A supercritical circulating fluidized bed boiler parameter prediction method and system based on attention guide fuzzy fusion

Abstract

This invention provides a method and system for predicting parameters of a supercritical circulating fluidized bed boiler based on attention-guided fuzzy fusion. The method includes: performing multi-scale segmentation and feature embedding on the time series of collected circulating fluidized bed boiler unit operating parameters to obtain feature representations; based on the feature representations, employing a dual attention mechanism to obtain local time-dependent features and cross-scale attention features; based on the local time-dependent features and the cross-scale attention features, using an element-level interaction and residual feature encoding network for reliability assessment to obtain confidence parameters and fuzzy measure values; based on the confidence parameters and the fuzzy measure values, employing a nonlinear fusion method based on fuzzy measure to obtain fused feature representations; and based on the fused feature representations, obtaining the time series prediction results of the circulating fluidized bed boiler unit operating parameters. This invention improves the accuracy and robustness of predicting bed temperature, power generation, and steam flow rate.

Description

Technical Field

[0001] This invention belongs to the field of smart environmental protection technology, specifically relating to a method and system for predicting parameters of supercritical circulating fluidized bed boilers based on attention-guided fuzzy fusion, which is applicable to the prediction of key parameters of supercritical circulating fluidized bed boilers in time series. Background Technology

[0002] Circulating fluidized bed (CFB) combustion technology occupies a central position in modern thermal power generation systems due to its advantages such as strong fuel adaptability, high combustion efficiency, and low pollutant emissions. Accurate prediction of key combustion-related variables is not only crucial for operational optimization but also a core prerequisite for ensuring the stable, reliable, and low-emission operation of the power generation system.

[0003] With the continuous improvement of industrial process control, energy system operation, and complex equipment condition monitoring, key parameter prediction based on time series data is playing an increasingly important role in safe operation, energy efficiency optimization, and intelligent decision-making. Industrial objects, such as large generator sets, chemical plants, and complex energy systems, often exhibit operating states that are multivariable, strongly coupled, nonlinear, and non-stationary time series data, placing high demands on the modeling capabilities and robustness of prediction models.

[0004] In practical engineering applications, industrial time series typically contain both rapid dynamic changes at short timescales and slow evolutionary trends at long timescales. The physical processes and operating mechanisms reflected at different timescales differ significantly, making it difficult for single-timescale modeling methods to fully characterize the system's true dynamic behavior. Therefore, in recent years, multi-scale time series forecasting methods have gained increasing attention. These methods introduce data representations with different time resolutions to provide forecasting models with multi-perspective temporal information.

[0005] However, multi-scale time series prediction still faces many challenges: on the one hand, there are significant differences in sequence length, temporal resolution and information density between different time scales, and direct fusion can easily introduce redundant or conflicting information; on the other hand, the system's dependence on information at each time scale is not fixed at different stages of operation, and if the model cannot adaptively adjust the importance of features at each scale, the prediction accuracy and stability will be limited.

[0006] To address the above problems, existing technologies mainly focus on the following categories of methods: (1) Time series prediction methods based on statistical models or traditional machine learning. These methods usually rely on fixed time windows or manually constructed features for modeling, and have limited ability to characterize time dependencies, making it difficult to capture both short-term fluctuations and long-term trends simultaneously; when facing multi-scale time series, it is often necessary to manually set feature combination rules, resulting in insufficient model generalization ability. (2) Sequence modeling methods based on deep learning. Recurrent neural networks, long short-term memory networks and their variants can model time dependencies to a certain extent, but their ability to model long sequences is limited, and in multi-scale scenarios, they usually use simple splicing or weighting to fuse features of different scales, making it difficult to fully explore the complex relationships between scales. (3) Time series prediction methods based on attention mechanisms. In recent years, the Transformer structure has achieved certain results in time series modeling with its self-attention mechanism, and can achieve adaptive information aggregation between time positions through attention weights. However, most existing methods focus on time dependency modeling within a single time scale, and are insufficient in characterizing the global correlation between different time scales. Although some multi-scale Transformer methods introduce cross-scale modeling mechanisms, they often fail to distinguish between local dependencies within a scale and global dependencies between scales, resulting in insufficient multi-scale feature interactions.

[0007] Furthermore, existing multi-scale attention methods generally share a common problem: the lack of an explicit modeling mechanism for the predictive reliability of features at different scales. In actual operation, features at different time scales are affected to varying degrees by noise interference, missing data, or changes in operating conditions, resulting in significant uncertainty in their contribution to the prediction results. Existing methods typically employ fixed weights or implicit weighting based on attention scores for fusion, which fails to accurately reflect the reliability of features at each scale in the current prediction task and is prone to amplifying errors when there is conflict in multi-scale information.

