Rainstorm waterlogging early warning method and system based on double propagation hidden state decoupling

Abstract

The application discloses a rainstorm waterlogging early warning method and system based on double propagation hidden state decoupling, relates to the technical field of intelligent waterlogging early warning, and provides the following scheme, which comprises the following steps: acquiring node water level, rainfall time sequence, pipe network topology and hydraulic properties, generating a candidate propagation hidden state, obtaining an inbound incremental response and a relay incremental response through reservation and shielding trial, forming a propagation connection condition, identifying an unconnected state according to the propagation connection condition, constructing an unconnected state memory and a propagation broken chain, and determining a waterlogging early warning area and a grade. The application solves the problem that node changes are still continuously written according to the original propagation relationship after downstream transportation and drainage are limited, makes the writing of the candidate propagation hidden state be constrained by the cross-hop connection relationship, and converts the state that loses the connection basis into an unconnected hidden state and a propagation broken edge.

Description

Technical Field

[0001] This invention relates to the field of intelligent early warning technology for urban flooding, and more specifically, this application relates to a method and system for early warning of urban flooding caused by rainstorms based on dual propagation hidden state decoupling. Background Technology

[0002] Urban rainstorm flood warnings typically target the drainage network comprised of manholes, storm drains, and drainage pipes. They require consideration of node water levels, rainfall patterns, and hydraulic conditions such as pipe elevation and diameter to assess flood risk. In the early stages of rainfall or when drainage capacity is sufficient, changes in node status caused by upstream water flow usually propagate downstream along the pipe network, with a relatively stable sequential relationship between adjacent nodes. However, as rainfall intensifies, downstream water levels rise, pipes become full, or there is localized backflow, this propagation relationship gradually changes.

[0003] Current methods for predicting urban flooding during heavy rain often employ hydrodynamic models, spatiotemporal neural networks, or graph neural networks. These methods predict subsequent water levels or risk states based on historical water levels at nodes, rainfall, and pipe network topology. Graph neural networks typically propagate node states along connecting edges to adjacent nodes, adjusting the propagation intensity through edge weights, attention, or temporal features. While these methods can describe the spatiotemporal correlation between nodes when drainage relationships are relatively stable, they generally assume that state changes at a particular node can still be propagated as continuous information for subsequent node updates, lacking further distinction regarding whether the same upstream change maintains its corresponding relationship in adjacent propagation stages.

[0004] When downstream drainage becomes increasingly restricted, upstream changes may still cause the water level at the current node to continue to rise, while the corresponding response at further downstream nodes may be significantly weakened, delayed, or even disappear. In this case, the continuous changes at the current node may come from normal water inflow, or may gradually include the impact of local accumulation due to downstream obstruction. If the network state is continuously updated according to the original propagation relationship, it is easy to continue to use the state that has lost the basis for continuous downstream response as normal propagation information, thereby masking the process of drainage restriction developing from downstream to upstream, making it difficult to reflect the location and development trend of urban flooding risk in a timely manner. Therefore, a rainstorm urban flooding early warning method and system based on dual propagation hidden state decoupling is proposed to solve this problem. Summary of the Invention

[0005] To address the aforementioned technical problems, this paper provides a method and system for early warning of rainstorms and urban flooding based on dual-propagation hidden state decoupling. This technical solution solves the problems mentioned in the background section.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows:

[0007] Firstly, this application provides a method for early warning of rainstorm-induced urban flooding based on dual-propagation hidden state decoupling, the method comprising:

[0008] Obtain node water levels and rainfall sequence, pipeline topology and hydraulic properties during the rainstorm, and determine the previous and next hop edges formed by sequentially connecting the sending node, receiving node and downstream node according to the pipeline topology;

[0009] The head difference, full pipe state and overflow margin determined along the preset drainage direction are obtained from the node water level and hydraulic properties. The node water level and rainfall time sequence are encoded as node hidden states. Candidate propagation hidden states are generated along the previous hop edge, and the propagation source is marked by the sending node and the previous hop edge.

[0010] For each candidate propagation hidden state, the hidden state is retained and masked respectively. The states of the receiving node and downstream node are calculated in turn. The inbound incremental response is obtained by the difference between the two receiving node states, and the retransmission incremental response is obtained by the difference between the two downstream node states. They are then paired according to the propagation source.

[0011] The continuity delay range is determined based on the consistency of direction, morphological differences, and time offset of the continuously paired incremental responses at the inlet and the subsequent incremental responses. The propagation continuity conditions are then formed by combining the variation range of head difference and overflow margin with the full pipe condition.

