decision-making reliability of the combustion control strategy. Summary of the invention

[0004] In view of the shortcomings of the prior art, the present invention provides a deep learning method and system for biomass power generation combustion parameters. The present invention addresses the phase mismatch problem of multi-source heterogeneous time series data caused by sampling frequency differences and clock reference offsets in the biomass combustion process, and proposes to eliminate the time domain aliasing effect through a dynamic feature alignment mechanism to improve the combustion efficiency prediction model's ability to characterize the dynamic coupling relationship between parameters and enhance the decision-making reliability of the combustion control strategy.

[0005] To solve the above technical problems, the specific technical solutions of the present invention are as follows: In a first aspect, the deep learning method for biomass power generation combustion parameters provided by the present invention includes: Obtain multi-source time-series data of the biomass combustion system, where the multi-source time-series data includes high-frequency sampling data and low-frequency control signals; Perform clock synchronization processing on the multi-source time-series data, align the time-series information of different-frequency data through a dynamic time compensation mechanism, and generate synchronized time-series data; Based on the synchronized time-series data, perform phase correction, and use a dynamic path planning algorithm to match the time-series correlation characteristics between multiple parameters to generate phase-aligned data; Construct a multi-modal time-series analysis model, where the multi-modal time-series analysis model includes a convolution module for extracting local fluctuation features and a recurrent neural network module for capturing long-period dependence relationships; Input the phase-aligned data into the multi-modal time-series analysis model to obtain combustion state prediction parameters; Based on the combustion state prediction parameters, construct a multi-objective optimization function to generate constraint conditions for combustion control parameters; According to the constraint conditions, generate an operation instruction set through a fuzzy inference engine, where the fuzzy inference engine is based on a fuzzy control rule base, and send the operation instruction set to the actuator, and at the same time, collect the combustion system feedback parameters in real time to form a closed-loop control; Deploy a simulation verification platform to simulate abnormal working conditions to generate a training data set, perform parameter iterative update on the multi-modal time-series analysis model according to the feedback parameters, and compensate for abnormal sensor data through a two-level fault tolerance strategy.

[0006] Further, in the deep learning method for biomass power generation combustion parameters of the present invention, the clock synchronization processing includes: Attach a unified time reference to the high-frequency sampling data and low-frequency control signals through the IEEE 1588 protocol; During the sliding window caching process, superimpose an exponentially weighted moving average algorithm on the high-frequency sampling data to compensate for the clock offset between devices, and constrain the interpolation mutation amplitude based on the physical characteristics of the fuel flow; When performing cubic spline interpolation resampling on the low-frequency control signal, eliminate non-physical jump points in combination with the continuity requirement of the combustion system pressure change.

[0007] Further, in the deep learning method for biomass power generation combustion parameters of the present invention, the phase matching includes: Based on the synchronized timing data, calculate the optimal alignment path of the combustion temperature and flue gas oxygen content sequences through the dynamic time warping algorithm, and constrain the path search range with their correlation. Stack a residual connection branch on the data channel after completing the time-domain translation correction, and fuse the original timing data with the corrected features through a 1×1 convolutional layer. According to the time-frequency characteristics of the sudden change signal of the furnace pressure, dynamically adjust the fusion weight of the original data and the corrected features in the residual branch.

[0008] Furthermore, for the deep learning method for biomass power generation combustion parameters of the present invention, the construction of the hybrid model includes: In the parallel multi-scale one-dimensional convolutional structure, the first convolutional layer extracts the transient fluctuation characteristics of the furnace pressure with a stride of 3, and the second convolutional layer extracts the slow change trend characteristics of the flue gas oxygen content with a stride of 15. At the output end of the time series features of the bidirectional long short-term memory network, activate the corresponding gated attention weights according to the combustion stage recognition results, and the combustion stages include ignition, steady state, and load adjustment. After splicing the output features of the first convolutional layer and the second convolutional layer, dynamically adjust the fusion ratio of features at different scales through an adaptive normalization layer.

[0009] Furthermore, for the deep learning method for biomass power generation combustion parameters of the present invention, the multi-objective optimization function includes: Based on the combustion state prediction parameters, construct a Pareto front optimization model for combustion efficiency and nitrogen oxide emissions. According to the load status signal output by the power plant operation mode recognition module, dynamically adjust the weight allocation ratio of the efficiency and emission targets. Introduce a historical working condition similarity screening mechanism in the constraint condition generator, and when the similarity between the prediction result and the historical data is lower than the set threshold, limit the search space of the control strategy.

[0010] Furthermore, for the deep learning method for biomass power generation combustion parameters of the present invention, the construction of the fuzzy control rule base includes: Define the membership function of the air volume deviation based on the expert knowledge base, where the partition threshold of the steam flow rate change rate is dynamically adjusted according to the combustion stage. Optimize the parameters of the membership function through the genetic algorithm to generate the correlation matrix of the premise and conclusion of the fuzzy rules. During the operation instruction set generation stage, superimpose the inertia delay model of the actuator to pre-compensate the control quantity.

[0011] Furthermore, for the deep learning method for biomass power generation combustion parameters of the present invention, the two-level fault tolerance mechanism includes: When a single sensor data interruption is detected, adjacent sensor data is fused based on the unscented Kalman filter algorithm to generate a compensation value; When the multi-parameter time series misalignment exceeds a preset threshold, the control strategy with the minimum dynamic time warping distance is matched from the historical similar working condition library; During the clock synchronization failure period, the current timestamp is compensated for deviation according to the historical synchronization error mean value.

[0012] Furthermore, in the deep learning method for biomass power generation combustion parameters of the present invention, the digital twin platform includes: A fuel calorific value fluctuation simulation module that generates a disturbance signal conforming to the biomass fuel component distribution law through a random generator; A sensor noise injection module that superimposes communication delays and measurement errors conforming to the characteristics of the industrial site in the simulation data stream; A model adaptation module that dynamically adjusts the regularization constraint strength of the multi-modal time series analysis model according to the prediction error statistical characteristics of the enhanced training data set.

[0013] Furthermore, in the deep learning method for biomass power generation combustion parameters of the present invention, the closed-loop feedback control includes: The operation instruction set is transmitted to the control interface of the blower frequency converter and the PLC controller of the feeder through the OPC protocol; The moving average of the furnace pressure gradient change rate is calculated in real time as the control loop stability criterion; When the deviation between the steam flow monitoring value and the set threshold continuously exceeds the time window, the historical control strategy backtracking loading mechanism is triggered.

[0014] In a second aspect, the present invention provides a deep learning system for biomass power generation combustion parameters, which is applied to the deep learning method for biomass power generation combustion parameters, and includes: A data acquisition module, connected to a temperature sensor, a flue gas composition analyzer, and a pressure transmitter, to obtain high-frequency sampling data and low-frequency control signals; A clock synchronization module, with its input end connected to the data acquisition module, built-in with a sliding window buffer unit and a cubic spline interpolation unit, and outputs synchronized timing data; A phase correction module, with its input end receiving the synchronized timing data, integrating the dynamic time warping algorithm and a residual connection branch, and outputs phase-aligned data; A hybrid modeling module, with its input end connected to the phase correction module, includes a parallel multi-scale one-dimensional convolutional network and a bidirectional long short-term memory network. The convolutional network extracts transient fluctuation features and slow-varying trend features, and the recurrent neural network outputs long-period time series dependence relationships; The optimization control module receives the prediction results of the hybrid modeling module at its input end, and has a built-in multi-objective optimization function generator and a fuzzy inference engine to generate constraint conditions for the air volume and the feeding ratio. The fault-tolerant execution module is connected to the optimization control module and the data acquisition module at its input end, deploys a digital twin platform to simulate extreme working conditions, and integrates two-level fault-tolerant compensation units. The feedback learning module receives the combustion efficiency monitoring data and the simulation results of the digital twin platform at its input end, and is connected to the hybrid modeling module at its output end to dynamically adjust the model parameters.