[0008] To introduce uncertainty modeling capabilities, some studies have attempted to correct prediction results using probabilistic modeling or simple fuzzy logic methods. However, these methods mostly focus on output layer adjustments and fail to deeply integrate attention representations of the feature layer for collaborative modeling, making it difficult to characterize the nonlinear interaction relationships between multi-scale features.

[0009] In summary, existing time series prediction methods still have the following shortcomings in multi-scale scenarios: First, they fail to simultaneously consider local time dependency modeling within a scale and global dependency modeling between scales; second, they lack an explicit quantification mechanism for the reliability of multi-scale feature prediction; and third, the multi-scale feature fusion methods are mostly linear or weakly nonlinear, making it difficult to express the cooperative, redundant, and complementary relationships between scales.

[0010] Therefore, there is an urgent need for a prediction method that can jointly model intra-scale and inter-scale time dependencies in multi-scale time series prediction and adaptively evaluate the reliability of features at each scale, thereby achieving nonlinear fusion of multi-scale features to improve prediction accuracy, stability and adaptability to complex dynamic conditions. Summary of the Invention

[0011] To address the problems existing in the prior art, this invention provides a method and system for predicting parameters of supercritical circulating fluidized bed boilers based on attention-guided fuzzy fusion. It aims to focus on the time series prediction of key operating variables (including power generation, bed temperature, and main steam flow) of CFB boiler generator sets. A fuzzy-aware multi-scale adaptive Transformer is proposed to form a robust and expressive framework for modeling the complex, nonlinear, and uncertain dynamic characteristics of CFB combustion systems.

[0012] To achieve the above objectives, the present invention provides the following solution: A parameter prediction method for supercritical circulating fluidized bed boilers based on attention-guided fuzzy fusion, the method comprising: The time series of operating parameters of the collected circulating fluidized bed boiler unit are divided into multiple scales and embedded with features to obtain the feature representations corresponding to each scale. Based on the aforementioned feature representation, a dual attention mechanism is employed to obtain local temporal-dependent features and cross-scale attention features; Based on the local time-dependent features and the cross-scale attention features, an element-level interactive operation and feature encoding network are used to carry out scale reliability quantification and relative importance modeling to obtain confidence parameters and fuzzy measure values. Based on the confidence parameter and the fuzzy measure value, a nonlinear fusion method based on fuzzy measure is used to obtain the fused feature representation; Based on the fusion feature representation, time series prediction results of the operating parameters of the circulating fluidized bed boiler unit are obtained.

[0013] Preferably, based on the feature representation, the method for obtaining local temporal-dependent features and cross-scale attention features using a dual attention mechanism includes: The feature representation is time-aligned and mapped to a unified time space. Within a unified and standardized time and space, the feature representation is processed using the in-scale attention mechanism within the dual attention mechanism to obtain local time-dependent features; Within a unified and standardized time and space, the feature representation is processed using a cross-scale attention mechanism within a dual attention mechanism to obtain cross-scale attention features.

[0014] Preferably, the method for obtaining confidence parameters and fuzzy measure values ​​by performing reliability assessment using an attention interaction mechanism and a feature encoding network based on the local time-dependent features and the cross-scale attention features includes: ; ; ; ; ; in, For interactive features, For the first i Time scale Scaling dot product self-attention on, This represents the element-level Hadamard product operation. For cross-scale attention features, For the input time series, This indicates a feature concatenation operation. This indicates the residual coding module. For parameters Learnable nonlinear transformations of control, This represents the residual encoding depth feature at the i-th time scale. This represents the comprehensive splicing features at the i-th time scale. Indicates the first i A confidence parameter for reliability over a time scale. Representing encoded features To confidence parameter The mapping function, This indicates a global pooling operation. and This represents a parameter-independent convolutional projection operator used to aggregate channel-level information. To represent the fuzzy measure of the relative importance of the i-th time scale, This represents the sigmoid function.

[0015] Preferably, the method for obtaining the fused feature representation based on the confidence parameter and the fuzzy measure value using a nonlinear fuzzy measure-based fusion approach includes: Computational fusion feature representation: ; in, For feature representation, This indicates the number of rows after sorting in descending order of confidence level. i Each time scale weight, and let , This represents the maximum number of input sequences at a specific time scale. It represents the fuzzy measure value of the set composed of the first i sorted time scales.