[0012] When the propagation continuation conditions are met, the candidate propagation hidden state is written into the hidden state of the receiving node. When the propagation continuation conditions are not met, and the head difference of the next hop edge continues to decrease, and a full pipe state or the overflow margin of the receiving node continues to decrease, it is recorded as an uncontinuation hidden state, and the next hop edge is recorded as a propagation break edge.

[0013] The unconnected hidden state is written back to the sending node according to the propagation source, and recursively merged with the historical unconnected hidden states of the same source into an unconnected state memory, and together with the hidden state of the sending node, it generates subsequent candidate propagation hidden states of the same source.

[0014] Based on the pipeline topology and the sequence of occurrence of adjacent propagation fracture edges to form a propagation fracture chain, the range of continuous nodes on this chain where unconnected state memories continuously occur and merge recursively, and the overflow margin continuously decreases, is determined as the warning area, and the warning level is determined according to the overflow margin.

[0015] Secondly, this application provides a rainstorm flooding early warning system based on dual-propagation hidden state decoupling, used to implement the aforementioned rainstorm flooding early warning method based on dual-propagation hidden state decoupling, including:

[0016] The topology configuration module is used to obtain the node water level and rainfall sequence, pipeline topology and hydraulic properties during rainstorms, and to determine the previous and next hop edges formed by sequentially connecting the sending node, receiving node and downstream node according to the pipeline topology.

[0017] The state encoding module is used to obtain the head difference, full pipe state and overflow margin determined along the preset drainage direction from the node water level and hydraulic properties, encode the node water level and rainfall time sequence as the node hidden state, generate candidate propagation hidden states along the previous hop edge, and mark the propagation source with the sending node and the previous hop edge.

[0018] The incremental calculation module is used to retain and mask the hidden state of each candidate propagation hidden state, and to calculate the state of the receiving node and downstream node in turn. The difference between the two receiving node states is used to obtain the inbound incremental response, and the difference between the two downstream node states is used to obtain the retransmission incremental response. The modules are then matched according to the propagation source.

[0019] The continuity determination module is used to determine the continuity delay range based on the consistency of direction, morphological differences and time offset of the continuously paired inbound incremental response and the subsequent incremental response, and to form the propagation continuity conditions by combining the variation range of head difference and overflow margin and the full pipe state.

[0020] The breakage identification module is used to write the candidate propagation hidden state into the hidden state of the receiving node when the propagation continuation conditions are met. When the propagation continuation conditions are not met, and the head difference of the next hop side continues to decrease, and a full pipe state or the overflow margin of the receiving node continues to decrease, it is recorded as an unconnected hidden state, and the next hop side is recorded as a propagation breakage side.

[0021] The memory write-back module is used to write back the unconnected hidden state to the sending node according to the propagation source, and recursively merge it with the historical unconnected hidden states of the same source into the unconnected state memory, and generate subsequent candidate propagation hidden states of the same source together with the hidden state of the sending node.

[0022] The urban flooding early warning module is used to form a propagation fracture chain by connecting adjacent propagation fracture edges in series according to the pipeline topology and the order of occurrence. The range of continuous nodes on the chain where the unconnected state memory continuously occurs and the overflow margin continuously decreases is determined as the early warning area, and the early warning level is determined according to the overflow margin.

[0023] Thirdly, this application provides a computer device, the computer device including a memory and a processor, the memory storing code, and the processor being configured to acquire the code and execute the above-described method for early warning of rainstorms and urban flooding based on dual-propagation hidden state decoupling.

[0024] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for early warning of rainstorms and urban flooding based on dual-propagation hidden state decoupling.

[0025] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0026] This application solves the problem of continuously writing node changes according to the original propagation relationship after downstream transmission and discharge is restricted by retaining and shielding the candidate propagation hidden state of the same propagation source, verifying the cross-jump connection relationship with the paired inbound incremental response and the continued transmission incremental response. This makes the writing of candidate propagation hidden state constrained by the cross-jump connection relationship, and transforms the state that loses the connection basis into the unconnected hidden state and the propagation break edge, thereby revealing the propagation interruption and its development position.

[0027] This application solves the problem of using the same historical propagation relationship for succession judgment after hydraulic state migration by associating the two-step succession reference state with the reference hydraulic state and selecting, updating or reforming the corresponding reference based on the current hydraulic state. This allows the propagation response formed under different hydraulic states to take on the corresponding judgment reference, avoiding the mixing of historical responses under different working conditions and thus masking the real changes in propagation relationship.

