Waste Incineration Control Decision-Making Method and System Using LSTM Networks

Abstract

This invention relates to the fields of artificial intelligence and machine learning, specifically providing a method and system for waste incineration control decision-making using LSTM networks. The method involves acquiring runtime sequence information generated during the continuous operation of a waste incineration system, inputting it into a main LSTM network, and constructing a global correlation link across stages through its memory unit reconstruction processing. This global correlation link information is then distributed to multiple slave LSTM networks. Each slave LSTM network performs adaptation processing for its corresponding incineration sub-stage, outputting a sub-stage adaptation request. All sub-stage adaptation requests from the slave LSTM networks are fed back to the main LSTM network. Based on the global correlation link information, the sub-stage adaptation requests undergo time-series linkage calibration processing to obtain the time-series linkage calibration result. The main LSTM network then generates control decisions adapted to the operating states of each incineration sub-stage. This invention can dynamically adapt to the operating states of each incineration sub-stage, improving the coordination and accuracy of the waste incineration system control process.

Description

Technical Field

[0001] This invention relates to the field of machine learning, and more specifically, to a waste incineration control decision-making method and system using LSTM networks. Background Technology

[0002] In the field of waste incineration, real-time monitoring and data analysis of the operational status of each stage of the incineration system can generate appropriate control commands to adjust various parameters during the incineration process. Traditional waste incineration control decision-making methods often fail to fully capture the dynamic relationships between multiple stages in the incineration system. The adjustment requests of each sub-stage lack coordinated calibration based on the global operational status, resulting in insufficient adaptability of the generated control decisions to the overall system operational status. This can easily lead to conflicts between local optimization and global objectives, affecting the stability and control accuracy of the waste incineration process. Summary of the Invention

[0003] This invention provides a waste incineration control decision-making method and system using LSTM networks.

[0004] In a first aspect, embodiments of the present invention provide a waste incineration control decision-making method using an LSTM network. The method includes: acquiring runtime sequence information generated by the continuous operation of a waste incineration system, the runtime sequence information including records of continuously changing operating states of each incineration stage and records of interactions and correlations between stages; inputting the runtime sequence information into a main LSTM network, constructing a global correlation link across stages through memory unit reconstruction processing of the main LSTM network, and obtaining global correlation link information characterizing the operational correlation relationships of each stage; distributing the global correlation link information to multiple slave LSTM networks, each slave LSTM network performing adaptation processing for its corresponding incineration sub-stage, and outputting a sub-stage adaptation request corresponding to each slave LSTM network; feeding back all sub-stage adaptation requests output by the slave LSTM networks to the main LSTM network, and performing time-series linkage calibration processing on the sub-stage adaptation requests based on the global correlation link information of the main LSTM network, obtaining a time-series linkage calibration result; and generating a control decision to guide the operation of the waste incineration system through the main LSTM network based on the time-series linkage calibration result, the control decision being adapted to the operating states of each incineration sub-stage.

[0005] Secondly, embodiments of the present invention provide a computer system, comprising: a memory storing a computer program; and a processor for loading the computer program to implement the waste incineration control decision-making method using an LSTM network as described above.

[0006] This invention acquires runtime sequence information, including records of continuously changing operational states at each stage of incineration and records of interactions between these stages. The main LSTM network constructs global correlation links across stages through memory unit reconstruction processing, extending the LSTM network's temporal memory function from single-stage to multi-stage correlation modeling. This achieves a global correlation expression of the operational states of each stage, rather than local independent modeling, effectively capturing the dynamic dependencies between multiple stages in the waste incineration system. Employing a distributed architecture of a main LSTM network and multiple slave LSTM networks, the global correlation link information generated by the main LSTM guides each slave LSTM network in performing sub-stage adaptation processing, thus avoiding efficiency bottlenecks in handling complex global problems. Furthermore, it addresses the limitation of lack of coordination in independent networks at each stage, forming a collaborative mechanism of global guidance and local adaptation. Sub-stage adaptation requests output from the LSTM network are fed back to the main LSTM network for timing-linked calibration. The main LSTM network dynamically adjusts the adaptation requests based on global correlation link information, correcting potential timing conflicts or parameter contradictions in sub-stage requests, achieving a dynamic balance between local sub-stage needs and global system coordination. Finally, the control decisions generated based on the timing-linked calibration results integrate global correlation information and the real-time status of sub-stages, enabling control decisions to dynamically adapt to the operating status of each incineration sub-stage, thus improving the coordination and accuracy of the waste incineration system control process. Attached Figure Description

[0007] Figure 1 This is a flowchart of a waste incineration control decision-making method using an LSTM network, provided by an embodiment of the present invention.

[0008] Figure 2 This is a schematic diagram of the composition of a computer system provided in an embodiment of the present invention. Detailed Implementation

[0009] Please see Figure 1 This is a flowchart of a waste incineration control decision-making method using an LSTM network, provided by an embodiment of the present invention. The method can be executed by a computer system and may include the following steps: Step S100: Obtain the runtime sequence information generated by the continuous operation of the waste incineration system. The runtime sequence information includes the continuously changing operation status records of each incineration stage and the interaction and correlation records between each stage.

[0010] Operational sequence information is a collection of information generated chronologically during the continuous operation of a waste incineration system. It includes continuously changing operational status records for each stage of the incineration process, such as waste feeding, incineration, waste heat recovery, and exhaust gas treatment, at different times. For example, the operational status record for the waste feeding stage might include feeding speed and feed volume; the operational status record for the incineration stage might include incineration temperature and oxygen content. Inter-stage interaction records reflect the interactions and influences between different stages, such as how changes in feeding speed affect the temperature and efficiency of the incineration stage, and how the state of the incineration stage affects waste heat recovery and exhaust gas treatment. Operational sequence information can be obtained by installing corresponding sensors at various key locations within the waste incineration system. For example, a flow sensor can be installed at the feed inlet to obtain data on feeding speed and feed volume; temperature and gas sensors can be installed inside the incinerator to measure data such as incineration temperature and oxygen content; and corresponding sensors can be installed on the waste heat recovery unit and exhaust gas treatment equipment to collect relevant data in real time. These sensors store the collected data in chronological order, thus forming runtime sequence information.

[0011] Step S200: Input the runtime sequence information into the main LSTM network, and construct the global correlation link across stages through the memory unit reconstruction processing of the main LSTM network to obtain the global correlation link information representing the operational correlation relationship of each stage.

[0012] In one implementation, step S200 may specifically include the following steps S210 to S260: Step S210: Perform stage state separation processing on the runtime sequence information, and split it into stage state sequences corresponding to different processing stages according to the functional attributes of each stage of incineration. Each stage state sequence contains the running parameter records and time stamps of continuous sampling time.

[0013] Stage state separation is a preprocessing operation performed on runtime sequence information. Its purpose is to break down runtime sequence information, which contains information about multiple stages, according to the functional attributes of each stage. The functional attributes of each incineration stage refer to the specific function each stage performs during waste incineration. For example, the function of the waste feeding stage is to transport waste to the incinerator; the function of the incineration stage is to burn the waste; the function of the waste heat recovery stage is to recover the heat generated during incineration; and the function of the exhaust gas treatment stage is to purify the exhaust gas produced during incineration. The stage state sequence is the information sequence corresponding to each stage after decomposition, containing the operational parameter records of that stage at continuous sampling times and the corresponding timestamps. The operational parameter records refer to the specific operational state data of that stage at different times, such as the feeding rate of the waste feeding stage, and the temperature and pressure of the incineration stage. The timestamps are used to specify the specific time corresponding to each operational parameter record.

[0014] Step S220: Input the state sequence of each stage into the input layer interface of the main LSTM network, perform vector transformation on the running parameter record at each sampling time, and generate a stage state vector that meets the network input dimension requirements. The stage state vector contains the feature information of the running parameters at that time and the stage category identifier.

[0015] The input layer interface of the main LSTM network is the entry point for the network to receive input data. After inputting the state sequences of each stage into this interface, the operating parameter records at each sampling time need to be vectorized. Vector transformation converts the operating parameter records from the original data format into a vector form suitable for network processing. The feature information of the operating parameters refers to the key information in the operating parameters that reflects the operating state of that stage, such as the feeding rate of the waste feeding stage and the temperature of the incineration stage. Stage category identifiers are used to distinguish different stages, for example, using different numbers or codes to represent the waste feeding stage, incineration stage, waste heat recovery stage, and exhaust gas treatment stage. The dimension of the stage state vector needs to meet the input requirements of the main LSTM network so that the network can process it correctly. Feature extraction and encoding methods can be used when performing vector transformation. For example, for numerical data in the operating parameters, they can be directly used as part of the vector; for categorical data, such as stage category identifiers, one-hot encoding can be used to convert them into vector form. Then, these feature information and stage category identifiers are combined into a vector, which serves as the stage state vector at that sampling time.

[0016] Step S230: The link state vector at the current sampling time and the hidden state vector at the previous sampling time are correlated and calculated through the input gate unit of the main LSTM network to generate the input correlation vector at the current time. The input correlation vector is used to represent the dynamic dependency relationship between the current state and the historical state.

[0017] The input gate units of the main LSTM network control the degree of input information. The stage state vector at the current sampling moment reflects the stage's operational state at that moment, while the hidden state vector at the previous sampling moment is the hidden state output by the main LSTM network at the previous moment, containing historical information from previous moments. Association computation involves the input gate units performing mathematical operations on the current stage state vector and the previous hidden state vector to determine the degree of association between the current and historical states. The input association vector is the result of the association computation, reflecting the dynamic dependency between the current and historical states. In a waste incineration system, for example, the current incineration temperature may be affected by the feed rate and incineration state at the previous moment. The input gate units generate the input association vector by performing weighted summation and nonlinear transformation on the input vector. Specifically, the input gate unit weights the current stage state vector and the previous hidden state vector according to its internal weight parameters, and then performs a nonlinear transformation through an activation function (such as the sigmoid function) to obtain the input association vector.

[0018] Step S240: Selectively filter the cell state vector at the previous sampling time through the forget gate unit of the main LSTM network, retain the historical state components that are strongly related to the current incineration process, and obtain the simplified historical cell state vector.

[0019] The forget gate unit in the main LSTM network determines which historical information should be retained and which should be forgotten. The cell state vector at the previous sampling time step is the cell state stored by the network at the previous time step, containing all historical information from previous time steps. Selective filtering involves the forget gate unit evaluating each component in the cell state vector based on the current input and the hidden state at the previous time step, judging its relevance to the current incineration process. Historical state components with strong relevance to the current incineration process refer to historical information that has a significant impact on the current incineration process, such as the incineration temperature and feed rate at the previous time step. The simplified historical cell state vector is the cell state vector obtained after filtering by the forget gate unit, retaining only historical information with strong relevance to the current incineration process, thereby reducing unnecessary information interference and improving the network's processing efficiency. The forget gate unit also implements selective filtering through weighted summation and activation functions (such as the sigmoid function).

[0020] Step S250: Input the input association vector and the simplified historical cell state vector into the cell state update unit of the main LSTM network, perform time-by-time state fusion, and generate the updated cell state vector at the current sampling time through nonlinear transformation. The updated cell state vector contains the running features of the current stage and the historical association features.

[0021] The cell state update unit of the main LSTM network is responsible for updating the cell states. After inputting the input correlation vector and the simplified historical cell state vector, time-by-time state fusion is performed. Time-by-time state fusion integrates the current input information (input correlation vector) and historical information (simplified historical cell state vector) to obtain a more comprehensive cell state representation. Nonlinear transformation is a processing operation performed by the cell state update unit on the fused information, which can be implemented using an activation function (such as the Tanh function). The purpose of nonlinear transformation is to increase the network's expressive power, enabling it to better capture complex relationships in the data. The updated cell state vector is the result of fusion and nonlinear transformation, containing the operational characteristics of the current stage and historical correlation characteristics. The operational characteristics of the current stage refer to the specific operational state of that stage at the current time, such as the feeding rate in the waste feeding stage or the temperature in the incineration stage. Historical correlation characteristics refer to historical information related to the current state, such as the influence of the incineration state at the previous time step on the current time step. In the cell state update unit, the input correlation vector and the simplified historical cell state vector are first multiplied element-wise to obtain an intermediate result. This intermediate result is then added to the cell state vector from the previous time step to obtain the fused state vector. Finally, the fused state vector is processed using a nonlinear transformation function to obtain the updated cell state vector for the current sampling time step.