[0016] Preferably, the method for obtaining time series prediction results of operating parameters of a circulating fluidized bed boiler unit based on the fused feature representation includes: ; in, For time series prediction results, This represents a prediction network used to map fused features to a target prediction space. This is a local time-dependent feature.

[0017] The present invention also provides a parameter prediction system for supercritical circulating fluidized bed boilers based on attention-guided fuzzy fusion. The system is used to implement the aforementioned method and includes: a multi-scale partitioning module, a dual attention module, a reliability assessment module, a fuzzy fusion module, and a prediction module. The multi-scale segmentation module is used to perform multi-scale segmentation and feature embedding on the time series of collected operating parameters of the circulating fluidized bed boiler unit to obtain the feature representation corresponding to the scale. The dual attention module is used to obtain local time-dependent features and cross-scale attention features based on the feature representation using a dual attention mechanism; The reliability assessment module is used to perform reliability assessment based on the local time-dependent features and the cross-scale attention features, using element-level interactive operations and feature encoding networks to obtain confidence parameters and fuzzy measure values. The fuzzy fusion module is used to obtain a fusion feature representation based on the confidence parameter and the fuzzy measure value using a nonlinear fusion method based on fuzzy measure. The prediction module is used to obtain time series prediction results of the operating parameters of the circulating fluidized bed boiler unit based on the fused feature representation.

[0018] Preferably, the dual attention module includes: a time-aligned unit, a local unit, and a global unit; The time alignment unit is used to perform time alignment processing, mapping the feature representation to a unified and standardized time space; The local unit is used to apply the intra-scale attention mechanism in the dual attention mechanism to the feature representation within a unified and standardized time space to obtain local time-dependent features. The global unit is used to apply the cross-scale attention mechanism in the dual attention mechanism to the feature representation within a unified time and space framework to obtain cross-scale attention features.

[0019] Preferably, the reliability assessment module, based on the local time-dependent features and the cross-scale attention features, employs an attention interaction mechanism and a feature encoding network to perform reliability assessment, and obtains the confidence parameter and fuzzy measure value through the following methods: ; ; ; ; ; in, For interactive features, For the first i Time scale Scaling dot product self-attention on, This represents the element-level Hadamard product operation. For cross-scale attention features, For the input time series, This indicates a feature concatenation operation. This indicates the residual coding module. For parameters Learnable nonlinear transformations of control, This represents the residual encoding depth feature at the i-th time scale. This represents the comprehensive splicing features at the i-th time scale. Indicates the first i A confidence parameter for reliability over a time scale. Representing encoded features To confidence parameter The mapping function, This indicates a global pooling operation. and This represents a parameter-independent convolutional projection operator used to aggregate channel-level information. To represent the fuzzy measure of the relative importance of the i-th time scale, This represents the sigmoid function.

[0020] Preferably, the fuzzy fusion module obtains the fused feature representation by employing a nonlinear fusion method based on fuzzy measure, based on the confidence parameter and the fuzzy measure value, including: Computational fusion feature representation: ; in, For feature representation, This indicates the number of rows after sorting in descending order of confidence level. i Each time scale weight, and let , This represents the maximum number of input sequences at a specific time scale. It represents the fuzzy measure value of the set composed of the first i sorted time scales.

[0021] Preferably, the method for the prediction module to obtain time series prediction results of the operating parameters of the circulating fluidized bed boiler unit based on the fused feature representation includes: ; in, For time series prediction results, This represents a prediction network used to map fused features to a target prediction space. This is a local time-dependent feature.

[0022] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. A dual-attention multi-scale Transformer framework is proposed, which can simultaneously capture local temporal dependencies and global cross-scale interactions, realizing scale-aware temporal representation; 2. An attention-guided fuzzy fusion strategy is designed to quantify the reliability of scale-specific representations using evidence-based confidence scores and model their relative importance using learned fuzzy measures. This strategy integrates temporal and cross-scale attention through element-wise interaction, residual encoding, and a Choquet integral-inspired aggregation method, achieving uncertainty-aware nonlinear fusion of multi-scale features. This results in a more reliable representation for accurate prediction of target time series. 3. Extensive experiments were conducted on the CFB boiler generator set dataset to verify the effectiveness of the proposed method in accurately predicting key operating variables and modeling the dynamic characteristics of heterogeneous combustion at multiple time scales. Attached Figure Description

[0023] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a schematic diagram of the supercritical circulating fluidized bed boiler parameter prediction method based on attention-guided fuzzy fusion according to an embodiment of the present invention. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0027] The concepts related to Transformer architecture, fuzzy measure, and fuzzy integral are as follows: I. Multi-scale time series forecasting Multi-scale time series forecasting aims to capture the heterogeneous dynamic characteristics of series data by learning forecast mappings at multiple time resolutions. Let... (1) This represents historical input presented at different time scales, where Corresponding to time scale The input sequence on, This indicates the length of time at that scale. This indicates the dimension of the input variable. Here, Indicates the number of scale divisions. This represents the maximum number of input sequences at a specific time scale.

