A sludge pyrolysis gasification waste gas emission prediction method based on a space-time coupling graph neural network

Abstract

The present application provides a sludge pyrolysis gasification waste gas emission prediction method based on a space-time coupling graph neural network, comprising: obtaining the equipment operation data and waste gas emission monitoring data of multiple parallel production lines of sludge pyrolysis gasification and cleaning; according to the material and energy transmission sequence between the equipment and the convergence relationship of the multiple production lines, a directed process topology graph is constructed; the optimal time lag step of each node input feature relative to the prediction target is calculated and time lag alignment is performed; based on the directed topology graph, the graph neural network is used for directed message transmission between adjacent nodes, combined with the device type cyclic neural network, the node's own time evolution and the cross-node space topology influence are coupled; the hierarchical attention mechanism is introduced to gradually aggregate the intra-line and inter-line operation states to predict the waste gas emission concentration change. The present application can depict the directed space coupling relationship between the equipment and the feature level time lag effect, and keep the prediction accuracy stable under the condition of sensor abnormality and severe working condition fluctuation.

Description

Technical Field

[0001] This invention belongs to the field of environmental engineering technology, specifically relating to a method for predicting waste gas emissions from sludge pyrolysis gasification based on a spatiotemporally coupled graph neural network. Background Technology

[0002] Sludge pyrolysis and gasification technology is an important technical approach for sludge reduction and resource utilization. Subject to strict industry emission standards, it aims to reduce SO2 and NO in the exhaust gas. X Maintaining particulate matter and other indicators within the specified limits is a prerequisite for the large-scale application of this technology. While traditional end-of-pipe treatment facilities can effectively reduce emissions, they significantly increase operating costs. Therefore, achieving high-precision, real-time emission prediction and proactive control at the source is a crucial breakthrough that balances environmental compliance and cost control.

[0003] The sludge pyrolysis gasification process is a highly interconnected and complex industrial system. Its main production line consists of a gasifier, a combustion furnace, a waste heat boiler, and a baghouse dust collector connected in series. Dewatered sludge sequentially passes through the drying, pyrolysis, and reduction zones of the gasifier to generate combustible syngas. The syngas then undergoes vigorous combustion in the combustion furnace, converting chemical energy into high-temperature flue gas. The flue gas then passes through the waste heat boiler for heat recovery and the baghouse dust collector for particulate matter removal before being discharged in compliance with standards. Fluctuations in the operating conditions of equipment along the process flow have a significant cross-equipment transmission effect and time lag effect on end-point emissions; upstream disturbances are propagated downstream along the process chain. Furthermore, actual production facilities typically have multiple parallel production lines, with the flue gas from each line ultimately merging and being emitted from the same monitoring point. While distributed control systems can collect a large number of electromechanical operating parameters, these parameters are susceptible to random noise interference and lack a direct physical correspondence with the internal chemical reactions and molecular dynamics of end-point emissions.

[0004] Existing emission prediction technologies mainly fall into two categories: mechanistic modeling and data-driven modeling. In terms of mechanistic modeling, thermodynamic equilibrium models can predict syngas composition but neglect internal time delays and mass transfer processes; kinetic models can characterize changes in intermediate variables but rely on rigid linearization assumptions, making them difficult to adapt to strongly nonlinear operating conditions; computational fluid dynamics models, while able to reconstruct transport phenomena relatively accurately, have excessive computational burdens, making it difficult to meet the timeliness requirements of real-time control. Simply relying on the idealized assumptions of mechanistic modeling cannot adapt to continuous industrial disturbances such as feed moisture content fluctuations and equipment wear, nor can it handle high-dimensional, noisy field operating data, thus limiting its online prediction capabilities.

[0005] In terms of data-driven approaches, machine learning and deep learning methods have been widely used for industrial exhaust emission prediction. However, existing frameworks generally suffer from limitations in empirical mapping from raw operating indicators to emission concentrations: black-box models based on static features or simple sliding windows do not embed chemical reaction mechanisms or specific device configuration information, resulting in a sharp decline in prediction performance when deviating from steady-state conditions; while simple time-series networks can capture time dependencies, they ignore the spatial topology between devices and the convergence paths of multiple production lines, easily leading to prediction lags during sudden increases or decreases in emissions; directly predicting absolute pollutant concentrations often introduces severe autocorrelation bias. Although some studies have attempted to introduce graph neural networks to characterize the relationships between devices, they typically treat each device node as an unweighted, isotropic node, failing to encode the directional relationships of material and energy transfer between devices, and neglecting the differentiated time-lag characteristics of different operating parameters relative to emission indicators. Furthermore, existing interpretability methods mainly rely on statistical attribution tools, making it difficult to reconstruct the physical quality and energy transfer paths in the process, and also failing to reveal the spatial response mechanisms of the model's internal evolution over time, thus limiting the deployment and reliability of data-driven models in practical engineering.

[0006] Therefore, there is an urgent need for a waste gas emission prediction method that can simultaneously characterize the directed spatial coupling relationship and characteristic-level time delay effect between sludge pyrolysis and gasification process equipment, and has both sensor anomaly robustness and implicit process mechanism interpretability, so as to make up for the shortcomings of existing pure time-series black box models and traditional graph neural networks in terms of structural consistency, dynamic robustness and mechanism interpretability. Summary of the Invention

[0007] The present invention aims to at least partially solve one of the technical problems in the related art.

[0008] Therefore, the first objective of this invention is to propose a method for predicting waste gas emissions from sludge pyrolysis gasification based on spatiotemporal coupling modeling of directed process topology graphs and graph neural networks.

[0009] The second objective of this invention is to provide a computer device.

[0010] A third objective of this invention is to provide a non-transitory computer-readable storage medium.

[0011] To achieve the above objectives, a first aspect of the present invention proposes a method for predicting sludge pyrolysis gasification waste gas emissions based on spatiotemporal coupling modeling of directed process topology graphs and graph neural networks, comprising: a. Obtain equipment operation data and exhaust gas emission data from multiple parallel production lines of the sludge pyrolysis and gasification target plant, perform validity verification and outlier processing on the data, and obtain continuous time series data for modeling. b. Based on the order of material and energy transfer between equipment in the sludge pyrolysis and gasification process, and the actual connection relationship of multiple production lines converging to the same exhaust gas emission monitoring point, a directed process topology graph is constructed. Each piece of equipment is abstracted as a node in the directed process topology graph, and the transfer relationship between equipment is abstracted as a directed connection edge between nodes. c. For the operating parameter characteristics of the equipment corresponding to each node in the directed process topology graph, calculate the cross-correlation function between the characteristic and the target exhaust gas emission index, determine the optimal time delay step, and perform time delay alignment on the historical sequence of the characteristic according to the optimal time delay step to obtain the node input characteristics after time delay compensation. d. Based on the directed process topology graph, a graph neural network is used to perform directed message passing between adjacent nodes to obtain the spatial context message of each node; the node input features after time delay alignment are concatenated with the spatial context message and input into the recurrent neural network unit specific to the type of the device to update the hidden state of the node. e. Introducing a hierarchical attention mechanism, first, the hidden states of each equipment node within the same production line are aggregated with intra-line attention weighting, and then the production line representation vectors of each production line are aggregated with inter-line attention weighting to obtain a global perception vector; based on the global perception vector, the change increment of the exhaust gas emission concentration at the current moment is predicted, and the change increment is superimposed with the exhaust gas emission concentration at the previous moment to obtain the predicted exhaust gas emission concentration at the current moment.