[0028] This application solves the problem of unconnected history of different downstream branches under the same propagation source being uniformly merged by forming a branch unconnected state memory according to the next hop edge, and verifying its subsequent connection relationship by the memory inbound impact response and the memory continuation impact response. This enables the unconnected state memory to constrain subsequent candidate propagation according to the actual mismatched branch, and to be gradually released after it regains the cross-hop connection relationship under its own influence. Attached Figure Description

[0029] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. Wherein:

[0030] Figure 1 The flowchart shows the rainstorm flooding early warning method based on dual-propagation hidden state decoupling proposed in this invention.

[0031] Figure 2 This is a structural block diagram of the rainstorm and urban flooding early warning system based on dual-propagation hidden state decoupling proposed in this invention. Detailed Implementation

[0032] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0033] For ease of understanding, this embodiment uses a tree-like drainage branch line with nodes A, B, C, and D arranged sequentially from upstream to downstream as an example for comprehensive explanation. The pipes from node A to node B, node B to node C, and node C to node D are connected sequentially along a predetermined drainage direction. This predetermined drainage direction is the gravity flow drainage direction determined during the pipeline network design and construction, and can be directly read from the upstream and downstream manhole numbers and pipe bottom elevation fields of each pipe segment in the pipeline network geographic information system. Level gauges are installed in the manholes at each node, reporting water levels according to a predetermined sampling period (typically 1 minute); regional rain gauges or nearby rain gauges report rainfall intensity according to the same sampling period.

[0034] Reference Figure 1 As shown, this application proposes a rainstorm flooding early warning method based on dual-propagation hidden state decoupling, including:

[0035] Obtain node water levels and rainfall sequence, pipeline topology and hydraulic properties during the rainstorm, and determine the previous and next hop edges formed by sequentially connecting the sending node, receiving node and downstream node according to the pipeline topology;

[0036] It should be noted that the node water level time series comes from the real-time reports of the level gauges in each inspection well. After being aligned with the rainfall time series according to a unified timestamp, it forms a sequence with equal sampling intervals. Missing points are filled by linear interpolation of adjacent valid samples. Nodes with consecutive missing data for more than a preset duration (typically 5 minutes) are marked as unusable data in the current prediction period and are temporarily excluded from the propagation judgment. The pipeline network topology and hydraulic attributes come from the pipeline network geographic information system. The topology is represented by node numbers and upstream and downstream well numbers of the pipe segment. The hydraulic attributes include static fields such as pipe diameter, pipe length, pipe bottom elevation, node ground elevation, and well cross-sectional area.

[0037] The sending node, receiving node, and downstream node are roles relative to a single propagation determination, not fixed entities: for any three-node sequence that is continuous along a preset drainage direction, the middle node is the receiving node, its upstream adjacent node is the sending node, its downstream adjacent node is the downstream node, the pipe segment from the sending node to the receiving node constitutes the previous jump edge, and the pipe segment from the receiving node to the downstream node constitutes the next jump edge.

[0038] Taking the branch in this embodiment as an example, when determining the propagation at node B, node A is the sending node, node B is the receiving node, and node C is the downstream node; the same node C acts as a downstream node when propagating at node B and as a receiving node when propagating at node C, and its role is determined as the determination object rolls.

[0039] The head difference, full pipe state and overflow margin determined along the preset drainage direction are obtained from the node water level and hydraulic properties. The node water level and rainfall time sequence are encoded as node hidden states. Candidate propagation hidden states are generated along the previous hop edge, and the propagation source is marked by the sending node and the previous hop edge.

[0040] It should be noted that head difference, full pipe condition, and overflow margin are all hydraulic state quantities that can be directly converted from actual measurements. They are obtained as follows: head difference The difference between the water levels at both ends of the next jump edge and the corresponding node bottom elevation is calculated using the piezometric head. This difference is subtracted from the downstream end along the preset drainage direction, in meters. A positive value indicates that there is still a driving head in the drainage direction; a negative value indicates that the downstream water level has surpassed the upstream level and has a backwater tendency. (Full pipe state) The overflow margin is determined by whether the water level at the receiving node reaches or exceeds the elevation of the top of the next side pipe. If it reaches the elevation, it is recorded as a full pipe; otherwise, it is recorded as a partially full pipe. The top pipe elevation is calculated by adding the bottom pipe elevation to the pipe diameter. The difference between the ground elevation of the node and the current water level of the node, in meters, represents the remaining space from the surface overflow. The smaller the value, the closer to the overflow.