[0022] Step S260: Based on the updated cell state vector sequence of continuous sampling time, extract the state evolution trajectory of each link in the time dimension, combine the interaction and correlation records between links, and construct a global correlation link information containing state nodes, time connections and link interaction directions. The nodes in the global correlation link information correspond to the running state of each link at a specific time, and connect the state influence relationship between the corresponding links.

[0023] In one implementation, step S260 may specifically include the following steps S261 to S266: Step S261: Group the updated cell state vector sequence according to the functional attributes of the incineration stage to obtain the state vector subsequence specific to each stage. Each state vector subsequence contains the updated cell state vector at all sampling times of that stage.

[0024] The functional attributes of the incineration process refer to the specific functions performed by each stage during waste incineration, such as waste feeding, incineration, waste heat recovery, and exhaust gas treatment. Grouping the updated cell state vector sequence according to functional attributes involves extracting the updated cell state vectors belonging to the same stage and forming a state vector subsequence specific to that stage. Each state vector subsequence contains the updated cell state vectors of that stage at all sampling times, recording the stage's operational state at different times. In practice, the updated cell state vectors can be grouped based on the stage category identifier within them. For example, by traversing the updated cell state vector sequence, vectors with the same stage category identifier can be extracted and formed into a new sequence, serving as the state vector subsequence for that stage.

[0025] Step S262: Extract time dimension features for each state vector subsequence, calculate the similarity of the cell state vectors updated at adjacent sampling times, and generate the state similarity sequence for this stage. The state similarity sequence is used to represent the degree of continuous change of the state within the stage.

[0026] Temporal feature extraction involves analyzing the state vector subsequence along the time dimension to uncover its feature information. The similarity of cell state vector updates at adjacent sampling times refers to the degree of similarity between the updated cell state vectors at two adjacent times. Similarity can be calculated using various methods, such as Euclidean distance and cosine similarity. The state similarity sequence is obtained by calculating the similarity of cell state vector updates at adjacent sampling times, reflecting the degree of continuous change of the state within that stage over time. A high similarity indicates that the state of that stage changes little between adjacent times, indicating a relatively stable state; a low similarity indicates that the state of that stage changes significantly between adjacent times, indicating an unstable state. Taking Euclidean distance as an example, when calculating the similarity of cell state vector updates at adjacent sampling times, the Euclidean distance between the two vectors is first calculated, and then the similarity is determined based on the distance. The smaller the distance, the higher the similarity; the larger the distance, the lower the similarity. Arranging the similarity values ​​of each adjacent time step sequentially yields the state similarity sequence for that stage.

[0027] Step S263: Identify the state inflection points of each stage based on the state similarity sequence. When the similarity value of two consecutive sampling times is lower than the pre-configured judgment standard, it is marked as a state inflection point, and the state vector sub-sequence is divided into multiple state stable segments.

[0028] A state inflection point is a moment when the state of a component undergoes a significant change, which can be identified using a state similarity sequence. A pre-configured judgment criterion is a set threshold used to determine if the similarity value is too low. When the similarity value at two consecutive sampling moments is lower than this criterion, it indicates that the state of the component has undergone a significant change at those two moments, and this moment is marked as a state inflection point. A state-stable segment is each relatively stable state interval obtained after segmentation. Within each state-stable segment, the state change of the component is small, and the similarity is high. In practice, the state similarity sequence can be traversed, comparing the relationship between two adjacent similarity values ​​and the judgment criterion. If two consecutive similarity values ​​are found to be lower than the judgment criterion, this moment is recorded as a state inflection point. Then, based on these state inflection points, the state vector subsequence is segmented into multiple state-stable segments.

[0029] Step S264: Extract the start time marker, end time marker, and average state vector of each stable segment, and use them as the basic nodes of the global associated links. The basic nodes include the link identifier, time interval, and state feature description.

[0030] The start and end timestamps of a stable segment represent the specific times when the segment begins and ends, respectively. The average state vector is the average of all updated cell state vectors within the stable segment, representing the average operational state of the segment. The basic node of the global association link is the fundamental element for constructing the global association link, containing a segment identifier, a time interval, and a state feature description. The segment identifier distinguishes different segments, the time interval is determined by the start and end timestamps, and the state feature description is represented by the average state vector. The start and end timestamps can be determined based on the timestamp information in the state vector subsequences. The average state vector is calculated by summing the state vectors of all updated cells within the stable segment element-wise and then dividing by the number of vectors. Combining this information yields a basic node.

[0031] Step S265: Parse the inter-stage interaction and association records in the runtime sequence information, determine the influence relationship between the stable state segments of different stages, and mark the interaction influence when the end time marker of the stable state segment of one stage overlaps with the start time marker of the stable state segment of another stage.

[0032] The inter-stage interaction records in the runtime sequence information are obtained in step S100 and describe the mutual influence relationships between different stages. Parsing these records involves analyzing and processing the information to determine the influence relationships between the stable state segments of different stages. Time overlap refers to the intersection of the end time marker of a stable state segment of one stage and the start time marker of a stable state segment of another stage in time. When time overlap exists, it indicates that the state changes of these two stages may be related, and they are marked as having an interactive influence. In actual operation, the stable state segments of different stages can be traversed and their time intervals compared. If a time overlap is found between the stable state segments of two stages, an interactive influence between these two stages is recorded, and relevant information such as stage identifier and time overlap interval is marked.

[0033] Step S266: Based on the influence direction markings between basic nodes and links and the time order of stable state segments, construct global association link information containing node attributes, connection directions and time associations. The connection weights in the global association link information are jointly determined by the time overlap of the interaction influence and the state similarity. The time overlap is the ratio of the duration of the overlapping part to the total duration of the time interval of the relevant stable state segments.

[0034] In one implementation, step S266 may include the following steps S2661 to S2666: Step S2661: Sort all basic nodes in ascending order according to the start time marker of the time interval to obtain a node sequence with a time sequence relationship. Each node corresponds to a unique time position index in the sequence.

[0035] The starting time marker of a time interval is the beginning time of the time interval contained in the base nodes. Sort all base nodes in ascending order according to their starting time markers, which means arranging the base nodes in chronological order according to their start times. The sequence of nodes with chronological order is the sorted sequence of nodes, reflecting the chronological order of the base nodes. Each node corresponds to a unique time position index in the sequence, which is used to identify the node's position in the sequence. In practice, the quicksort algorithm can be used to sort the base nodes according to the starting time marker of the time interval, and then assign a unique time position index to each node.

[0036] Step S2662: Calculate the time interval overlap of adjacent nodes in the node sequence. When the time intervals of two nodes overlap, the overlap is the ratio of the duration of the overlapping part to the total duration of the time intervals of the two nodes, and generate the time correlation parameter between nodes.

[0037] The time interval overlap of adjacent nodes in a node sequence refers to the degree of overlap between the time intervals of two adjacent nodes. When the time intervals of two nodes overlap, the overlap is calculated by dividing the duration of the overlapping portion by the total duration of the two node time intervals. The time correlation parameter is obtained by calculating the time interval overlap of adjacent nodes and reflects the degree of temporal correlation between adjacent nodes. The higher the overlap, the closer the temporal correlation between the two nodes; the lower the overlap, the looser the temporal correlation between the two nodes. When calculating the time interval overlap, it is first necessary to determine the time intervals of the two nodes and then find their overlapping portion. For example, if the time interval of node A is [t1, t2] and the time interval of node B is [t3, t4], and t1 <= t3 <= t2, then the duration of the overlapping portion is min(t2, t4) - t3. Dividing the duration of the overlapping portion by the total duration of the two node time intervals ((t2 - t1) + (t4 - t3)) gives the time interval overlap. By sequentially arranging the time interval overlap of each adjacent node, a sequence of time correlation parameters between nodes is generated.

[0038] Step S2663: Combine the influence direction markers between links with the time correlation parameter to construct a node connection matrix. The row index of the matrix corresponds to the preceding node, the column index corresponds to the following node, and the element value of the matrix is ​​the product of the time correlation parameter and the influence direction marker.

[0039] The influence direction marker between links indicates the direction of influence between different links; for example, 0 or 1 indicates whether an influence exists. The temporal correlation parameter reflects the degree of temporal correlation between adjacent nodes. The node connection matrix is ​​a two-dimensional matrix used to represent the connection relationship and strength between nodes. The row index of the matrix corresponds to the preceding node, and the column index corresponds to the following node. The matrix element value is the product of the temporal correlation parameter and the influence direction marker. If there is no influence between two nodes (influence direction marker is 0), then the matrix element value is 0; if there is an influence, the matrix element value is the temporal correlation parameter. When constructing the node connection matrix, first create a matrix of size (number of nodes × number of nodes), and then fill the matrix elements according to the influence direction marker between links and the temporal correlation parameter. For example, for nodes i and j, if there is an influence between them, and the temporal correlation parameter is r, and the influence direction marker is 1, then the element in the i-th row and j-th column of the matrix has the value r; if there is no influence, the influence direction marker is 0, and the matrix element value is 0.

[0040] Step S2664: Sparsify the node connection matrix, retaining the node connection relationships whose matrix element values ​​are greater than the pre-configured judgment criteria, to obtain the simplified node connection matrix.

[0041] Sparsity reduction is a process of optimizing the node connection matrix. Its purpose is to reduce the number of non-zero elements in the matrix, thereby improving processing efficiency. A pre-configured criterion is a set threshold used to determine whether matrix element values ​​are sufficiently large. When a matrix element value is greater than this threshold, the node connection is considered important and needs to be retained; when a matrix element value is less than or equal to the threshold, the node connection is considered unimportant and can be ignored. The simplified node connection matrix is ​​the matrix obtained after sparsification, retaining only the important node connections. During sparsification, the node connection matrix can be traversed, setting elements with values ​​less than or equal to the threshold to 0.

[0042] Step S2665: Based on the simplified node connection matrix and the time position index of the node sequence, generate the topology data of the global associated links. The topology data includes the time position of the nodes, link identifiers, state feature descriptions, and connection relationships between nodes.

[0043] The topology data of globally interconnected links describes the structure of these links, including node time positions, link identifiers, state characteristic descriptions, and connection relationships between nodes. Node time positions can be determined using the time position index of the node sequence. Link identifiers and state characteristic descriptions are information contained in the base nodes, and the connection relationships between nodes can be represented by a simplified node connection matrix. When generating the topology data, firstly, time position information is assigned to each node based on its time position index. Then, the link identifiers and state characteristic descriptions of the base nodes are added to the topology data. Finally, the connection relationships between nodes are determined based on the simplified node connection matrix and added to the topology data.

[0044] Step S2666: Perform loop detection on the topology data. When a node connection with a time loop is detected, adjust the connection weight of the relevant node and reduce the weight value of the loop connection so that the global associated link conforms to the time unidirectional flow characteristic, and obtain the global associated link information that represents the operational association relationship of each link.

[0045] Loop detection examines topology data to identify node connections with temporal loops. A temporal loop occurs when a circular path exists within node connections, allowing time to flow backward, contradicting the unidirectional flow of time. When a temporal loop is detected, the connection weights of the relevant nodes are adjusted, reducing the weight of the looping connection to eliminate the loop and restore the unidirectional flow of time in the global network. Loop detection can utilize algorithms such as depth-first search (DFS) or topology sorting. For example, a DFS can traverse the topology data, marking the access status of each node. If a node is found to have been visited and a path returns from that node to itself, a temporal loop is identified. Upon detecting a loop, the connection weights of the relevant nodes are reduced based on the connections within the loop.

[0046] Step S300: Distribute the global association link information to multiple slave LSTM networks. Each slave LSTM network performs adaptation processing for the corresponding incineration sub-link and outputs the sub-link adaptation request corresponding to each slave LSTM network.