[0012] In one embodiment of the present invention, the data validity verification and outlier handling include: The quantile gradient jump detection method is used to scan the lower and upper intervals of the quantile distribution curve of the exhaust gas emission monitoring index. The inflection point of the gradient abnormal change is determined as the physical operating boundary of the index, and values ​​that exceed the boundary are eliminated. Map integer hardware fault values ​​that exceed the preset quantile range to missing values; mark intervals where the sensor continuously outputs zero values ​​or readings are stuck as invalid data; When the proportion of invalid data for the same operating parameter on any production line exceeds a preset threshold, the parameter is removed from all production lines. For the remaining valid intervals, the mean, median, and standard deviation within the window are calculated based on a sliding window. Data points that deviate from the median within the window by more than a preset multiple of the standard deviation are identified as outliers and corrected using a linear interpolation method.

[0013] In one embodiment of the invention, a sample-level quality mask for labeling the credibility of data is also included: For continuous abnormal data intervals whose correction duration does not exceed a preset duration threshold, a sliding window interpolation method is used to fill them, and the corresponding mask is marked as reliable. For consecutive abnormal data intervals whose correction duration exceeds the preset duration threshold, the global mean of the corresponding indicator is used to fill the intervals, and the corresponding mask is marked as untrustworthy.

[0014] In one embodiment of the present invention, the method for constructing the directed process topology graph includes: On each production line of the sludge pyrolysis gasification process, the gasifier, combustion furnace, waste heat boiler and bag filter are abstracted as nodes in sequence. According to the actual process sequence of material and energy being transferred in sequence along the gasifier, combustion furnace, waste heat boiler and bag filter, directed connection edges are established between adjacent nodes. The outputs of the corresponding nodes of the bag filters of each parallel production line are connected to the same exhaust gas emission monitoring node to represent the actual connection relationship of the exhaust gases of multiple production lines converging at the emission monitoring point.

[0015] In one embodiment of the present invention, the method for determining the node input features further includes a feature selection step: For the original operating parameters of the equipment corresponding to each node in the directed process topology graph, the mutual information and Pearson correlation coefficient between the parameter and the target exhaust gas emission index are calculated respectively. Only parameters with mutual information greater than a preset threshold or absolute value of Pearson correlation coefficient greater than a preset threshold are retained. Combined with the understanding of the process mechanism of material and energy transfer of each equipment unit, parameters with causal relationship with the target exhaust gas emission index are further screened out to form the optimal feature subset for constructing the input features of each node.

[0016] In one embodiment of the present invention, the graph neural network uses a scaled dot product attention mechanism to calculate the confidence coefficient: Using learnable weight matrices corresponding to the device types of upstream and downstream nodes, the historical hidden states of upstream and downstream nodes are mapped to message key vectors and information query vectors, respectively. The message key vectors and information query vectors are then multiplied by a dot product and mapped using a Sigmoid function to obtain the confidence coefficients transmitted from upstream to downstream nodes. The message key vectors of upstream nodes are then scaled using these confidence coefficients to obtain the spatial messages transmitted from upstream to downstream nodes.

[0017] In one embodiment of the present invention, a set of recurrent neural network parameters is maintained independently for different types of device nodes in the directed process topology graph, so that each type of device has its own independent hidden state transition characteristics.

[0018] In one embodiment of the invention, robust design for sensor anomalies and sudden changes in operating conditions is also included: Construct a dual-objective loss function that includes the mean square error of the predicted concentration change increment and the prediction error of the superimposed absolute concentration; By combining the constructed sample-level quality mask, the gradient contribution of data points marked as untrusted is truncated during the backpropagation process, preventing the model from learning false patterns introduced by manually filled data. This enables the method to maintain stable prediction of exhaust emission trends by relying on the learned process inertia under extreme conditions such as local sensor failure or abnormal freezing.

[0019] In one embodiment of the present invention, online deployment is further performed using a rolling prediction mechanism: The prediction period is divided into several consecutive fixed-duration intervals. At the beginning of each interval, the hidden state of the recurrent neural network is initialized with the measured values ​​of the exhaust gas emission monitoring points. In subsequent time intervals within this interval, autoregressive rolling inference is performed based on the prediction results of the previous time interval to output the predicted sequence of exhaust gas emissions within this interval.

[0020] In one embodiment of the present invention, the target exhaust gas emission indicators include SO2 concentration and NO. X Concentration and particulate matter concentration; the attention weights within and between production lines extracted by the hierarchical attention mechanism are used to identify key production lines and key equipment units that dominate the fluctuations in exhaust gas emissions at different times, so as to achieve interpretable mining of the implicit reaction mechanism of sludge pyrolysis gasification process.

[0021] In one embodiment of the present invention, the recurrent neural network unit dedicated to the exhaust gas emission monitoring point takes the global perception vector and the exhaust gas emission target value of the previous moment as input, updates its hidden state and predicts the change increment of the exhaust gas emission concentration at the current moment, and then superimposes the change increment with the exhaust gas emission concentration of the previous moment to output the predicted exhaust gas emission concentration at the current moment.