[0015] Advantages of the present invention The advantages of the present invention are reflected in improving the combustion control accuracy through the collaborative mechanism of multi-level time series processing and intelligent modeling: using the IEEE 1588 protocol and the dynamic time compensation algorithm to eliminate the clock offset between devices, and combining the interpolation resampling technology with combustion physical constraints to solve the time domain aliasing problem of multi-frequency data and generate high-quality inputs with consistent time series; correcting the phase deviation caused by communication delay through the dynamic time warping algorithm, and the residual feature fusion mechanism retains the true correlation of transient signals, effectively restoring the dynamic coupling relationship of multi-source data; the parallel multi-scale convolution and gated attention mechanism of the hybrid modeling module cooperate to extract local fluctuation and long-period trend features, and combine the adaptive normalization layer to dynamically optimize the feature weight distribution, enhancing the model's ability to represent the non-linear relationship of combustion conditions; the closed-loop feedback control and the fault-tolerant strategy driven by digital twin form a self-optimizing system, maintaining control stability through historical strategy matching and parameter dynamic adjustment under abnormal working conditions, and finally realizing the dual optimization of combustion efficiency and environmental protection emission indicators. Description of the drawings

[0016] In order to more clearly illustrate the technical solutions of the present invention, the drawings required for the embodiments will be briefly introduced below. Obviously, for those of ordinary skill in the art, other drawings can be obtained according to the drawings without creative efforts.

[0017] Figure 1 It is a flowchart of the deep learning method for biomass power generation combustion parameters provided by the embodiment of the present invention. Specific implementation manners

[0018] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with specific embodiments of the present invention and the corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts belong to the scope of protection of the present invention. The following will detail the technical solutions provided by each embodiment of the present invention with reference to the drawings. To better understand the objectives of the present invention, the present invention will be further described in detail below.

[0019] In a first aspect, please refer to Figure 1 , the deep learning method for biomass power generation combustion parameters provided by the present invention includes: Step S101: Obtain multi-source time-series data of the biomass combustion system, where the multi-source time-series data includes high-frequency sampling data and low-frequency control signals; Step S102: Perform clock synchronization processing on the multi-source time-series data, align the time-series information of different-frequency data through a dynamic time compensation mechanism, and generate synchronized time-series data; Step S103: Perform phase correction based on the synchronized time-series data, use a dynamic path planning algorithm to match the time-series correlation features between multiple parameters, and generate phase-aligned data; Step S104: Construct a multi-modal time-series analysis model, where the multi-modal time-series analysis model includes a convolutional module for extracting local fluctuation features and a recurrent neural network module for capturing long-term dependency relationships; Step S105: Input the phase-aligned data into the multi-modal time-series analysis model to obtain combustion state prediction parameters; Step S106: Construct a multi-objective optimization function based on the combustion state prediction parameters to generate constraint conditions for combustion control parameters; Step S107: Generate an operation instruction set through a fuzzy inference engine according to the constraint conditions, where the fuzzy inference engine is based on a fuzzy control rule base, and send the operation instruction set to the actuator, and at the same time, collect the feedback parameters of the combustion system in real time to form a closed-loop control; Step S108: Deploy a simulation verification platform to simulate abnormal working conditions to generate a training data set, perform parameter iterative update on the multi-modal time-series analysis model according to the feedback parameters, and compensate for abnormal sensor data through a two-level fault tolerance strategy.

[0020] The deep learning method for biomass power generation combustion parameters provided by the present invention realizes the optimal control of the combustion process through the following steps: When obtaining multi-source time-series data of a biomass combustion system, high-frequency sampling data and low-frequency control signals are collected in real time by connecting temperature sensors, flue gas component analyzers, and pressure transmitters. The high-frequency sampling data includes millisecond-level temperature sequences and pressure fluctuation signals, and the low-frequency control signals cover minute-level feed rate and air volume adjustment commands. The acquisition of multi-source data provides the original input for subsequent time-series alignment and feature modeling.

[0021] When performing clock synchronization processing on multi-source time-series data, a unified time reference is added to the high-frequency data and low-frequency signals based on the IEEE 1588 protocol to eliminate the clock reference offset between devices. For high-frequency sampling data, a sliding window mechanism is used to segment and cache the data stream, and the exponential weighted moving average algorithm is used to dynamically compensate for the clock offset, combined with the physical continuity constraint of the fuel flow to interpolate the mutation amplitude during the process; for low-frequency control signals, cubic spline interpolation is used for time-domain resampling, and the continuity check of the pressure change in the combustion system is introduced during the interpolation process to filter out non-physical jump points, and finally synchronized time-series data is generated.

[0022] When performing phase correction based on the synchronized time-series data, the dynamic time warping algorithm is used to match the time-series correlation features of the combustion temperature and flue gas oxygen content sequences. Adjacent time interval data segments are extracted through a sliding window, and the correlation constraint between temperature and oxygen content is combined to search the range of the path, and local time offsets caused by communication delays or uneven sampling intervals are identified. A residual connection branch is superimposed on the corrected data channel, and a 1×1 convolutional layer is used to fuse the original time-series data and the corrected features. According to the transient characteristics of the furnace pressure sudden change signal, the weight distribution ratio of the residual branch is dynamically adjusted to generate phase-aligned data to retain the true transient features.

[0023] When constructing a multi-modal time-series analysis model, a parallel multi-scale one-dimensional convolutional structure and a bidirectional long short-term memory network are designed. Among them, the first convolutional layer extracts the transient fluctuation features of the furnace pressure with a stride of 3, and the second convolutional layer captures the slow-changing trend features of the flue gas oxygen content with a stride of 15; the bidirectional long short-term memory network analyzes the lag effect of fuel calorific value fluctuations and pollutant generation, and a combustion stage recognition module is set at the output end to dynamically activate the corresponding gated attention weights according to the ignition, steady state, and load adjustment stages. The fusion ratio of multi-scale features is dynamically adjusted through an adaptive normalization layer to enhance the model's ability to represent time-series correlations.

[0024] After inputting the phase-aligned data into the multi-modal time-series analysis model, the model outputs the prediction parameters of the combustion efficiency and nitrogen oxide emissions. A Pareto front optimization model is constructed based on the prediction results to dynamically adjust the objective weights of efficiency maximization and emission minimization. A historical operating condition similarity screening mechanism is introduced into the constraint condition generator. When the similarity between the prediction results and historical data is lower than the threshold, the search space of the control strategy is restricted, and the boundary constraint conditions of the air volume and feed ratio are generated.

[0025] When generating an operation instruction set through a fuzzy inference engine according to the constraint conditions, a fuzzy membership function of the blast volume deviation and the steam flow rate change rate is defined based on the expert knowledge base, and a genetic algorithm is used to offline optimize the premise and conclusion correlation parameters of the fuzzy rules. In the instruction set generation stage, an inertia delay model of the actuator is superimposed to pre-compensate the response delay of the control quantity transmitted to the blower and the feeder, improving the instruction execution accuracy.

[0026] Deploy a simulation verification platform to simulate abnormal working conditions such as sudden changes in fuel calorific value and sensor failures, and generate an enhanced training data set to optimize the model robustness. Inject noise and communication delay signals conforming to the characteristics of the industrial site through the digital twin platform, and dynamically adjust the regularization constraint strength of the multi-modal time series analysis model. Real-time collect combustion efficiency and emission data and feedback them to the model to trigger parameter iterative updates, forming an adaptive learning mechanism.

[0027] In the closed-loop control process, the operation instructions are sent to the actuator through the OPC protocol, and the change rate of the furnace pressure gradient is monitored in real time as the stability criterion. When the monitored value of the steam flow deviates from the set threshold and the duration exceeds the preset window, trigger the backtracking loading mechanism of the historical control strategy, and switch to the optimized instruction set under similar working conditions to ensure the continuity and reliability of the system control.

[0028] Through the synergistic effect of the above steps, the system realizes the full-process technical chain from data synchronization, feature correction, dynamic modeling to closed-loop optimization, effectively solves the problem of decreased prediction accuracy caused by the time series mismatch of biomass combustion parameters, and improves the decision-making reliability of combustion efficiency and control strategies.