[0047] In one implementation, step S300 may specifically include the following steps S310 to S360: Step S310: Parse the topology data of the global associated link information and extract the dedicated node set corresponding to each incineration sub-stage. The dedicated node set includes all basic nodes belonging to the sub-stage and other basic nodes of the sub-stage that have a direct connection with the sub-stage.

[0048] The topology data of the globally associated link information includes the time location of nodes, link identifiers, status characteristic descriptions, and connection relationships between nodes. Parsing the topology data involves analyzing and processing this information to extract the dedicated node set corresponding to each incineration sub-link. The dedicated node set is the set of nodes corresponding to each incineration sub-link, containing all basic nodes belonging to that sub-link and basic nodes of other links directly connected to that sub-link. Basic nodes belonging to that sub-link are those with the same link identifier as that sub-link; basic nodes of other links directly connected to that sub-link are those directly connected to the basic nodes of that sub-link through the node connection matrix. When parsing the topology data, the dedicated node set for each sub-link can be extracted by traversing the nodes and connection relationships in the topology data, based on the link identifier and connection matrix.

[0049] Step S320: Configure an independent LSTM network processing channel for each incineration sub-stage to convert the topology data of the dedicated node set into an input data format that can be recognized by the LSTM network, including node time intervals, state feature descriptions, and connection relationship encoding.

[0050] The LSTM network processing channel is a dedicated processing path configured for each incineration sub-stage, used to receive and process the topology data of the dedicated node set for that sub-stage. Converting the topology data of the dedicated node set into an input data format recognizable by the LSTM network is crucial for its proper processing. Node time intervals represent the temporal location information of nodes in the topology data, state feature descriptions are the state feature information contained in the basic nodes, and connection relationship encoding converts the connection relationships between nodes into a digital encoded form. Configuring the LSTM network processing channel can be achieved through software programming or hardware circuitry. The conversion of the topology data of the dedicated node set can employ encoding and formatting methods. For example, node time intervals can be converted into timestamps, state feature descriptions can be normalized, and connection relationships can be represented using binary encoding. This information is then combined into a vector or matrix as input data to the LSTM network.

[0051] Step S330: Divide the converted input data into time windows and adjust the window size according to the typical operating cycle of the sub-process. Each time window contains the topology data of multiple consecutive basic nodes so that the nodes in the window cover a complete state change cycle of the sub-process.

[0052] Time window partitioning involves dividing the transformed input data along the time dimension into multiple time windows. The typical operating cycle of a sub-process refers to the time required for that sub-process to complete a full cycle. Adjusting the window size based on the typical operating cycle of the sub-process ensures that each time window encompasses a complete state change cycle of that sub-process. Each time window contains topological data of multiple consecutive basic nodes, recording the operating state of the sub-process within a given time period. When partitioning time windows, the window size and step size can be determined based on the typical operating cycle of the sub-process. For example, if the typical operating cycle of a sub-process is T, the window size can be set to T, and the step size can be set to T / 2. This ensures that each time window covers a complete state change cycle of the sub-process, and that adjacent time windows overlap to improve data utilization.

[0053] Step S340: Input the divided time window data into the hidden layer of the LSTM network in chronological order. Through multiple hidden units, perform time-series dependency modeling on the state feature description and connection relationship encoding of the nodes in the window, and extract the dynamic operation features of the sub-links in the time window. The dynamic operation features include the state change trend and the interaction response characteristics of the links.

[0054] In one implementation, step S340 may specifically include the following steps S341 to S346: Step S341: Convert the node data within the time window into a sequence of input vectors from the LSTM network in chronological order. Each node data corresponds to an input vector at a time step. The input vector contains a numerical vector describing the node state features and a binary vector encoding the connection relationship.

[0055] The node data within a time window consists of the topological structure data of the basic nodes contained within the divided time window. Converting this node data into a sequence of input vectors for the LSTM network in chronological order is to transform the data into a format suitable for network processing. Each node data point corresponds to an input vector for a time step, containing two parts: a numerical vector describing the node's state features and a binary vector encoding the connectivity relationships. The numerical vector describing the node's state features converts these features into numerical values, such as converting temperature and pressure. The binary vector encoding the connectivity relationships represents the connections between nodes using binary encoding. During the conversion process, the node data can be processed according to its specific content. For example, the node's state feature descriptions can be normalized and converted into numerical vectors; the connectivity relationships can be encoded using one-hot encoding or other binary encoding methods to obtain binary vectors encoding the connectivity relationships. These two vectors are then combined into a single input vector, which serves as the input for that time step.

[0056] Step S342: Input the input vector sequence into the first hidden layer of the LSTM network, perform a linear transformation on the input vector at each time step to generate an initial hidden state vector, the dimension of which is the same as the dimension of the input vector.

[0057] The first hidden layer of an LSTM network is the first processing layer, receiving and processing the input vector sequence. A linear transformation of the input vector at each time step involves matrix multiplication of the input vector with the weight matrix of the first hidden layer and adding a bias vector. The purpose of this linear transformation is to extract and transform features from the input vector to uncover its latent information. The initial hidden state vector is the result of this linear transformation, with the same dimension as the input vector, ensuring smooth processing in subsequent steps. In practice, the weight matrix and bias vector of the first hidden layer are learned during network training. For each time step's input vector, it is multiplied by the weight matrix and then the bias vector is added to obtain the initial hidden state vector.

[0058] Step S343: Input the initial hidden state vector into the second hidden layer of the LSTM network, perform element-wise multiplication and addition of the second hidden layer state vector from the previous time step, and then fuse them to strengthen the feature dependencies across time steps and generate a fused hidden state vector.

[0059] The second hidden layer of the LSTM network is the second processing layer, receiving the initial hidden state vector and performing further processing. The initial hidden state vector is fused with the second hidden layer state vector from the previous time step through element-wise multiplication and summation. This fusion process strengthens the feature dependencies across time steps. Cross-time step feature dependencies refer to the association between the state at the current time step and the states at past time steps. The fused hidden state vector is the result of this fusion process, containing information from both the current and past time steps, and better reflects the dynamic characteristics of the sub-processes. During the fusion process, the initial hidden state vector is multiplied element-wise with the corresponding elements of the second hidden layer state vector from the previous time step, and then the results are summed.

[0060] Step S344: Perform batch normalization on the fused hidden state vector to obtain a standardized hidden state vector.

[0061] Batch normalization accelerates network training and improves network stability and generalization ability. Batch normalization of the fused hidden state vector standardizes it to have a mean of 0 and a variance of 1. The standardized hidden state vector, obtained after batch normalization, has better distribution characteristics, making it easier for the network to learn features from the data. During batch normalization, the mean and variance of the fused hidden state vector are first calculated, then the mean is subtracted and divided by the square root of the variance to obtain the standardized vector. Simultaneously, to ensure the network's expressive power, learnable scaling and bias factors are introduced to further adjust the standardized vector.

[0062] Step S345: Input the standardized hidden state vector into the third hidden layer of the LSTM network, and enhance the vector through a non-linear activation function to highlight the key feature components of the sub-stage's operating state, thereby generating an enhanced hidden state vector.

[0063] The third hidden layer of the LSTM network is the third processing layer, receiving the standardized hidden state vector and performing further processing. Non-linear activation functions include ReLU (Rectified Linear Unit) and the Sigmoid function. Feature enhancement of the standardized hidden state vector using non-linear activation functions highlights the key feature components of the sub-process's operational state, enabling the network to better capture important information from the data. The enhanced hidden state vector is the result of feature enhancement, containing the key feature components of the sub-process's operational state and more accurately reflecting the dynamic characteristics of the sub-process. In practice, the standardized hidden state vector is input into a non-linear activation function, such as ReLU. When the input value is greater than 0, the output value equals the input value; when the input value is less than or equal to 0, the output value is 0.

[0064] Step S346: Perform average pooling on the time dimension on the enhanced hidden state vectors of all time steps within the time window to generate the average hidden state vector of the time window. Use this as the dynamic running feature of the sub-stage within the time window. The dynamic running feature includes the state change trend of the sub-stage and the interaction response characteristics of the stage.

[0065] Average pooling is a method for reducing the dimensionality of data in the time dimension. It averages the enhanced hidden state vectors of all time steps within a time window to obtain an average vector. This average vector, serving as the average hidden state vector for that time window, comprehensively reflects the overall operational status of sub-processes within that time window. Dynamic operational characteristics include the state change trends of sub-processes and the interactive response characteristics of processes; the average hidden state vector obtained through average pooling can effectively reflect these characteristics. During average pooling, the enhanced hidden state vectors of all time steps within the time window are summed element-wise, and then divided by the number of time steps to obtain the average hidden state vector.

[0066] Step S350: Retrieve the historical benchmark running feature library of this sub-process from the storage unit of the LSTM network. The historical benchmark running feature library contains a set of dynamic running feature samples of this sub-process under normal operating conditions.

[0067] The storage unit in an LSTM network is the part of the network used to store data. The historical benchmark operating feature library is pre-stored in the storage unit and contains a set of dynamic operating feature samples for that sub-process under normal operating conditions. The dynamic operating feature sample set is a collection of multiple dynamic operating feature samples, each recording information such as the state change trend and interaction response characteristics of that sub-process under normal operating conditions. In practice, the historical benchmark operating feature library can be built and updated during network training. When needed, the corresponding historical benchmark operating feature library is read from the storage unit through a programming interface. For example, for the waste feeding sub-process, the historical benchmark operating feature library for that sub-process is read from the storage unit; this library contains dynamic operating feature samples for that sub-process under different normal operating conditions.

[0068] Step S360: Compare the dynamic operating features of the current time window with the sample features in the historical benchmark operating feature library dimension by dimension, calculate the feature deviation value of each dimension, generate the adjustment requirement description of the sub-link based on the feature deviation value, convert the adjustment requirement description into a standardized data format containing the adjustment direction, target parameters and time interval, and obtain the adaptation request of each sub-link corresponding to the LSTM network.

[0069] In one implementation, step S360 may specifically include the following steps S361 to S366: Step S361: Select sample features with the same duration as the current time window from the historical benchmark running feature library to form a benchmark feature set. Each sample feature contains feature components of the same dimension as the dynamic running features.

[0070] The historical benchmark operation feature library contains dynamic operation feature samples of this sub-process at different time lengths. Selecting sample features from this library with the same duration as the current time window ensures the accuracy and effectiveness of the comparison. The benchmark feature set is the collection of selected sample features, each containing feature components of the same dimension as the dynamic operation features of the current time window, allowing for dimension-by-dimensional comparison. In practice, sample features in the historical benchmark operation feature library can be filtered based on time length information. For example, if the current time window duration is T, sample features with duration T are selected from the historical benchmark operation feature library and combined into the benchmark feature set.

[0071] Step S362: Calculate the distance between the current dynamic running feature and each sample feature in the benchmark feature set, and select the sample feature with the smallest distance as the matching benchmark feature. The matching benchmark feature is the historical normal state feature that is most similar to the current state.

[0072] The distance between the current dynamic operating features and each sample feature in the baseline feature set is an indicator of their similarity. The smaller the distance, the more similar the two features are. Various distance calculation methods can be used, such as Euclidean distance and Manhattan distance. The matching baseline features are historical normal state features selected from the baseline feature set that are most similar to the current state. They can be used as a reference to assess whether the current sub-process is operating normally. When calculating the distance, for each sample feature, it is compared with the current dynamic operating features dimension by dimension, and the distance between them is calculated.

[0073] Step S363: Calculate the difference between the current dynamic running feature and the matching benchmark feature in each dimension to obtain the feature deviation value for each dimension. The deviation value is the current feature component value minus the corresponding component value of the matching benchmark feature.

[0074] Dimensional difference calculation involves subtracting the feature components of the current dynamic operating feature and the matching benchmark feature in each dimension. The feature deviation value is the calculated difference, reflecting the difference between the current sub-process's operating state and the historical normal state in each dimension. In practice, the feature deviation value is obtained by subtracting the feature components of the corresponding dimensions of the current dynamic operating feature and the matching benchmark feature.