[0022] The present invention provides a method for predicting waste gas emissions from sludge pyrolysis gasification based on a spatiotemporally coupled graph neural network, which has the following beneficial effects: 1. This invention discloses a method for predicting emissions from sludge pyrolysis gasification waste gas based on a spatiotemporally coupled graph neural network. By constructing a directed process topology graph reflecting the direction of material and energy transfer between equipment and the convergence relationship of multiple production lines, and employing a graph neural network for directed message passing, it can realistically depict the spatial coupling relationship and signal propagation direction between equipment such as the gasifier, combustion furnace, waste heat boiler, and bag filter. Compared to traditional graph neural network methods that treat equipment nodes as unweighted, isotropic nodes, this method can more accurately represent the actual topological structure of the sludge pyrolysis gasification process, significantly improving emission prediction accuracy. 2. This invention discloses a method for predicting emissions from sludge pyrolysis gasification waste gas based on a spatiotemporally coupled graph neural network. It determines the optimal time delay step size for each feature based on the operating parameter characteristics of each node in the topology graph and performs time delay alignment, overcoming the problem of differentiated lags in relative emission indicators for different process stages and operating parameters. Compared to methods that uniformly apply a fixed lag step size to all features, this method significantly reduces prediction bias during operating condition transitions and emission peak stages, improving the prediction stability of transient processes. 3. This invention discloses a method for predicting emissions from sludge pyrolysis gasification waste gas based on a spatiotemporally coupled graph neural network. This method couples the spatial message passing of the graph neural network with the temporal evolution of a recurrent neural network categorized by equipment type, and introduces a hierarchical attention mechanism to aggregate the operating states within and between production lines. Compared to prediction methods that rely solely on historical time series or spatial topology, this method can simultaneously capture the equipment's own reaction inertia and the transmission of disturbances across equipment, achieving more accurate and timely predictions during transient processes such as sudden increases and decreases in emissions. 4. This invention discloses a method for predicting waste gas emissions from sludge pyrolysis gasification based on a spatiotemporally coupled graph neural network. Through a robust design combining a dual-objective loss function and a sample-level quality mask, the gradient contribution of unreliable data points during backpropagation is truncated. Compared to methods that train directly using all original data, this method can maintain stable predictions of waste gas emission trends under extreme conditions such as local sensor failure and abnormal freezing, relying on learned process inertia, significantly improving the reliability of the model deployment in complex industrial sites. 5. The present invention discloses a method for predicting waste gas emissions from sludge pyrolysis gasification based on a spatiotemporally coupled graph neural network. The extracted hierarchical attention weights within and between production lines can adaptively identify key production lines and key equipment units that dominate waste gas emission fluctuations at different times without prior mechanism labels. This method can uncover the implicit reaction mechanism of the sludge pyrolysis gasification process and provide interpretable decision-making basis for process diagnosis, source control, and the construction of digital twin systems.

[0023] To achieve the above objectives, a second aspect of this application provides a computer device, including a processor and a memory; wherein the processor runs a program corresponding to the executable program code stored in the memory, in order to implement a method for predicting sludge pyrolysis gasification emissions based on a spatiotemporally coupled graph neural network as described in the first aspect embodiment.

[0024] To achieve the above objectives, a third aspect of this application provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements a method for predicting sludge pyrolysis gasification emissions based on a spatiotemporally coupled graph neural network as described in the first aspect embodiment.

[0025] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0026] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart of a method for predicting waste gas emissions from sludge pyrolysis gasification based on a spatiotemporally coupled graph neural network according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the directional topology of the sludge pyrolysis gasification process according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the spatiotemporal coupled network architecture and hierarchical attention mechanism according to an embodiment of the present invention. (A) is a schematic diagram of graph neural network message passing and LSTM spatiotemporal coupled update of device nodes, and (B) is a schematic diagram of hierarchical attention aggregation and end-of-pipe emission increment prediction. Figure 4 The spatiotemporal coupling model of LightGBM, GNN (spatial modeling), LSTM (temporal modeling), and the LSTM-GNN of this invention is used to evaluate SO2 concentration and NO. X Comparison of the goodness-of-fit of predictions for ablation experiments using SO2 concentration and particulate matter concentration: (A)-(D) represent the predicted SO2 concentration, and (E)-(H) represent the predicted NO concentration. X Concentration prediction results, (I)-(L) are the particulate matter concentration prediction results; Figure 5 For different ablation models, SO2 concentration and NO on the test set X Time-series comparison curves of predicted and actual concentrations of SO2 and particulate matter: (A) SO2 concentration, (B) NO concentration. X Concentration, (C) represents particulate matter concentration; Figure 6 Verification diagrams for the feature-level optimal time-delay alignment strategy and the robustness mechanism of the quality mask: (A) and (C) represent SO2 concentration and NO concentration, respectively. X The peak time prediction curves of concentration under fixed time delay and optimal time delay strategy are compared. (B) and (D) are box plots of absolute error distribution in the corresponding high concentration range. (E) and (F) are the effects of the masking mechanism on the prediction trajectory under local abnormal pressure test of sensor. (G) and (H) are fitting plots comparing the prediction accuracy of SO2 concentration and particulate matter concentration before and after the introduction of the masking mechanism. Figure 7The spatiotemporal interpretability analysis diagrams for the hierarchical attention mechanism are as follows: (A) Comparison of the stacked area diagram of production line-level attention weights and the predicted trajectory of SO2 concentration; (B) Spatiotemporal heat map of equipment-level global attention weights; (C) Equipment-level attention propagation trajectory diagram of production line 2. Figure 8 The following plots illustrate the equipment-level average attention weight distribution and the model's prediction accuracy on the independent-date external validation set: (A) and (B) represent the average attention weight percentage of each device in the SO2 and particulate matter concentration prediction tasks, respectively; (C) and (D) are the stacked area plots showing the evolution of equipment attention weights on production line 2 over time; and (E) shows the SO2 and NO concentrations on the independent-date external validation set. X Goodness-of-fit plot for concentration and particulate matter concentration prediction; Figure 9 It is a computer device according to an embodiment of the present invention. Detailed Implementation

[0027] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0028] To enable those skilled in the art to better understand the present invention, 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. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0029] To address the shortcomings of existing emission prediction models, such as difficulty in characterizing directed spatial topological relationships and feature-level time delay effects between equipment, insufficient robustness under sensor anomalies or drastic fluctuations in operating conditions, and lack of interpretability for implicit process mechanisms, this invention aims to provide a method for predicting sludge pyrolysis gasification waste gas emissions based on a spatiotemporally coupled graph neural network. By constructing a directed process topology graph reflecting the direction of material and energy transfer between equipment and the convergence relationship of multiple production lines, feature-level optimal time delay alignment is performed on the input features of each node based on a cross-correlation function. An architecture coupling graph neural network and recurrent neural networks based on equipment type is adopted to achieve joint modeling of spatial topological message passing and temporal response inertia. Simultaneously, a hierarchical attention mechanism is introduced to aggregate the operating states within and between production lines at each level, achieving high-precision prediction, robust anti-interference prediction, and implicit process mechanism mining of sludge pyrolysis gasification waste gas emissions.

