A multi-objective prediction method and system for biomass gas coupling

Abstract

The present application relates to the technical field of industrial boiler, in particular to a kind of multi-objective prediction method and system coupled with biomass gas.The method collects multi-source operation data, carries out time alignment and resampling to generate mode characterization vector;Based on historical characterization vector clustering, divide K operation modes, and train corresponding expert model to constitute multi-objective prediction expert group respectively;Current characterization vector is input into gate network to obtain basic gate weight, and mode migration jitter is inhibited by double threshold hysteresis judgment and shortest residence time constraint, and final gate weight is generated by weight smoothing;The final gate weight is used to weight and fuse the multi-objective results output by each expert, to obtain continuous and stable prediction results.The method can improve the prediction accuracy and stability under complex working conditions, enhance the adaptability to working condition change and mode evolution, and has good engineering application value.

Description

Technical Field

[0001] This invention relates to the field of industrial boiler technology, specifically to a multi-objective prediction method and system for biomass gas coupling. Background Technology

[0002] Biomass gasification coupled with coal-fired boilers is a power generation method that converts biomass into syngas, mainly composed of CO and H2, in a gasifier, and then feeds it into a coal-fired boiler for co-combustion with coal. This method can achieve carbon reduction and efficiency improvement without drastically modifying existing coal-fired power facilities. However, in engineering operation, biomass feedstocks are characterized by dispersion, seasonality, and heterogeneity of composition. Fluctuations in moisture, ash, and alkali metals can cause changes in gasification temperature, syngas calorific value, and composition, thereby affecting boiler combustion stability, the risk of slagging and corrosion on heating surfaces, and pollutant generation behavior. This results in a system exhibiting characteristics such as multivariability, strong nonlinearity, strong coupling, and large time lag. The biomass gasification or combustion process is a complex system with multivariable, nonlinear, and large time lag. Traditional control and experience-based parameter tuning are insufficient to balance multiple objectives such as efficiency and environmental protection. There is an urgent need to improve operational optimization capabilities by leveraging intelligent prediction and data-driven methods.

[0003] In existing technologies, modeling and prediction of biomass gasification and coupled combustion typically employ mechanistic models or data-driven models, and a hybrid modeling framework integrating mechanism and data has been gradually developed to compensate for the shortcomings of single models in terms of extrapolation capability, interpretability, or accuracy. However, coupled systems can exhibit multiple operating modes under different fuel batches, blending ratios, loads, and air distribution strategies, with significant differences in data distribution across these modes. Furthermore, the diverse sources of field data, inconsistent sampling frequencies, and varying quality, along with insufficient data standardization and sharing, can all affect the reliability of model training and online applications.

[0004] To improve prediction accuracy, the industry often employs strategies such as segmented modeling based on operating conditions or different models, or multi-model fusion. However, near model boundaries, model selection and fusion weights are susceptible to frequent switching due to noise and short-term fluctuations, resulting in prediction output jitter, spikes, or short-term distortion. Existing methods often lack hysteresis determination, dwell constraints, and smooth transition design to address model migration jitter, making it difficult to guarantee the continuity and stability of multi-objective prediction results under dynamic operating conditions. Furthermore, insufficient sample size for a few models and the occurrence of operating condition drift / new sub-models can also cause model degradation, reducing the robustness of long-term online prediction.

[0005] Therefore, a technical solution is needed that can adapt to multiple data sources and multiple operating modes, suppress mode switching jitter, and at the same time take into account the stability of multi-target prediction. Summary of the Invention

[0006] The technical problem this invention aims to solve is to address the shortcomings of existing technologies by providing a multi-objective prediction method and system for biomass gasification coupling. This method involves collecting multi-source operational data from biomass gasification systems and coal-fired boiler systems, performing time alignment and resampling to generate pattern representation vectors. Based on historical representation vectors, K operational modes are clustered, and corresponding expert models are trained to form a multi-objective prediction expert group. The current representation vector is input into a gating network to obtain basic gating weights. Double-threshold hysteresis judgment and shortest residence time constraints are used to suppress mode migration jitter, and the weights are smoothed to generate the final gating weights. The final gating weights are then used to weightedly fuse the multi-objective results output by each expert to obtain continuous and stable prediction results. This method can improve prediction accuracy and stability under complex operating conditions, enhance adaptability to changes in operating conditions and pattern evolution, and has good engineering application value.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A multi-objective prediction method coupled with biomass gas, the method comprising:

[0009] Collect multi-source operational data, perform time alignment and resampling on the multi-source operational data, and generate pattern representation vectors;

[0010] The pattern representation vectors of historical multi-source operation data are divided into patterns to generate an operation pattern set. Based on each operation pattern in the operation pattern set, the corresponding expert model is trained to construct a multi-objective prediction expert group.

[0011] The current mode representation vector is input into a preset gating network, and the final gating weights are output. The preset gating network generates the basic gating weights of the multi-objective prediction expert group based on the current mode representation vector, and performs hysteresis correction on the basic gating weights. The hysteresis correction suppresses mode migration jitter through dual threshold judgment and minimum residence time constraint, and generates the final gating weights.

[0012] The current mode representation vector is input into each expert model of the multi-objective prediction expert group. The multi-objective prediction results output by each expert model are weighted and fused through the final gating weight to generate the final multi-objective prediction vector.

[0013] The process involves dividing the pattern representation vectors of historical multi-source operational data into patterns to generate an operational pattern set. Based on each operational pattern in the operational pattern set, a corresponding expert model is trained to construct a multi-objective prediction expert group, including:

[0014] Cluster analysis is performed on the pattern representation vectors of historical multi-source operation data to divide the pattern representation vectors into K operation modes, thus obtaining a set of operation modes;

[0015] For each operating mode in the set of operating modes, extract the historical training samples corresponding to that operating mode and train the expert model for that operating mode, wherein the historical training samples include the mode representation vector corresponding to the operating mode;

[0016] The K trained expert models are combined to construct the multi-objective prediction expert group.

[0017] For each operating mode in the set of operating modes, extract the historical training samples corresponding to that operating mode, and train an expert model for that operating mode, including:

[0018] For each operating mode in the set of operating modes, an expert model for that operating mode is constructed. The expert model is based on a neural network, with the input being the mode representation vector within the time window and the output being the multi-objective prediction result. The structure of the neural network includes a shared feature extraction layer and K mode output layers. The shared feature extraction layer is used by all operating modes to extract common features of the input data, and each mode output layer corresponds to one operating mode in the set of operating modes.

[0019] The number of historical training samples corresponding to each operating mode in the set of operating modes is counted. When the number of historical training samples of the target operating mode is less than the preset sample threshold, a loss weight higher than the benchmark value is set for the training samples of the target operating mode.

[0020] The historical training samples corresponding to each operating mode are input into the shared feature extraction layer for feature extraction. The extracted features are then input into the corresponding mode output layer to output the multi-target prediction results.

[0021] The weighted loss value is calculated based on the loss weight and the actual label value, and the parameters of the shared feature extraction layer and each mode output layer are updated based on the weighted loss value.

[0022] The basic gating weights are then subjected to hysteresis correction. This hysteresis correction suppresses mode migration jitter through dual threshold determination and minimum dwell time constraints, generating the final gating weights, including:

[0023] For each operating mode in the set of operating modes, an entry threshold and an exit threshold are set, wherein the entry threshold is greater than the exit threshold;

[0024] Obtain the current operating mode at the current moment, and perform a mode transition determination based on the weight value corresponding to each operating mode in the basic gating weight to determine whether a mode transition is triggered.

[0025] When a mode transition is triggered, a minimum mode dwell time is set, and mode transitions are prohibited from being executed again within the minimum mode dwell time.

[0026] The basic gating weights are smoothed to generate the final gating weights.

[0027] Based on the weight values ​​corresponding to each operating mode in the basic gating weights, a mode transition determination is performed to determine whether a mode transition is triggered, including:

[0028] Extract the first weight value corresponding to the current operating mode from the basic gating weights, and determine whether the first weight value is less than or equal to the exit threshold of the current operating mode;

[0029] When the first weight value is less than or equal to the exit threshold, count the number of exit sampling periods that continuously satisfy the condition that the first weight value is less than or equal to the exit threshold.

[0030] When the number of exit sampling cycles reaches the preset number of exit duration cycles, the candidate running mode with the largest weight value is selected from the basic gating weights, and it is determined whether the second weight value corresponding to the candidate running mode is greater than or equal to the entry threshold of the candidate running mode.

[0031] When the second weight value is greater than or equal to the entry threshold, the number of entry sampling periods that continuously satisfy the second weight value being greater than or equal to the entry threshold is counted.

[0032] When the number of sampling cycles enters reaches the preset number of continuous cycles, a mode transition from the current operating mode to the candidate operating mode is triggered.

[0033] Smoothing is performed on the basic gating weights to generate the final gating weights, including:

[0034] When no mode transition is triggered, the base gating weight at the current moment is weighted and averaged with the final gating weight at the previous moment to generate the final gating weight at the current moment.

[0035] When a mode transition is triggered, the basic gating weights corresponding to the current operating mode and the candidate operating modes are retained, the basic gating weights corresponding to the other operating modes are set to zero, the retained basic gating weights are normalized, and transition weights are generated.

[0036] The transition weights are weighted and averaged with the final gating weights of the previous time step to generate the final gating weights of the current time step.

[0037] The step of inputting the current mode representation vector into each expert model of the multi-objective prediction expert group, and then weighting and fusing the multi-objective prediction results output by each expert model through the final gating weights to generate the final multi-objective prediction vector includes:

[0038] The current mode representation vector is input into the K expert models in the multi-objective prediction expert group to obtain the multi-objective prediction results output by each expert model, wherein each multi-objective prediction result contains m predicted target values.

[0039] Extract the predicted weight values ​​corresponding to each expert model from the final gating weights;

[0040] For each of the m prediction targets, the prediction target value output by the K expert models is multiplied by the corresponding prediction weight value and then summed to obtain the fused prediction value of the prediction target.

[0041] The fused prediction values ​​of m predicted targets are combined to generate the final multi-target prediction vector.