[0029] Specifically, for the deep learning method of biomass power generation combustion parameters described in the present invention, the clock synchronization process includes: Attach a unified time reference to the high-frequency sampling data and the low-frequency control signal through the IEEE 1588 protocol; During the sliding window caching process, superimpose an exponentially weighted moving average algorithm on the high-frequency sampling data to compensate for the clock offset between devices, and constrain the interpolation mutation amplitude based on the physical characteristics of the fuel flow rate; When performing cubic spline interpolation resampling on the low-frequency control signal, eliminate non-physical jump points in combination with the continuity requirement of the combustion system pressure change.

[0030] The clock synchronization process described in the present invention realizes the time series alignment of multi-source data through the following technical means: When establishing a unified time reference for high-frequency sampling data and low-frequency control signals through the IEEE 1588 protocol, a master clock node is deployed in the distributed data acquisition system, which periodically broadcasts synchronization messages to temperature sensors, pressure transmitters, and control signal generation units. Each slave device calculates the local clock offset based on the message transmission delay, attaches a microsecond-level timestamp to the original timing data, eliminates the initial clock deviation between devices, and provides a unified time scale system for subsequent multi-frequency data fusion.

[0031] For high-frequency sampling data, a sliding window mechanism is adopted to segment and cache the data stream according to the time series. The window length is dynamically adjusted according to the fluctuation period of the fuel calorific value. During the window sliding process, the exponential weighted moving average algorithm is applied to calculate the clock offset between adjacent windows, and dynamic compensation is achieved by recursively updating the weight coefficient. At the same time, based on the physical continuity law of the fuel flow rate, a flow rate change rate threshold is set during the interpolation process to limit the phenomenon of over-limit mutation caused by device response delay and maintain the physical rationality of the evolution of combustion process parameters.

[0032] When performing cubic spline interpolation resampling on low-frequency control signals, interpolation constraint conditions are constructed based on the continuity characteristics of the pressure change in the combustion system. A pressure gradient change threshold is introduced at the interpolation nodes. When the pressure change rate between adjacent sampling points exceeds the design parameters of the combustion chamber, the interpolation curve smoothing correction mechanism is automatically triggered. By forcing the second derivative of the interpolation function to be continuous, the non-physical jump points caused by the step change of the control command are eliminated, and the time-domain evolution trend of the low-frequency signal is ensured to conform to the combustion dynamics characteristics.

[0033] In the above processing process, the dynamic compensation of high-frequency data and the interpolation constraint of low-frequency signals form a synergistic effect. The time series caching of the sliding window provides a data buffer interval for the exponential weighted algorithm, and the physical constraint conditions of the interpolation nodes rely on the accurate time scale established by the unified time reference. For the data stream after synchronization processing, the time axis deviation is controlled within the time resolution required for combustion process feature extraction, effectively solving the problem of multi-frequency data time-domain aliasing and providing input data with time series consistency for subsequent phase correction.

[0034] Specifically, in the deep learning method for combustion parameters of biomass power generation described in the present invention, the phase matching includes: Based on the synchronized timing data, the optimal alignment path of the combustion temperature and flue gas oxygen content sequences is calculated through the dynamic time warping algorithm, and the path search range is constrained by their correlation; A residual connection branch is superimposed on the data channel after the time domain translation correction, and the original timing data is fused with the corrected features through a 1×1 convolutional layer; According to the time-frequency characteristics of the sudden change signal of the furnace pressure, the fusion weight of the original data and the corrected features in the residual branch is dynamically adjusted.

[0035] The phase matching in the present invention achieves precise alignment of timing features through the following technical means: When processing synchronous timing data based on the dynamic time warping algorithm, a sliding window is used to extract data segments from the combustion temperature and flue gas oxygen content sequences. By constructing a correlation coefficient matrix of combustion temperature and oxygen content, the physical correlation relationship between the two in the combustion reaction is quantified as a constraint condition for path search, restricting the alignment path to perform dynamic programming calculations within the region where the correlation coefficient is higher than the set value. After identifying the local time offset according to the optimal alignment path, the time domain translation correction is performed on the flue gas oxygen content sequence to eliminate the phase mismatch caused by the sensor sampling delay.

[0036] When superimposing a residual connection branch on the corrected data channel, the original timing data and the corrected features are concatenated at the channel level, and cross-channel feature fusion is achieved through a 1×1 convolutional layer. The convolutional kernel parameter sharing mechanism enables the fusion process to adaptively retain the high-frequency components of transient signals such as sudden pressure changes, while suppressing the noise interference introduced during the correction process. The introduction of the residual branch forms a dual-path feature parallel processing structure, avoiding the over-smoothing of the dynamic characteristics of the original data by a single correction path.

[0037] According to the time-frequency analysis results of the furnace pressure sudden change signal, wavelet transform is used to detect the mutation points and frequency band distribution characteristics of the pressure waveform. Based on the time distribution density of the mutation points, the fusion weight of the original data and the corrected features in the residual branch is dynamically adjusted. The weight allocation ratio of the original data is increased within the time interval when the pressure sudden change occurs, strengthening the ability to retain transient features. The weight adjustment parameters are continuously updated through an online learning module, and the weight allocation strategy is optimized in combination with the combustion stage identification results, enabling the fusion process to adapt to the signal characteristic changes under different working conditions.

[0038] In the above processing process, the path constraint of dynamic time warping and the weight adjustment of the residual branch form a collaborative mechanism. The restriction of the path search range ensures that the timing alignment conforms to the physical laws of the combustion process, and the residual fusion compensates for the feature loss of a single correction path. The weight dynamic adjustment module optimizes the fusion ratio through the feedback of the pressure sudden change characteristics, forming a closed-loop feature enhancement mechanism. After the phase matching process, the timing correlation and feature integrity of the data stream are significantly improved, providing high-confidence input data for subsequent hybrid modeling.

[0039] Specifically, for the deep learning method of biomass power generation combustion parameters described in the present invention, the construction of the hybrid model includes: In the parallel multi-scale one-dimensional convolutional structure, the first convolutional layer extracts the transient fluctuation characteristics of the furnace pressure with a stride of 3, and the second convolutional layer extracts the slow-changing trend characteristics of the flue gas oxygen content with a stride of 15; At the time-series feature output end of the bidirectional long short-term memory network, corresponding gated attention weights are activated according to the combustion stage recognition result, and the combustion stage includes ignition, steady state, and load adjustment; After splicing the output features of the first convolutional layer and the second convolutional layer, the fusion ratio of features at different scales is dynamically adjusted through an adaptive normalization layer.

[0040] The construction of the hybrid model according to the present invention realizes the effective extraction and fusion of multi-scale time-series features through the following technical means: In the parallel multi-scale one-dimensional convolutional structure, the first convolutional layer is set to use a sliding window with a stride of 3. Its narrow receptive field design adapts to the millisecond-level transient fluctuation characteristics of the furnace pressure, and captures the waveform details of sudden pressure changes through local feature focusing. The second convolutional layer uses a sliding window with a stride of 15. Its wide receptive field covers the minute-level change cycle of the oxygen content in the flue gas, and extracts the slow-changing trend characteristics of the oxygen concentration in the combustion reaction. The feature maps output by the dual-channel convolution respectively characterize the fast dynamic response and slow thermodynamic evolution characteristics of the combustion process, providing complementary spatial feature inputs for subsequent time-series modeling.

[0041] When the bidirectional long short-term memory network processes the time-series features of the convolution output, a combustion stage recognition module is embedded in the hidden state propagation path. This module constructs a combustion stage determination rule based on the furnace temperature gradient and the steam flow rate change rate, and identifies the rapid temperature rise interval in the ignition stage, the constant temperature interval in the steady-state operation, and the dynamic transition interval in the load adjustment stage. According to different combustion stages, corresponding gated attention weights are activated. In the ignition stage, the feature weight of the fuel supply rate is strengthened, in the steady state stage, the correlation weight between the oxygen concentration and the emission parameters is increased, and in the load adjustment stage, the time-series dependence weight of the pressure fluctuation feature is increased, realizing the dynamic allocation of feature importance.

[0042] After splicing the feature maps output by the dual-channel convolution and the time-series dependence features of the bidirectional long short-term memory network, they are input into the adaptive normalization layer. This normalization layer dynamically calculates the scaling factor and bias of features at different scales based on the variance distribution of the feature maps and the combustion stage labels. In the time interval with significant pressure transient fluctuations, the fusion ratio of the narrow convolution features is increased, and in the slow-changing stage of the oxygen concentration, the weight ratio of the wide convolution features is increased. The feature fusion strategy is optimized through an online learning mechanism, so that the multi-scale feature combination adapts to the dynamic changes of the combustion conditions.