[0030] This invention discloses a method for predicting waste gas emissions from sludge pyrolysis gasification based on a spatiotemporally coupled graph neural network. The method acquires and cleans equipment operation data from multiple parallel production lines in a target sludge pyrolysis gasification plant, as well as waste gas emission data collected from the same monitoring point. A directed process topology graph is constructed based on the material and energy transfer order between equipment and the convergence relationship of multiple production lines, abstracting equipment as nodes and the transfer relationship between equipment as directed connecting edges. For the operating parameter characteristics of each node in the topology graph, the optimal time delay step size relative to the prediction target is calculated based on the cross-correlation function and time delay alignment is performed. Based on the directed process topology graph, a graph neural network is used for directed message passing between adjacent nodes, combined with recurrent neural network units of different equipment types to couple the node's own temporal evolution with the spatial topological influence across nodes. A hierarchical attention mechanism is introduced, first performing attention-weighted aggregation on each equipment node within the production line, and then performing attention-weighted aggregation on the representation vectors of multiple production lines to obtain the global perception vector of the waste gas emission monitoring point. Based on this, the incremental change in waste gas emission concentration is predicted, achieving spatiotemporally coupled dynamic prediction of waste gas emissions from sludge pyrolysis gasification.

[0031] The following description, with reference to the accompanying drawings, describes a method and apparatus for compressing and storing historical data of a hydropower monitoring system according to an embodiment of the present invention.

[0032] This embodiment takes a large-scale urban sewage sludge pyrolysis and gasification treatment system as the research object. The system includes three parallel sludge pyrolysis and gasification production lines. Each production line consists of a gasifier, a combustion furnace, a waste heat boiler, and a bag filter. The flue gas generated by the three production lines is finally combined and monitored by the same continuous emission monitoring system (CEMS) to measure SO2 concentration and NO. X Online monitoring of indicators such as concentration and particulate matter concentration. In this embodiment, approximately six months of historical operating data from the distributed control system (DCS) and emission monitoring data from the CEMS were collected. A spatiotemporal coupled graph neural network model that integrates directed process topology and feature-level optimal time delay alignment was used to predict changes in exhaust gas emission concentration. Furthermore, the implicit process mechanisms affecting exhaust gas emissions were uncovered using the hierarchical attention mechanism embedded in the model.

[0033] like Figure 1 As shown in the figure, this embodiment discloses a method for predicting waste gas emissions from sludge pyrolysis gasification based on a spatiotemporally coupled graph neural network. The specific implementation steps are as follows: S1: Data Acquisition from DCS and CEMS of Multiple Production Lines Operating parameters of the gasifiers, combustion furnaces, waste heat boilers, and bag filters of three parallel production lines were collected, along with exhaust gas emission data collected from CEMS monitoring points after the flue gas from the three production lines was combined. Among these: The operating parameters of the gasifier include: bottom blower pressure, flow rate, and temperature; steam pressure, steam temperature, and steam flow rate; gasifying agent pressure and temperature; vapor-liquid separator level, pressure, and temperature; gasifier outlet pressure and outlet temperature; average slag layer temperature; average temperature of the dry distillation zone; average temperature of the drying zone; feed level; slag discharge speed; slag lock temperature; and weight of the top hopper. The operating parameters of the combustion furnace include: pressure and flow rate of the main fire blower (combustion aid blower), biogas pressure, upper furnace temperature, lower furnace temperature, furnace negative pressure, outlet oxygen content, and outlet temperature. The operating parameters of the waste heat boiler include: boiler drum pressure, boiler drum liquid level, steam flow rate, boiler inlet pressure, inlet temperature, urea flow rate, boiler outlet pressure, and outlet temperature. The operating parameters of a bag filter dust collector include: outlet pressure, outlet temperature, and induced draft fan frequency; Exhaust gas emission data include: SO2 concentration, NO X Concentration, particulate matter (PM) concentration.

[0034] S2: Data Cleaning and Sample-Level Quality Mask Construction The raw operational and emission data acquired by S1 are preprocessed sequentially as follows to obtain continuous time-series data for constructing directed topological node features: (1) Using the quantile gradient jump detection method, scan the lower end (0.1%-10%) and upper end (90%-99.9%) intervals of the quantile distribution curve of each exhaust gas emission monitoring index, determine the inflection point of the gradient abnormal acceleration change as the physical operating boundary of the index, and eliminate the values ​​that exceed the boundary. (2) Map integer outliers (such as -4.0, -500.0, -5000.0, etc., obvious hardware fault codes) that exceed the 1%-99th percentile range in each original DCS data file to missing values; at the same time, identify the range of sensor failure (reading is always zero for 2 consecutive hours) or hardware stagnation (variance of 60-minute sliding window is always zero for 2 consecutive hours), replace them with missing values ​​and mark them as invalid (Flag=False); when the invalid ratio of the same indicator on any production line exceeds 30%, remove the feature column from all three production lines. (3) For intervals marked as valid (Flag=True), calculate the window mean, median and standard deviation based on a 30-minute sliding window. Data points that deviate from the window median by more than 3 times the standard deviation are judged as outliers and corrected by linear interpolation. Those located at the window boundary are filled with the window median. (4) Define a sample-level quality mask to distinguish the credibility of the data: for the part of the continuous abnormal interval whose correction time does not exceed the preset threshold, use sliding window interpolation to fill it and mark the mask as credible; for the continuous abnormal interval whose correction time exceeds the preset threshold, use the global mean of the index as a placeholder and mark the mask as untrustworthy, so as to explicitly prompt the model of the unreliability of the data in the interval while maintaining the continuity of the time series. (5) All operating parameters and emission indicators after cleaning are Z-score standardized. The mean and standard deviation statistics required for standardization are extracted only from the training set to avoid data leakage.

[0035] S3: Feature selection and feature-level optimal time delay alignment like Figure 2 As shown, in order to avoid introducing noise and increasing the computational burden by directly inputting all the original operating parameters into the model, this embodiment adopts a dual screening strategy that combines statistical indicators and process mechanisms to extract the optimal feature subset. Specifically, for the target exhaust emission indicators, the mutual information and Pearson correlation coefficient between each original operating parameter and the target indicator are calculated. Following statistical convention, 0.1 is used as the lower limit for determining a significant statistical relationship, and only parameters with mutual information greater than 0.1 or absolute Pearson correlation coefficients greater than 0.1 are retained. Simultaneously, the process mechanism of reaction efficiency and exhaust gas fluctuations in each equipment unit is verified. For example, in the gasifier, key variables characterizing the intake volume and gasification efficiency, such as the flow rate and temperature of the bottom fan, are retained, while redundant electromechanical indicators (such as frequency and current) are eliminated. In the combustion furnace, parameters directly related to reaction kinetics, such as the pressure and flow rate of the main combustion fan, biogas pressure, furnace temperature, and furnace negative pressure, are retained. In the waste heat boiler and bag filter, parameters affecting heat recovery and purification efficiency, such as boiler drum pressure, liquid level, steam flow rate, and induced draft fan frequency, are retained. Finally, these parameters are aggregated to form the optimal feature subset used to construct the input features of each node in the directed topology graph.