[0042] The m predicted targets include at least two of the following: nitrogen oxide emission concentration, carbon monoxide emission concentration, main steam temperature deviation, flue gas temperature, and boiler efficiency proxy value.

[0043] The method further includes:

[0044] Obtain the actual measurement value corresponding to the final multi-objective prediction vector, calculate the prediction residual, and generate the health index of the current operating mode.

[0045] When the health index is continuously lower than the preset health threshold within a preset time period, the multi-source operation data and pattern representation vector within the preset time period are marked as new sub-pattern candidate data.

[0046] Based on the new sub-mode candidate data, incremental training is performed to update the expert model corresponding to the current operating mode.

[0047] A multi-objective prediction system for biomass-gas coupling is provided to implement the aforementioned multi-objective prediction method for biomass-gas coupling. The system includes a data acquisition module, a model training module, a gating correction module, and a multi-objective prediction module, wherein:

[0048] The data acquisition module is used to acquire multi-source operational data, perform time alignment and resampling on the multi-source operational data, and generate a pattern representation vector.

[0049] The model training module is used to divide the pattern representation vector of historical multi-source operation data into patterns, generate a set of operation patterns, train the corresponding expert model based on each operation pattern in the set of operation patterns, and construct a multi-objective prediction expert group.

[0050] The gating correction module is used to input the current mode representation vector into a preset gating network and output the final gating weights. The preset gating network generates the basic gating weights of the multi-objective prediction expert group based on the current mode representation vector and performs hysteresis correction on the basic gating weights. The hysteresis correction suppresses mode migration jitter through dual threshold judgment and minimum residence time constraint to generate the final gating weights.

[0051] The multi-objective prediction module is used to input the current mode representation vector into each expert model of the multi-objective prediction expert group, and to perform weighted fusion of the multi-objective prediction results output by each expert model through the final gating weight to generate the final multi-objective prediction vector.

[0052] Compared with the prior art, the beneficial effects of the present invention are:

[0053] 1. This invention obtains K operating modes based on historical pattern representation vector clustering, and trains corresponding expert models for each mode, avoiding the decrease in accuracy caused by compromise fitting of a single model under multimodal data distribution, and enhancing the ability to characterize the differences in operating conditions such as different fuel batches, blending ratios, loads and air distribution strategies.

[0054] 2. After the gating network outputs the basic gating weights, this invention introduces a dual-threshold hysteresis judgment and a minimum dwell time constraint to avoid frequent switching at the mode boundary caused by noise and short-term fluctuations; and performs smoothing processing on the weights to achieve "soft switching" of the weights, thereby significantly reducing the spikes, jitter and short-term distortion of the prediction results. Attached Figure Description

[0055] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0056] Figure 1 This is a flowchart illustrating a multi-objective prediction method for biomass gas coupling according to an embodiment of this application;

[0057] Figure 2 This is a schematic diagram of the structure of the multi-objective prediction expert group model in an embodiment of this application;

[0058] Figure 3 This is a schematic diagram of the gating hysteresis correction process in an embodiment of this application;

[0059] Figure 4 This is a schematic diagram of the structure of a biomass gas coupled multi-objective prediction system according to an embodiment of this application. Detailed Implementation

[0060] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0061] Example 1

[0062] Please see Figure 1 This invention provides an embodiment: a multi-objective prediction method for biomass gas coupling, applied to a biomass gas coupled coal-fired boiler comprising a biomass gasification system and a coal-fired boiler system. The specific steps of the method are as follows:

[0063] S1: Collect multi-source operational data, perform time alignment and resampling on the multi-source operational data, and generate a pattern representation vector.

[0064] Specifically, the biomass gasification coupled coal-fired boiler system consists of a biomass gasification system and a coal-fired boiler system. These two systems differ significantly in their process flow, control strategies, and measurement point layout. Their operational data exhibits diverse sources, inconsistent sampling frequencies, and complex time lag relationships. Directly modeling the raw, multi-source operational data can easily lead to misjudgments due to inconsistent time scales or instantaneous fluctuation noise, thus affecting the stability and accuracy of subsequent operational mode division and multi-objective prediction. Therefore, this embodiment first performs systematic preprocessing and structured characterization of the multi-source operational data to construct a pattern representation vector that reflects the overall operational status of the system.

[0065] Furthermore, considering the time lag effect in the physical coupling between the biomass gasification system and the coal-fired boiler system—such as the time required for changes in biomass gas composition to be transmitted to the combustion state in the furnace—this embodiment introduces a unified time reference in the data preprocessing stage to perform time alignment processing on multi-source operating data. Specifically, using a preset unified sampling period as the target time axis, interpolation, aggregation, or downsampling operations are performed on the data from each source, so that data with different sampling frequencies are mapped to the same time series index, ensuring that features of each dimension correspond to the same system operating state.

[0066] Furthermore, to suppress the interference of field measurement noise and short-term random disturbances on pattern recognition, resampling processing is performed on multi-source operating data after time alignment. A sliding time window method is preferred, in which the mean, weighted mean or statistical feature extraction is performed on the corresponding variables within each window to obtain representative steady-state or quasi-steady-state feature expressions, effectively weakening the impact of occasional spikes or measurement anomalies on pattern representation.

[0067] After completing time alignment and resampling, the operating variables of different systems and with different physical meanings are concatenated according to the preset feature order to construct a unified-dimensional pattern representation vector, which is used to characterize the multi-dimensional operating features at the current moment. This provides a unified, stable and discriminative feature basis for operating mode division, gating network weight calculation and multi-objective prediction expert model input.

[0068] Through the above steps, this embodiment achieves structured modeling of the complex operating state of biomass gas coupled with coal-fired boilers at the data level, transforming multi-source operating data into pattern representation vectors that reflect the differences in system operating modes, thereby improving the reliability and engineering applicability of pattern recognition and multi-objective prediction processes.

[0069] S2: Divide the pattern representation vectors of historical multi-source operation data into patterns to generate a set of operation patterns. Train the corresponding expert model based on each operation pattern in the set of operation patterns to construct a multi-objective prediction expert group.

[0070] Specifically, in the actual operation of biomass gas coupled with coal-fired boilers, the system's operating state frequently switches depending on factors such as the batch and blending ratio of biomass feedstock. The coupling relationships of various operating variables differ significantly under different operating conditions. Using a single, unified model to model all historical data is prone to declining model fitting ability due to multi-peak data distribution, making it difficult to guarantee prediction accuracy and stability during extreme operating conditions or mode switching phases. On the one hand, modeling without distinguishing operating modes as a whole results in insufficient characterization of individual operating states due to the trade-offs between various operating mechanisms in the model parameters. On the other hand, relying solely on coarse-grained classification of operating conditions based on manual experience makes it difficult to capture implicit operating modes and easily misses representative operating sub-states. Therefore, this embodiment proposes a data-driven mode segmentation method based on mode representation vectors from historical multi-source operating data to achieve adaptive identification and structured modeling of complex operating states.

[0071] In this embodiment, the pattern representation vectors corresponding to historical multi-source operational data are first clustered and analyzed. Based on their distribution characteristics in the feature space, they are divided into K operational modes, forming a set of operational modes. Each operational mode corresponds to an operational state interval with similar characteristics, ensuring high consistency in statistical distribution and physical meaning of data within the same operational mode. After completing the operational mode division, historical sample data is extracted for each operational mode, and the corresponding pattern representation vector is used as input to train an expert model. In this way, each expert model can focus on characterizing the system behavior features under the corresponding operational mode, improving the targeting and accuracy of multi-objective prediction.

[0072] In this way, multiple expert models trained for different operating modes are combined into a multi-objective prediction expert group, so that the prediction task under different operating modes is completed by the most suitable expert model, laying the foundation for dynamic weight allocation and multi-model fusion of gating networks.

[0073] S3: Input the current mode representation vector into a preset gating network and output the final gating weights. The preset gating network generates the basic gating weights of the multi-objective prediction expert group based on the current mode representation vector and performs hysteresis correction on the basic gating weights. The hysteresis correction suppresses mode migration jitter through dual threshold judgment and minimum residence time constraint to generate the final gating weights.

[0074] Specifically, the operating state frequently evolves between different modes, especially under conditions of load fluctuations, where the mode representation vector changes rapidly within a short period. If expert model weights are assigned solely based on the current moment's vector, they are susceptible to noise or disturbances, leading to frequent weight switching and reducing the stability and reliability of the prediction results. Therefore, this embodiment introduces a hysteresis correction mechanism combined with continuous-time constraints into the gating network to achieve a dynamic balance between sensitive response and stable output. The basic gating weights are obtained by mapping and normalizing the current mode representation vector through a feedforward neural network from a preset gating network. Preferably, the basic gating weights can be output through a multi-layer fully connected network combined with a Softmax function to represent the instantaneous matching degree between the current operating condition and each operating mode.

[0075] This embodiment first inputs the current mode representation vector into the gating network and outputs the basic gating weights for each operating mode, reflecting the degree of matching between the vector and the mode and the theoretical adaptation priority of the expert model. However, these weights have no time continuity constraint and may fluctuate in the short term. To address this, this embodiment introduces hysteresis correction processing on the basic gating weights, sets entry and exit thresholds for each operating mode, and constructs a dual-threshold judgment interval to form a hysteresis band. When the weight is in the hysteresis band, the original mode is maintained, avoiding unnecessary migration.

[0076] Furthermore, after determining that the mode transition conditions are met, a shortest dwell time constraint mechanism is introduced. After the switch is completed, further transition determinations are prohibited within a preset time, ensuring the stability and continuity of the new mode and avoiding high-frequency back-and-forth switching, which conforms to practical engineering principles. After completing the dual threshold determination and time constraint, the basic gating weights are smoothed to generate the final gating weights. The smoothing process uses a weighted fusion of the current and previous time-stamped weights to ensure continuous change in the expert model weights, suppressing abrupt changes and guaranteeing a stable output of the prediction results.

[0077] Through the above mechanism, a dynamic weight adjustment process is constructed that takes into account both pattern recognition sensitivity and prediction output stability. The weights remain consistent during the stable or slowly changing phases, and the transition is smooth when the conditions are met during pattern transition, avoiding non-physical abrupt changes.