[0043] During the above processing, the multi-scale convolution structure and the gated attention mechanism form a spatial-temporal feature collaborative extraction network. The differential design of narrow / wide convolution strides covers the time-varying characteristic range of combustion parameters, and the stage-dependent attention weight allocation enhances the feature representation ability under key operating conditions. The adaptive normalization layer balances the contribution degrees of multi-source features through a dynamic adjustment mechanism, avoiding the feature suppression problem caused by fixed-weight fusion. The hybrid model effectively improves the prediction accuracy of combustion efficiency and pollutant generation through multi-level feature interaction.

[0044] Specifically, for the deep learning method for biomass power generation combustion parameters described in the present invention, the multi-objective optimization function includes: Based on the combustion state prediction parameters, a Pareto front optimization model for combustion efficiency and nitrogen oxide emissions is constructed; According to the load status signal output by the power plant operation mode recognition module, the weight allocation ratio of efficiency and emission targets is dynamically adjusted; A historical operating condition similarity screening mechanism is introduced into the constraint condition generator. When the similarity between the prediction result and historical data is lower than the set threshold, the search space of the control strategy is restricted.

[0045] The multi-objective optimization function described in the present invention realizes the dynamic optimization of combustion control parameters through the following technical means: When constructing the Pareto front optimization model based on the combustion state prediction parameters, maximizing combustion efficiency and minimizing nitrogen oxide emissions are used as the dual-objective functions, and the non-dominated sorting algorithm is used to generate the optimal solution set in the multi-dimensional objective space. By analyzing the correlation relationships among combustion temperature, flue gas oxygen content, and fuel calorific value, coupling constraint conditions between the objective functions are established, and feasible solutions that meet the steam flow stability are screened from the solution set to form a dynamically optimized strategy space.

[0046] According to the load status signal output by the power plant operation mode recognition module, a dynamic weight allocation strategy is designed. During the load climbing stage, the weight coefficient of the nitrogen oxide emission target is increased to give priority to meeting environmental protection constraints; during the steady-state operation stage, the weight ratio of combustion efficiency is increased to optimize the calorific value utilization rate; during the load sudden drop stage, the dual-objective weights are balanced by combining historical operating condition data. The weight adjustment parameters are iteratively updated through the online learning module in combination with the real-time combustion efficiency feedback, so that the optimization strategy adapts to the changes in the power plant operation state.

[0047] When introducing a historical operating condition similarity screening mechanism in the constraint generator, the dynamic time warping distance is used to calculate the similarity between the current prediction result and the historical database. When the similarity is lower than the preset threshold, the current operating condition is determined to be abnormal or unrecorded, and the optimization space limitation strategy is triggered: narrowing the blast volume adjustment range to a safe interval, locking the fluctuation range of the feed ratio, and enabling the standby control strategy library. The similarity threshold is dynamically adjusted according to the combustion stage, with a lower threshold set in the ignition stage to improve fault tolerance and a higher threshold in the steady state stage to enhance optimization accuracy.

[0048] In the above processing process, the Pareto front model provides the theoretical optimal solution set, the dynamic weight allocation realizes the adjustment of the target priority, and the historical similarity screening ensures the feasibility and safety of the control strategy. The real-time adjustment of the weight coefficient affects the selection tendency of the Pareto solution, and the change of the similarity threshold further restricts the boundary conditions of the solution space. The three work together to form a closed-loop optimization mechanism, which dynamically approaches the theoretical optimal value of the combustion efficiency while meeting the strict constraints of environmental protection emissions, achieving the balance between economy and environmental protection in the biomass power generation process.

[0049] Specifically, for the deep learning method of biomass power generation combustion parameters described in the present invention, the construction of the fuzzy control rule base includes: Defining the membership function of the blast volume deviation based on the expert knowledge base, where the partition threshold of the steam flow rate change rate is dynamically adjusted according to the combustion stage; Optimizing the parameters of the membership function through the genetic algorithm to generate the correlation matrix of the premise and conclusion of the fuzzy rules; In the operation instruction set generation stage, the inertial delay model of the actuator is superimposed to pre-compensate the control quantity.

[0050] The construction of the fuzzy control rule base described in the present invention realizes the optimized generation of the control strategy through the following technical means: When defining the membership function of the blast volume deviation based on the expert knowledge base, the deviation interval is divided according to the thermodynamic characteristics of the combustion system. A wide membership interval is set in the ignition stage to accommodate the fluctuating characteristics of the rapid temperature rise, a narrow interval is adopted in the steady-state operation stage to improve the control accuracy, and the interval range is dynamically expanded according to the steam flow rate change rate in the load adjustment stage. The partition threshold of the steam flow rate change rate is linked with the combustion stage. When the load climbs, the upper limit of the change rate threshold is increased to tolerate the feed fluctuation, and when the load drops suddenly, the lower limit of the threshold is tightened to prevent pressure instability, forming a dynamically adaptive fuzzification processing mechanism.

[0051] When optimizing the parameters of the membership function through a genetic algorithm, a parameter population including the center point of the membership degree, the interval width, and the overlap rate is initialized. The fitness function constructs a multi-objective evaluation system by synthesizing the combustion efficiency deviation, the frequency of emission exceeding the standard, and the oscillation amplitude of the control instruction. Through selection, crossover, and mutation operations, an optimal parameter combination is iteratively generated. The optimized parameters are mapped to the correlation matrix of the premise and conclusion of the fuzzy rules, establishing a non-linear correspondence between the blast volume deviation interval and the correction amount of the feeding rate, and enhancing the adaptability of the rule base to complex working conditions.

[0052] In the operation instruction set generation stage, an inertial delay model of the actuator is constructed to characterize the hysteresis characteristics of the frequency conversion response of the blower and the mechanical transmission of the feeder. Based on the time difference analysis of the historical control instructions and the execution feedback data, a prediction model of the delay time under different load conditions is established. The delay parameters are retrieved according to the real-time load condition, and the control quantity is compensated in advance during the instruction generation, pre-adjusting the instruction amplitude and action time to offset the influence of the dynamic response hysteresis of the actuator on the combustion control stability.

[0053] In the above processing process, the dynamic membership function provides adaptive input features for fuzzy reasoning, the genetic algorithm optimization strengthens the working condition coverage ability of the rule base, and the inertial delay pre-compensation improves the instruction execution accuracy. The parameter optimization of expert experience and data-driven forms a complementarity, and the delay model compensation mechanism makes up for the inherent response hysteresis defect of fuzzy control. The technical collaboration of each link in the rule base construction enables the control strategy to achieve a dynamic balance of efficiency and emission indicators while maintaining combustion stability.

[0054] Specifically, for the deep learning method of biomass power generation combustion parameters described in the present invention, the two-level fault tolerance mechanism includes: When a single sensor data interruption is detected, the adjacent sensor data is fused based on the unscented Kalman filter algorithm to generate a compensation value; When the multi-parameter time series misalignment exceeds a preset threshold, the control strategy with the minimum dynamic time warping distance is matched from the historical similar working condition library; During the clock synchronization failure period, the current timestamp is compensated for deviation according to the historical synchronization error mean.

[0055] The two-level fault tolerance mechanism described in the present invention realizes reliable control under abnormal working conditions of the system through the following technical means: When a data interruption occurs in the temperature or pressure sensor, based on the spatial correlation characteristics of the sensor network, adjacent area oxygen content sensors of the flue gas and non-interrupted pressure sensors are selected as data sources. An observation model of the combustion state is constructed through the unscented Kalman filter algorithm, and the oxygen content data and the historical pressure change law are non-linearly fused to reconstruct the missing sensor data sequence. During the filtering process, the process noise covariance matrix is dynamically adjusted, and the fluctuation range of the compensation value is adaptively controlled according to the characteristics of the combustion stage to generate virtual sensor data that conforms to the thermodynamic law.