[0036] After feature selection, for each feature in the optimal feature subset, the cross-correlation function (CCF) is used to calculate the cross-correlation coefficient between the feature and the target exhaust emission index sequence at different time offsets. The time offset that maximizes the absolute value of the cross-correlation coefficient is determined as the optimal time lag step size for that feature relative to the predicted target. The maximum lag search range is preset based on process experience. By performing the above time lag compensation on each feature separately, the model input features and the end-of-pipe emission targets are strictly aligned on the physical response time scale, thereby overcoming the problem of differentiated lags between different process stages and parameters relative to emission indicators.

[0037] S4: Directed Process Topology Graph Construction and Spatiotemporal Coupling Modeling of Graph Neural Networks like Figure 3As shown in (A), to collaboratively characterize the temporal dynamics and spatial topological relationships in the sludge pyrolysis gasification process, this embodiment constructs a spatiotemporally coupled prediction architecture that integrates Long Short-Term Memory (LSTM) and Graph Neural Network (GNN). In terms of the process, the gasifier, combustion furnace, waste heat boiler, and bag filter constitute a unidirectional material and energy cascade transfer network. The operating status of the upstream equipment directly determines the composition of the downstream flue gas, exhibiting a directional spatial dependency that is difficult to characterize using traditional neural networks. To embed this physical process boundary, this embodiment constructs a directed process topology graph based on an adjacency matrix, abstracting each equipment unit as a discrete node in the topology graph. The graph neural network aggregates the operating status of the upstream equipment and propagates this spatial information to the downstream nodes.

[0038] Specifically, a scaled dot product attention mechanism is adopted to adapt to the heterogeneous node types along the process line: For upstream and downstream nodes, the historical hidden states are mapped to the shared embedding space using learnable weight matrices corresponding to their respective device categories to obtain message key vectors and information query vectors; the downstream node obtains a confidence coefficient that quantifies its acceptance of information from the upstream node by performing a scaled dot product operation on the message key vector and the information query vector and mapping it through the Sigmoid activation function; the upstream node's message key vector is scaled using this confidence coefficient to obtain the directed message passed from the upstream node to the downstream node; in order to integrate the concurrent influence from multiple adjacent upstream devices, all directed messages received by the target node are summed and aggregated to obtain the spatial context message of the node at the current moment.

[0039] For the four types of equipment deployed on the production line, the temporal state evolution of each node is driven by a dedicated LSTM unit. To adapt to the unique physical behavior and operating characteristics of different equipment, this embodiment maintains an independent set of LSTM parameters for each type of equipment, enabling each type of equipment to have its own exclusive hidden state transition dynamics. At each time step, the node concatenates the input feature sequence aligned by S3 time delay with the spatial neighbor messages obtained by the graph neural network layer, and inputs it into the LSTM unit dedicated to that equipment type to update the node's hidden state, thereby achieving a deep fusion of spatial topological information and local temporal evolution: the node can not only dynamically update its own state based on the current input and historical memory, but also perceive the disturbances of upstream equipment in real time, collaboratively characterizing the dynamic process of the entire process flow.

[0040] S5: Hierarchical Attention Convergence and End-of-Line Emission Prediction like Figure 3As shown in (B), after obtaining the hidden states of each equipment node, a hierarchical attention mechanism is introduced to reconstruct the physical process of the three parallel production lines converging at the same CEMS monitoring port, giving the framework multi-scale interpretability. Specifically, for the l-th production line, the model calculates the attention weights within the production line, weights and aggregates the hidden states of the four core equipment units on the production line to generate a production line representation vector that characterizes the overall operating state of the production line. Analyzing the dynamic distribution of these attention weights can determine whether emission disturbances originate from the front-end gasification stage or the back-end purification stage. To take into account the differentiated loads and operational contributions of the three parallel production lines under actual operating conditions, the CEMS monitoring node performs inter-line attention weighted aggregation on the representation vectors of each production line to obtain a global perception vector that characterizes the overall operating state of the system. This weight can clearly indicate the key production line that dominates the end-of-pipe emission trend at the current moment. Based on the global perception vector and the emission target value at the previous moment, the dedicated CEMS node LSTM unit updates its hidden state to capture the temporal dynamics of the emission sequence.

[0041] To avoid the autocorrelation bias and prediction lag introduced by directly predicting the absolute concentration of pollutants, this embodiment uses a fully connected layer to map the hidden state of the CEMS node to the predicted emission concentration change increment. During the stepwise forward inference process, the absolute concentration value of the previous moment is superimposed with the predicted change increment of the current moment to calculate the final emission concentration prediction value of the current moment.

[0042] S6: Robust Design, Model Training, and Rolling Prediction Deployment The continuous time series data were divided into training and test sets in an 8:2 ratio, with 10% allocated from the training set as a validation set. During training, the model employed a bi-objective loss function, balancing the mean squared error of the predicted concentration increments with the summed absolute concentration prediction error. The weight hyperparameter in the loss function was optimized to a value of 0.5 to balance trend capture and absolute baseline prediction. To filter unreliable manually imputed data, a sample-level quality mask constructed using S2 was used to truncate the gradient contribution of untrusted sample points during backpropagation, preventing the model from learning spurious patterns introduced by the manually imputed data. Training was automatically terminated early when the validation set loss improved by less than 1% over five consecutive training epochs, and the model weights that performed best on the validation set were rolled back.

[0043] During the testing phase, a 5-minute rolling prediction architecture was used to simulate online deployment under real-world conditions: the test sequence was divided into several consecutive 5-minute intervals. At the beginning of each interval, the LSTM hidden state of each node was initialized with the measured values ​​of CEMS. In subsequent time intervals, the model performed autoregressive forward inference based only on the prediction results of the previous time interval, and output the exhaust emission prediction sequence for that interval. This design can suppress error accumulation drift and simulate the real-time status feedback process of the factory.

[0044] S7: Ablation experiments verify the effectiveness of the spatiotemporal coupling architecture To systematically evaluate the independent contributions of each core module in the spatiotemporal coupling architecture proposed in this invention, this embodiment sets up multiple ablation control models under the principle of strictly controlling variables: each variant shares the same input features, prediction output layer, loss function, and training strategy, differing only in the hidden state update mechanism of device nodes. First, LightGBM is used as the baseline model, flattening the features of all nodes after time-delay alignment into a single vector input, completely destroying temporal evolution and spatial topology; second, a GNN ablation model that retains only spatial information is constructed, where nodes update their state by concatenating the instantaneously aligned features with the upstream neighbor messages propagated by the graph neural network at each time step; then, an LSTM ablation model that retains only temporal information is constructed, setting all upstream spatial messages received by each node to zero; finally, the complete LSTM-GNN spatiotemporal coupling model of this invention is presented, simultaneously integrating explicit historical windows and graph message passing.