[0078] S4: Input the current mode representation vector into each expert model of the multi-objective prediction expert group, and use the final gating weight to weight and fuse the multi-objective prediction results output by each expert model to generate the final multi-objective prediction vector.

[0079] Specifically, the combustion mechanism, heat transfer characteristics, and pollutant generation patterns of the system differ significantly under different operating modes, and these differences are reflected in multiple prediction targets, such as nitrogen oxide emission concentration. Relying solely on the prediction results of a single expert model or a particular operating mode for the final output is prone to prediction distortion at mode boundaries or during operational transitions, failing to balance the continuity and accuracy of multi-target predictions. Therefore, this embodiment proposes a weighted fusion mechanism based on final gating weights, allowing each expert model to maintain its specialization while jointly participating in the generation of the final prediction result.

[0080] In this embodiment, the m predicted targets include at least two of the following: nitrogen oxide emission concentration, carbon monoxide emission concentration, main steam temperature deviation, flue gas temperature, and boiler efficiency proxy value. The multi-source operating data includes a first type of operating data for the biomass gasification system and a second type of operating data for the coal-fired boiler system. The first type of operating data includes biomass gas flow rate, biomass gas pressure, biomass gas temperature, valve opening, biomass gas composition, and gasifier operating parameters. The second type of operating data includes coal quantity, primary air volume, secondary air volume, furnace oxygen content, furnace carbon monoxide concentration, furnace nitrogen oxide concentration, main steam temperature, main steam pressure, reheat steam temperature, flue gas temperature, and load command. The multi-source operating data is typically collected by different monitoring and control systems, with sampling periods ranging from seconds to minutes or longer.

[0081] Furthermore, the current mode representation vector is first input into K expert models of the multi-objective prediction expert group. Each expert model outputs a set of multi-objective prediction results, ensuring the complete preservation of prediction information under different operating modes. Then, prediction weight values ​​corresponding to each expert model are extracted from the final gating weights. These prediction weight values ​​reflect the degree of matching between each operating mode and the current mode representation vector, and are stable weight results after processing, avoiding interference from short-term fluctuations on the fusion results. For each prediction objective, the corresponding prediction objective value output by the K expert models is weighted and summed with the prediction weight value to obtain the fused prediction value. Weighted fusion for each objective ensures that each prediction objective shares a consistent weight adjustment logic, maintaining the physical consistency and interpretability of the multi-objective prediction results. Finally, the fused prediction values ​​of m prediction objectives are combined to generate the final multi-objective prediction vector. This vector balances advantages in both time and space, and can more realistically reflect the overall operating performance of the boiler.

[0082] Through the aforementioned weighted fusion mechanism, this embodiment constructs a "soft switching" multi-model collaborative prediction process. When the operating mode is stable, the corresponding expert model dominates the prediction results; during the transition of operating states or the mixed mode stage, multiple expert models participate in the prediction together with different weights, so that the prediction output transitions smoothly, avoids non-physical abrupt changes, conforms to the continuous evolution law of the actual engineering system operating state, and improves the stability, robustness and engineering application value of multi-objective prediction results under complex dynamic conditions.

[0083] Please refer to Figure 2 The structural diagram of the multi-objective prediction expert group model provided in this application embodiment is shown below. The specific steps of S2 are as follows:

[0084] S2.1: Perform cluster analysis on the pattern representation vectors of historical multi-source operation data, divide the pattern representation vectors into K operation modes, and obtain a set of operation modes.

[0085] Specifically, during long-term operation of biomass gas coupled with coal-fired boilers, the diverse sources and fluctuating composition of biomass feedstock, along with frequent adjustments to boiler load, air distribution, and co-firing ratios, result in a multimodal distribution of system operation. The combustion organization and thermodynamic response mechanisms of the same equipment differ at different times. If historical operating states are not distinguished and modeling is directly performed using mixed-mode representation vectors, operational characteristics are easily interfered with, making it difficult for the model to accurately depict the real system behavior. Therefore, this embodiment introduces a clustering analysis method based on data distribution characteristics to perform unsupervised partitioning of historical pattern representation vectors, achieving objective identification of operating modes.

[0086] In this embodiment, the pattern representation vector comprehensively reflects the statistical characteristics and dynamic change characteristics of multi-source operating data within the time window. The pattern representation vectors corresponding to different operating modes usually exhibit a relatively concentrated distribution in the feature space.

[0087] In a preferred embodiment, a clustering analysis algorithm based on the center partitioning idea is used to cluster the pattern representation vectors. This clustering process uses a preset number of running modes K as the clustering target number, and iteratively optimizes the position of each cluster center, assigning each pattern representation vector to the running mode corresponding to the cluster center closest to its features. By repeatedly executing the iterative process of sample allocation and center update, the intra-cluster feature distribution gradually becomes more compact, and the inter-cluster differences gradually increase until the clustering results converge.

[0088] The K operating patterns obtained from cluster analysis constitute an operating pattern set, which is used to characterize the main state types of boiler historical operation. The number of operating patterns K can be preset or adaptively determined based on factors such as the scale of historical data, in order to balance the granularity of the patterns with the sufficiency of the samples.

[0089] S2.2: For each operating mode in the set of operating modes, extract the historical training samples corresponding to the operating mode and train the expert model for the operating mode, wherein the historical training samples include the mode representation vector corresponding to the operating mode.

[0090] Specifically, due to the instability of biomass feedstock composition and gasification product quality, as well as frequent adjustments to boiler load, the system's operating characteristics vary significantly across different time periods, reflected in the range of individual operating variables, the coupling relationship of multi-source operating data, and dynamic response characteristics. Mixing data from different operating modes for unified model training would cause mutual interference during parameter learning, making it impossible to accurately characterize the specific mode's mechanism and maintain predictive stability under mode switching or boundary conditions. Therefore, this embodiment, after dividing the historical mode representation vector into operating modes, extracts corresponding historical training samples for each operating mode to train a dedicated expert model. By separating samples and modeling according to operating modes, each expert model focuses on the data distribution characteristics of the corresponding mode, enabling a more refined learning of the nonlinear coupling relationship between biomass gasification and the coal-fired boiler system under that mode.

[0091] In this embodiment, historical training samples are constructed in units of time windows, including multi-source operational data and pattern representation vectors within that window. The pattern representation vectors reflect the overall operational state of the system, enabling the expert model to perceive the background of operational patterns during training and enhancing its ability to express the inherent structural features of these patterns. Furthermore, different operational patterns occur at different frequencies, and some patterns have fewer historical samples. To avoid bias in expert model training towards patterns with more samples, this embodiment adjusts the training sample loss weights based on the sample size of each pattern during training, improving the expert model's prediction robustness under non-mainstream patterns and extreme conditions.

[0092] This embodiment achieves effective decoupling of different operating modes by independently extracting samples and training expert models for each operating mode. Each expert model can accurately model the operating mode it is good at, providing a high-quality and complementary prediction output basis for the dynamic weighted fusion of multi-expert prediction results by the gating network, so that the multi-objective prediction results maintain high accuracy and continuity in complex working conditions and mode switching.

[0093] S2.3: Combine the K trained expert models to construct the multi-objective prediction expert group.

[0094] Specifically, expert models corresponding to different operating modes have high prediction accuracy in their respective data distribution areas, but a single expert model can only reflect local features and is difficult to adapt to mode transitions, superpositions, and boundary condition changes in actual system operation. Therefore, this embodiment combines expert models corresponding to multiple operating modes to construct a multi-objective prediction expert group, providing a unified and stable model foundation for dynamic fusion based on gating weights.

[0095] In this embodiment, the multi-objective prediction expert group is not simply a stack of multiple expert models, but rather each expert model is regarded as a professional expression of the same prediction problem in different operating mode subspaces. The input of each expert model is a mode representation vector within a time window, and the output is the multi-objective prediction result, ensuring composability and scalability at the expert group level.

[0096] Furthermore, the construction of a multi-objective prediction expert group enables the system to simultaneously obtain prediction results from multiple operating mode perspectives during the prediction phase. Each expert model independently responds to the current mode representation vector, forming candidate multi-objective prediction outputs, providing an information basis for the gating network to allocate prediction weights. This design avoids rigid mode selection at the model level and eliminates the problem of frequent model switching caused by mode misjudgment or short-term fluctuations.

[0097] From a system perspective, the construction of the multi-objective prediction expert group introduces a mode redundancy protection mechanism. When the operating condition is in a typical region of a certain operating mode, the corresponding expert model plays a dominant role; when the operating condition is in a mode transition region or has mixed mode characteristics, multiple expert models participate in the prediction, creating conditions for smooth weighted fusion of the gating network. This multi-expert parallel and weighted post-decision decision-making structure improves the continuity and stability of multi-objective prediction results in dynamic operation.

[0098] The specific steps of S2.2 are as follows:

[0099] S2.2.1: For each operating mode in the set of operating modes, construct an expert model for that operating mode. The expert model is based on a neural network, with the input being the mode representation vector within a time window and the output being a multi-objective prediction result. The structure of the neural network includes a shared feature extraction layer and K mode output layers. The shared feature extraction layer is used by all operating modes to extract common features of the input data. Each mode output layer corresponds to one operating mode in the set of operating modes.

[0100] This embodiment employs a combination of a shared feature extraction layer and a pattern output layer in its expert model structure design. The shared feature extraction layer is used by all operating modes, performing feature transformation and abstract representation on the pattern representation vector within the input time window to extract common feature information. By sharing network parameters, the model can utilize data samples from different modes to learn basic coupled features and improve its overall feature representation capability.

[0101] In this embodiment, the pattern representation vector serves as input, reflecting the overall state of the multi-source operational data. A shared feature extraction layer performs multi-layer nonlinear mapping on this data, transforming the original data into high-level semantic features and providing a unified feature foundation for predicting different patterns. In this embodiment, the shared feature extraction layer preferably employs a multi-layer feedforward fully connected neural network structure, including an input layer, at least two hidden feature extraction layers, and a feature output layer connected sequentially. The number of neurons in the input layer is consistent with the dimension of the pattern representation vector, and it is used to receive the pattern representation vector within the time window as network input.