[0056] In the multi-parameter time series misalignment detection link, the dynamic time warping distances of the combustion temperature, oxygen content of the flue gas, and furnace pressure are calculated in real time. When the misalignment distance exceeds the allowable range of the combustion chamber design parameters, the historical working condition matching mechanism is triggered: according to the current load rate and fuel calorific value interval label, similar working condition records with the minimum dynamic time warping distance are retrieved from the historical database, and the optimized control strategy parameter set under this working condition is extracted. During the matching process, secondary screening is carried out in combination with the combustion stage label, and the control strategy in the same stage is preferentially selected to achieve smooth switching.

[0057] During the clock synchronization failure period, the historical synchronization error statistics module is enabled to extract the clock offset data sequence within the most recent 24 hours. The sliding window mean algorithm is used to calculate the timestamp deviation compensation amount, and the window length is dynamically adjusted according to the clock stability before the failure. A hysteresis correction mechanism is introduced in the compensation execution stage to predict the future timestamp deviation trend according to the change trend of the fuel feeding rate, and a trend extrapolation correction amount is superimposed on the basis of the deviation mean compensation to maintain the time series data alignment accuracy within the time resolution required for combustion feature extraction.

[0058] In the above processing process, sensor data compensation, historical strategy matching, and clock deviation correction form a hierarchical fault tolerance system. The unscented Kalman filter realizes local data repair, historical working condition matching provides a global control strategy backup, and clock compensation maintains the stability of the system time reference. The three are linked through the combustion stage identification module, focusing on the rapid response of data compensation in the ignition stage, strengthening the accuracy of historical strategy matching in the steady state stage, and improving the dynamic adaptability of clock compensation in the load adjustment stage, forming a multi-level fault tolerance mechanism adapted to the combustion working condition.

[0059] Specifically, for the deep learning method of biomass power generation combustion parameters described in the present invention, the digital twin platform includes: A fuel calorific value fluctuation simulation module that generates a disturbance signal that conforms to the biomass fuel component distribution law through a random generator; A sensor noise injection module that superimposes communication delays and measurement errors that conform to the characteristics of the industrial site in the simulation data stream; The model adaptation module dynamically adjusts the regularization constraint strength of the multimodal time series analysis model according to the prediction error statistical characteristics of the enhanced training dataset.

[0060] The digital twin platform described in the present invention realizes the robustness improvement and adaptive optimization of the combustion model through the following technical means: The fuel calorific value fluctuation simulation module constructs a probability distribution model based on the component database of biomass fuels, analyzes the calorific value fluctuation range and component ratio correlation of raw materials such as straw and wood chips. A perturbation signal that conforms to the actual fuel mixing law is generated by the Markov chain Monte Carlo method, and extreme conditions such as sudden increase and sudden drop of fuel calorific value are simulated in the simulation environment. The amplitude change rate of the perturbation signal is linked with the design parameters of the combustion chamber heat load, the perturbation frequency is increased in the ignition stage to test the rapid response ability of the model, and the perturbation amplitude is increased in the steady state stage to verify the combustion stability boundary.

[0061] The sensor noise injection module collects the historical communication delay data of industrial field devices and constructs a Poisson distribution model of time delay. A variable delay buffer is inserted into the simulation data stream transmission path to simulate communication jitter under different network load conditions. According to the measurement error characteristics of the temperature sensor and the gas analyzer, a superimposed model of Gaussian white noise and impulse noise is designed respectively, and the noise intensity is dynamically adjusted according to the sensor calibration period, so that the simulation data stream retains the real error characteristics of industrial equipment.

[0062] The model adaptation module real-time statistically analyzes the prediction error distribution characteristics of the enhanced training dataset, and uses a sliding window to calculate the mean and variance of the prediction errors of combustion efficiency and nitrogen oxide emissions. The regularization constraint strength is adjusted according to the skewness characteristics of the error distribution: when the error shows a positive shift, L2 regularization is enhanced to suppress overfitting, and when the error shows a negative shift, the Dropout ratio is increased to prevent the local optimal solution from being locked. The adjustment amplitude of the regularization parameter is linked with the combustion stage. A larger adjustment step size is set in the load adjustment stage to accelerate model update, and a smaller step size is used in the steady state stage to maintain prediction stability.

[0063] In the above processing process, the fuel calorific value perturbation generates input variations of industrial typical working conditions, the sensor noise injection restores the real data acquisition environment, and the model adaptation mechanism realizes the dynamic tuning of algorithm parameters. The output of the calorific value fluctuation module is used as the basic data source for noise injection, and the simulation data after superimposing noise drives the model training process. The prediction error statistical results inversely regulate the regularization strength to form a closed-loop optimization. Through the collaborative action of multi-level simulation elements, the platform effectively improves the generalization ability and anti-interference performance of the multimodal time series analysis model for complex working conditions.

[0064] Specifically, for the deep learning method for biomass power generation combustion parameters described in the present invention, the closed-loop feedback control includes: Transfer the operation instruction set to the control interface of the blower frequency converter and the PLC controller of the feeder through the OPC protocol; Calculate the moving average of the change rate of the furnace pressure gradient in real time as the stability criterion of the control loop; When the deviation between the monitored steam flow value and the set threshold continuously exceeds the time window, trigger the backtracking loading mechanism of the historical control strategy.

[0065] The closed-loop feedback control described in the present invention realizes the adaptive regulation of the combustion process through the following technical means: When transmitting operation instructions through the OPC protocol, deploy protocol conversion middleware between the control interface of the blower frequency converter and the PLC controller of the feeder to parse the instruction set generated by the fuzzy inference engine into device-executable instructions. For the frequency conversion speed regulation characteristics of the blower, design the instruction encoding format and verification mechanism, and use cyclic redundancy check codes to ensure the integrity of the transmitted data; the control instructions of the feeder coordinate the action timing of multiple actuators through the timestamp alignment mechanism to eliminate the problem of control instruction mismatch caused by communication delay.

[0066] When calculating the change rate of the furnace pressure gradient in real time, use a sliding window to calculate the moving average, and the window length is dynamically adjusted according to the combustion stage: set a shorter window in the ignition stage to capture rapid pressure fluctuations, and extend the window in the steady state stage to smooth random noise. Compare the trend of the moving average with the designed pressure curve of the combustion chamber. When the trend deviation exceeds the safety threshold for three consecutive sampling periods, trigger the mechanism for generating fine-tuning instructions for the air volume, and send correction parameters to the frequency converter through the OPC protocol to suppress pressure oscillation.

[0067] When the deviation between the monitored steam flow value and the set threshold continuously exceeds the time window, start the control strategy backtracking mechanism. The length of the time window is adaptively adjusted according to the load change rate. Shorten the window in the load climbing stage to enhance the response sensitivity, and extend the window in the load stable stage to avoid false triggering. When retrieving the historical similar working condition library, perform double matching based on the dynamic time warping distance and the current combustion stage label. After loading the optimal historical strategy, smoothly transition to the new strategy parameter set through linear interpolation to avoid system disturbances caused by step changes in the control quantity.

[0068] In the above processing process, the OPC protocol realizes the accurate issuance of control instructions, the pressure gradient monitoring provides real-time stability feedback, and the strategy backtracking mechanism ensures the control continuity under abnormal working conditions. The pressure correction instruction and the strategy switching instruction are coordinated and executed by the priority arbitration module, and the pressure stability regulation is preferentially executed when triggered simultaneously. The interaction and update of the monitoring data and the historical strategy library form a closed-loop learning mechanism, continuously optimizing the adaptive ability of the control system and realizing the efficient and stable operation of the combustion process.

[0069] Second aspect, the present invention provides a deep learning system for biomass power generation combustion parameters, which is applied to the deep learning method for biomass power generation combustion parameters, and includes: A data acquisition module, connected to a temperature sensor, a flue gas composition analyzer and a pressure transmitter, to obtain high-frequency sampling data and low-frequency control signals; A clock synchronization module, with its input end connected to the data acquisition module, and built-in with a sliding window buffer unit and a cubic spline interpolation unit, to output synchronized timing data; A phase correction module, with its input end receiving the synchronized timing data, integrating a dynamic time warping algorithm and a residual connection branch, to output phase-aligned data; A hybrid modeling module, with its input end connected to the phase correction module, including a parallel multi-scale one-dimensional convolutional network and a bidirectional long short-term memory network. The convolutional network extracts transient fluctuation features and slow-varying trend features, and the recurrent neural network outputs long-period time series dependence relationships; An optimization control module, with its input end receiving the prediction results of the hybrid modeling module, and built-in with a multi-objective optimization function generator and a fuzzy inference engine, to generate constraint conditions for the air volume and the feed ratio; A fault-tolerant execution module, with its input end connected to the optimization control module and the data acquisition module, deploying a digital twin platform to simulate extreme working conditions, and integrating two-level fault-tolerant compensation units; A feedback learning module, with its input end receiving combustion efficiency monitoring data and the simulation results of the digital twin platform, and its output end connected to the hybrid modeling module, to dynamically adjust the model parameters.