[0045] like Figure 4 As shown in Table 1, the LightGBM baseline model affects SO2 concentration and NO concentration. X The goodness of fit (R²) for concentration prediction was only 31.40% and 64.52%, respectively, with the R² for particulate matter concentration prediction even dropping to -477.78%. This indicates that, ignoring the temporal evolution and spatial topology of the process, conventional machine learning models struggle to establish an effective mapping between high-dimensional, non-stationary operating signals and emission concentrations. GNN models that only incorporate spatial topology information cannot effectively map SO2 and NO... X The R² values ​​for SO2 and NO2 increased to 64.74%, 60.50%, and 72.79%, respectively, indicating that even without explicitly introducing multi-step time inputs, the upstream equipment status information propagated through the process topology can still provide effective prediction signals. The LSTM model retaining only the time dimension showed a significant leap in prediction accuracy, particularly for SO2 and NO2. X The R² values ​​for SO2 and particulate matter reached 84.22%, 70.76%, and 75.49%, respectively. The R² for SO2 was improved by more than 30 percentage points compared to the GNN model, indicating significant physicochemical inertia in the sludge pyrolysis and gasification process, with historical trend characteristics being more important than pure spatial interactions. The complete LSTM-GNN spatiotemporal coupling model of this invention achieved optimal prediction accuracy for all predicted targets, including SO2 and NO. XThe R² values ​​for particulate matter reached 91.18%, 80.37%, and 78.09%, respectively. Compared to the LSTM model based solely on the time dimension, the R² values ​​for the three pollutants, after incorporating spatial topological information, showed consistent improvements of 6.95, 9.61, and 2.6 percentage points, respectively. This indicates that spatial interaction is a necessary supplement to temporal memory, and only through collaborative modeling can the dynamic process of the entire process be accurately depicted. The corresponding root mean square error (RMSE) also verified the above conclusions: the predicted RMSE for SO2 concentration decreased successively from 638.37 mg / m³ in LightGBM to 457.65 mg / m³ in GNN, 306.13 mg / m³ in LSTM, and finally to 228.92 mg / m³ in the LSTM-GNN model of this invention.

[0046] Table 1. Comparison of the goodness of fit of four ablation models for predicting three waste gas emission indicators.

[0047] S8: Prediction Stability and Residual Analysis like Figure 5 As shown, in SO2 concentration prediction, the LSTM-GNN model of this invention tracks the actual fluctuation trend more closely than other control models: the LightGBM baseline model cannot distinguish emission peaks and troughs at all; the time-only LSTM model significantly overestimates emissions during the transition from peak to trough, because it relies solely on historical trends, and historical high emission trajectories make it difficult to respond promptly to changes in upstream equipment parameters; the spatial-only GNN model is prone to underestimation because it directly maps the aligned operating indicators of the current moment to the emission concentration, without considering the inherent lag in the reaction process. Local magnification results show that the model of this invention achieves a better fit to emission peaks and can provide a more forward-looking early warning for SO2 emissions. This advantage is also evident in NO... X This is more pronounced in the prediction of concentration and particulate matter concentration.

[0048] To further verify statistical robustness and rule out systematic biases, this embodiment performs distribution analysis on the prediction residuals, and the results are shown in Table 2: SO2, NO X The average residuals of particulate matter concentrations were 28.08 mg / m³, -6.84 mg / m³, and -0.01 mg / m³, respectively, which are negligible relative to the baseline concentration levels of each pollutant. The corresponding skewnesses were -0.57, 0.44, and 0.61, respectively, indicating that the prediction error is highly symmetrically distributed near the zero error baseline. The residual distribution exhibits a symmetrical bell-shaped curve, which is highly consistent with the theoretical Gaussian distribution. This shows that the LSTM-GNN model proposed in this invention has fully learned the deterministic spatiotemporal dynamic laws contained in the DCS signal and has not introduced systematic bias.

[0049] Table 2 Statistical characteristics of the prediction residuals of the spatiotemporal coupling model of the present invention

[0050] S9: Sensitivity Analysis of Feature-Level Optimal Delay Alignment Strategy To verify the effectiveness of the feature-level optimal time-delay alignment strategy proposed in S3, this embodiment sets up a control experiment, uniformly setting the lag step size of all features to a fixed value τ_fix, and comparing it with the optimal time-delay τ_best configuration of this invention. Overall, the difference in overall prediction accuracy at the macroscopic level is not significant: SO2 and NO... X The test set R² is reduced by about 4% compared to the optimal time lag configuration, and the particulate matter concentration prediction accuracy is almost unaffected. This phenomenon stems from the physical inertia of the continuous and closed operation of the sludge pyrolysis gasification system—once the system enters a stable operating state, the emission concentration exhibits a strong time dependence. For a specific operating parameter, even if the moment with the most information is at the optimal lag step, the historical sliding window can still capture the changing trend of the parameter over the past few time steps. In addition, the CEMS node status update integrates the global perception vector and the emission target value of the previous moment, thus avoiding a significant decline in macroscopic prediction accuracy.

[0051] However, as Figure 6 As shown in (A)-(D), a detailed examination of specific emission peak periods reveals that introducing dynamic optimal time-delay alignment can significantly improve the prediction stability during the operating condition transition phase: within the emission fluctuation range from 17:00 to 20:00 on a certain day in the test set, the prediction curve using the optimal time-delay configuration tracks the actual emission changes more closely than the fixed time-delay configuration, which exhibits greater fluctuations and a significant prediction lag; within this range, the mean absolute error (MAE) and root mean square error (RMSE) of SO2 concentration prediction under the optimal time-delay configuration are 138.9 mg / m³ and 176.56 mg / m³, respectively, representing reductions of 51.5% and 48.2% compared to the fixed time-delay configuration. Further statistical analysis of the absolute errors for high-concentration emission points above the 75th quantile shows that the median, 75th quantile, and maximum error of the fixed time-delay group are significantly higher than those of the optimal time-delay group, confirming that characteristic-level optimal time-delay alignment can effectively alleviate prediction bias during transient emission peaks.

[0052] S10: Robustness verification of the sample-level quality masking mechanism To verify the contribution of the quality masking mechanism constructed by S2 to the model's robustness, this embodiment compares the performance of models with and without the masking mechanism. Figure 6As shown in (G) and (H), the goodness of fit R² for SO2 concentration prediction is approximately 86% and 90%, respectively, with mean absolute errors of approximately 207.26 mg / m³ and 167.69 mg / m³. For particulate matter concentration prediction, the R² is approximately 75% and 73%, respectively, with mean absolute errors of approximately 0.08 mg / m³ and 0.06 mg / m³. There is no significant difference in accuracy between the two configurations, indicating that the masking mechanism can eliminate the gradient contribution of damaged samples and improve robustness without sacrificing the overall prediction accuracy.