[0028] The prediction task is defined as a mapping relationship: (2) This results in future predictions: (3) in Indicates the length of the historical input. Indicates the prediction time domain, Indicates the predicted sequence. This specifies the dimension of the output variable.

[0029] II. Transformer Architecture Transformer is a sequence modeling framework based on self-attention mechanism, which can effectively model long-term temporal dependencies.

[0030] 1) Self-attention mechanism: Based on the historical input sequence X, the self-attention mechanism projects the input through linear projection. Mapped to query vector, key vector, and value vector matrices: (4) Among them, and It is a learnable projection matrix. The attention output is calculated as follows: (5) Wherein, the dimension of the key vector is .

[0031] 2) Multi-head self-attention mechanism: To enhance representation capabilities, this mechanism is implemented by executing multiple attention operations in parallel. (6) in, , and It is the first The projection matrices of each head are used. The outputs of all heads are concatenated and then linearly transformed. (7) in Indicates the number of attention heads. This is a learnable output matrix.

[0032] III. Fuzzy Measure and Fuzzy Integral Fuzzy measures generalize classical additive measures by relaxing the additivity requirement to monotonicity, thus providing a flexible framework for modeling uncertainty and interaction relationships among multiple information sources. Given an input sequence... Fuzzy measure is defined as a set function: (8) It satisfies the following properties: Boundary conditions: (9) Monotonicity: For any, : (10) Based on fuzzy measures, the Choquet fuzzy integral provides a nonlinear aggregation operator. Let the features be ordered as follows: (11) and Then the Choquet fuzzy integral is defined as: (12) This can be equivalently represented as: (13) in, express The measure function, Let L represent the i-th input, and L be the maximum number of inputs.

[0033] Example 1 This invention provides a method for predicting parameters of a supercritical circulating fluidized bed boiler based on attention-guided fuzzy fusion, comprising: The time series of operating parameters of the collected circulating fluidized bed boiler unit are divided into multiple scales and embedded with features to obtain the feature representations corresponding to each scale. Based on the aforementioned feature representation, a dual attention mechanism is employed to obtain local temporal-dependent features and cross-scale attention features; Based on the local time-dependent features and the cross-scale attention features, an attention interaction mechanism and a feature encoding network are used to evaluate reliability and obtain confidence parameters and fuzziness measures. Based on the confidence parameter and the fuzzy measure value, a nonlinear fusion method based on fuzzy measure is used to obtain the fused feature representation; Based on the fusion feature representation, time series prediction results of the operating parameters of the circulating fluidized bed boiler unit are obtained.

[0034] like Figure 1 As shown, the specific implementation process of the present invention is as follows: This invention proposes an attention-guided fuzzy multi-scale Transformer framework for time series prediction. This framework employs a dual-attention mechanism consisting of intra-temporal attention and cross-scale attention to jointly capture fine-grained local temporal dependencies within a single scale and coarse-grained global dependencies between heterogeneous time scales. Based on attention-enhanced representations, this invention introduces an attention-guided confidence estimation method to quantify the reliability of scale-specific features. Subsequently, these confidence cues are integrated into a fuzzy fusion strategy based on learned fuzzy measures and Choquet integrals, achieving adaptive nonlinear aggregation of multi-scale information to complete the prediction task. The following describes the process in two steps: A. Dual attention mechanism for multi-scale temporal modeling 1) Multi-scale partitioning: Based on the multi-scale partitioning input defined in formula (1), each scale-specific sequence is generated. The features are independently projected into the latent feature space, i.e., feature embedding is performed, which is represented as: (14) in, The parameter matrix is ​​a learnable embedding layer that projects the input features to... A 3D latent space. The resulting feature representations. While encoding the original observations into a shared latent space, the time scale is preserved. Time resolution .