[0075] Step S364: Calculate the comprehensive deviation index based on the characteristic deviation values ​​of each dimension and the importance weight of the corresponding incineration parameters. The comprehensive deviation index is the sum of the squares of the products of the deviation values ​​of each dimension and the importance weights. The importance weights are pre-configured according to the degree of influence of the incineration parameters on the system operation.

[0076] The characteristic deviation values ​​for each dimension reflect the difference between the current operating status of the sub-stage and the historical normal status in each dimension. The importance weights of incineration parameters are pre-configured weight values ​​based on the degree of influence of incineration parameters on the operation of the waste incineration system, representing the importance of the characteristic deviation value for each dimension in the comprehensive evaluation. The comprehensive deviation index is calculated by summing the squares of the products of the deviation values ​​for each dimension and their importance weights. It comprehensively considers both the characteristic deviation values ​​and importance weights of each dimension, enabling a more comprehensive assessment of the difference between the operating status of the sub-stage and the historical normal status. When calculating the comprehensive deviation index, the characteristic deviation value for each dimension is first multiplied by its corresponding importance weight, then the result is squared, and finally the results for all dimensions are summed.

[0077] Step S365: When the comprehensive deviation index is greater than the pre-configured judgment standard, identify the incineration parameters corresponding to the dimensions with non-zero deviation values, use them as parameters to be adjusted, and determine the adjustment direction based on the sign of the deviation value.

[0078] The pre-configured judgment criterion is a set threshold used to determine whether the operating status of a sub-process deviates from the normal state. When the comprehensive deviation index is greater than this judgment criterion, it indicates that the operating status of the sub-process differs significantly from the historical normal state, requiring adjustment. The incineration parameters corresponding to dimensions with non-zero deviation values ​​refer to the incineration parameters where the current dynamic operating characteristics differ from the matching benchmark characteristics in these dimensions; these are considered parameters to be adjusted. The adjustment direction is determined based on the sign of the deviation value. If the deviation value is positive, it means the current feature component value is greater than the corresponding component value of the matching benchmark feature, and the parameter needs to be decreased; if the deviation value is negative, it means the current feature component value is less than the corresponding component value of the matching benchmark feature, and the parameter needs to be increased. In actual operation, by comparing the comprehensive deviation index and the judgment criterion, when the comprehensive deviation index is greater than the judgment criterion, the feature deviation values ​​of each dimension are traversed to find the dimensions with non-zero deviation values; the corresponding incineration parameters are the parameters to be adjusted. The adjustment direction is determined based on the sign of the deviation value.

[0079] Step S366: Based on the parameters to be adjusted, the adjustment direction, the time interval of the current time window, and the comprehensive deviation index, generate a structured adjustment requirement description, including parameter identifiers, adjustment direction codes, target parameter values, and execution time intervals. Convert the structured description into a standardized data format that conforms to the system communication protocol to obtain the sub-link adaptation request.

[0080] The parameters to be adjusted are the incineration parameters that need to be adjusted. The adjustment direction is determined by the sign of the deviation value. The current time window refers to the time range corresponding to the current operating state of the sub-stage. The comprehensive deviation index reflects the degree of difference between the operating state of the sub-stage and its historical normal state. The structured adjustment requirement description integrates this information to form a structured description containing parameter identifiers, adjustment direction codes, target parameter values, and execution time ranges. The parameter identifier is used to uniquely identify each parameter to be adjusted. The adjustment direction code is the digital representation of the adjustment direction. The target parameter value is the parameter value that is expected to be achieved after adjustment, which can be calculated based on the matching benchmark features and feature deviation values. The execution time range refers to the time range within which the adjustment operation needs to be performed. Converting the structured description into a standardized data format that conforms to the system communication protocol is to ensure that the adjustment requirements can be accurately received and executed by the waste incineration system.

[0081] Specifically, the physical adjustment amount of each parameter to be adjusted can be calculated based on the characteristic deviation value corresponding to that parameter and the preset characteristic value-physical quantity mapping relationship of that parameter. The mapping relationship defines the change in physical quantity corresponding to the unit change of characteristic value. The target parameter value is calculated based on the physical adjustment amount and the current actual value of the parameter. Based on the parameter to be adjusted, the adjustment direction, the target parameter value, the time interval of the current time window, and the comprehensive deviation index, a structured adjustment requirement description is generated, which includes parameter identifier, adjustment direction code, target parameter value, and execution time interval. The structured description is converted into a standardized data format that conforms to the system communication protocol to obtain the sub-link adaptation request.

[0082] Step S400: Feed back all sub-link adaptation requests output from the LSTM network to the main LSTM network. The main LSTM network performs timing linkage calibration on the sub-link adaptation requests based on the global associated link information to obtain the timing linkage calibration result.

[0083] In one implementation, step S400 may specifically include the following steps S410 to S460: Step S410: Collect all sub-stage adaptation requests output from the LSTM network, sort them in ascending order according to the request generation time stamp, and obtain the adaptation request sequence. Each request includes the sub-stage identifier, adjustment parameters, adjustment direction, target parameter value, and execution time interval.

[0084] The sub-stage adaptation requests output from the LSTM network are adjustment requirements generated by each LSTM network for its corresponding sub-stage. After collecting all these requests, they are sorted in ascending order according to their generation timestamps to arrange them chronologically for subsequent processing. The adaptation request sequence is the sorted sequence of requests, with each request containing information such as a sub-stage identifier, adjustment parameters, adjustment direction, target parameter value, and execution time interval. The sub-stage identifier distinguishes different sub-stages, the adjustment parameters are the parameters to be adjusted, the adjustment direction is determined based on the feature deviation value, the target parameter value is the desired parameter value after adjustment, and the execution time interval refers to the time range within which the adjustment operation needs to be performed. In practice, the sub-stage adaptation requests output from the LSTM network can be collected and sorted using data storage and sorting algorithms. For example, these requests can be stored in a database, and then sorted in ascending order according to their generation timestamps to obtain the adaptation request sequence.

[0085] Step S420: Parse the global association link information, extract the upstream and downstream association relationships of each sub-link in the link, and when there is a time overlap between the end time mark of the execution time interval of the first sub-link and the start time mark of the execution time interval of the second sub-link, mark the second sub-link as the downstream associated sub-link of the first sub-link.

[0086] Global correlation link information describes the operational relationships between various stages of a waste incineration system. Parsing this information involves analyzing and processing the data to extract the upstream and downstream relationships of each sub-stage within the link. Upstream and downstream relationships refer to the sequence and mutual influence between sub-stages during operation. When the end time marker of the execution time interval of the first sub-stage overlaps with the start time marker of the execution time interval of the second sub-stage, it indicates that the operation of the second sub-stage may be affected by the first sub-stage, and the second sub-stage is marked as a downstream related sub-stage of the first sub-stage. When parsing global correlation link information, the upstream and downstream relationships of each sub-stage can be determined based on the node connection relationships and time information in the link. For example, by traversing the nodes and connection relationships in the global correlation link information, pairs of sub-stages with overlapping execution time intervals can be identified and marked as upstream and downstream related sub-stages.

[0087] Step S430: Divide the adaptation request sequence into multiple related request groups based on the upstream and downstream relationship. Each related request group contains multiple sub-link adaptation requests that have direct upstream and downstream relationships, so that the sub-links of the requests within the group have a link connection relationship.

[0088] In a waste incineration system, there are close upstream and downstream connections between the various sub-stages, reflecting the logical sequence and inherent relationships of the waste incineration process. For example, the waste feeding sub-stage is upstream of the incineration sub-stage because the subsequent incineration operation can only proceed after the waste feeding is completed; the incineration sub-stage, in turn, is upstream of the waste heat recovery sub-stage and the exhaust gas treatment sub-stage, as the heat and exhaust gas generated by incineration are the inputs for waste heat recovery and exhaust gas treatment.

[0089] When grouping requests, the sequence of adaptation requests is first traversed. For each sub-stage adaptation request, based on the previously parsed upstream and downstream relationships, other sub-stage adaptation requests with direct upstream and downstream connections are searched. For example, if the current processing involves adaptation requests for the waste feeding sub-stage, then adaptation requests for the incineration sub-stage are searched and grouped together. This division aims to ensure that the sub-stages within each related request group are interconnected and mutually influential in actual operation. Adjustments to sub-stage adaptation requests within a related request group must consider their dependencies to ensure the coordinated operation of the entire system.

[0090] Step S440: Perform execution time interval overlap detection on the adaptation requests in each associated request group. If the execution time intervals of two requests overlap and the adjustment parameters are the same associated parameters, mark them as having a time conflict and record the sub-stage identifier, adjustment direction and time overlap interval of the conflicting request pair.

[0091] Within a group of related requests, adaptation requests from different sub-stages may overlap in execution time, and these requests may adjust the same related parameter. Execution time interval overlap detection aims to identify such potential conflicts. Related parameters are parameters that affect the operating status of multiple sub-stages; for example, incineration temperature affects not only the efficiency of the incineration sub-stage but also the waste heat recovery and exhaust gas treatment sub-stages. During detection, the execution time intervals of each pair of adaptation requests within each group are compared one by one. If two requests are found to have overlapping execution time intervals and their adjustment parameter is the same related parameter, then a time conflict is considered to exist between the two requests. For example, if the adaptation requests for the waste feeding sub-stage and the incineration sub-stage both adjust the related parameter of the feeding rate within the same time period, a time conflict will occur. Once a time conflict is detected, detailed information about the conflicting request pair needs to be recorded, including the sub-stage identifier to identify which sub-stage's request is conflicting; the adjustment direction to understand whether each request increases or decreases the value of the related parameter; and the time overlap interval for subsequent conflict resolution.

[0092] Step S450: Based on the connection weights of sub-links within the associated request group in the global associated link information, prioritize request pairs with time conflicts. Adaptation requests corresponding to sub-links with larger connection weights will retain their original execution time intervals in the event of a conflict, while adaptation requests corresponding to sub-links with smaller connection weights will have their execution time intervals adjusted.

[0093] In one implementation, step S450 may specifically include the following steps S451 to S456: Step S451: Extract the connection weight value of each sub-link within the association request group from the global association link information. The connection weight value is the average of the connection weights of the sub-link and all upstream and downstream sub-links, reflecting the overall association importance of the sub-link in the link.

[0094] The global interconnected link information records the connection weights between each sub-link and its upstream and downstream sub-links. These connection weights reflect the degree of mutual influence and the tightness of the connection between the sub-links. To measure the overall interconnected importance of each sub-link in the entire link, it is necessary to calculate the average of all connection weights between that sub-link and its upstream and downstream sub-links.

[0095] When extracting connection weight values, the process begins by identifying each sub-link within the associated request group. Then, the connection weights of each sub-link with its upstream and downstream sub-links are retrieved from the global association link information. For a given sub-link, its connection weights are summed with those of all its upstream and downstream sub-links, and then divided by the number of connections to obtain its connection weight value. This connection weight value comprehensively considers the upstream and downstream associations of the sub-link in the link, providing a more complete reflection of its importance within the entire system. For example, if the incineration sub-link is connected to the waste feeding sub-link, the waste heat recovery sub-link, and the exhaust gas treatment sub-link, the connection weight value obtained by summing and averaging the weights of these connections can be used to assess the importance of the incineration sub-link in the overall link.

[0096] Step S452: For request pairs with time conflicts, compare the connection weight values ​​of the sub-stages to which the two requests belong, mark the request corresponding to the sub-stage with the larger connection weight value as a high-priority request, and retain its original execution time range.

[0097] After determining the connection weight values ​​of each sub-stage within the associated request group, for request pairs with time conflicts, it is necessary to compare the connection weight values ​​of their respective sub-stages. The purpose of this comparison is to determine which request has a higher priority, enabling reasonable decisions in conflict resolution. During the comparison, the connection weight values ​​of the sub-stages to which the two requests belong are directly compared. If the connection weight value of one request's sub-stage is greater than that of the other, the request corresponding to the former is marked as a high-priority request. This means that this request is more critical to the operation of the entire waste incineration system and has a greater impact on its normal operation; therefore, its original execution time range needs to be preserved to ensure that the adjustment operations of this sub-stage can proceed smoothly as planned. For example, when a time conflict occurs between the waste feeding sub-stage and the adaptation request of a minor sub-stage, if the connection weight value of the waste feeding sub-stage is larger, then the adaptation request of the waste feeding sub-stage will be marked as a high-priority request, and its original execution time range will be preserved.