[0070] In the specific implementation of the deep learning system for biomass power generation combustion parameters, first, a distributed data acquisition network is used to connect a temperature sensor, a flue gas composition analyzer and a pressure transmitter to obtain high-frequency sampling data and low-frequency control signals in real time. The temperature sensor adopts a thermocouple array for multi-point layout to collect millisecond-level furnace temperature fluctuation data; the flue gas composition analyzer integrates an infrared spectrum detection unit to output second-level flue gas oxygen content sequences; the pressure transmitter monitors the furnace pressure change to form a high-frequency pressure time series flow. The low-frequency control signals include minute-level feeder speed commands and blower frequency conversion signals, which are transmitted to the data acquisition module through an industrial bus. The signal conditioning circuit performs noise reduction filtering on the high-frequency pulse signals and amplitude normalization on the low-frequency signals to generate a standardized multi-source time series data stream.

[0071] For the clock synchronization problem of multi-source data, a distributed clock synchronization system based on the IEEE 1588 protocol is deployed. The master clock node periodically broadcasts synchronization messages to each slave device. The temperature sensor and pressure transmitter correct the local clock according to the message transmission delay and attach a microsecond-level timestamp to the original data. High-frequency data is segmented and cached using a sliding window mechanism. The window length is dynamically adjusted to 50 - 200 milliseconds according to the fluctuation period of the fuel calorific value. The clock offset between windows is calculated by the exponentially weighted moving average algorithm, and the interpolation mutation amplitude is combined with the physical continuity constraint of the fuel flow not exceeding 5% of the design flow. The low-frequency control signal is resampled using cubic spline interpolation. A pressure gradient change threshold is introduced at the interpolation nodes. When the pressure change rate between adjacent sampling points exceeds 0.2 MPa / s, the interpolation curve is forced to be smoothly corrected to eliminate non-physical jump points, and finally, synchronized timing data aligned with the time axis is generated.

[0072] When the phase correction module processes the synchronization data, it uses the dynamic time warping algorithm to match the combustion temperature and flue gas oxygen content sequences. By calculating the correlation coefficient matrix of the two sequences, the path search is limited to the area where the correlation coefficient is greater than 0.8 to identify the local phase deviation caused by sampling delay. After performing time-domain translation correction on the flue gas oxygen content sequence, the residual connection branch is superimposed to fuse the original data and the corrected features through a 1×1 convolutional layer. The convolutional kernel parameters are dynamically initialized according to the frequency-domain energy distribution of the furnace pressure sudden change signal. When the proportion of the frequency-domain energy of the pressure signal in the 100 - 500 Hz interval exceeds 30%, the fusion weight of the original data is increased to 0.7 to enhance the ability to retain transient features.

[0073] The hybrid modeling module adopts a combined architecture of parallel multi-scale one-dimensional convolutional network and bidirectional long short-term memory network. The first convolutional layer is configured with a narrow convolutional kernel with a stride of 3 to extract the millisecond-level transient fluctuation features of the furnace pressure signal; the second convolutional layer uses a wide convolutional kernel with a stride of 15 to capture the minute-level trend changes of the flue gas oxygen content. The bidirectional long short-term memory network embeds a combustion stage recognition unit in the hidden layer. According to the steam flow rate change rate threshold, the ignition stage (flow growth rate > 5% / min), steady state stage (volatility < 1%), and load adjustment stage are divided, and the corresponding gated attention weights are activated. The adaptive normalization layer dynamically adjusts the fusion ratio based on the feature variance. When the pressure fluctuation variance exceeds 0.05, the weight of the narrow convolutional feature is increased to 60%, and when the oxygen concentration variance exceeds 0.02, the weight ratio of the wide convolution increases to 55%.

[0074] The optimization control module constructs a double-objective Pareto front model for combustion efficiency and nitrogen oxide emissions, and the real-time load status signal triggers dynamic weight allocation. During the load ramping stage, the emission weight coefficient is set to 0.7, and the efficiency weight increases to 0.6 during the steady state stage. The historical operating condition similarity screening uses a dynamic time warping distance threshold. When the similarity is lower than 0.65, the adjustment range of the air flow rate is restricted to 80%-120% of the design value, and the feeding ratio fluctuation is locked within the range of ±3%. The fuzzy inference engine calls the rule base optimized by the genetic algorithm. The membership function of the air flow rate deviation expands to ±15% during the ignition stage and shrinks to ±5% during the steady state stage, and the inertial delay model is superimposed to perform a 50-200 ms lead compensation on the blower command.

[0075] When the digital twin platform simulates the fuel calorific value fluctuation, a calorific value disturbance signal is generated based on the biomass component database, and the disturbance amplitude randomly varies according to the Poisson distribution within the range of ±20%. The sensor noise injection module superimposes a communication delay of 0.5-2 ms and random noise of 0.1%-0.5% of the range on the simulation data. The model adaptive module adjusts the regularization strength according to the prediction error distribution. When the error mean exceeds 2%, the L2 regularization coefficient is enhanced to 0.01, and when the error variance exceeds 0.1, the Dropout ratio is increased to 0.3. The feedback learning module updates the model parameters every 30 minutes, and mixes and trains the digital twin simulation data and the real-time monitoring data at a ratio of 7:3.

[0076] During the closed-loop control process, the OPC protocol middleware encodes the control instructions into the ModbusRTU format dedicated to the blower frequency converter and the Profinet instructions of the feeder PLC. The moving average value of the furnace pressure gradient is monitored in real time. When it deviates from the design curve by more than 10% for three consecutive sampling periods, a fine-tuning instruction for the air flow rate is triggered. When the steam flow deviation continuously exceeds the set threshold for 5 minutes, the historical operating condition library is retrieved to load the control strategy with the highest similarity, and it transitions to the new strategy parameters within 10 seconds through linear interpolation. The two-level fault tolerance mechanism runs in parallel: when the temperature sensor fails, the temperature sequence is reconstructed by fusing the data of the adjacent three oxygen content sensors in the flue gas; during the abnormal clock synchronization period, the sliding window mean value is used to compensate for the timestamp deviation, and the window length is automatically adjusted to 5-15 minutes according to the clock stability in the previous hour.

[0077] Explanation of the technical feature terms of the present invention: IEEE 1588 Protocol: A high-precision network clock synchronization protocol (Precision Time Protocol, PTP) used to achieve microsecond-level time synchronization in a distributed system. By the master clock node broadcasting synchronization messages, the slave devices calculate the transmission delay and correct the local clock offset, providing a unified time reference for multi-source data.

[0078] Dynamic Time Compensation Algorithm: An algorithm based on the sliding window mechanism and Exponentially Weighted Moving Average (EWMA). By segmentally caching high-frequency data streams, dynamically adjusting the clock offset compensation weights, and combining the physical characteristics constraints of fuel flow to interpolate the mutation amplitude, it suppresses the clock cumulative error between devices.

[0079] Cubic Spline Interpolation Resampling: For the non-uniform sampling problem of low-frequency control signals, a cubic polynomial is used to construct a smooth interpolation curve. The interpolation nodes are set according to the continuity requirements of the combustion system pressure change to eliminate non-physical jump points and achieve the continuous reconstruction of low-frequency signals on a unified time axis.

[0080] Dynamic Time Warping (DTW): A time series alignment algorithm that identifies local phase deviations caused by communication delays or uneven sampling intervals by calculating the optimal alignment path between the combustion temperature and flue gas oxygen content sequences. The path search range is constrained by their correlation to ensure that the alignment result conforms to the physical relevance of the combustion process.