[0035] To facilitate attention-based interaction and fusion across different timescales, this invention further introduces a time alignment step, projecting scale-specific representations onto a unified standard space. Specifically, each embedded representation is linearly projected along the time dimension to a space of dimension 1. Public extended space: (15) in, This is a learnable temporal projection matrix shared across all scales. The resulting feature representation... Aligning scale-specific representations to a unified standard time space enables consistent attention-based interactions across heterogeneous timescales.

[0036] 2) Intra-scale attention mechanism: Local dependency modeling The intra-scale temporal attention module aims to independently model fine-grained, short-term temporal dependencies within each timescale. Since the attention mechanism operates in a self-attention manner within a single scale, all attention computations are restricted to the embedding sequence corresponding to that scale.

[0037] The feature representation after feature alignment is defined in Equation (15). As a common input source for the projection of queries, keys, and values. Following the standard self-attention form shown in Equation (4), the projection of intra-scale temporal attention is calculated as follows: (16) in, This is a learnable projection matrix specific to the intra-scale temporal attention module. Accordingly, All located in In space.

[0038] According to equations (5)-(6), the scaling dot product self-attention output at each scale can be calculated as follows: (17) Among them, the attention mechanism enables each time step to selectively aggregate information from other time locations within the same scale.

[0039] Following the multi-head attention modeling approach in equation (7), the outputs of all attention heads are concatenated and linearly transformed to obtain the final intra-scale temporal attention representation: (18) in This is the learnable output projection matrix. The calculated... By achieving effective information aggregation at each time scale, locally enhanced temporal feature representations are captured.

[0040] 3) Cross-scale attention mechanism: global dependency modeling To capture global dependencies and interactions across different time scales, this invention introduces a cross-scale attention module that operates on scale-level representations.

[0041] According to the input time series defined in equation (1) This invention first projects it onto a unified latent feature space through a linear transformation: (19) in It is a learnable projection matrix. The projection results in... While maintaining the original temporal resolution, the original observations are embedded into a semantically aligned feature space.

[0042] based on The query, key, and value matrix is ​​calculated as follows: (20) in, It is a learnable projection matrix.

[0043] Finally, the global attention output across scales is calculated as follows: (twenty one) B. Attention-guided fuzzy fusion While the proposed dual-attention mechanism enables adaptive feature selection across time dimensions and scales, it does not explicitly quantify the reliability of representations at specific scales. To address this limitation, this invention proposes an attention-guided fuzzy fusion strategy: treating scale confidence as a reference and modeling its relative importance using learned fuzzy measures. The final fusion process borrows from the Choquet fuzzy integral form to achieve nonlinear interaction between multi-scale representations.

[0044] For each time scale This invention first introduces cross-scale attention as a global modulation factor, and interacts with the attention head element-wise within the corresponding time period: (twenty two) in, Used to characterize the fine-grained consistency between scale-specific attention and global attention.

[0045] The original attention features and their interactions are then concatenated and input into the residual block for joint encoding: (twenty three) (twenty four) in This indicates the residual coding module. For parameters The controllable learnable nonlinear transformation is specifically implemented by stacked convolutional layers and their associated normalization and activation functions. This represents the residual encoding depth feature at the i-th time scale. This represents the comprehensive splicing features at the i-th time scale.

[0046] Based on the encoded representation, a further evidence function is learned to quantize the first... Confidence level at each time scale: (25) in, Representing encoded features To confidence parameter The mapping function, This represents the convolutional projection operator used to aggregate channel-level information. This represents a global pooling operation used to aggregate spatial features. For the first i A confidence parameter for reliability at each time scale, used to reflect the time scale. Overall reliability of temporal attention representations at lower scales.

[0047] At the same time, this invention learns a single-element fuzzy measure to characterize the relative importance at different time scales: (26) in, This represents the sigmoid function. Representing encoded features To single-element fuzzy measure The mapping function.

[0048] Finally, the confidence-aware time-scale representation is fused using Choquet fuzzy integrals: (27) in, This indicates the number of rows after sorting in descending order of confidence level. i Each scale weight, is a virtual confidence parameter, with a fixed value of 0, used to simplify Choquet fuzzy integral calculation; M represents the total number of time scales for multi-scale input. This represents the fuzzy measure value of the set formed by the first i time scales after sorting in descending order of confidence parameters.

[0049] Based on the fused reliability-aware representation, the final prediction sequence is obtained as follows: (28) in, For time series prediction results, This represents a prediction network used to map fused features to a target prediction space. This is a local time-dependent feature.