[0098] Step S453: Mark the requests corresponding to the sub-links with smaller connection weight values ​​as low-priority requests, calculate the overlap between their original execution time intervals and the execution time intervals of high-priority requests, and extend the execution time intervals of low-priority requests by the overlap interval to obtain the adjusted execution time intervals.

[0099] After identifying high-priority and low-priority requests, the execution time interval of low-priority requests needs to be adjusted to avoid time conflicts with high-priority requests. First, the overlap between the original execution time interval of the low-priority request and the execution time interval of the high-priority request is calculated. This overlap reflects the degree of time conflict between the two requests. The overlap duration is determined by comparing the start and end times of the two execution time intervals. If the execution time intervals of the low-priority and high-priority requests intersect, the start and end times of the intersection are identified, and the duration between them is calculated; this is the overlap duration. Then, the execution time interval of the low-priority request is shifted forward by the overlap duration. Specifically, the overlap duration is added to both the start and end times of the low-priority request's execution time interval to obtain the adjusted execution time interval. For example, if the original execution time interval of the low-priority request is a specific time period, and the execution time interval of the high-priority request partially overlaps with this time period, after calculating the overlap duration, the entire execution time interval of the low-priority request is shifted forward by this duration, thus avoiding time conflicts with the high-priority request.

[0100] Step S454: Check whether the adjusted execution time interval has new overlap with the execution time interval of other high-priority requests in the associated request group. If so, continue to extend the same overlapping part of the time until there is no overlap or the pre-configured maximum extension number is reached.

[0101] After adjusting the execution time interval for low-priority requests, it's necessary to check whether the adjusted interval will create new time conflicts with the execution time intervals of other high-priority requests within the same request group. This is because an adjustment might only resolve a conflict with one high-priority request without considering other high-priority requests, thus requiring further checks.

[0102] During the check, the adjusted execution time interval is compared with the execution time intervals of other high-priority requests within the associated request group that have not yet been checked. If new overlaps are found, it indicates that time conflicts still exist, and the execution time intervals of low-priority requests need to be adjusted again. The adjustment method is to continue to extend the overlap by the same amount of time.

[0103] This process repeats until the adjusted execution time interval no longer overlaps with the execution time intervals of all high-priority requests within the associated request group, or until the pre-configured maximum number of postponements is reached. The pre-configured maximum number of postponements is to prevent infinite postponements from causing the adjusted execution time interval to exceed a reasonable range and affecting the normal operation of the system.

[0104] Step S455: When the adjusted execution time interval of a low-priority request exceeds the running cycle of the sub-process, it is split into multiple sub-requests. The execution time interval of each sub-request does not exceed the running cycle and does not overlap with the high-priority request. The target values ​​of the adjustment parameters of the sub-requests are allocated according to the split ratio.

[0105] After repeatedly adjusting the execution time interval of low-priority requests, the adjusted execution time interval may exceed the runtime of the sub-process. The runtime of a sub-process refers to the time required for the sub-process to complete one full cycle. Adjustments exceeding the runtime may cause abnormalities in the sub-process and affect the stability of the system.

[0106] When this occurs, low-priority requests need to be split into multiple sub-requests. The principle of splitting is that the execution time interval of each sub-request does not exceed the runtime cycle of the sub-process, and there is no overlap with high-priority requests within the associated request group. During the splitting process, the target values ​​of the original request's adjustment parameters need to be allocated to each sub-request according to the splitting ratio. For example, if the original request is split into two sub-requests, the target values ​​of the original request's adjustment parameters should be allocated accordingly based on the ratio of the execution time intervals of the two sub-requests to ensure the overall adjustment effect.

[0107] Step S456: Record the priority markers of the conflict request pairs, the execution time intervals before and after adjustment, and the splitting situation, and generate conflict handling results as a component of the timing linkage calibration results.

[0108] After processing request pairs with time conflicts, relevant processing information needs to be recorded to generate conflict resolution results. These results include priority markers for the conflicting request pairs, clearly identifying which requests are high-priority and which are low-priority. This helps in understanding and analyzing subsequent system adjustments.

[0109] The report also records the execution time intervals before and after the adjustment. A comparison clearly shows how the execution time intervals for low-priority requests were adjusted and whether the adjusted time intervals are reasonable. Furthermore, for low-priority requests that underwent splitting, the report records the splitting details, including the number of sub-requests, the execution time interval for each sub-request, and the target values ​​of the adjustment parameters. The generated conflict resolution results will serve as a component of the timing-based calibration results, providing crucial reference for subsequent system control decisions. By analyzing the information in the report, the effectiveness of conflict resolution can be evaluated, lessons learned can be summarized, and similar time conflict situations can be better addressed in future operations, thereby improving the operational efficiency and stability of the waste incineration system.

[0110] Step S460: Merge the adjusted associated request group back into a calibration request sequence, and perform parameter consistency verification on all adaptation requests in the calibration request sequence to ensure that the target values ​​of the adjusted parameters of the upstream and downstream sub-links conform to the interaction rules in the global associated link information, so as to obtain the time-series linkage calibration result containing the execution time interval after calibration and the target parameter value.

[0111] In one implementation, step S460 may specifically include the following steps S461 to S466: Step S461: Sort all the adjusted adaptation requests of the associated request groups in ascending order according to the start time marker of the execution time interval to generate a preliminary calibration request sequence.

[0112] After re-merging the adjusted associated request groups, to facilitate subsequent parameter consistency verification, all adaptation requests need to be sorted in ascending order by the start time marker of their execution time interval. Ascending order sorting allows the calibration request sequence to be arranged chronologically, facilitating sequential checking and processing of adaptation requests in upstream and downstream sub-stages.

[0113] During the sorting process, the start and end timestamps of the execution time interval for each adaptation request are first extracted. Then, a sorting algorithm is used to rank these adaptation requests. In the resulting preliminary calibration request sequence, the adaptation requests are arranged in chronological order of execution time. This more clearly reflects the adjustment needs of each sub-stage of the waste incineration system at different points in time. For example, adaptation requests executed earlier are listed first, and those executed later are listed last, facilitating consistency verification of parameters according to time sequence and ensuring that system adjustments are performed in a reasonable chronological order.

[0114] Step S462: Parse the interaction rule base in the global association link information, extract the association constraints of the adjustment parameters of upstream and downstream sub-links. When there is an interaction relationship between the adjustment parameters of the upstream sub-link and the downstream sub-link, and the target value is a certain value, the target value of the corresponding adjustment parameter of the downstream sub-link must maintain a preset ratio and a fixed difference with the target value of the adjustment parameter of the upstream sub-link, which is predefined by the interaction rule base.

[0115] The interaction rule base in the global correlation link information records the correlation constraints of parameter adjustment between upstream and downstream sub-links. These constraints reflect the inherent connections and mutual influences between the various sub-links in the waste incineration system, and are an important basis for ensuring the coordinated operation of the system.

[0116] When parsing the interaction rule base, it is necessary to extract the correlation constraints of the adjustment parameters of upstream and downstream sub-stages. For adjustment parameters that are interactively related between upstream and downstream sub-stages, the interaction rule base predefines the relationships between them. For example, when the waste feeding sub-stage sets a target value for the feed rate adjustment parameter to a certain value, the target value of the temperature adjustment parameter in the incineration sub-stage needs to maintain a preset proportional relationship and a fixed difference with the feed rate target value. This preset proportional relationship and fixed difference are determined based on the physical characteristics and operational experience of the waste incineration system, ensuring that the adjustment operations of upstream and downstream sub-stages can cooperate with each other to achieve the best operating results.

[0117] By parsing the interaction rule base, these associated constraints are extracted, providing clear rules and standards for subsequent parameter consistency verification.

[0118] Step S463: Traverse each adaptation request in the preliminary calibration request sequence, identify the adaptation requests of its downstream related sub-links, check whether the target value of the adjustment parameter of the downstream request meets the related constraint conditions, and mark it as a parameter conflict if it does not meet the constraint conditions.

[0119] After obtaining the preliminary calibration request sequence and parsing the associated constraints in the interaction rule base, it is necessary to traverse the preliminary calibration request sequence and check the parameter consistency between each adaptation request and the adaptation requests of the downstream associated sub-links.

[0120] For each adaptation request in the initial calibration request sequence, the downstream related sub-link adaptation requests are first identified based on the global association link information. Then, based on the association constraints extracted from the interaction rule base, it is checked whether the target values ​​of the adjustment parameters of the downstream requests meet these conditions. For example, for an adaptation request of a certain upstream sub-link, after its target value of adjustment parameters is determined, it is checked whether the target value of the adjustment parameters of the corresponding downstream sub-link adaptation request maintains a preset proportional relationship and a fixed difference with the target value of the upstream sub-link.

[0121] If the target value of the adjustment parameter in a downstream request is found to be inconsistent with the association constraints, the upstream and downstream request pair is marked as having a parameter conflict. Marking parameter conflicts helps in subsequent handling of these conflicts, ensuring that the target values ​​of the adjustment parameters in the upstream and downstream sub-stages conform to the interaction rules, and guaranteeing the stable operation of the waste incineration system.

[0122] Step S464: For upstream and downstream request pairs marked as having parameter conflicts, recalculate the target value of the adjustment parameter of the downstream request based on the proportional relationship and fixed difference in the interaction rule base, and update the target parameter value of the downstream request.

[0123] Once upstream and downstream request pairs marked as having parameter conflicts are identified, the target values ​​of the adjustment parameters for the downstream request need to be recalculated and updated based on the proportional relationships and fixed differences in the interaction rule base. During recalculation, the target values ​​of the adjustment parameters for the downstream request are calculated based on the target values ​​of the adjustment parameters in the upstream sub-link and the proportional relationships and fixed differences specified in the interaction rule base. For example, if the interaction rule base specifies a certain proportional relationship and fixed difference between adjustment parameter A in the upstream sub-link and adjustment parameter B in the downstream sub-link, and the target value of adjustment parameter A in the upstream sub-link is known, then the target value of adjustment parameter B in the downstream sub-link can be calculated according to this rule. Then, the newly calculated target parameter values ​​are updated in the downstream request.

[0124] Step S465: Recheck whether there are new parameter conflicts between the updated downstream request and other related requests. If so, repeat the above calculation process until all parameter conflicts are resolved or the pre-configured maximum number of iterations is reached.

[0125] After updating the target parameter values ​​of the downstream request in a pair of upstream and downstream requests marked as having parameter conflicts, it is necessary to re-check whether the updated downstream request has new parameter conflicts with other related requests. This is because an update may affect the parameter consistency of other related requests, leading to new conflicts. During the re-check, following the previous checking method, the updated downstream request is checked against the adjustment parameter target values ​​of other related requests based on the association constraints in the interaction rule base. If new parameter conflicts are found, the adjustment parameter target values ​​of the relevant requests need to be recalculated and updated again based on the proportional relationships and fixed differences in the interaction rule base. This process will be repeated until all parameter conflicts are resolved or the pre-configured iteration limit is reached. The pre-configured iteration limit is to avoid getting stuck in an infinite loop and ensure that parameter conflicts are resolved as much as possible within a reasonable number of iterations.

[0126] Step S466: Integrate the final calibration request sequence and parameter conflict handling records into a time-series linkage calibration result. The time-series linkage calibration result includes the post-calibration execution time interval, target parameter value, and parameter adjustment basis of each sub-stage adaptation request.

[0127] After completing parameter consistency verification and conflict resolution for all adaptation requests in the initial calibration request sequence, the final calibration request sequence and parameter conflict resolution records are integrated into a time-series linkage calibration result. The final calibration request sequence has undergone time adjustment and parameter consistency verification. Each adaptation request has a calibrated execution time interval and target parameter values, ensuring the time and parameter coordination of adjustment operations in each sub-stage of the waste incineration system. The parameter conflict resolution record documents parameter conflicts discovered during the parameter consistency verification process, along with the measures and basis for resolving these conflicts. The time-series linkage calibration result includes the calibrated execution time interval, target parameter values, and parameter adjustment basis for each sub-stage adaptation request.