[0102] In a preferred embodiment, the shared feature extraction layer includes 3 to 5 hidden layers, with the number of neurons in each hidden layer set to 128–512, 64–256, and 32–128 respectively, for layer-by-layer compression and abstraction of input features; each hidden layer is preferably constructed by a combination of linear transformation and nonlinear activation function, wherein the activation function can be selected as ReLU, Leaky ReLU, or ELU function to enhance the network's ability to express nonlinear relationships.

[0103] Furthermore, K mode output layers are set after the shared feature extraction layer, each corresponding to a running mode. Based on the shared features, each mode output layer learns the unique response relationship and prediction preference of that mode, outputting multi-target prediction results. The K mode output layers are respectively connected to the feature output layer of the shared feature extraction layer, each corresponding to a running mode, and each mode output layer is constructed using an independent multi-layer fully connected prediction sub-network structure. Specifically, in one embodiment, each mode output layer includes 1 to 3 hidden layers with 32 to 128 neurons. The last layer is the output layer, with the number of neurons matching the number m of multi-target prediction targets, used to simultaneously output continuous values ​​for m prediction targets. The output layer preferably uses a linear activation function to meet the requirements of multi-target regression prediction. This design of sharing underlying features and outputting in different modes preserves mode differences while avoiding resource waste.

[0104] Through the aforementioned neural network structure, this embodiment achieves the separation of common feature learning and pattern-specific modeling: the shared feature extraction layer characterizes general patterns, while the pattern output layer describes personalized features. This structure improves sample utilization efficiency and training stability, provides suitable input for dynamic weighted fusion of the gating network, and enhances the accuracy and robustness of multi-objective prediction under complex conditions.

[0105] S2.2.2: Count the number of historical training samples corresponding to each operating mode in the set of operating modes. When the number of historical training samples of the target operating mode is less than the preset sample threshold, set a loss weight higher than the benchmark value for the training samples of the target operating mode.

[0106] Specifically, the number of historical training samples for each operating mode in the set of operating modes is first counted and compared with a preset sample threshold. If the number of historical training samples for a target operating mode is less than the sample threshold, it is determined to be a mode with insufficient samples. For target operating modes with insufficient samples, a loss weight higher than the baseline value is set for their corresponding training samples during training, so that the prediction error of this operating mode accounts for a higher proportion in the overall loss function. By increasing the loss weight, the impact of the samples of this operating mode on the update of model parameters is amplified during backpropagation, prompting the model to learn more fully the characteristics of operating modes with fewer but representative samples.

[0107] In a preferred embodiment, the preset sample threshold is set according to the ratio of the total number of historical training samples to the number of operating modes K. Specifically, the total number of historical training samples is divided by the number of operating modes K to obtain the average number of samples for a single operating mode, and 0.3 to 0.8 times this average number of samples is used as the preset sample threshold. In this way, operating modes with a sample number significantly lower than the average level are identified as insufficient sample modes, while operating modes with a sample number close to or higher than the average level participate in training with normal weights.

[0108] Furthermore, for operating modes where the number of historical training samples reaches or exceeds a preset sample threshold, the loss weight of these training samples is maintained at a baseline value, allowing them to participate in parameter updates in a normal proportion. Through this differentiated loss weight setting, this embodiment achieves a dynamic balance of sample contributions from different operating modes without altering the model structure.

[0109] S2.2.3: Input the historical training samples corresponding to each operating mode into the shared feature extraction layer for feature extraction, input the extracted features into the corresponding mode output layer, and output the multi-target prediction results.

[0110] In this embodiment, historical training samples corresponding to each operating mode are first uniformly input into a shared feature extraction layer. This layer performs multi-layer nonlinear mapping on the mode representation vectors within the time window, extracting high-level feature representations reflecting the general operating laws of biomass gasification systems and coal-fired boiler systems, such as the basic coupling features of fuel supply and furnace combustion, and the overall trend features of thermal parameters changing with load. Through the unified processing of the shared feature extraction layer, training samples from different operating modes are mapped to common feature representations with consistent semantic meaning in the feature space, ensuring the consistency of feature expressions for different operating modes and avoiding redundant learning of common features. After the shared feature extraction is completed, the high-level feature representations are input into the mode output layer of the corresponding operating mode. Each mode output layer, based on the shared features, further learns the unique mapping relationship of multi-objective variables under that operating mode and outputs the corresponding multi-objective prediction results. Setting independent output layers for different operating modes enables the model to form targeted prediction responses for different situations, avoiding prediction bias between different operating modes.

[0111] S2.2.4: Calculate the weighted loss value based on the loss weight and the actual label value, and update the parameters of the shared feature extraction layer and each mode output layer based on the weighted loss value.

[0112] Specifically, when constructing historical training samples, for the pattern representation vector corresponding to each time window, the multi-source operational measurement value at the time corresponding to the preset prediction lag after that time window is selected as the actual label value of that sample. For each historical training sample, its multi-objective prediction result is first compared with the actual label value to calculate the prediction error of each prediction objective; then, the error is weighted according to the loss weight of its respective operational mode to obtain the weighted loss value. Introducing operational mode-related loss weights allows operational modes with fewer samples or higher engineering importance to account for a higher proportion of the overall loss.

[0113] In this embodiment, the weighted loss value is used to drive the update of neural network parameters. Through backpropagation, the weighted loss is passed layer by layer to the shared feature extraction layer and the corresponding mode output layer, simultaneously updating the parameters of both layers. Specifically, the parameter update of the shared feature extraction layer comprehensively considers the weighted loss of all operating mode samples, enabling it to stably learn general operating rules; the parameter update of each mode output layer is mainly affected by the weighted loss of the corresponding operating mode samples, making it more focused on characterizing specific mapping relationships.

[0114] Please see Figure 3 The flowchart of the gated hysteresis correction provided in this application embodiment shows the specific steps of S3 as follows:

[0115] S3.1: For each operating mode in the set of operating modes, set an entry threshold and an exit threshold, wherein the entry threshold is greater than the exit threshold.

[0116] Specifically, due to factors such as load regulation, biomass gas quality fluctuations, and measurement noise, the basic gating weights calculated based on the current model representation vector often exhibit continuous fluctuation characteristics. If the operating mode is determined solely based on a single threshold, repeated fluctuations in the gating weights around the threshold can easily lead to frequent switching between adjacent operating modes, resulting in problems such as unstable expert model selection and fluctuating prediction results. Therefore, this embodiment introduces a dual-threshold determination mechanism with hysteresis characteristics in the gating correction stage.

[0117] In this embodiment, for each operating mode in the set of operating modes, a corresponding entry threshold and exit threshold are set, and the entry threshold is explicitly limited to be greater than the exit threshold. By introducing a threshold interval between the mode entry and mode exit determinations, a determination hysteresis interval is formed, so that the switching of operating modes no longer depends on a single instantaneous weight judgment, but has a certain stability margin.

[0118] In a preferred embodiment, for each operating mode, the basic gating weight distribution interval corresponding to that operating mode under typical stable operating conditions is first statistically analyzed based on historical training data. The upper quantile of this distribution interval is used as the entry threshold, and the lower quantile of this distribution interval is used as the exit threshold. For example, the 70% to 90% quantile of the weight distribution is used as the entry threshold, and the 30% to 50% quantile of the weight distribution is used as the exit threshold, thereby forming a judgment interval with hysteresis characteristics between the two.

[0119] In this embodiment, when a certain operating mode is not yet activated, the determination to enter the operating mode is only allowed when its corresponding basic gating weight exceeds the entry threshold of the operating mode; while when the operating mode is already activated, the determination to exit the operating mode is only allowed when its corresponding basic gating weight drops to or below the exit threshold of the operating mode. Since the entry threshold is higher than the exit threshold, the operating mode remains unchanged within the weight range formed between the two, thereby effectively suppressing frequent mode transitions caused by short-term fluctuations or noise.

[0120] S3.2: Obtain the current operating mode at the current moment, and perform mode transition determination according to the weight value corresponding to each operating mode in the basic gating weight to determine whether to trigger mode transition.

[0121] Specifically, during the online operation phase of multi-objective prediction, a pre-defined gating network outputs basic gating weights corresponding to each operating mode based on the current mode representation vector, thereby characterizing the degree of matching between the current operating condition and each operating mode. However, due to the obvious continuity and inertia characteristics of the operation process of biomass gas coupled coal-fired boilers, and the susceptibility of basic gating weights to short-term disturbances, measurement noise, and fine-tuning of operating conditions, directly switching operating modes based solely on the weight magnitude at a single moment can easily lead to frequent fluctuations in mode determination, resulting in instability in expert model selection and jitter in prediction results. Therefore, this embodiment introduces a mode transition determination mechanism based on the current operating mode and basic gating weights during the gating correction phase.

[0122] In this embodiment, the current operating mode is first obtained and used as a reference benchmark for judgment. Then, the weight values ​​corresponding to each operating mode in the basic gating weight are comprehensively compared. Combined with the setting of entry threshold and exit threshold, it is determined whether it is necessary to transition from the current operating mode to other operating modes.

[0123] S3.3: When a mode transition is triggered, a minimum mode dwell time is set, and mode transitions are prohibited from being executed again within the minimum mode dwell time.

[0124] Specifically, changes in operating conditions exhibit significant time inertia. The impacts of biomass gas composition, load adjustments, and air distribution changes on combustion state, thermal parameters, and pollutant generation require a certain amount of time to be reflected in multi-source operating data. If a new mode is determined immediately after a mode transition, short-term fluctuations in the basic gating weights or model response lags can easily lead to repeated mode switching, resulting in discontinuous or fluctuating multi-objective prediction results. Therefore, this embodiment introduces a minimum mode dwell time constraint after detecting a mode transition. After determining the transition from the current operating mode to the target operating mode, a corresponding minimum dwell time is set. During this period, the current operating mode remains unchanged, and mode transition determinations are prohibited to ensure the new mode is stable and continuous, allowing the expert model prediction results to reflect the system's operating characteristics.