[0050] In summary, this invention discloses a multi-scale time series prediction method for key parameters of supercritical circulating fluidized bed boilers based on attention-guided fuzzy fusion, applicable to the prediction of operating parameters such as bed temperature, power generation, and steam flow rate of 350MW supercritical circulating fluidized bed units. This method first divides the unit's operating time series into multiple scales and embeds features, then maps features from different time scales to a unified, standardized time space through time alignment. Subsequently, it performs intra-scale attention modeling within each time scale to extract local time-dependent features, and obtains global correlations between different time scales through cross-scale attention modeling. 。 Building upon this foundation, an attention interaction mechanism and a feature encoding network are introduced to evaluate the reliability of multi-scale features (acquired local time-dependent features and cross-scale attention features), generating confidence parameters and fuzzy measure values ​​corresponding to each time scale. Furthermore, a nonlinear fusion method based on fuzzy measure is employed to fuse the multi-scale features, obtaining a fused feature representation, and outputting the predicted results of the unit's operating parameters accordingly. This invention can comprehensively characterize the dynamic characteristics of circulating fluidized bed units at multiple time scales during operation, improving the accuracy and robustness of bed temperature, power generation, and steam flow prediction, providing technical support for unit operation optimization and safety control.

[0051] Example 2 This embodiment uses specific experimental data to illustrate the specific implementation process and effects of the method described in the foregoing embodiments.

[0052] The dataset used in this embodiment was collected from a 350 MW supercritical circulating fluidized bed (CFB) power plant in a certain city. This dataset contains 60,000 samples recorded from the plant's distributed control system (DCS), with a fixed sampling interval of 10 seconds. The data covers various operating conditions, including load changes, fuel characteristic fluctuations, and routine disturbances, realistically reflecting the actual operating dynamics of the CFB boiler. The input vector contains 12 key operating variables related to the combustion and heat transfer processes, while the output vector corresponds to three key performance indicators: bed temperature, steam pressure, and power output. All variables are time-synchronized. An input matrix is ​​constructed. Output matrix To maintain temporal causality and avoid information leakage, the dataset was divided into a training set (70%), a validation set (10%), and a test set (20%) in chronological order.

[0053] This invention conducts multi-time-domain prediction experiments on three core operating parameters—bed temperature, steam flow rate, and unit load—using actual operating datasets of circulating fluidized bed (CFB) boilers. The core evaluation indicators are the mean absolute error (MAE, with lower values ​​indicating higher accuracy) and the root mean square error (RMSE, with lower values ​​indicating higher accuracy). Experimental results show that the model of this invention achieves high-precision predictions in all prediction time domains. The core accuracy data are as follows: Bed temperature (unit: °C): 1-step short-term forecast MAE is 0.62, RMSE is 0.95; 24-step medium-term forecast MAE is 2.33, RMSE is 3.36; 48-step medium-to-long-term forecast MAE is 3.97, RMSE is 5.96; Steam flow rate (unit: kg / s): 1-step short-term forecast MAE is 2.65, RMSE is 4.25; 24-step medium-term forecast MAE is 8.50, RMSE is 14.74; 48-step medium-to-long-term forecast MAE is 13.60, RMSE is 25.75; Unit load (unit: MW): 1-step short-term forecast MAE is 0.70, RMSE is 1.25; 24-step medium-term forecast MAE is 2.07, RMSE is 3.96; 48-step medium-to-long-term forecast MAE is 3.46, RMSE is 7.21.

[0054] The above data fully demonstrates that the model of this invention has extremely high accuracy in the short, medium and long term prediction of the core operating parameters of circulating fluidized bed boilers, can accurately capture the complex nonlinear dynamic characteristics of boiler combustion, and meet the parameter prediction needs of actual industrial production.

[0055] Example 3 Based on the same inventive concept, the present invention also provides a supercritical circulating fluidized bed boiler parameter prediction system based on attention-guided fuzzy fusion, used to implement the method described in the foregoing embodiments. The system includes: a multi-scale partitioning module, a dual attention module, a reliability assessment module, a fuzzy fusion module, and a prediction module. The multi-scale segmentation module is used to perform multi-scale segmentation and feature embedding on the time series of collected operating parameters of the circulating fluidized bed boiler unit to obtain the feature representation corresponding to the scale. The dual attention module is used to obtain local time-dependent features and cross-scale attention features based on the feature representation using a dual attention mechanism; The reliability assessment module is used to perform reliability assessment based on the local time-dependent features and the cross-scale attention features, using an attention interaction mechanism and a feature encoding network, to obtain confidence parameters and fuzzy measure values. The fuzzy fusion module is used to obtain a fusion feature representation based on the confidence parameter and the fuzzy measure value using a nonlinear fusion method based on fuzzy measure. The prediction module is used to obtain time series prediction results of the operating parameters of the circulating fluidized bed boiler unit based on the fused feature representation.