[0128] Step S500: Based on the time-series linkage calibration results, control decisions are generated through the main LSTM network to guide the operation of the waste incineration system. The control decisions are adapted to the operating status of each incineration sub-stage.

[0129] In one implementation, step S500 may specifically include the following steps S510-S560: Step S510: Analyze the timing linkage calibration results, extract the calibration execution time interval, target parameter value, adjustment direction and parameter adjustment basis of each sub-link adaptation request, classify and integrate them according to the sub-link identifier, and obtain the exclusive decision parameter set for each sub-link. The exclusive decision parameter set includes the target value, execution time, adjustment range and related constraint description of all parameters to be adjusted in the sub-link.

[0130] The timing-linkage calibration results are analyzed to extract key information about the adaptation requests for each sub-stage. The post-calibration execution time interval clarifies the time period during which each adaptation request should be executed. The target parameter value indicates the desired parameter standard after adjustment, the adjustment direction explains whether to increase or decrease the parameter value, and the basis for parameter adjustment records why such adjustments are made. This basis may come from previous data analysis and conflict resolution results.

[0131] Categorizing and integrating by sub-stage identifier involves summarizing and organizing information from all adaptation requests belonging to the same sub-stage. This generates a unique set of decision parameters for each sub-stage. This set includes the target values, execution times, adjustment ranges, and related constraints for all parameters to be adjusted within that sub-stage. For example, the unique set of decision parameters for the waste feeding sub-stage would include the target value for feeding speed, the execution time of the feeding operation, the adjustment range for the feeding speed, and related constraints with other sub-stages, such as the relationship between feeding speed and incineration temperature.

[0132] Step S520: Input the dedicated decision parameter set into the decision generation layer of the main LSTM network. The decision generation layer contains parameter mapping units corresponding to each sub-stage. Each parameter mapping unit converts the dedicated decision parameter set into a control parameter sequence for that sub-stage according to the device characteristics of the sub-stage. The control parameter sequence contains the parameter setting value and parameter change rate at each sampling time.

[0133] The decision generation layer of the main LSTM network is specifically designed to transform the set of proprietary decision parameters into a sequence of specific control parameters. This layer contains parameter mapping units corresponding to each sub-stage, with each unit responsible for processing the proprietary decision parameter set for a particular sub-stage. The equipment characteristics of a sub-stage refer to the specific performance and operational requirements of the equipment used in each sub-stage, such as the feeding capacity of a waste feeding machine or the heating efficiency of an incineration machine. Based on these equipment characteristics, the parameter mapping unit transforms the proprietary decision parameter set. It converts information such as the target value, execution time, and adjustment range from the proprietary decision parameter set into the parameter setpoints and parameter change rates for that sub-stage at each sampling time.

[0134] Step S530: The global coordination unit of the main LSTM network performs cross-link correlation verification on the control parameter sequences of all sub-links. When the parameter setting values ​​and parameter change rates of the control parameter sequences of two related sub-links do not conform to the interaction rules in the global correlation link information in the same time interval, they are marked as coordination conflicts, and the sub-link identifiers, time intervals and parameter difference details involved in the conflict are recorded.

[0135] In one implementation, step S530 may specifically include the following steps S531 to S536: Step S531: Extract all sub-link pairs with direct connections from the global associated link information to form an associated sub-link pair list. Each sub-link pair includes the upstream sub-link identifier, the downstream sub-link identifier, and the connection weight value.

[0136] The global association link information records the connection relationships and connection weights between each sub-link. From this information, all sub-link pairs with direct connections are extracted and compiled into a list of associated sub-link pairs. Each sub-link pair includes the upstream sub-link identifier, the downstream sub-link identifier, and the connection weight value between them.

[0137] Upstream and downstream sub-stage identifiers are used to clearly identify the two sub-stages in a sub-stage pair, while the connection weight value reflects the degree of correlation between these two sub-stages. For example, the waste feeding sub-stage and the incineration sub-stage have a direct connection, forming a sub-stage pair, where the waste feeding sub-stage is the upstream sub-stage and the incineration sub-stage is the downstream sub-stage. The connection weight value between them reflects the degree of influence of waste feeding on the incineration process. By creating a list of associated sub-stage pairs, all sub-stage pairs with direct connections can be easily processed and checked.

[0138] Step S532: For each associated sub-link pair, extract its corresponding control parameter sequence, align it according to the time sampling time, and obtain the parameter set value sequence and parameter change rate sequence of the two sub-links in the same time interval.

[0139] For each sub-link pair in the list of associated sub-link pairs, its corresponding control parameter sequence needs to be extracted. The control parameter sequence is generated in step S520 and contains the parameter settings and parameter change rates of each sub-link at different sampling times.

[0140] Alignment by sampling time involves arranging the control parameter sequences of two sub-stages according to the same sampling time point. This allows for comparison of the parameter setpoints and rates of change of the two sub-stages within the same time interval. Alignment yields the parameter setpoint sequences and rate of change sequences for both sub-stages within the same time interval. For example, for a sub-stage pair consisting of a waste feeding sub-stage and an incineration sub-stage, aligning their control parameter sequences by sampling time clearly shows the feed rate setpoint and rate of change for the waste feeding sub-stage, as well as the temperature setpoint and rate of change for the incineration sub-stage, at the same time point, facilitating subsequent correlation verification.

[0141] Step S533: Parse the interaction rules in the global associated link information to determine the parameter association method of the associated sub-link pair. The parameter association method includes the proportional relationship of the parameter setting values ​​of the upstream and downstream sub-links, the synchronization relationship of the parameter change rate, and the time delay compensation mechanism, which are defined in detail by the link interaction rule library.

[0142] The interaction rules in the global association link information define the parameter association methods between each sub-link. For each pair of associated sub-links, these interaction rules need to be parsed to determine their parameter association methods.

[0143] Parameter association methods encompass multiple aspects. The proportional relationship between upstream and downstream sub-stage parameter settings describes the mathematical relationship between them; for example, there might be a proportional relationship between the feeding rate in the waste feeding sub-stage and the temperature in the incineration sub-stage. The synchronization of parameter change rates involves whether the parameter change rates in upstream and downstream sub-stages need to be consistent. The time delay compensation mechanism considers the potential time delay in the transmission of parameter changes between different sub-stages, requiring appropriate compensation. These parameter association methods are defined in detail by the link interaction rule base, providing specific rule bases for subsequent parameter comparison and verification.

[0144] Step S534: Substitute the upstream sub-link parameter setpoint sequence and parameter change rate sequence into the parameter correlation method to calculate the downstream sub-link theoretical parameter setpoint sequence and theoretical parameter change rate sequence in the same time interval.

[0145] After determining the parameter association method for the related sub-links, the parameter setpoint sequence and parameter change rate sequence of the upstream sub-links are substituted into it. Based on the rules such as the proportional relationship, synchronization relationship, and time delay compensation mechanism specified in the parameter association method, the theoretical parameter setpoint sequence and theoretical parameter change rate sequence of the downstream sub-links in the same time interval are calculated.

[0146] For example, based on the feed rate setpoint sequence and parameter change rate sequence of the waste feeding sub-stage, and their correlation with the temperature parameters of the incineration sub-stage, the theoretical temperature setpoint sequence and theoretical temperature change rate sequence of the incineration sub-stage are calculated for the same time interval. These theoretical values ​​are calculated based on interaction rules, providing a reference standard for subsequent comparisons between actual and theoretical values.

[0147] Step S535: Compare the calculated theoretical parameter setpoint sequence and theoretical parameter change rate sequence with the parameter setpoint sequence and parameter change rate sequence in the actual control parameter sequence of the downstream sub-link at each time step, and calculate the parameter setpoint deviation and parameter change rate deviation at each time step.

[0148] The calculated theoretical parameter setpoint sequence and theoretical parameter change rate sequence of the downstream sub-stage are compared time-by-time with the actual control parameter sequence of the downstream sub-stage. Through this comparison, the parameter setpoint deviation and parameter change rate deviation at each time point can be calculated.

[0149] Parameter setpoint deviation refers to the difference between the actual parameter setpoint and the theoretical parameter setpoint, while parameter change rate deviation refers to the difference between the actual parameter change rate and the theoretical parameter change rate. For example, at a certain point in time, the difference between the actual temperature setpoint of the incineration sub-stage and the theoretical temperature setpoint calculated based on the waste feed rate is the parameter setpoint deviation at that moment; similarly, the difference between the actual temperature change rate and the theoretical temperature change rate is the parameter change rate deviation at that moment. These deviation values ​​reflect the degree of difference between the actual control parameters and the theoretical control parameters.

[0150] Step S536: When the absolute value of the parameter setpoint deviation and the absolute value of the parameter change rate deviation are both greater than the pre-configured judgment criteria at multiple consecutive time points, mark that there is a coordination conflict in the control parameter sequence of the associated sub-link pair, and record the conflict start time, end time, parameter setpoint deviation sequence, parameter change rate deviation sequence, and the sub-link identifier involved.

[0151] The pre-configured judgment criterion is a set threshold used to determine whether the deviations in parameter setpoints and parameter change rates are too large. When the absolute values ​​of the parameter setpoint deviations and parameter change rate deviations are all greater than this judgment criterion for multiple consecutive time points, it indicates that the control parameter sequences of the associated sub-links do not conform to the interaction rules, and a coordination conflict exists. Once a coordination conflict is discovered, detailed information about the conflict needs to be recorded, including the start and end times of the conflict, specifying the time period in which the conflict occurred; the parameter setpoint deviation sequence and parameter change rate deviation sequence should be recorded, describing the changes in the deviations in detail; and the identifiers of the sub-links involved should be recorded to determine which sub-links are in conflict.

[0152] Step S540: For control parameter sequence pairs marked as coordination conflicts, calculate the parameter adjustment amount in different time intervals based on the connection weights and interaction rules in the global association link information, and perform phased fine-tuning on the control parameter sequences of sub-links with smaller weights so that the adjusted parameter settings and parameter change rates conform to the interaction rules, thereby generating the coordinated control parameter sequence.

[0153] In one implementation, step S540 may specifically include the following steps S541 to S546: Step S541: Extract the connection weight values ​​of the associated sub-link pairs from the global associated link information, and calculate the sum of the connection weight values ​​of the upstream and downstream sub-links respectively. The sum of the connection weight values ​​is the cumulative value of the connection weights of the sub-link with all associated sub-links, reflecting the importance of the sub-link in the global link.

[0154] The global association link information records the connection weights between each sub-link and other associated sub-links. For associated sub-link pairs marked as having coordination conflicts, their connection weight values ​​need to be extracted from the global association link information.

[0155] Calculate the sum of the connection weights for the upstream and downstream sub-stages separately. The sum of connection weights is the cumulative value of the connection weights between the current sub-stage and all related sub-stages, comprehensively reflecting the importance of the sub-stage in the global chain. For example, for a pair of related sub-stages consisting of the waste feeding sub-stage and the incineration sub-stage, calculate the sum of the connection weights of the waste feeding sub-stage with all other related sub-stages (such as the pretreatment sub-stage, conveying sub-stage, etc.), and the sum of the connection weights of the incineration sub-stage with all other related sub-stages (such as the waste heat recovery sub-stage, exhaust gas treatment sub-stage, etc.). By comparing these two sums of connection weights, it can be determined which sub-stage is more important in the global chain.

[0156] Step S542: The sub-link with the smaller sum of connection weight values ​​is identified as the sub-link to be adjusted, and its control parameter sequence is the parameter sequence to be adjusted. The control parameter sequence of the sub-link with the larger sum of connection weight values ​​is the benchmark parameter sequence. The parameter settings and change rates of the benchmark parameter sequence are retained.

[0157] After calculating the sum of the connection weights of the associated sub-links to the upstream and downstream sub-links, the sub-links with smaller sums of connection weights are identified as those requiring adjustment. This is because sub-links with smaller sums of connection weights have relatively lower importance in the global link, and adjusting their control parameter sequences has a relatively smaller impact on the entire system.