[0125] In this embodiment, the minimum residence time is set based on the dynamic response characteristics of the biomass gas coupled coal-fired boiler, the data sampling period, and the risk level of the operating mode switching project. For example, in modes with stable operation and allowing for rapid response, the minimum residence time can be set to 5–10 sampling periods, corresponding to 5–10 minutes; for modes requiring high predictive stability, such as startup, low load, or abnormal operating conditions, the minimum residence time can be set to 20–60 sampling periods, corresponding to 20–60 minutes, to enhance the robustness of mode determination. This setting does not change the mode transition determination conditions, but only constrains the frequency of mode switching time.

[0126] By introducing a minimum residence time constraint, this embodiment suppresses mode switching behavior in the time dimension, making the changes in operating modes more consistent with the evolution of boiler operating states. This method, in conjunction with a hysteresis determination mechanism, suppresses mode migration jitter from both the determination conditions and time constraints perspectives, providing a stable operating mode foundation for basic gating weight smoothing.

[0127] S3.4: Perform smoothing processing on the basic gating weights to generate the final gating weights.

[0128] Specifically, during online multi-objective prediction, even with dual threshold judgment and minimum dwell time constraints suppressing frequent transitions in operating modes, the basic gating weights may still fluctuate instantaneously due to input data volatility, model uncertainty, or noise. Directly using these basic gating weights for weighted fusion of multi-expert prediction results would introduce short-term jitter into the prediction output, affecting the continuity and engineering usability of the prediction results. Therefore, this embodiment performs smoothing processing on the basic gating weights during the gating correction stage to generate the final gating weights.

[0129] When no mode transition is triggered, a time-continuous weighted average method is used to smooth the gating weights. This involves averaging the current basic gating weight with the previous final gating weight to obtain the current final gating weight, while incorporating the previous final gating weight as a historical constraint to reduce instantaneous fluctuations in the basic gating weight. The weighting coefficient between the current basic gating weight and the previous final gating weight in the weighted average method is used to balance the system's response speed to the latest operating condition changes with the time smoothness of the prediction results. Its value is determined comprehensively based on the data sampling period, the fluctuation amplitude of the basic gating weight, and the stability requirements of the operating mode.

[0130] When triggering a mode transition, to avoid interference from multiple modes in the prediction results, the basic gating weights are first constrained, retaining only the weights corresponding to the current operating mode and candidate operating modes, while setting the rest to zero. After normalization, the transition weights are obtained, reflecting the relative weight relationship between the current operating mode and candidate operating modes. Subsequently, the transition weights are weighted and averaged with the final gating weights of the previous time step to obtain the final gating weights of the current time step, ensuring a continuous transition in mode weight changes and avoiding drastic changes in prediction fusion weights during mode switching.

[0131] The specific steps of S3.2 are as follows:

[0132] S3.2.1: Extract the first weight value corresponding to the current operating mode from the basic gating weight, and determine whether the first weight value is less than or equal to the exit threshold of the current operating mode.

[0133] Specifically, the pre-defined gating network outputs the basic gating weights corresponding to each operating mode based on the current mode representation vector, which are used to reflect the degree of matching between the current operating condition and each operating mode. Based on this, this embodiment first clarifies the current operating mode at the current moment and uses this operating mode as the starting reference object for mode transition determination.

[0134] In this embodiment, a first weight value corresponding to the current operating mode is extracted from the basic gating weights to characterize the matching strength between the current operating condition and the operating mode. Subsequently, the first weight value is compared with an exit threshold pre-set for the operating mode to determine whether the current operating mode is still suitable for describing the current operating condition.

[0135] When the first weight value is greater than the exit threshold, it indicates that the current working condition still maintains a high degree of matching with the current operating mode. Even if the basic gating weight of other operating modes increases in a short period of time, it is not immediately considered that a mode switch is needed, thereby avoiding frequent fluctuations in the operating mode between similar working conditions.

[0136] When the first weight value is less than or equal to the exit threshold, it indicates that the matching degree between the current operating condition and the current operating mode has significantly decreased, the current operating mode's ability to describe the system state has weakened, and there is a possibility of exiting the operating mode. In this case, this embodiment does not immediately trigger a mode transition, but uses the result as a prerequisite for subsequent exit duration statistics and candidate operating mode determination, thereby ensuring that the mode switching decision is based on the continuous trend of weight changes rather than a single instantaneous fluctuation.

[0137] S3.2.2: When the first weight value is less than or equal to the exit threshold, count the number of exit sampling cycles that continuously satisfy the condition that the first weight value is less than or equal to the exit threshold.

[0138] Specifically, when the basic gating weights change over time, the weight value corresponding to the current operating mode may momentarily decrease due to short-term operating condition disturbances, measurement noise, or model estimation errors. If the mode switching is triggered solely based on whether the weight value at a single sampling moment is lower than the exit threshold, it is easy to misjudge short-term fluctuations as substantial changes in the operating mode, leading to premature exit of the operating mode and affecting the continuity and stability of the prediction results. Therefore, this embodiment introduces a continuous sampling period statistical mechanism after determining that the current operating mode meets the exit conditions. When the first weight value corresponding to the current operating mode is detected to be less than or equal to the exit threshold, it is not immediately determined that exit is necessary, but rather the continuous sampling period is statistically analyzed to assess whether the weight decrease continues.

[0139] In this embodiment, the exit sampling period count refers to the number of times the first weight value continuously satisfies the condition of being less than or equal to the exit threshold across multiple consecutive sampling moments. When the first weight value rises above the exit threshold again at a certain sampling moment, the exit sampling period count is reset to zero or recounted, ensuring that the exit sampling period count is only accumulated when the weight is continuously below the exit threshold. By counting the number of sampling periods that continuously meet the exit condition, this embodiment changes the operation mode exit determination from instantaneous judgment to time consistency judgment, which can effectively distinguish between short-term random fluctuations and continuous operating condition changes, avoid misjudgments caused by single-point anomalies, and provide a reliable basis for determining whether the preset exit duration period has been reached.

[0140] S3.2.3: When the number of exit sampling cycles reaches the preset number of exit duration cycles, select the candidate running mode with the largest weight value from the basic gating weights, and determine whether the second weight value corresponding to the candidate running mode is greater than or equal to the entry threshold of the candidate running mode.

[0141] Specifically, when the first weight value of the current operating mode is continuously less than or equal to the exit threshold for multiple consecutive sampling periods, and the number of exit sampling periods reaches the preset number of exit duration periods, it indicates that the matching degree between the current working condition and the current operating mode is continuously decreasing, and the current operating mode is no longer suitable as the main description mode of the system state. At this time, the candidate operating mode screening and entry condition judgment stage is entered.

[0142] In a preferred embodiment, the preset number of exit duration periods is set based on the data sampling period as the basic time unit, that is, it is set to an integer multiple of a certain number of consecutive sampling periods. For example, when the system sampling period is 1 minute, the number of exit duration periods can be set to 3 to 10 sampling periods, corresponding to 3 to 10 minutes; when the sampling period is 5 minutes, the number of exit duration periods can be set to 2 to 6 sampling periods. In this way, the exit determination is based on a continuous time trend, avoiding false exits triggered by short-term fluctuations in a single sampling period.

[0143] In this embodiment, based on the basic gating weights output by the gating network at the same time, the candidate operating mode with the largest basic gating weight value is selected from the non-current operating modes. This mode has the highest relative matching degree at the current time and can optimally reflect the characteristics of the current operating condition, avoiding blind switching or introducing unnecessary candidate objects. Subsequently, the second weight value corresponding to the candidate operating mode is extracted and compared with a pre-set entry threshold to determine whether it meets the entry conditions. If the second weight value is greater than or equal to the entry threshold, it indicates that the matching degree between the current operating condition and the candidate operating mode is acceptable, and there is a possibility of entering the candidate operating mode. If the second weight value does not reach the entry threshold, even if the current operating mode meets the exit conditions, the mode transition is not triggered immediately, but the current operating mode is maintained, waiting for the candidate operating mode weight to change in subsequent sampling periods. In this way, forced switching is avoided when there is no suitable alternative mode, and frequent switching of operating modes between low matching modes is prevented.

[0144] Through the aforementioned candidate operating mode screening and initial entry threshold determination steps, this embodiment introduces an "exit-entry" dual-condition constraint mechanism in mode transition, requiring both that the current operating mode be inapplicable and that the candidate operating mode have sufficient matching credibility. This method, combined with exit duration statistics, ensures the robustness of mode transition determination from both the perspectives of mode applicability and substitution rationality, providing a reliable premise for subsequent duration statistics and final mode transition decision-making.

[0145] S3.2.4: When the second weight value is greater than or equal to the entry threshold, count the number of entry sampling cycles that continuously satisfy the condition that the second weight value is greater than or equal to the entry threshold.

[0146] Specifically, after completing the screening of candidate operating modes and confirming that their second weight value has reached the entry threshold, it is necessary to determine whether the matching relationship between the mode and the current operating condition is stable over time. Due to factors such as fuel fluctuations in the operation of biomass gas coupled with coal-fired boilers, the weight value of a candidate operating mode may increase in a short period of time, but this does not mean that the system has stably switched to that mode. Therefore, this embodiment introduces a continuous statistical mechanism for the number of sampling cycles after the candidate operating mode meets the entry threshold condition.

[0147] In this embodiment, the number of sampling periods refers to the number of times that the second weight value of the candidate operating mode continuously meets the condition of being greater than or equal to the entry threshold during multiple consecutive sampling times. If the second weight value is lower than the entry threshold at any sampling time, the number of sampling periods is reset to zero or recounted to ensure that the number of sampling periods is accumulated only when the weight is continuously in the high matching range.

[0148] By continuously counting the number of sampling cycles, this embodiment expands the entry determination of candidate operating modes from a weight judgment at a single moment to a comprehensive judgment based on time consistency, avoiding the problem of false entry caused by the instantaneous peak of weight, and ensuring that the candidate operating modes stably adapt to the current working conditions within a continuous time.

[0149] S3.2.5: When the number of entry sampling cycles reaches the preset number of entry duration cycles, determine to trigger a mode transition from the current operating mode to the candidate operating mode.