[0081] Residual Connection Branch: A feature fusion structure in deep learning that performs cross-channel fusion of the original time series data through a 1×1 convolutional layer and the corrected features. By retaining transient signals (such as sudden pressure changes) in the uncorrected data, it prevents the loss of high-frequency features during the phase correction process.

[0082] Parallel Multi-Scale 1D CNN: A parallel structure composed of convolutional layers with different strides: Narrow convolutional kernel (stride 3): Captures the millisecond-level transient fluctuation characteristics of the furnace pressure; Wide convolutional kernel (stride 15): Extracts the minute-level slow-changing trend characteristics of the flue gas oxygen content.

[0083] Bidirectional Long Short-Term Memory Network (Bi-LSTM): A recurrent neural network structure that analyzes the long-term dependencies of fuel calorific value fluctuations and pollutant generation through forward and backward time series propagation. A combustion stage identification module is embedded at the output end to dynamically activate the gated attention weights in different stages.

[0084] Gated Attention Mechanism: A module that dynamically assigns feature weights according to the combustion stage (ignition / steady state / load adjustment). For example, it strengthens the fuel supply rate feature during the ignition stage and enhances the correlation weight between oxygen concentration and emission parameters during the steady state stage.

[0085] Adaptive Normalization Layer: Dynamically adjusts the multi-scale feature fusion ratio based on the feature variance distribution and combustion stage labels. For example, it increases the weight of narrow convolution features when the pressure fluctuation is significant and increases the proportion of wide convolution features when the oxygen concentration changes slowly.

[0086] Digital Twin Platform: Integrates a combustion kinetics simulation model, generates biomass component perturbation signals through a fuel calorific value fluctuation simulation module, and superimposes industrial-level communication delays and measurement errors through a sensor noise injection module for model robustness training and extreme condition verification.

[0087] Multi-Objective Optimization Function Generator: Constructs a Pareto Front solution set for maximizing combustion efficiency and minimizing nitrogen oxide emissions, dynamically adjusts the objective weights in combination with real-time load status signals, and screens feasible solutions that meet the steam flow stability.

[0088] Fuzzy Inference Engine: Based on a fuzzy rule base optimized by a genetic algorithm, maps the blast volume deviation and steam flow change rate to control instructions. Pre-compensates the response lag of the actuator through an inertia delay model to improve the instruction execution accuracy.

[0089] The technical solution of the present invention collaboratively solves the phase mismatch problem through the following logical chain: Time Series Alignment: The IEEE 1588 protocol and dynamic time compensation algorithm eliminate the clock offset between devices, and cubic spline interpolation resampling aligns low-frequency signals to generate synchronous time series data; Phase Correction: The DTW algorithm corrects the local phase deviation caused by communication delays, and the residual connection branch retains transient features to avoid information loss during the correction process; Feature Modeling: Multi-scale convolution extracts spatio-temporal features, Bi-LSTM analyzes long-term dependencies, and gated attention and adaptive normalization dynamically optimize feature weights; Closed-Loop Optimization: The digital twin platform enhances the model generalization ability, multi-objective optimization and fuzzy inference generate control strategies, and a two-level fault tolerance mechanism ensures control continuity under abnormal conditions.

[0090] Through the above technical chain, the present invention effectively restores the dynamic coupling relationship between multi-source data, and improves the prediction accuracy of combustion efficiency and the decision-making reliability of control strategies.

[0091] The system described in the present invention realizes the deep optimization control of biomass combustion parameters through the cooperation of the following modules: The data acquisition module is connected to a temperature sensor, a flue gas composition analyzer, and a pressure transmitter through an industrial bus interface. Among them, the temperature sensor adopts a multi-point distributed layout of a thermocouple array, and the flue gas composition analyzer integrates an infrared spectroscopy and an electrochemical sensing unit. The module is built with a signal conditioning circuit to denoise the millisecond-level temperature pulse signal, and at the same time normalize the amplitude of the minute-level feed control signal, and output the original time series of high-frequency sampling data and low-frequency control signals.

[0092] After the clock synchronization module receives the original time series, the sliding window cache unit dynamically adjusts the window length according to the fuel calorific value fluctuation frequency, and uses a short window to capture the fast-changing characteristics during the load climbing stage. The cubic spline interpolation unit constructs interpolation nodes for the low-frequency control signal based on the continuity constraint of the combustion chamber pressure change, and eliminates the non-physical distortion caused by the step jump of the control instruction. The time axis deviation of the output synchronous time series data is controlled within the time resolution threshold required for combustion feature extraction.

[0093] When the phase correction module processes the synchronous time series data, the dynamic time warping algorithm combines the correlation coefficient matrix of combustion temperature and oxygen content in the flue gas to limit the search space of the optimal alignment path. The residual connection branch performs cross-channel fusion of the original data stream and the corrected features through a 1×1 convolutional kernel, and the convolutional kernel parameters are dynamically initialized based on the frequency domain characteristics of the furnace pressure sudden change signal. The output phase-aligned data retains the dual characteristics of transient fluctuations and slow-changing trends, providing high-quality input for hybrid modeling.

[0094] In the hybrid modeling module, the first layer of the parallel multi-scale one-dimensional convolutional network uses a narrow convolutional kernel to capture the pressure transient waveform, and the second layer uses a wide convolutional kernel to extract the oxygen concentration trend feature. The bidirectional long short-term memory network embeds a combustion stage marking unit in the hidden layer, identifies the current operating stage according to the steam flow rate change rate, and activates the corresponding gated attention weight distribution. The adaptive normalization layer dynamically adjusts the fusion ratio of convolutional and recurrent features based on the feature variance distribution, and outputs a combined representation of long-period time series dependencies and multi-scale spatial features.

[0095] After the optimization control module receives the model prediction results, the multi-objective optimization function generator constructs the Pareto front solution set of the combustion efficiency and emission indicators, and combines the real-time load status tags to screen the feasible solution space. The fuzzy inference engine calls the rule association matrix optimized by the genetic algorithm, converts the constraint conditions into the variable frequency instruction of the blower and the stroke control quantity of the feeder, and adapts to the device interface protocols of different manufacturers through the OPC protocol conversion middleware.

[0096] The digital twin platform of the fault-tolerant execution module integrates the combustion kinetics simulation model, and the fuel calorific value disturbance generator simulates the biomass component fluctuation based on the Markov chain. The two-stage fault-tolerant unit starts the unscented Kalman filter data compensation when the sensor fails, and switches to the historical synchronous error compensation mode when the clock is abnormal. The simulation results and the real-time data jointly drive the generation of the fault-tolerant strategy through the confidence weighted mechanism.

[0097] The feedback learning module establishes the error distribution map of the combustion efficiency monitoring data and the model prediction results, and updates the model regularization parameters by using the statistical features of the sliding window. After the simulation test results of the digital twin platform are verified by the credibility, they are injected into the online training data stream of the hybrid modeling module to form a closed loop of dynamic optimization of the model parameters. Each module realizes information interaction through the data bus and the event trigger mechanism, and maintains the adaptive optimization ability of the combustion control system.

[0098] The present invention solves the phase mismatch problem through the multi-level time series alignment and feature fusion technology. First, a unified time reference is established based on the IEEE1588 protocol, the clock offset of the high-frequency sampling data is dynamically compensated by using the sliding window mechanism, and the clock cumulative error between devices is suppressed by the exponentially weighted moving average algorithm; for the low-frequency control signal, combined with the pressure continuity constraint of the combustion system, cubic spline interpolation resampling is implemented to eliminate the time domain aliasing effect and generate the multi-source data stream with time series synchronization. This step aligns the data with different sampling frequencies to the unified time axis and solves the time series misalignment problem caused by the initial clock reference offset.

[0099] Secondly, the dynamic time warping algorithm is used for phase correction, and the physical correlation between the combustion temperature and the oxygen content in the flue gas is used to constrain the path search range, and the local phase deviation caused by communication delay or uneven sampling interval is identified. The relative offset between the data sequences is eliminated by time domain translation correction, and at the same time, the residual connection branch is superimposed to retain the transient features in the original time series data. The 1×1 convolutional layer is used to fuse the features before and after correction to prevent the loss of key dynamic information during the phase adjustment process, and solve the problem of misalignment of the feature correlation of asynchronous data streams.