[0158] The control parameter sequence of the sub-stage to be adjusted is the parameter sequence to be adjusted, while the control parameter sequence of the sub-stage with the larger sum of connection weights is used as the baseline parameter sequence. The parameter setpoints and rates of change of the baseline parameter sequence are retained to ensure that the operation of important sub-stages is not affected. For example, if the sum of connection weights in the waste feeding sub-stage is less than the sum of connection weights in the incineration sub-stage, then the waste feeding sub-stage is identified as the sub-stage to be adjusted, and its control parameter sequence is the parameter sequence to be adjusted. The control parameter sequence of the incineration sub-stage is the baseline parameter sequence, and the parameter setpoints and rates of change of the incineration sub-stage are retained.

[0159] Step S543: Based on the parameter association method and the baseline parameter sequence, calculate the theoretical target parameter setpoint sequence and theoretical target parameter change rate sequence of the sub-link to be adjusted in the conflict time interval, and use them as the target reference for adjustment.

[0160] Based on the parameter correlation method and the baseline parameter sequence, the theoretical target parameter setpoint sequence and the theoretical target parameter change rate sequence for the sub-stage to be adjusted are calculated for each time interval. For example, based on the parameter correlation method between the waste feeding sub-stage and the incineration sub-stage and the baseline parameter sequence of the incineration sub-stage, the theoretical target feeding rate setpoint sequence and the theoretical target feeding rate change rate sequence for the waste feeding sub-stage are calculated for each different time interval.

[0161] Step S544: Calculate the difference between the parameter setpoint sequence in the parameter sequence to be adjusted and the theoretical target parameter setpoint sequence in the conflict time interval, and use it as the parameter setpoint adjustment amount sequence. Calculate the difference between the parameter change rate sequence in the parameter sequence to be adjusted and the theoretical target parameter change rate sequence, and use it as the parameter change rate adjustment amount sequence.

[0162] After obtaining the theoretical target parameter setpoint sequence and the theoretical target parameter change rate sequence for the sub-stage to be adjusted, the parameter setpoint sequence in the parameter setpoint sequence to be adjusted is compared with the theoretical target parameter setpoint sequence within the conflict time interval. The difference sequence between them is calculated; this difference sequence is the parameter setpoint adjustment amount sequence. Similarly, the parameter change rate sequence in the parameter setpoint sequence to be adjusted is compared with the theoretical target parameter change rate sequence, and the difference sequence is calculated as the parameter change rate adjustment amount sequence. For example, for the parameter setpoint sequence to be adjusted in the waste feeding sub-stage, the difference between its feeding speed setpoint sequence and the theoretical target feeding speed setpoint sequence is calculated at each time point within the conflict time interval to obtain the parameter setpoint adjustment amount sequence; the difference between the feeding speed change rate sequence and the theoretical target feeding speed change rate sequence is calculated to obtain the parameter change rate adjustment amount sequence. These adjustment amount sequences clarify the magnitude and direction of the parameter adjustments that need to be made.

[0163] Step S545: Perform phased smoothing on the parameter setpoint adjustment sequence and the parameter change rate adjustment sequence, and eliminate abrupt changes in the adjustment amount through multi-segment linear transition to generate a smoothed parameter setpoint adjustment sequence and parameter change rate adjustment sequence.

[0164] To avoid abrupt changes during parameter adjustment that could lead to system instability, it is necessary to perform phased smoothing on the sequence of parameter setpoint adjustments and the sequence of parameter change rate adjustments. Phased smoothing eliminates abrupt changes in the adjustment amounts through a multi-segment linear transition method, making the adjustment process smoother.

[0165] For example, a sequence of parameter setpoint adjustments can be divided into multiple time periods, with adjustments made using a linear transition within each period to avoid sudden, large adjustments at any given point in time. This generates a smoothed sequence of parameter setpoint adjustments and a sequence of parameter change rate adjustments. The smoothed adjustment sequence makes the control parameter sequence of the sub-process to be adjusted more stable during the adjustment process, reducing the impact on the system.

[0166] Step S546: Add the smoothed parameter setpoint adjustment sequence to the parameter setpoint sequence in the parameter sequence to be adjusted time by time, and add the smoothed parameter change rate adjustment sequence to the parameter change rate sequence in the parameter sequence to be adjusted time by time to obtain the adjusted control parameter sequence. Perform correlation verification between the adjusted control parameter sequence and the reference parameter sequence again. If there is still a conflict, repeat the above adjustment process until the conflict is resolved or the pre-configured upper limit of the number of adjustments is reached, and generate the coordinated control parameter sequence.

[0167] The smoothed parameter setpoint adjustment sequence is added to the parameter setpoint sequence in the parameter sequence to be adjusted time-by-time, and the smoothed parameter change rate adjustment sequence is added to the parameter change rate sequence in the parameter sequence to be adjusted time-by-time to obtain the adjusted control parameter sequence. This allows the adjustment amounts to be applied to the parameter sequence to achieve parameter adjustment. After obtaining the adjusted control parameter sequence, its correlation is checked again with the baseline parameter sequence to check for any remaining coordination conflicts. If conflicts still exist, it indicates that the adjustment is insufficient, and the above adjustment process is repeated, including recalculating the adjustment amounts and performing smoothing. This process is repeated continuously until the conflict is resolved or the pre-configured adjustment limit is reached. Finally, a coordinated control parameter sequence is generated, ensuring that the control parameter sequences of the correlated sub-links conform to the interaction rules and guaranteeing the stable operation of the waste incineration system.

[0168] Step S550: Integrate all coordinated control parameter sequences into a global control parameter matrix according to the time sampling time. The matrix row index corresponds to the time sampling time, the column index corresponds to the sub-stage identifier, the matrix element value is the parameter setting value of the sub-stage at that time, and the matrix additional rows contain the change rate information of the parameters of each sub-stage.

[0169] All coordinated control parameter sequences, obtained after adjustments in step S540, conform to the interaction rules in the global correlation link information, ensuring coordinated operation between sub-links. Integrating these coordinated control parameter sequences into a global control parameter matrix according to the time sampling sequence is to present the control parameter information of all sub-links in a structured manner, facilitating subsequent processing and execution.

[0170] The matrix row indices correspond to the time sampling moments, meaning each row represents a specific point in time. The column indices correspond to sub-stage identifiers, indicating that each column represents a specific sub-stage. Matrix element values ​​are the parameter settings for that sub-stage at that specific time; for example, the element value in a row and column represents the parameter setting for that sub-stage at that particular time. Additional rows contain information on the rate of change of parameters for each sub-stage, which helps to further understand the dynamic changes in the parameters of each sub-stage. In this way, the global control parameter matrix can comprehensively display the parameter settings and rates of change of each sub-stage in the waste incineration system at different time points.

[0171] Step S560: Based on the global control parameter matrix and the interface specifications of the execution equipment of each sub-link, generate a set of control instructions containing parameter setpoints, execution time, equipment identification and parameter change rate. Convert the set of control instructions into a standardized instruction format that conforms to the communication protocol of the waste incineration system to obtain control decisions for guiding the operation of the waste incineration system.

[0172] The global control parameter matrix contains the parameter setpoints and change rates of each sub-stage of the waste incineration system at different time points. The interface specifications for the execution devices in each sub-stage define the instruction formats and requirements that the execution devices used in each sub-stage can receive and process.

[0173] Based on the global control parameter matrix and the execution device interface specifications of each sub-stage, a set of control instructions is generated, including parameter setpoints, execution times, device identifiers, and parameter change rates. Specifically, the parameter setpoints define the parameter standards that each sub-stage needs to achieve at a specific point in time; the execution time determines when the instruction should be executed; the device identifier specifies the receiving device for the instruction; and the parameter change rate describes how the parameters change over time.

[0174] Converting the set of control commands into a standardized command format that conforms to the communication protocol of the waste incineration system is crucial to ensuring the accurate and reliable transmission and execution of these commands within the system. For example, the command set is encoded and formatted according to the requirements of the communication protocol, enabling it to be correctly interpreted and executed by sensors, controllers, and other devices within the system. The resulting control decisions, guiding the operation of the waste incineration system, precisely control the execution equipment in each sub-stage, ensuring the system operates stably and efficiently as intended.

[0175] Please see Figure 2 , Figure 2This is a schematic diagram of a computer system provided in an embodiment of the present invention. The computer system includes at least a processor 101, a communication interface 102, and a memory 103. The processor 101, communication interface 102, and memory 103 can be connected via a bus or other means. The processor 101 (or Central Processing Unit, CPU) is the computing and control core of the computer system, capable of parsing various instructions and processing various data within the computer system. The communication interface 102 may optionally include a standard wired interface or a wireless interface (such as Wi-Fi, mobile communication interface, etc.), and can be used to send and receive data under the control of the processor 101; the communication interface 102 can also be used for data transmission and interaction within the computer system. The memory 103 is a storage device in the computer system used to store programs and data. It is understood that the memory 103 here can include the computer system's built-in memory, or it can include extended memory supported by the computer system. The memory 103 provides storage space, which stores the computer system's operating system; this invention does not limit this storage space.

[0176] In one embodiment, the processor 101 executes the waste incineration control decision-making method using LSTM networks provided above in the embodiments of the present invention by running a computer program in the memory 103.

Claims

1. A method for waste incineration control decision using an LSTM network, characterized by, The method comprises: obtaining running timing information generated by continuous operation of a waste incineration system, wherein the running timing information comprises records of continuously changing running states of each link of incineration and records of interaction between links; inputting the running timing information into a main LSTM network, constructing a global correlation link across links through memory unit reconstruction processing of the main LSTM network, and obtaining global correlation link information representing the correlation relationship between links; distributing the global correlation link information to multiple slave LSTM networks, each slave LSTM network performing adaptive processing on a corresponding incineration sub-link, and outputting a sub-link adaptive request corresponding to each slave LSTM network; feeding back the sub-link adaptive requests output by all slave LSTM networks to the main LSTM network, performing timing linkage calibration processing on the sub-link adaptive requests based on the global correlation link information through the main LSTM network, and obtaining a timing linkage calibration result; based on the timing linkage calibration result, generating a control decision for guiding the operation of the waste incineration system through the main LSTM network, wherein the control decision is adapted to the running state of each incineration sub-link.

2. The method of claim 1, wherein, The inputting of the running timing information into the main LSTM network, the construction of the global correlation link across links through the memory unit reconstruction processing of the main LSTM network, and the obtaining of the global correlation link information representing the correlation relationship between links comprise: performing link state separation processing on the running timing information, splitting each link state sequence corresponding to different processing stages according to the functional attributes of each link of incineration, and each link state sequence comprising running parameter records and time markers at consecutive sampling times; inputting each link state sequence into the input layer interface of the main LSTM network, performing vector conversion on the running parameter records at each sampling time, generating a link state vector meeting the network input dimension requirement, and the link state vector comprising feature information of the running parameters at this time and a link category identifier; performing correlation calculation on the link state vector at the current sampling time and the hidden state vector at the previous sampling time through the input gate unit of the main LSTM network, generating an input correlation vector at the current time, and the input correlation vector being used to represent the dynamic dependence relationship between the current state and the historical state; performing selective filtering on the cell state vector at the previous sampling time through the forget gate unit of the main LSTM network, retaining historical state components with strong correlation to the current incineration process, and obtaining a simplified historical cell state vector; inputting the input correlation vector and the simplified historical cell state vector into the cell state update unit of the main LSTM network, performing state fusion at each time, generating an updated cell state vector at the current sampling time through nonlinear conversion, and the updated cell state vector comprising running features of the current link and historical correlation features; The sequence of updated cell state vectors based on continuous sampling time points is used to extract the state evolution trajectory of each link in the time dimension, and the global association link information including state nodes, time connections and link interaction directions is constructed by combining the interaction association records between links.