[0150] Specifically, when the second weight value corresponding to the candidate operating mode meets the entry threshold condition for multiple consecutive sampling periods, and the number of entry sampling periods reaches the preset entry duration period, it indicates that the current operating condition has stably evolved towards the candidate operating mode in the time dimension. At this time, the candidate operating mode can continuously and reliably characterize the current operating characteristics of the biomass gas coupled coal-fired boiler. Therefore, in this embodiment, after meeting the entry duration period condition, the trigger mode transition is formally confirmed, the current operating mode is switched to the candidate operating mode, and it is determined as the new current operating mode for gating weight correction and multi-objective prediction processes.

[0151] In a preferred embodiment, the number of consecutive entry periods is set based on the data sampling period as the basic time unit, i.e., it is set to an integer multiple of a certain number of consecutive sampling periods. For example, when the system sampling period is 1 minute, the number of consecutive entry periods can be set to 3 to 10 sampling periods, corresponding to 3 to 10 minutes; when the sampling period is 5 minutes, the number of consecutive entry periods can be set to 2 to 6 sampling periods. This method ensures that the weights of candidate operating modes remain in a high matching range, avoiding false entry due to short-term weight peaks.

[0152] In this embodiment, the setting of the number of entry duration cycles, the number of exit duration cycles, the entry threshold, and the exit threshold constitute a complete dual-threshold hysteresis determination system. This ensures that the switching of operating modes not only depends on the weight relationship but is also subject to time consistency constraints, effectively avoiding erroneous switching when weights cross, there are critical fluctuations, or when the weights of multiple modes are close, and preventing the operating mode from frequently going back and forth between adjacent modes.

[0153] The specific steps in S3.4 are as follows:

[0154] S3.4.1: When no mode transition is triggered, the current basic gating weight and the previous final gating weight are weighted and averaged to generate the current final gating weight.

[0155] Specifically, when no mode transition is triggered, the current operating mode remains unchanged, and the system's operating state is stable or evolves slowly. However, the basic gating weights output by the gating network at this stage may fluctuate between adjacent sampling times due to noise from multi-source operating data, measurement errors, and model estimation uncertainties. If the basic gating weights at the current time are directly used as prediction fusion weights, short-term fluctuations will be amplified, affecting the continuity and smoothness of the prediction output. Therefore, in this embodiment, when no mode transition is triggered, a time smoothing mechanism is introduced for the gating weights, that is, the weighted average of the basic gating weights at the current time and the final gating weights at the previous time is taken to obtain the final gating weights at the current time. The final gating weights at the previous time reflect the historical stable mode weight distribution, while the basic gating weights at the current time reflect the impact of the latest operating condition information on the mode matching relationship.

[0156] By introducing historical weight constraints through weighted averaging, the gating weights evolve continuously over time, preventing drastic changes in the prediction fusion weights caused by sudden weight mutations within a single sampling period. The weighting coefficients can be set according to the system's requirements for response speed and smoothness, ensuring responsiveness to changes in operating conditions while suppressing high-frequency fluctuations.

[0157] S3.4.2: When a mode transition is triggered, the basic gating weights corresponding to the current operating mode and the candidate operating modes are retained, the basic gating weights corresponding to the other operating modes are set to zero, the retained basic gating weights are normalized, and transition weights are generated.

[0158] Specifically, when a mode transition from the current operating mode to a candidate operating mode is triggered after determining the aforementioned dual threshold, entering and exiting duration statistics, and minimum dwell time constraints, the operating state is in a critical stage of transitioning from the old mode to the new mode. During this stage, if all operating modes are still allowed to participate in gating weight allocation simultaneously, it can easily lead to superimposed interference from multiple expert models during prediction fusion, causing instability or distortion in the prediction results at the moment of mode switching. Therefore, this embodiment performs targeted constraint processing on the basic gating weights when triggering a mode transition. Specifically, only the basic gating weights corresponding to the current operating mode and the candidate operating mode are retained, while the basic gating weights corresponding to the remaining operating modes are uniformly set to zero, thereby clearly limiting the range of operating modes participating in weight allocation during the mode switching process.

[0159] After weight selection, the basic gating weights corresponding to the two retained operating modes are normalized to ensure that their sum meets predetermined constraints, generating transition weights for the mode switching phase. These transition weights reflect the relative importance of the current operating mode and the candidate operating modes at the moment of switching, providing a clear and controllable weight foundation for the smooth fusion of gating weights.

[0160] By introducing the aforementioned transition weight generation mechanism during the mode transition phase, this embodiment transforms the operation mode switching process from multi-mode competition constraint to dual-mode transition, making the direction of gating weight change and the participating objects clearer.

[0161] S3.4.3: Calculate the weighted average of the transition weight and the final gating weight of the previous time step to generate the final gating weight of the current time step.

[0162] Specifically, at the decision level, the operation mode has completed the switch from the current operation mode to the candidate operation mode. However, at the weight level, it is in a critical stage of smoothly transitioning from the old mode to the new mode. If the transition weight is directly used as the final gating weight at the current moment, it is easy to introduce abrupt changes in the prediction fusion, causing the multi-objective prediction results to change abruptly at the moment of mode switching, affecting the continuity of output and engineering usability. Therefore, this embodiment introduces a weight smoothing fusion mechanism based on time continuity in the mode transition stage, which weights and averages the transition weight generated in this step with the final gating weight of the previous moment to obtain the final gating weight at the current moment.

[0163] Specifically, the final gating weight at the previous moment reflects the weight distribution during the stable operation phase of the system before mode transition, while the transition weight characterizes the relative importance of the current and candidate operating modes at the moment of switching. Through weighted fusion, the gating weights evolve smoothly and continuously during the mode transition phase, avoiding drastic changes during mode switching. The weighting coefficients can be adjusted according to the system's requirements for smoothness and response speed, ensuring timely mode switching and suppressing unnecessary weight oscillations.

[0164] Through the above processing, this embodiment achieves a gradual transition of gating weights from historical states to new states during the mode transition stage, ensuring the continuity of weight allocation in the multi-objective prediction expert model over time. This step, in conjunction with the shortest dwell time constraint, dual-threshold hysteresis determination, and transition weight generation mechanism, jointly suppresses mode transition jitter from the levels of determination, weight selection, and smooth fusion.

[0165] For example, the following numerical example is provided to illustrate how to perform hysteresis correction on the basic gating weights output by the gating network through a dual-threshold decision to obtain the final gating weights for expert model fusion. This example is only used to explain the relationship between the computational link and the units; the selected parameters and values ​​are illustrative and do not represent actual calibration results or engineering recommendations.

[0166] It should be noted that, in the context of this application, the basic gating weights are output by the preset gating network based on the current mode representation vector, reflecting the "degree of matching between the current working condition and each operating mode / expert model"; then, hysteresis correction is used to suppress mode migration jitter, and the weights are smoothed to generate the final gating weights, which are used for weighted fusion of multiple expert outputs.

[0167] In one specific embodiment, it is assumed that the historical operating patterns are divided into three main operating patterns, denoted as Pattern A, Pattern B, and Pattern C, corresponding to three multi-objective prediction expert models. The gating network outputs three basic gating weights in each sampling period to reflect the instantaneous matching degree between the current pattern representation vector and each operating pattern.

[0168] To suppress frequent switching near the mode boundary caused by short-term fluctuations in weights, this embodiment sets an entry threshold and an exit threshold for each operating mode, with the entry threshold being 0.60 and the exit threshold being 0.40. Simultaneously, the exit duration is set to 2 sampling periods, the entry duration to 2 sampling periods, and the minimum dwell time to 3 sampling periods. When no mode transition is triggered, the weighted average of the current basic gating weight and the final gating weight at the previous moment is used as the final gating weight.

[0169] Assume at time... At this time, the current operating mode is mode A, and the corresponding final gating weight at the previous moment is:

[0170]

[0171] At any moment The basic gating weights output by the gating network for:

[0172]

[0173] The current operating mode A has a weight value of 0.45, which is greater than the exit threshold of 0.40. Since the exit condition has not yet been met, the mode transition decision is not triggered; only the weight is smoothed. The current mode representation vector is input into a preset gating network, and after forward computation and normalization of the output score, the basic gating weights corresponding to each operating mode are obtained. The weighted average of the current basic gating weights and the final gating weights from the previous time step is calculated as follows:

[0174]

[0175] For example, take =0.3, then the time is obtained. final gating weight for:

[0176]

[0177] It can be seen that although the basic weight of mode B is close to that of mode A, the final gating weight still maintains mode A as the dominant factor, thus avoiding mode switching errors caused by a single fluctuation.

[0178] At any moment The basic gating weights output by the gating network for:

[0179]

[0180] At this point, the weight value corresponding to the current mode A is 0.39, which is less than the exit threshold of 0.40 for the first time, satisfying the exit condition once, and the count of exit duration cycles begins. However, since the preset exit duration cycle number of 2 has not yet been reached, the mode transition is not triggered, and weight smoothing continues to obtain the time. Ultimate gate control for:

[0181]

[0182] As can be seen, although the base weight of mode A is lower than the exit threshold, the final gating weight is still allocated to mode A at a relatively high proportion, thus suppressing short-term fluctuations.

[0183] At any moment The basic gating weights output by the gating network for:

[0184]

[0185] If the weight corresponding to the current mode A is less than the exit threshold for the second consecutive period, and the number of periods for exiting reaches 2, the exit judgment condition is met. At this time, mode B with the largest basic gating weight is selected as the candidate running mode from the non-current modes, and its corresponding weight is 0.62.

[0186] Since 0.62 is greater than the entry threshold of 0.60, the count of entry duration cycles for candidate mode B begins to be calculated. However, since the count of entry duration cycles has not yet reached 2, mode transition is not triggered for the time being. Instead, weight smoothing continues to be performed to obtain the time. final gating weight for:

[0187]

[0188] At this point, it can be observed that the weight ratios of mode A and mode B in the final gating weights gradually approach each other, but no hard switch has yet occurred.

[0189] At any moment The basic gating weights output by the gating network for:

[0190]

[0191] When the weight corresponding to mode B is greater than the entry threshold of 0.60 for the second consecutive period, and the number of consecutive periods reaches 2, the mode transition from mode A to mode B is officially triggered.