[0100] Finally, multi-scale feature parsing is achieved through the hybrid modeling module. The parallel multi-scale one-dimensional convolutional network extracts the features of pressure transient fluctuations and the slow-changing trend of oxygen concentration respectively, and the bidirectional long short-term memory network analyzes the long-period time series dependence relationship. The adaptive normalization layer dynamically adjusts the feature fusion weights according to the combustion stage, enabling the model to accurately represent the dynamic coupling mechanism between multi-source data. Through the synergistic effect of time series alignment, phase correction, and feature fusion, this technical chain effectively restores the true causal relationship between combustion parameters and improves the decision-making reliability of control strategies.

Claims

1. A deep learning method for biomass power generation combustion parameters, characterized in that: include: Acquire multi-source time series data of a biomass combustion system, wherein the multi-source time series data includes high-frequency sampling data and low-frequency control signals; Performing clock synchronization processing on the multi-source time series data, aligning the time series information of data with different frequencies through a dynamic time compensation mechanism, and generating synchronized time series data; Performing phase correction based on the synchronous timing data, using a dynamic path planning algorithm to match timing correlation features between multiple parameters, and generating phase alignment data; Constructing a multimodal time series analysis model, wherein the multimodal time series analysis model includes a convolution module for extracting local fluctuation features and a recurrent neural network module for capturing long-term dependencies; Inputting the phase alignment data into the multi-modal timing analysis model to obtain combustion state prediction parameters; Constructing a multi-objective optimization function based on the combustion state prediction parameters to generate constraints for combustion control parameters; Generate an operation instruction set through a fuzzy inference engine according to the constraint conditions, wherein the fuzzy inference engine is based on a fuzzy control rule base, and issue the operation instruction set to an actuator, while collecting feedback parameters of the combustion system in real time to form a closed-loop control; A simulation verification platform is deployed to simulate abnormal working conditions to generate a training data set, the multimodal timing analysis model is iteratively updated according to the feedback parameters, and sensor data anomalies are compensated through a two-level fault-tolerant strategy.

2. The biomass power generation combustion parameter deep learning method according to claim 1 is characterized in that: The clock synchronization process includes: Adding a unified time reference to the high-frequency sampling data and the low-frequency control signal through the IEEE 1588 protocol; In the sliding window caching process, the exponentially weighted moving average algorithm is superimposed on the high-frequency sampling data to compensate for the clock offset between devices, and the interpolation mutation amplitude is constrained based on the physical characteristics of the fuel flow; When the low-frequency control signal is resampled by cubic spline interpolation, non-physical jump points are eliminated in combination with the continuity requirement of the combustion system pressure change.

3. The biomass power generation combustion parameter deep learning method according to claim 1 is characterized in that: The phase matching includes: Based on the synchronized time series data, the optimal alignment path of the combustion temperature and smoke oxygen content sequence is calculated by a dynamic time warping algorithm, and the path search range is constrained by the correlation between the two; The residual connection branch is superimposed on the data channel after the time domain translation correction, and the original time series data is fused with the corrected features through a 1×1 convolution layer; According to the time-frequency characteristics of the furnace pressure sudden change signal, the fusion weight of the original data and the correction feature in the residual branch is dynamically adjusted.

4. The biomass power generation combustion parameter deep learning method according to claim 1 is characterized in that: The construction of the hybrid model includes: In the parallel multi-scale one-dimensional convolution structure, the first convolution layer extracts the transient fluctuation characteristics of the furnace pressure with a step length of 3, and the second convolution layer extracts the slow-changing trend characteristics of the flue gas oxygen content with a step length of 15; At the temporal feature output end of the bidirectional long short-term memory network, a corresponding gated attention weight is activated according to the combustion stage recognition result, wherein the combustion stage includes ignition, steady state and load adjustment; After concatenating the output features of the first convolutional layer and the second convolutional layer, the fusion ratio of features of different scales is dynamically adjusted through an adaptive normalization layer.

5. The biomass power generation combustion parameter deep learning method according to claim 1 is characterized in that: The dynamic optimization objective function is a multi-objective optimization function, comprising the following steps: Based on the combustion state prediction parameters, a Pareto frontier optimization model of combustion efficiency and nitrogen oxide emissions is constructed; According to the load status signal output by the power plant operation mode recognition module, the weight distribution ratio of efficiency and emission targets is dynamically adjusted; A historical operating condition similarity screening mechanism is introduced into the constraint condition generator. When the similarity between the prediction result and the historical data is lower than a set threshold, the search space of the control strategy is restricted.

6. The biomass power generation combustion parameter deep learning method according to claim 1 is characterized in that: The construction of the fuzzy control rule base includes: The blast volume deviation membership function is defined based on the expert knowledge base, in which the partition threshold of the steam flow rate change rate is dynamically adjusted according to the combustion stage; Optimizing the parameters of the membership function through a genetic algorithm to generate a correlation matrix of premises and conclusions of fuzzy rules; During the operation instruction set generation stage, the inertia delay model of the superimposed actuator is used to pre-compensate the control variable.

7. The biomass power generation combustion parameter deep learning method according to claim 1 is characterized in that: The multi-level fault-tolerance strategy includes a two-level fault-tolerance mechanism, specifically including: When a single sensor data interruption is detected, the adjacent sensor data are fused based on the unscented Kalman filter algorithm to generate a compensation value; When the multi-parameter timing misalignment exceeds the preset threshold, the control strategy with the smallest dynamic time warping distance is matched from the historical similar operating condition library; During the period of clock synchronization failure, the current timestamp is compensated for deviation according to the mean value of historical synchronization errors.

8. The biomass power generation combustion parameter deep learning method according to claim 1 is characterized in that: The digital twin platform includes: The fuel calorific value fluctuation simulation module generates disturbance signals that conform to the distribution law of biomass fuel components through a random generator; The sensor noise injection module superimposes communication delays and measurement errors that conform to the characteristics of industrial sites in the simulated data stream; The model adaptation module dynamically adjusts the regularization constraint strength of the multimodal time series analysis model according to the prediction error statistical characteristics of the enhanced training data set.

9. The biomass power generation combustion parameter deep learning method according to claim 1, characterized in that: The closed-loop feedback control comprises: The operation instruction set is transmitted to the control interface of the blower inverter and the PLC controller of the feeder through the OPC protocol; Real-time calculation of the moving average of the furnace pressure gradient change rate as a criterion for the stability of the control loop; When the deviation between the steam flow monitoring value and the set threshold value continues to exceed the time window, the backtracking loading mechanism of the historical control strategy is triggered.

10. A biomass power generation combustion parameter deep learning system, applied to the biomass power generation combustion parameter deep learning method according to any one of claims 1 to 9, characterized in that: include: Data acquisition module, connected to temperature sensor, flue gas composition analyzer and pressure transmitter, to obtain high-frequency sampling data and low-frequency control signals; A clock synchronization module, the input end of which is connected to the data acquisition module, has a built-in sliding window buffer unit and a cubic spline interpolation unit, and outputs synchronous time series data; A phase correction module, the input end of which receives the synchronous timing data, integrates a dynamic time warping algorithm and a residual connection branch, and outputs phase alignment data; A hybrid modeling module, the input end of which is connected to the phase correction module, comprising a parallel multi-scale one-dimensional convolutional network and a bidirectional long short-term memory network, wherein the convolutional network extracts transient fluctuation characteristics and slow-changing trend characteristics, and the recurrent neural network outputs long-period temporal dependency; An optimization control module, whose input end receives the prediction result of the hybrid modeling module, has a built-in multi-objective optimization function generator and a fuzzy reasoning engine, and generates constraint conditions for the blast volume and the feed ratio; A fault-tolerant execution module, the input end of which is connected to the optimization control module and the data acquisition module, deploys a digital twin platform to simulate extreme working conditions, and integrates a two-level fault-tolerant compensation unit; A feedback learning module receives combustion efficiency monitoring data and simulation results of the digital twin platform at its input end, and is connected to the hybrid modeling module at its output end to dynamically adjust model parameters.

Images