[0056] Furthermore, in this embodiment, the dual attention module includes: a time alignment unit, a local unit, and a global unit; The time alignment unit is used to map the feature representation to a unified time space through time alignment. The local unit is used to apply the intra-scale attention mechanism in the dual attention mechanism to the feature representation within a unified and standardized time space to obtain local time-dependent features. The global unit is used to apply the cross-scale attention mechanism in the dual attention mechanism to the feature representation within a unified time and space framework to obtain cross-scale attention features.

[0057] Furthermore, in this embodiment, the reliability assessment module, based on the local time-dependent features and the cross-scale attention features, employs an attention interaction mechanism and a feature encoding network to perform reliability assessment, and obtains the confidence parameter and fuzzy measure value through the following method: ; ; ; ; ; in, For interactive features, For the first i Time scale Scaling dot product self-attention on, For cross-scale attention features, For the input time series, Indicates splicing, This indicates the residual coding module. For parameters Learnable nonlinear transformations of control, This represents the residual encoding depth feature at the i-th time scale. This represents the comprehensive splicing features at the i-th time scale. Indicates the first i A confidence parameter for reliability over a time scale. Representing encoded features To confidence parameter The mapping function, This indicates a global pooling operation. This represents the convolutional projection operator used to aggregate channel-level information. For fuzzy measure values, This represents the sigmoid function.

[0058] Furthermore, in this embodiment, the fuzzy fusion module obtains the fused feature representation by adopting a nonlinear fusion method based on fuzzy measure, based on the confidence parameter and the fuzzy measure value, including: Computational fusion feature representation: ; in, For feature representation, This indicates the number of rows after sorting in descending order of confidence level. i Each time scale weight, and let , This represents the maximum number of input sequences at a specific time scale. It represents the fuzzy measure value of the set composed of the first i sorted time scales.

[0059] Furthermore, in this embodiment, the method by which the prediction module obtains the time series prediction results of the operating parameters of the circulating fluidized bed boiler unit based on the fused feature representation includes: ; in, For time series prediction results, This represents a prediction network used to map fused features to a target prediction space. This is a local time-dependent feature.

[0060] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for predicting parameters of a supercritical circulating fluidized bed boiler based on attention-guided fuzzy fusion, characterized in that, The method includes: The time series of operating parameters of the collected circulating fluidized bed boiler unit are divided into multiple scales and embedded with features to obtain the feature representations corresponding to each scale. Based on the aforementioned feature representation, a dual attention mechanism is employed to obtain local temporal-dependent features and cross-scale attention features; Based on the local time-dependent features and the cross-scale attention features, an attention interaction mechanism and a feature encoding network are used to evaluate reliability and obtain confidence parameters and fuzziness measures. Based on the confidence parameter and the fuzzy measure value, a nonlinear fusion method based on fuzzy measure is used to obtain the fused feature representation; Based on the fusion feature representation, time series prediction results of the operating parameters of the circulating fluidized bed boiler unit are obtained.

2. The method according to claim 1, characterized in that: Based on the aforementioned feature representation, the method for obtaining local temporal-dependent features and cross-scale attention features using a dual attention mechanism includes: The feature representation is time-aligned and mapped to a unified and standardized temporal feature space. Within a unified and standardized time and space, the feature representation is processed using the in-scale attention mechanism within the dual attention mechanism to obtain local time-dependent features; Within a unified and standardized time and space, the feature representation is processed using a cross-scale attention mechanism within a dual attention mechanism to obtain cross-scale attention features.

3. The method according to claim 1, characterized in that, Based on the local time-dependent features and the cross-scale attention features, and through element-level interactive operations in conjunction with a residual feature encoding network, scale reliability quantification and scale relative importance modeling are carried out respectively, thereby obtaining the confidence parameters and fuzzy measure values ​​corresponding to each scale. Specifically, the steps include: ； ； ； ； ； in, For interactive features, For the first i Time scale Scaling dot product self-attention on, This represents the element-level Hadamard product operation. For cross-scale attention features, For the input time series, This indicates a feature concatenation operation. This indicates the residual coding module. For parameters Learnable nonlinear transformations of control, This represents the residual encoding depth feature at the i-th time scale. This represents the comprehensive splicing features at the i-th time scale. Indicates the first i A confidence parameter for reliability over a time scale. Representing encoded features To confidence parameter The mapping function, This indicates a global pooling operation. and This represents a parameter-independent convolutional projection operator used to aggregate channel-level information. To represent the fuzzy measure of the relative importance of the i-th time scale, This represents the sigmoid function.