[0041] The node hidden state and candidate propagation hidden state are encoded and generated by a graph neural network. Its input consists of aligned node water level sequences, rainfall sequences, and node static hydraulic properties, and its output is a fixed-length vector. The candidate propagation hidden state is not a direct reading of a physical quantity, but rather a state vector to be propagated from the sending node to the receiving node along the previous hop edge, whose subsequent propagation eligibility has not yet been determined. The propagation source is identified by a tuple of the sending node number and the previous hop edge number, allowing candidate propagation hidden states received by the same receiving node from different upstream branches to be tracked separately, providing an index for subsequent branch-based processing. A simple implementation of the graph neural network is as follows: node encoding uses two layers of gated recurrent units to encode the water level and rainfall sequences, concatenates them with static properties, and then passes through a fully connected layer to obtain the node hidden state. The candidate propagation hidden state is obtained by generating the candidate propagation hidden state from the sending node's hidden state and the features of the previous hop edge through a fully connected edge message generation function. .

[0042] For each candidate propagation hidden state, the hidden state is retained and masked respectively. The states of the receiving node and downstream node are calculated in turn. The inbound incremental response is obtained by the difference between the two receiving node states, and the retransmission incremental response is obtained by the difference between the two downstream node states. They are then paired according to the propagation source.

[0043] It should be noted that the counterfactual trial-and-error method, which employs the same graph neural network to perform forward inference once under both conditions of retaining and masking the same candidate propagation hidden state, and then subtracts the results from the two, aims to isolate the impact of the candidate propagation hidden state on the receiving node and downstream nodes, rather than directly reading the node state. Except for whether or not the candidate propagation hidden state is incorporated, the other inputs of the receiving node, the candidate propagation hidden states of other upstream branches, and the network parameters remain unchanged in both trials, thus ensuring the accuracy of the inbound incremental response. With incremental response It only reflects the marginal effect of the candidate propagation hidden state.

[0044] Inbound incremental response With incremental response It represents the influence of the candidate propagation hidden state on the state of the receiving node and the downstream node, rather than the existence of a component that can be extracted separately within the state of the receiving node or the downstream node. All subsequent processing based on these two responses is performed in the sense of "influence" and does not segment the node state based on this. The pairing uses the propagation source tuple as the key to bind the inbound incremental response and the follow-up incremental response generated by the same sending node and the same previous hop edge and their corresponding time to form a homogeneous response pair for subsequent time-accumulated comparison.

[0045] The continuity delay range is determined based on the consistency of direction, morphological differences, and time offset of the continuously paired incremental responses at the inlet and the subsequent incremental responses. The propagation continuity conditions are then formed by combining the variation range of head difference and overflow margin with the full pipe condition.

[0046] It should be noted that directional consistency refers to the inflow incremental response and the subsequent incremental response both increasing or decreasing within the corresponding segment; morphological difference refers to the magnitude of the difference in amplitude at each time step after alignment; time offset refers to the lag of the subsequent incremental response relative to the inflow incremental response on the time axis, its physical meaning being the propagation delay required for upstream changes to be transmitted from the receiving node to the downstream node. The continuity delay range is determined by taking the minimum and maximum values ​​of the time offsets of multiple consecutive response pairs that have historically satisfied continuity, respectively, as the lower and upper limits of the continuity delay, used to define the delay window for "normal continuity". Beyond the delay window, the propagation continuity conditions further incorporate the variation range of head difference and overflow margin, as well as the full pipe state, to ensure that the continuity determination is adapted to the current drainage conditions.

[0047] When the propagation continuation conditions are met, the candidate propagation hidden state is written into the hidden state of the receiving node. When the propagation continuation conditions are not met, and the head difference of the next hop edge continues to decrease, and a full pipe state or the overflow margin of the receiving node continues to decrease, it is recorded as an uncontinuation hidden state, and the next hop edge is recorded as a propagation break edge.

[0048] It should be noted that, when a continuation is achieved, the write operation will incorporate the candidate propagation hidden state into the update of the receiving node's hidden state, allowing it to participate in the subsequent node state evolution as valid propagation information. If the continuation is not achieved, it is not sufficient to determine a break. Hydraulic constraint evidence must also be satisfied, namely, the head difference of the next hop edge continuously decreases (indicating that the head driven by the drainage direction is declining) and there is a full pipe or the overflow margin of the receiving node continuously decreases (indicating that water is accumulating and constraining at that location). Only then will the candidate propagation hidden state be recorded as an uncontinuation hidden state and the next hop edge be recorded as a propagation break edge.