[0053] To further verify the role of the masking mechanism in maintaining process inertia, this embodiment simulates an extreme stress test of local sensor failure: The core operating parameters of the gasifiers on three production lines, including the pressure, flow rate, and temperature of the furnace bottom blowers, as well as the steam pressure and temperature, were artificially replaced with the historical maximum values ​​of the corresponding parameters from the training set during the interval of 2:00 AM to 3:00 AM on a certain day in the test set, to simulate the condition of local sensor failure. The results show that, as... Figure 6 As shown in (E) and (F), the model without the masking mechanism rapidly deviates from the actual emission trajectory and exhibits violent oscillations, demonstrating overfitting to the abnormal DCS signal and failing to identify the failure state. In contrast, the model with the masking mechanism can still accurately track the actual emission change trend, indicating that the mechanism not only truncates the gradient backpropagation of damaged samples during the training phase, but also prompts the graph neural network and recurrent neural network units to rely more on historical effective information and spatial topological relationships to learn the real process dynamics. This enables the model to accurately capture short-term process inertia even under critical data anomalies, and this capability has important practical value for highly variable industrial engineering sites.

[0054] S11: Spatiotemporal interpretability analysis of hierarchical attention mechanisms To overcome the black-box nature of model predictions and improve process transparency, this embodiment utilizes a hierarchical attention mechanism to dynamically trace the sources of end-of-pipe emission fluctuations. By extracting production line-level and equipment-level attention weights, the model can autonomously identify the key production lines and key equipment units that contribute most significantly to end-of-pipe emissions at any given time, without requiring prior mechanistic labels.

[0055] like Figure 7As shown in Figure (A), the production line-level analysis results indicate that during the operation period when SO2 concentration is at a high level, the attention weight allocated to production line 2 is absolutely dominant: around 9:15 on a certain day, the weight of production line 2 rapidly climbed to over 45%, and remained at a high level of over 40% during the subsequent period of continuous high emissions from 13:00 to 19:00; after 21:00, the weight of production line 2 dropped, and the weight of production line 1 became dominant, accompanied by a synchronous and steady decrease in the end SO2 concentration from a high pollution level to the baseline level, indicating that the abnormal end emissions on that day were mainly triggered by the operating condition fluctuations of production line 2.

[0056] like Figure 7 As shown in (B), the equipment-level analysis results further indicate that among all equipment units in the three production lines, the gasifier consistently exhibits the strongest attention weight intensity, making it the core source dominating changes in terminal SO2 concentration. A horizontal comparison reveals that the gasifier in production line 2 has the highest attention activity, followed by production line 3, while production line 1 contributes little before 21:00 but then rapidly increases, maintaining a high degree of temporal consistency with terminal concentration fluctuations. Figure 7 As shown in Figure (C), the analysis of the attention trajectory of production line 2 during the period from 9:00 to 10:00 on a certain day further reveals the spatiotemporal propagation mechanism of equipment-level attention before the emission peak: during this period, there were two significant surges in end-of-pipe SO2 emissions. The model's perception of the gasifier's operating parameters was significantly ahead of the actual detection at the CEMS end-of-pipe monitoring point before each surge in emissions. Subsequently, the growth trend of attention weight strictly followed the physical sequence of syngas flow, which was transmitted from the gasifier to the combustion furnace, waste heat boiler, and bag filter in sequence. This dynamic propagation mode was most evident 5 to 8 minutes before the emission peak reached the CEMS monitoring port.

[0057] S12: Heterogeneity of Device-Level Attention Allocation and Model Generalization Validation As shown in Table 3 and Figure 8 As shown in (A) and (B), the distribution of average attention weights at the equipment level differs significantly under different pollutant prediction tasks: In SO2 concentration prediction, attention is highly concentrated on the upstream gasifier (65.46%) and combustion furnace (27.17%), with a combined share of 92.63%, mainly corresponding to the kinetics of syngas generation and oxidation reaction; the weight of the end-stage bag filter is only 0.42%, which is highly consistent with the fact that dry filtration cannot intercept gaseous pollutants. In particulate matter concentration prediction, the model shifts to a full-process collaborative attention mode, with attention distributed more evenly among the gasifier (29.82%), combustion furnace (21.38%), waste heat boiler (23.45%), and bag filter (25.35%). Figure 8The temporal evolution shown in (C) and (D) also indicates that this allocation pattern remains stable at different times. This adaptive attention allocation phenomenon is highly consistent with the pollutant formation mechanism: SO2, NO X Gaseous pollutants are products of multi-stage chemical reactions upstream, and there is a significant spatiotemporal span between the reaction sites and the end-of-pipe monitoring points, thus exhibiting an upstream-dominated, lag-type response characteristic. In contrast, particulate matter concentration is a physical result jointly determined by the instantaneous operating states of each process unit; the lag in the propagation of dust in high-velocity flue gas is negligible, thus exhibiting a spatially coordinated response characteristic across the entire process. These results demonstrate that the spatiotemporal coupling architecture proposed in this invention can autonomously capture the inherent physicochemical dynamics and spatiotemporal correlations of complex engineering networks without prior mechanistic labels, endowing data-driven models with physical and engineering credibility.

[0058] Table 3. Distribution of equipment-level average attention weights in SO2 and particulate matter concentration prediction tasks.

[0059] To evaluate the predictive robustness of the proposed method under unseen operating conditions, this embodiment introduces new running date data, completely independent of the training and validation sets, for external validation. For example... Figure 8 As shown in Figure (E) and Table 4, the predicted results maintain a high goodness of fit with the actual emission trajectory under all unseen operating conditions: the predicted R², RMSE, and MAE for SO2 concentration are 82%, 224.5 mg / m³, and 146.18 mg / m³, respectively; NO X The predicted R², RMSE, and MAE for concentration were 95%, 50.7 mg / m³, and 32.17 mg / m³, respectively; the predicted R², RMSE, and MAE for particulate matter concentration were 90%, 0.04 mg / m³, and 0.03 mg / m³, respectively. These results demonstrate that the spatiotemporal coupled graph neural network method proposed in this invention still possesses excellent generalization ability when dealing with common disturbances in industrial settings such as production line failures, fluctuations in feed properties, and mechanical drift. It can provide support for source emission control, digital twin construction, and intelligent operation optimization of sludge pyrolysis gasification processes.

[0060] Table 4. Prediction accuracy on independent date external validation sets

[0061] The above detailed description further illustrates the purpose, technical solution, and beneficial effects of the invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0062] To implement the methods of the above embodiments, the present invention also provides a computer device, such as... Figure 9 As shown, the computer device 600 includes a memory 601 and a processor 602; wherein, the processor 602 reads executable program code stored in the memory 601 to run a program corresponding to the executable program code, so as to implement the various steps of the method described above.

[0063] To implement the above embodiments, this application also proposes a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the method described in the foregoing embodiments.