[0192] At the moment of triggering mode transition According to the gating correction strategy in this embodiment, only the basic gating weights corresponding to the current operating mode A and the candidate operating mode B are retained, the weights of the remaining modes are reset to zero, and normalization is performed to obtain the transition weights:

[0193]

[0194] Subsequently, the transition weight is weighted and averaged with the final gating weight of the previous time step to obtain the time step. final gating weight for:

[0195]

[0196] As can be seen, although mode B has become the new dominant mode, mode A still retains a certain weight ratio, achieving a continuous and smooth transition from mode A to mode B, rather than a sudden change in weight.

[0197] Simultaneously, the current operating mode is updated to Mode B, and a minimum dwell time counter is activated, prohibiting mode transitions from occurring again within the next three sampling periods. During the minimum dwell time, even if the basic gating weights experience some oscillation, for example at time [time value missing], [the system will prevent further transitions]. Basic gating weights for:

[0198]

[0199] At this point, the base weight of mode B is lower than the entry threshold of 0.60. However, due to the shortest dwell time constraint period, the system no longer performs mode transition determination, but only performs weight smoothing update to obtain the time. final gating weight :

[0200]

[0201] At any moment Continue to smoothly update to obtain the time. final gating weight :

[0202]

[0203] It can be seen that within the shortest dwell time, the final gating weight gradually and steadily concentrates on mode B, avoiding mode switching due to a short-term drop in the basic gating weight.

[0204] The specific steps for S4 are as follows:

[0205] S4.1: Input the current mode representation vector into the K expert models in the multi-objective prediction expert group, and obtain the multi-objective prediction results output by each expert model, wherein each multi-objective prediction result contains m predicted target values.

[0206] Specifically, given the significant differences in operating conditions under different operating modes, this embodiment does not directly use a single prediction model for output during the prediction phase. Instead, it inputs the current mode representation vector in parallel to K expert models in a multi-objective prediction expert group. These K expert models correspond to different operating modes within the set of operating modes, and each expert model has been specifically modeled for historical data of its corresponding operating mode during the training phase, thus demonstrating its adaptability to specific operating modes in terms of structural parameters and feature mapping relationships.

[0207] In this embodiment, the current mode representation vector is simultaneously input to the K expert models in the same input form. Each expert model performs forward inference and outputs the corresponding multi-objective prediction result. Each multi-objective prediction result contains m predicted target values, which are used to characterize the predicted output for multiple predicted targets under the assumption of the corresponding operating mode of the expert model. The predicted targets may include, but are not limited to, operating status indicators, performance indicators, or safety-related indicators, and their specific meanings can be set according to the application scenario.

[0208] S4.2: Extract the prediction weight values ​​corresponding to each expert model from the final gating weights.

[0209] Specifically, the final gating weights reflect the relative reliability of each operating mode at the current moment after comprehensively considering historical states, current operating conditions, and mode transition constraints. Since the K expert models in the multi-objective prediction expert group correspond to each operating mode in the set of operating modes, the final gating weights can be directly used as the weight basis for each expert model in the prediction fusion stage.

[0210] In this embodiment, the final gating weight is a weight vector of length K, with each component corresponding one-to-one with one of the K expert models in the multi-objective prediction expert group. During prediction fusion, the weight component at the corresponding position is read from the final gating weight vector according to the index order or mode identifier of the expert model, and used as the prediction weight value of that expert model to weight the multi-objective prediction results output by that expert model. In other words, each expert model is assigned a prediction weight value consistent with its corresponding operating mode, and this prediction weight value is used to characterize the credibility and contribution ratio of the expert model's output result in the current prediction scenario.

[0211] Specifically, the determination of prediction weights does not depend on the instantaneous performance of a single prediction result, but is based on the final gating weights after smoothing and mode transition constraints in the preceding steps. This makes the prediction weights continuous in the time dimension and stable during mode switching, effectively avoiding the problem of drastic changes in the weights of a certain expert model due to instantaneous operating condition fluctuations.

[0212] S4.3: For each of the m prediction targets, multiply the prediction target value output by the K expert models by the corresponding prediction weight value and sum them to obtain the fused prediction value of the prediction target.

[0213] Specifically, after acquiring the multi-objective prediction results of each expert model and extracting the prediction weights, it is necessary to fuse the outputs of multiple expert models for the same prediction objective to form the final prediction output. Since each expert model in this embodiment is modeled based on different operating modes, its prediction results have different reliability under different operating conditions. Therefore, by introducing prediction weights for weighted fusion, the prediction advantages of each expert model in its preferred operating mode can be fully utilized.

[0214] In this embodiment, for any one of the m prediction targets, firstly, K prediction target values ​​corresponding to the prediction target are extracted from the multi-target prediction results of the K expert models; then, the K prediction target values ​​are matched one by one with the corresponding prediction weight values, and a product operation is performed on each set of prediction target values ​​and prediction weight values; finally, all product results are summed to obtain the fused prediction value of the prediction target at the current time.

[0215] By independently performing the above weighted summation process for each prediction target, this embodiment can maintain the consistency of the multi-target prediction structure while avoiding the problem of weight interference between different prediction targets, so that the fusion result of each prediction target is strictly controlled by the prediction performance of each expert model on that target and its corresponding operating mode weight.

[0216] Furthermore, since the prediction weights are derived from the final gating weights after smoothing and mode transition constraints, the contribution ratio of the expert model is continuous over time and exhibits gradual changes during mode switching during the prediction result fusion process. This effectively avoids abrupt changes or oscillations in the prediction results during mode switching. This fusion method ensures the stability of the prediction results while also enhancing the responsiveness of the prediction output to changes in actual operating conditions.

[0217] S4.4: Combine the fused prediction values ​​of m predicted targets to generate the final multi-target prediction vector.

[0218] Specifically, the fused prediction values ​​of each prediction objective characterize multiple key aspects of the system's future behavior under the current operating state from different dimensions. In order to facilitate subsequent unified processing, output and application, the multiple fused prediction values ​​need to be structurally combined to form a unified expression of multi-objective prediction results.

[0219] In this embodiment, the m fused prediction values ​​obtained for m prediction targets are arranged and combined according to a pre-agreed order of prediction targets to form a prediction vector containing m elements, which is the final multi-target prediction vector. Each component in the prediction vector corresponds to a prediction target, and its value is the final prediction result of that prediction target after multi-expert weighted fusion at the current time.

[0220] Specifically, the final multi-target prediction vector maintains the semantic independence of each prediction target while achieving unified encapsulation of the multi-target prediction results, making the prediction output more compact in data structure and more standardized in expression. This prediction vector can be used as a whole as the output of the method, or one or more prediction target components can be read and used individually according to specific application requirements.

[0221] A multi-objective prediction method coupled with biomass gas also includes an expert model update step, comprising:

[0222] A1: Obtain the actual measurement value corresponding to the final multi-objective prediction vector, calculate the prediction residual, and generate the health index of the current operating mode.

[0223] Specifically, after obtaining the final multi-objective prediction vector, in order to quantitatively evaluate the current operating status, the prediction results need to be compared and analyzed with the actual measurement data at the same time or within the corresponding time window. The actual measurement values ​​and the final multi-objective prediction vector correspond one-to-one in the dimension of the prediction targets, each containing m measurement target values.

[0224] In this embodiment, the actual measurement value vector, which is time-aligned with the final multi-target prediction vector, is first obtained. For each prediction target, the prediction residual between the fused prediction value and the corresponding actual measurement value is calculated. The prediction residual characterizes the degree of deviation between the model prediction result and the actual operating state, reflecting whether the current operating state conforms to the normal evolution law.

[0225] Furthermore, after obtaining the prediction residuals of m predicted targets, the data are aggregated and processed to generate a health index characterizing the overall operational status of the current operating mode. This index can be constructed as needed using a weighted combination of prediction residuals, normalization processing, or statistical aggregation results, reflecting the system's operational health level as a single scalar or low-dimensional indicator.

[0226] By introducing a health index based on multi-objective prediction residuals, this embodiment transforms the outputs of multi-operation mode and multi-expert prediction models into a state assessment index with intuitive engineering significance. This allows the prediction results to be used for trend judgment, operational status monitoring, anomaly early warning, or maintenance decision-making. Furthermore, because the final multi-objective prediction vector is a fusion result under dynamic gating weight constraints, the prediction residuals are typically small when the operational mode is stable. However, when the operational status deviates or there are potential anomalies, the prediction residuals will systematically increase, making the health index highly sensitive to operational mode deterioration or abnormal evolution.

[0227] A2: When the health index is continuously lower than the preset health threshold within a preset time period, the multi-source operation data and pattern representation vector within the preset time period are marked as new sub-pattern candidate data.

[0228] Specifically, this embodiment introduces a time continuity constraint after obtaining the health index corresponding to the current operating mode to monitor its changing trend and avoid misjudgment due to short-term disturbances or occasional noise. When the health index is continuously lower than the preset health threshold for multiple consecutive sampling times or within a preset time period, it indicates that the current operating state has deviated from the normal prediction range for a long time. The preset time period limits the duration of abnormal health status. Only when the prediction residual is stable and continuously large will the new sub-mode candidate data marking operation be triggered to improve the reliability of mode expansion judgment. The health threshold distinguishes between acceptable operating deviations and potential structural changes, and the value is set based on historical operating statistical characteristics or engineering experience.

[0229] Furthermore, when the health index remains below a preset health threshold for a preset time period, it is considered that the current operating state may correspond to a new operating feature structure. The multi-source operating data collected within the preset time period and the corresponding pattern representation vectors are then uniformly labeled as candidate data for new sub-patterns. The multi-source operating data characterizes the original feature distribution of the actual operating state, while the pattern representation vectors reflect the overall expression characteristics of the operating state in the pattern space during that time period.

[0230] This embodiment achieves automatic discovery and data accumulation of potential new operating modes by uniformly marking data from abnormal time periods as candidate data for new sub-modes, eliminating the need for manual predefinition of all operating mode types. This data marking mechanism, triggered by health indicators, enables the system to perceive changes in the operating environment or working conditions over long-term operation.

[0231] A3: Based on the new sub-mode candidate data, perform incremental training and update the expert model corresponding to the current operating mode.