4. The method according to claim 3, characterized in that, Based on the confidence parameter and the fuzzy measure value, a method for obtaining fused feature representations using a fuzzy measure-based nonlinear fusion approach includes: Computational fusion feature representation: ; in, For feature representation, This indicates the number of rows after sorting in descending order of confidence level. i Each time scale weight, and let , This represents the maximum number of input sequences at a specific time scale. It represents the fuzzy measure value of the set composed of the first i sorted time scales.

5. The method according to claim 4, characterized in that, The method for obtaining time series prediction results of operating parameters of circulating fluidized bed boiler units based on the fused feature representation includes: ； in, For time series prediction results, This represents a prediction network used to map fused features to a target prediction space. This is a local time-dependent feature.

6. A parameter prediction system for a supercritical circulating fluidized bed boiler based on attention-guided fuzzy fusion, the system being used to implement the method described in any one of claims 1-5, characterized in that, The system includes: a multi-scale partitioning module, a dual attention module, a reliability assessment module, a fuzzy fusion module, and a prediction module; The multi-scale segmentation module is used to perform multi-scale segmentation and feature embedding on the time series of collected operating parameters of the circulating fluidized bed boiler unit to obtain the feature representation corresponding to the scale. The dual attention module is used to obtain local time-dependent features and cross-scale attention features based on the feature representation using a dual attention mechanism; The reliability assessment module is used to perform scale reliability quantification and relative importance modeling based on the local time dependency features and the cross-scale attention features, using element-level interactive operation and residual feature encoding network, to obtain confidence parameters and fuzzy measure values ​​for each scale. The fuzzy fusion module is used to obtain a fusion feature representation based on the confidence parameter and the fuzzy measure value using a fuzzy integral nonlinear fusion method based on fuzzy measure. The prediction module is used to obtain time series prediction results of the operating parameters of the circulating fluidized bed boiler unit based on the fused feature representation.

7. The system according to claim 6, characterized in that, The dual attention module includes: temporally aligned units, local units, and global units; The time alignment unit is used to perform time alignment processing on the feature representation, mapping the feature representation to a unified and standardized time space; The local unit is used to apply the intra-scale attention mechanism in the dual attention mechanism to the feature representation within a unified and standardized time space to obtain local time-dependent features. The global unit is used to apply the cross-scale attention mechanism in the dual attention mechanism to the feature representation within a unified time and space framework to obtain cross-scale attention features.

8. The system according to claim 6, characterized in that, The reliability assessment module, based on the local time-dependent features and the cross-scale attention features, employs an attention interaction mechanism and a feature encoding network to perform reliability assessment, and the method for obtaining the confidence parameter and fuzzy measure value includes: ； ； ； ； ； in, For interactive features, For the first i Time scale Scaling dot product self-attention on, This represents the element-level Hadamard product operation. For cross-scale attention features, For the input time series, This indicates a feature concatenation operation. This indicates the residual coding module. For parameters Learnable nonlinear transformations of control, This represents the residual encoding depth feature at the i-th time scale. This represents the comprehensive splicing features at the i-th time scale. Indicates the first i A confidence parameter for reliability over a time scale. Representing encoded features To confidence parameter The mapping function, This indicates a global pooling operation. and This represents a parameter-independent convolutional projection operator used to aggregate channel-level information. To represent the fuzzy measure of the relative importance of the i-th time scale, This represents the sigmoid function.

9. The system according to claim 8, characterized in that, The fuzzy fusion module, based on the confidence parameter and the fuzzy measure value, employs a nonlinear fusion method based on fuzzy measure to obtain the fused feature representation. The method includes: Computational fusion feature representation: ; in, For feature representation, This indicates the number of rows after sorting in descending order of confidence level. i Each time scale weight, and let , This represents the maximum number of input sequences at a specific time scale. It represents the fuzzy measure value of the set composed of the first i sorted time scales.

10. The system according to claim 9, characterized in that, The method for obtaining time series prediction results of circulating fluidized bed boiler unit operating parameters based on the fused feature representation includes: ； in, For time series prediction results, This represents a prediction network used to map fused features to a target prediction space. This is a local time-dependent feature.

Images