[0064] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0065] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

Claims

1. A method for predicting waste gas emissions from sludge pyrolysis gasification based on a spatiotemporally coupled graph neural network, characterized in that, include: Equipment operation data and exhaust gas emission data from multiple parallel production lines of the sludge pyrolysis and gasification target plant are obtained. The validity of the data is verified and outliers are processed to obtain continuous time series data for modeling. Based on the sequence of material and energy transfer between equipment in the sludge pyrolysis and gasification process, and the actual connection relationship of multiple production lines converging at the same exhaust gas emission monitoring point, a directed process topology graph is constructed. Each piece of equipment is abstracted as a node in the directed process topology graph, and the transfer relationship between equipment is abstracted as a directed connection edge between nodes. For the operating parameter characteristics of the equipment corresponding to each node in the directed process topology graph, the cross-correlation function between the characteristic and the target exhaust gas emission index is calculated respectively, the optimal time delay step is determined, and the historical sequence of the characteristic is time-delayed and aligned according to the optimal time delay step to obtain the node input characteristics after time delay compensation. Based on the directed process topology graph, a graph neural network is used to perform directed message passing between adjacent nodes to obtain the spatial context message of each node; the time-delay aligned node input features are concatenated with the spatial context message and input into the recurrent neural network unit specific to the type of the device to update the hidden state of the node. A hierarchical attention mechanism is introduced. First, the hidden states of each equipment node within the same production line are aggregated with intra-line attention weighting. Then, the production line representation vectors of each production line are aggregated with inter-line attention weighting to obtain a global perception vector. Based on the global perception vector, the change increment of the exhaust gas emission concentration at the current moment is predicted. The change increment is then superimposed with the exhaust gas emission concentration at the previous moment to obtain the predicted exhaust gas emission concentration at the current moment.

2. The method according to claim 1, characterized in that, The data validity verification and outlier handling include: The quantile gradient jump detection method is used to scan the lower and upper intervals of the quantile distribution curve of the exhaust gas emission monitoring index. The inflection point of the gradient abnormal change is determined as the physical operating boundary of the index, and values ​​that exceed the boundary are eliminated. Map integer hardware fault values ​​that exceed the preset quantile range to missing values; mark intervals where the sensor continuously outputs zero values ​​or readings are stuck as invalid data; When the proportion of invalid data for the same operating parameter on any production line exceeds a preset threshold, the parameter is removed from all production lines. For the remaining valid intervals, the mean, median, and standard deviation within the window are calculated based on a sliding window. Data points that deviate from the median within the window by more than a preset multiple of the standard deviation are identified as outliers and corrected using a linear interpolation method.

3. The method according to claim 2, characterized in that, It also includes constructing a sample-level quality mask to label the credibility of the data: For continuous abnormal data intervals whose correction duration does not exceed a preset duration threshold, a sliding window interpolation method is used to fill them, and the corresponding mask is marked as reliable. For consecutive abnormal data intervals whose correction duration exceeds the preset duration threshold, the global mean of the corresponding indicator is used to fill the intervals, and the corresponding mask is marked as untrustworthy.

4. The method according to claim 1, characterized in that, The method for constructing the directed process topology graph includes: On each production line of the sludge pyrolysis gasification process, the gasifier, combustion furnace, waste heat boiler and bag filter are abstracted as nodes in sequence. According to the actual process sequence of material and energy being transferred in sequence along the gasifier, combustion furnace, waste heat boiler and bag filter, directed connection edges are established between adjacent nodes. The outputs of the corresponding nodes of the bag filters of each parallel production line are connected to the same exhaust gas emission monitoring node to represent the actual connection relationship of the exhaust gases of multiple production lines converging at the emission monitoring point.

5. The method according to claim 1, characterized in that, The method for determining the node input features also includes a feature selection step: For the original operating parameters of the equipment corresponding to each node in the directed process topology graph, the mutual information and Pearson correlation coefficient between the parameter and the target exhaust gas emission index are calculated respectively. Only parameters with mutual information greater than a preset threshold or absolute value of Pearson correlation coefficient greater than a preset threshold are retained. Combined with the understanding of the process mechanism of material and energy transfer of each equipment unit, parameters with causal relationship with the target exhaust gas emission index are further screened out to form the optimal feature subset for constructing the input features of each node.

6. The method according to claim 1, characterized in that, The graph neural network uses a scaled dot product attention mechanism to calculate the confidence coefficient. Using learnable weight matrices corresponding to the device types of upstream and downstream nodes, the historical hidden states of upstream and downstream nodes are mapped to message key vectors and information query vectors, respectively. The message key vectors and information query vectors are then multiplied by a dot product and mapped using a Sigmoid function to obtain the confidence coefficients transmitted from upstream to downstream nodes. The message key vectors of upstream nodes are then scaled using these confidence coefficients to obtain the spatial messages transmitted from upstream to downstream nodes.

7. The method according to claim 1, characterized in that, For each type of device node in the directed process topology graph, a set of recurrent neural network parameters is maintained independently, so that each type of device has its own independent hidden state transition characteristics.

8. The method according to claim 3, characterized in that, It also includes robust designs for sensor anomalies and sudden changes in operating conditions: Construct a dual-objective loss function that includes the mean square error of the predicted concentration change increment and the prediction error of the superimposed absolute concentration; By combining the constructed sample-level quality mask, the gradient contribution of data points marked as untrusted is truncated during the backpropagation process, preventing the model from learning false patterns introduced by manually filled data. This enables the method to maintain stable prediction of exhaust emission trends by relying on the learned process inertia under extreme conditions such as local sensor failure or abnormal freezing.

9. The method according to claim 1, characterized in that, This also includes online deployment using a rolling prediction mechanism: The prediction period is divided into several consecutive fixed-duration intervals. At the beginning of each interval, the hidden state of the recurrent neural network is initialized with the measured values ​​of the exhaust gas emission monitoring points. In subsequent time intervals within this interval, autoregressive rolling inference is performed based on the prediction results of the previous time interval to output the predicted sequence of exhaust gas emissions within this interval.

10. The method according to claim 1, characterized in that, The target exhaust emission indicators include SO2 concentration and NO. X Concentration and particulate matter concentration; the attention weights within and between production lines extracted by the hierarchical attention mechanism are used to identify key production lines and key equipment units that dominate the fluctuations in exhaust gas emissions at different times, so as to achieve interpretable mining of the implicit reaction mechanism of sludge pyrolysis gasification process.

11. The method according to claim 1, characterized in that, The recurrent neural network unit dedicated to the exhaust gas emission monitoring point takes the global perception vector and the exhaust gas emission target value of the previous moment as input, updates its hidden state, predicts the change increment of the exhaust gas emission concentration at the current moment, and then superimposes the change increment with the exhaust gas emission concentration of the previous moment to output the predicted exhaust gas emission concentration at the current moment.

Images