[0232] Specifically, after obtaining candidate data for new sub-modes, this embodiment does not immediately introduce them as entirely new operating modes. Instead, it first performs incremental training on the expert model corresponding to the current operating mode based on the candidate data, so as to gradually absorb the new features that emerge during the evolution of the operating state and improve the model's adaptability to changes in complex operating conditions.

[0233] In this embodiment, the candidate data for the new sub-mode includes multi-source operational data and its corresponding mode representation vectors collected during a preset time period when the health index is continuously abnormal. By inputting the candidate data as supplementary training samples into the expert model corresponding to the current operational mode, the model parameters are slightly updated while keeping the original model parameter structure unchanged, thereby gradually expanding the feature coverage capability of the expert model within the semantic range of the original operational mode.

[0234] Unlike completely retraining the expert model, this embodiment uses incremental training to update the model, making limited adjustments to the model parameters only based on newly added candidate data to avoid forgetting existing normal operation characteristics. In this way, the expert model can learn newly emerging operational features while maintaining a stable characterization of the original operational patterns.

[0235] Furthermore, during incremental training, the aforementioned prediction residuals or health indicators can be combined to assign higher training weights to candidate data of new sub-modes, enabling the model to focus on sample regions with large original prediction errors. This allows for targeted correction of the model's prediction bias in this operating mode, which helps to accelerate the model's convergence speed to new operating features and improve update efficiency.

[0236] Example 2

[0237] Please see Figure 4 The present invention provides an embodiment of a biomass gas coupled multi-objective prediction system, the system comprising a data acquisition module, a model training module, a gating correction module, and a multi-objective prediction module, wherein:

[0238] The data acquisition module is used to collect multi-source operating data of the biomass gasification system and the coal-fired boiler system, perform time alignment and resampling on the multi-source operating data, and generate a pattern representation vector.

[0239] The model training module is used to divide the pattern representation vector of historical multi-source operation data into patterns, generate a set of operation patterns, train the corresponding expert model based on each operation pattern in the set of operation patterns, and construct a multi-objective prediction expert group.

[0240] The gating correction module is used to input the current mode representation vector into a preset gating network and output the final gating weights. The preset gating network generates the basic gating weights of the multi-objective prediction expert group based on the current mode representation vector and performs hysteresis correction on the basic gating weights. The hysteresis correction suppresses mode migration jitter through dual threshold judgment and minimum residence time constraint to generate the final gating weights.

[0241] The multi-objective prediction module is used to input the current mode representation vector into each expert model of the multi-objective prediction expert group, and to perform weighted fusion of the multi-objective prediction results output by each expert model through the final gating weight to generate the final multi-objective prediction vector.

[0242] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A multi-objective prediction method coupled with biomass gas, characterized in that, The method includes: Collect multi-source operational data, perform time alignment and resampling on the multi-source operational data, and generate pattern representation vectors; The pattern representation vectors of historical multi-source operation data are divided into patterns to generate an operation pattern set. Based on each operation pattern in the operation pattern set, the corresponding expert model is trained to construct a multi-objective prediction expert group. The current mode representation vector is input into a preset gating network, and the final gating weights are output. The preset gating network generates the basic gating weights of the multi-objective prediction expert group based on the current mode representation vector, and performs hysteresis correction on the basic gating weights. The hysteresis correction suppresses mode migration jitter through dual threshold judgment and minimum residence time constraint, and generates the final gating weights. The current mode representation vector is input into each expert model of the multi-objective prediction expert group. The multi-objective prediction results output by each expert model are weighted and fused through the final gating weight to generate the final multi-objective prediction vector. The basic gating weights are then subjected to hysteresis correction. This hysteresis correction suppresses mode migration jitter through dual threshold determination and minimum dwell time constraints, generating the final gating weights, including: For each operating mode in the set of operating modes, an entry threshold and an exit threshold are set, wherein the entry threshold is greater than the exit threshold; Obtain the current operating mode at the current moment, and perform a mode transition determination based on the weight value corresponding to each operating mode in the basic gating weight to determine whether a mode transition is triggered. When a mode transition is triggered, a minimum mode dwell time is set, and mode transitions are prohibited from being executed again within the minimum mode dwell time. Smoothing is performed on the basic gating weights to generate the final gating weights; Based on the weight values ​​corresponding to each operating mode in the basic gating weights, a mode transition determination is performed to determine whether a mode transition is triggered, including: Extract the first weight value corresponding to the current operating mode from the basic gating weight, and determine whether the first weight value is less than or equal to the exit threshold of the current operating mode; When the first weight value is less than or equal to the exit threshold, count the number of exit sampling periods that continuously satisfy the condition that the first weight value is less than or equal to the exit threshold. When the number of exit sampling cycles reaches the preset number of exit duration cycles, the candidate running mode with the largest weight value is selected from the basic gating weights, and it is determined whether the second weight value corresponding to the candidate running mode is greater than or equal to the entry threshold of the candidate running mode. When the second weight value is greater than or equal to the entry threshold, the number of entry sampling periods that continuously satisfy the second weight value being greater than or equal to the entry threshold is counted. When the number of sampling cycles enters reaches the preset number of continuous cycles, a mode transition from the current operating mode to the candidate operating mode is triggered.

2. The multi-objective prediction method for biomass gas coupling according to claim 1, characterized in that, The process involves dividing the pattern representation vectors of historical multi-source operational data into patterns to generate an operational pattern set. Based on each operational pattern in the operational pattern set, a corresponding expert model is trained to construct a multi-objective prediction expert group, including: Cluster analysis is performed on the pattern representation vectors of historical multi-source operation data to divide the pattern representation vectors into K operation modes, thus obtaining a set of operation modes; For each operating mode in the set of operating modes, extract the historical training samples corresponding to that operating mode and train the expert model for that operating mode, wherein the historical training samples include the mode representation vector corresponding to the operating mode; The K trained expert models are combined to construct the multi-objective prediction expert group.

3. The multi-objective prediction method for biomass gas coupling according to claim 2, characterized in that, For each operating mode in the set of operating modes, extract the historical training samples corresponding to that operating mode, and train an expert model for that operating mode, including: For each operating mode in the set of operating modes, an expert model for that operating mode is constructed. The expert model is based on a neural network, with the input being the mode representation vector within the time window and the output being the multi-objective prediction result. The structure of the neural network includes a shared feature extraction layer and K mode output layers. The shared feature extraction layer is used by all operating modes to extract common features of the input data, and each mode output layer corresponds to one operating mode in the set of operating modes. The number of historical training samples corresponding to each operating mode in the set of operating modes is counted. When the number of historical training samples of the target operating mode is less than the preset sample threshold, a loss weight higher than the benchmark value is set for the training samples of the target operating mode. The historical training samples corresponding to each operating mode are input into the shared feature extraction layer for feature extraction. The extracted features are then input into the corresponding mode output layer to output the multi-target prediction results. The weighted loss value is calculated based on the loss weight and the actual label value, and the parameters of the shared feature extraction layer and each mode output layer are updated based on the weighted loss value.

4. The multi-objective prediction method for biomass gas coupling according to claim 1, characterized in that, Smoothing is performed on the basic gating weights to generate the final gating weights, including: When no mode transition is triggered, the base gating weight at the current moment is weighted and averaged with the final gating weight at the previous moment to generate the final gating weight at the current moment. When a mode transition is triggered, the basic gating weights corresponding to the current operating mode and the candidate operating modes are retained, the basic gating weights corresponding to the other operating modes are set to zero, the retained basic gating weights are normalized, and transition weights are generated. The transition weights are weighted and averaged with the final gating weights of the previous time step to generate the final gating weights of the current time step.

5. The multi-objective prediction method for biomass gas coupling according to claim 1, characterized in that, The step of inputting the current mode representation vector into each expert model of the multi-objective prediction expert group, and then weighting and fusing the multi-objective prediction results output by each expert model through the final gating weights to generate the final multi-objective prediction vector includes: The current mode representation vector is input into the K expert models in the multi-objective prediction expert group to obtain the multi-objective prediction results output by each expert model, wherein each multi-objective prediction result contains m predicted target values. Extract the predicted weight values ​​corresponding to each expert model from the final gating weights; For each of the m prediction targets, the prediction target value output by the K expert models is multiplied by the corresponding prediction weight value and then summed to obtain the fused prediction value of the prediction target. The fused prediction values ​​of m predicted targets are combined to generate the final multi-target prediction vector.

6. The multi-objective prediction method for biomass gas coupling according to claim 5, characterized in that, The m predicted targets include at least two of the following: nitrogen oxide emission concentration, carbon monoxide emission concentration, main steam temperature deviation, flue gas temperature, and boiler efficiency proxy value.

7. The multi-objective prediction method for biomass gas coupling according to claim 1, characterized in that, The method further includes: Obtain the actual measurement value corresponding to the final multi-objective prediction vector, calculate the prediction residual, and generate the health index of the current operating mode. When the health index is continuously lower than the preset health threshold within a preset time period, the multi-source operation data and pattern representation vector within the preset time period are marked as new sub-pattern candidate data. Based on the new sub-mode candidate data, incremental training is performed to update the expert model corresponding to the current operating mode.

8. A biomass-gas coupled multi-objective prediction system, used to implement the biomass-gas coupled multi-objective prediction method as described in any one of claims 1-7, characterized in that, The system includes a data acquisition module, a model training module, a gating correction module, and a multi-objective prediction module, wherein: The data acquisition module is used to acquire multi-source operational data, perform time alignment and resampling on the multi-source operational data, and generate a pattern representation vector. The model training module is used to divide the pattern representation vector of historical multi-source operation data into patterns, generate a set of operation patterns, train the corresponding expert model based on each operation pattern in the set of operation patterns, and construct a multi-objective prediction expert group. The gating correction module is used to input the current mode representation vector into a preset gating network and output the final gating weights. The preset gating network generates the basic gating weights of the multi-objective prediction expert group based on the current mode representation vector and performs hysteresis correction on the basic gating weights. The hysteresis correction suppresses mode migration jitter through dual threshold judgment and minimum residence time constraint to generate the final gating weights. The multi-objective prediction module is used to input the current mode representation vector into each expert model of the multi-objective prediction expert group, and to perform weighted fusion of the multi-objective prediction results output by each expert model through the final gating weight to generate the final multi-objective prediction vector.

Images