3. The method of claim 2, wherein, The sequence of updated cell state vectors based on continuous sampling time points is used to extract the state evolution trajectory of each link in the time dimension, and the global association link information including state nodes, time connections and link interaction directions is constructed by combining the interaction association records between links. The sequence of updated cell state vectors based on continuous sampling time points is used to extract the state evolution trajectory of each link in the time dimension, and the global association link information including state nodes, time connections and link interaction directions is constructed by combining the interaction association records between links. The sequence of updated cell state vectors based on continuous sampling time points is used to extract the state evolution trajectory of each link in the time dimension, and the global association link information including state nodes, time connections and link interaction directions is constructed by combining the interaction association records between links. The sequence of updated cell state vectors based on continuous sampling time points is used to extract the state evolution trajectory of each link in the time dimension, and the global association link information including state nodes, time connections and link interaction directions is constructed by combining the interaction association records between links. The sequence of updated cell state vectors based on continuous sampling time points is used to extract the state evolution trajectory of each link in the time dimension, and the global association link information including state nodes, time connections and link interaction directions is constructed by combining the interaction association records between links. The sequence of updated cell state vectors based on continuous sampling time points is used to extract the state evolution trajectory of each link in the time dimension, and the global association link information including state nodes, time connections and link interaction directions is constructed by combining the interaction association records between links. The sequence of updated cell state vectors based on continuous sampling time points is used to extract the state evolution trajectory of each link in the time dimension, and the global association link information including state nodes, time connections and link interaction directions is constructed by combining the interaction association records between links.

4. The method of claim 3, wherein, The sequence of updated cell state vectors based on continuous sampling time points is used to extract the state evolution trajectory of each link in the time dimension, and the global association link information including state nodes, time connections and link interaction directions is constructed by combining the interaction association records between links. The sequence of updated cell state vectors based on continuous sampling time points is used to extract the state evolution trajectory of each link in the time dimension, and the global association link information including state nodes, time connections and link interaction directions is constructed by combining the interaction association records between links. The sequence of updated cell state vectors based on continuous sampling time points is used to extract the state evolution trajectory of each link in the time dimension, and the global association link information including state nodes, time connections and link interaction directions is constructed by combining the interaction association records between links. The sequence of updated cell state vectors based on continuous sampling time points is used to extract the state evolution trajectory of each link in the time dimension, and the global association link information including state nodes, time connections and link interaction directions is constructed by combining the interaction association records between links. By combining the influence direction markers between links with the time correlation parameter, a node connection matrix is ​​constructed. The row index of the matrix corresponds to the preceding node, the column index corresponds to the following node, and the element value of the matrix is ​​the product of the time correlation parameter and the influence direction marker. The node connection matrix is ​​sparsified, retaining the node connection relationships whose matrix element values ​​are greater than the pre-configured judgment criteria, to obtain a simplified node connection matrix. Based on the simplified node connection matrix and the time position index of the node sequence, the topology data of the globally associated links is generated. The topology data includes the time position of the nodes, link identifiers, state feature descriptions, and connection relationships between nodes. Loop detection is performed on the topology data. When a node connection with a time loop is detected, the connection weight of the relevant node is adjusted to reduce the weight value of the loop connection so that the global associated links conform to the time unidirectional flow characteristics, thereby obtaining global associated link information that represents the operational association relationship of each link.

5. The method of claim 1, wherein, The step of distributing the global association link information to multiple slave LSTM networks, with each slave LSTM network performing adaptation processing for its corresponding incineration sub-stage and outputting a sub-stage adaptation request for each slave LSTM network, includes: The topology data of the global associated link information is parsed, and the dedicated node set corresponding to each incineration sub-stage is extracted. The dedicated node set includes all basic nodes belonging to the sub-stage and other basic nodes of the sub-stage that have a direct connection with the sub-stage. Each incineration sub-stage is configured with an independent LSTM network processing channel to convert the topology data of the dedicated node set into an input data format recognizable by the LSTM network, including node time intervals, state feature descriptions, and connection relationship encoding. The transformed input data is divided into time windows, and the window size is adjusted according to the typical operating cycle of the sub-process. Each time window contains the topology data of multiple consecutive basic nodes, so that the nodes in the window cover a complete state change cycle of the sub-process. The divided time window data is input into the hidden layer of the LSTM network in chronological order. The state feature description and connection relationship encoding of the nodes in the window are modeled by multiple hidden units to perform time-series dependency modeling. The dynamic operation features of the sub-links in the time window are extracted. The dynamic operation features include state change trends and link interaction response characteristics. The historical benchmark operating feature library of this sub-process is retrieved from the storage unit of the LSTM network. The historical benchmark operating feature library contains a set of dynamic operating feature samples of this sub-process under normal operating conditions. The dynamic operating characteristics of the current time window are compared with the sample characteristics in the historical benchmark operating characteristic library dimension by dimension. The feature deviation value of each dimension is calculated. Based on the feature deviation value, the adjustment requirement description of the sub-link is generated. The adjustment requirement description is converted into a standardized data format that includes the adjustment direction, target parameters and time interval, and the adaptation request of each sub-link corresponding to the LSTM network is obtained.

6. The method of claim 5, wherein, The process involves inputting the divided time window data into the hidden layer of the LSTM network in chronological order. Through multiple hidden units, the state feature descriptions and connection relationship encodings of nodes within the window are used to perform temporal dependency modeling, extracting the dynamic operational features of sub-processes within that time window. This includes: The node data within the time window is converted into a sequence of input vectors from the LSTM network in chronological order. Each node data corresponds to an input vector at a time step. The input vector contains a numerical vector describing the node state features and a binary vector encoding the connection relationship. The input vector sequence is input into the first hidden layer of the LSTM network. A linear transformation is performed on the input vector at each time step to generate an initial hidden state vector. The dimension of the initial hidden state vector is the same as the dimension of the input vector. The initial hidden state vector is input from the second hidden layer of the LSTM network and fused with the second hidden layer state vector of the previous time step by multiplying element by element and then adding them together to strengthen the feature dependence across time steps and generate a fused hidden state vector. The fused hidden state vector is batch normalized to obtain a standardized hidden state vector; The standardized hidden state vector is input from the third hidden layer of the LSTM network, and the vector is enhanced by a nonlinear activation function to highlight the key feature components of the sub-stage's operating state, thereby generating an enhanced hidden state vector. The enhanced hidden state vectors of all time steps within the time window are subjected to average pooling in the time dimension to generate the average hidden state vector of the time window. This average hidden state vector is used as the dynamic operation feature of the sub-stage within the time window. The dynamic operation feature includes the state change trend of the sub-stage and the interaction response characteristics of the stage.

7. The method of claim 5, wherein, The process involves comparing the dynamic operating characteristics of the current time window with the sample characteristics in the historical benchmark operating characteristic library dimension by dimension, calculating the feature deviation value for each dimension, generating an adjustment requirement description for the sub-step based on the feature deviation value, converting the adjustment requirement description into a standardized data format containing the adjustment direction, target parameters, and time interval, and obtaining the adaptation request for each sub-step corresponding to the LSTM network, including: Select sample features with the same duration as the current time window from the historical benchmark operation feature library to form a benchmark feature set. Each sample feature contains feature components of the same dimension as the dynamic operation features. Calculate the distance between the current dynamic running feature and each sample feature in the benchmark feature set, and select the sample feature with the smallest distance as the matching benchmark feature. The matching benchmark feature is the historical normal state feature that is most similar to the current state. Calculate the difference between the current dynamic running feature and the matching benchmark feature in each dimension to obtain the feature deviation value of each dimension. The deviation value is the current feature component value minus the corresponding component value of the matching benchmark feature. Based on the characteristic deviation values ​​of each dimension and the importance weight of the incineration parameters corresponding to that dimension, a comprehensive deviation index is calculated. The comprehensive deviation index is the sum of squares of the products of the deviation values ​​of each dimension and the importance weights. The importance weights are pre-configured according to the degree of influence of the incineration parameters on the system operation. When the comprehensive deviation index is greater than the pre-configured judgment standard, the incineration parameter corresponding to the dimension with a non-zero deviation value is identified and used as the parameter to be adjusted. The adjustment direction is determined according to the sign of the deviation value. Based on the parameters to be adjusted, the adjustment direction, the time interval of the current time window, and the comprehensive deviation index, a structured adjustment requirement description is generated, including parameter identifiers, adjustment direction codes, target parameter values, and execution time intervals. The structured description is then converted into a standardized data format that conforms to the system communication protocol to obtain the sub-stage adaptation request.

8. The method of claim 1, wherein, The process involves feeding back all sub-link adaptation requests output from the LSTM network to the main LSTM network, and then using the main LSTM network to perform timing-linked calibration processing on the sub-link adaptation requests based on the global associated link information to obtain timing-linked calibration results, including: Collect all sub-stage adaptation requests output from the LSTM network, sort them in ascending order according to the time stamp generated by the requests, and obtain the adaptation request sequence. Each request includes the sub-stage identifier, adjustment parameters, adjustment direction, target parameter value, and execution time interval. The global association link information is parsed, and the upstream and downstream association relationships of each sub-link in the link are extracted. When the end time marker of the execution time interval of the first sub-link overlaps with the start time marker of the execution time interval of the second sub-link, the second sub-link is marked as the downstream associated sub-link of the first sub-link. Based on the upstream and downstream relationships, the adaptation request sequence is divided into multiple related request groups. Each related request group contains multiple sub-link adaptation requests that have direct upstream and downstream relationships, so that the sub-links of the requests within the group have a link connection relationship. For each group of related requests, the execution time interval overlap is detected. If the execution time intervals of two requests overlap and the adjustment parameter is the same related parameter, it is marked as a time conflict. The sub-stage identifier, adjustment direction and time overlap interval of the conflicting request pair are recorded. Based on the connection weights of sub-links within the associated request group in the global associated link information, priority is assigned to request pairs with time conflicts. The adaptation requests corresponding to sub-links with larger connection weights retain their original execution time intervals in the event of a conflict, while the adaptation requests corresponding to sub-links with smaller connection weights have their execution time intervals adjusted. The adjusted associated request groups are re-merged into a calibration request sequence. All adaptation requests in the calibration request sequence are checked for parameter consistency to ensure that the target values ​​of the adjusted parameters of the upstream and downstream sub-links conform to the interaction rules in the global associated link information, thus obtaining a time-series linkage calibration result that includes the execution time interval after calibration and the target parameter values.

9. The method of claim 8, wherein, The step of prioritizing request pairs with time conflicts based on the connection weights of sub-links within the associated request group in the global associated link information, prioritizing the adaptation requests corresponding to sub-links with larger connection weights in the conflict by retaining their original execution time intervals, and adjusting the execution time intervals of adaptation requests corresponding to sub-links with smaller connection weights, includes: The connection weight values ​​of each sub-link within the association request group are extracted from the global association link information. The connection weight value is the average of the connection weights of the sub-link and all upstream and downstream sub-links, reflecting the overall association importance of the sub-link in the link. For request pairs with time conflicts, compare the connection weight values ​​of the sub-stages to which the two requests belong, mark the request corresponding to the sub-stage with the larger connection weight value as a high-priority request, and retain its original execution time range; The requests corresponding to the sub-links with smaller connection weight values ​​are marked as low-priority requests. The overlap between the original execution time interval of the low-priority request and the execution time interval of the high-priority request is calculated. The execution time interval of the low-priority request is then extended by the overlap, resulting in the adjusted execution time interval. Check whether the adjusted execution time interval overlaps with the execution time intervals of other high-priority requests in the associated request group. If so, continue to extend the same overlapping portion of the time until there is no overlap or the pre-configured maximum extension number is reached. When the adjusted execution time interval of a low-priority request exceeds the running cycle of the sub-process, it is split into multiple sub-requests. The execution time interval of each sub-request does not exceed the running cycle and does not overlap with the high-priority request. The target values ​​of the adjustment parameters of the sub-requests are allocated according to the split ratio. Record the priority markers of conflicting request pairs, the execution time intervals before and after adjustment, and the splitting situation, and generate conflict handling results as part of the timing linkage calibration results.

10. A computer system, characterized by include: A memory, wherein a computer program is stored; A processor for loading the computer program to implement the waste incineration control decision method using an LSTM network as described in any one of claims 1-9.

Images