[0049] The reason for setting up this composite judgment is that: relying solely on discontinuous response may be caused by sensor noise, short-term disturbances, or model fluctuations. Judging a break based on this alone would generate a large number of false alarms. Introducing evidence of hydraulic limitation, which shows a continuous decrease in head difference accompanied by a continuous decrease in full pipe or overflow margin, as a parallel condition allows for cross-confirmation of the two types of evidence: "change in response relationship" and "definite drainage limitation." A propagation break is only recognized when both conditions are met, thereby suppressing false alarms. Among them, "continuous decrease" is defined as monotonically not increasing for multiple consecutive sampling periods and the cumulative decrease exceeding the preset range. The typical number of consecutive periods is 3, and the cumulative decrease threshold is set separately according to the dimensions of head difference and overflow margin.

[0050] The unconnected hidden state is written back to the sending node according to the propagation source, and recursively merged with the historical unconnected hidden states of the same source into an unconnected state memory, and together with the hidden state of the sending node, it generates subsequent candidate propagation hidden states of the same source.

[0051] It should be noted that write-back refers to returning the unconnected hidden state along the propagation source tuple to its sending node, so that the state that was originally sent downstream is now taken over by the sending node itself; recursive fusion refers to remembering the unconnected hidden state of this write-back along with the unconnected states accumulated by that source at the sending node, according to the same propagation source. The updates are merged and updated sequentially. The example implementation is to add the historical memory weighted by the forgetting factor to the current write-back state. The updated unconnected state memory and the hidden state of the sending node are used as the generation input to participate in the generation of candidate propagation hidden states of the source in subsequent moments. This means that the propagation state that has been rejected by the downstream is no longer simply discarded, but rather the memory influences which candidate propagation hidden states will be generated by the branch in the future.

[0052] This creates a cross-time feedback loop: once a branch remains unconnected, its unconnected state memory accumulates and strengthens over time, thereby changing the generation of subsequent candidate propagation hidden states for that branch. This allows the impact of downstream drainage constraints on upstream subsequent propagation to be gradually reflected within the network, rather than being forgotten after an isolated decision at a single moment.

[0053] Based on the pipeline topology and the sequence of occurrence of adjacent propagation fracture edges to form a propagation fracture chain, the range of continuous nodes on this chain where unconnected state memories continuously occur and merge recursively, and the overflow margin continuously decreases, is determined as the warning area, and the warning level is determined according to the overflow margin.

[0054] It should be noted that, in series connection refers to connecting adjacent propagation fracture edges along the preset drainage direction end to end according to the adjacency relationship of the pipeline network topology and the order of occurrence of each propagation fracture edge, forming a propagation fracture chain that represents the extension of drainage restriction along the pipeline network; among the nodes covered by this chain, the range of consecutive nodes where the unconnected state memory of the branch continuously occurs and is recursively merged (that is, the branch continuously generates and merges unconnected hidden states at multiple consecutive times, representing the continuous existence of restriction rather than instantaneous fluctuation) and the overflow margin continuously decreases is determined as the warning area; the warning level is determined by the overflow margin of the nodes in the area, the smaller the overflow margin, the closer to the surface overflow, and the higher the level. The fields of the warning result are: warning area (a continuous set of node numbers), propagation break chain (an ordered sequence of propagation break edges), overflow margin of each node in the area and the warning level mapped to it; for example, an output can be {warning area = [node B, node C], break chain = [edge BC, edge CD], overflow margin of node B 0.35 meters, overflow margin of node C 0.18 meters, warning level = level 2 (yellow) / level 3 (orange)}.

[0055] Through the above technical solution, this embodiment separates the inbound incremental response and the continuation incremental response of the same candidate propagation hidden state to the receiving node and the downstream node by counterfactual calculation under the conditions of retention and shielding. The basis for determining whether the upstream change still maintains the corresponding relationship in the adjacent propagation stage is whether the two are connected, rather than the amount of the whole message transmitted. The difference is that the determination of "propagation qualification" is placed at the level of the response correspondence of the same state across two hops, and the unconnected state is written back to the sending node and the memory is used to influence the subsequent generation of the branch. In this way, the process of downstream drainage restriction developing upstream along the pipeline can be gradually characterized within the network, rather than treating the state that has lost the basis of downstream continuous response as normal propagation information as in the way that only edge weights or attention are adjusted.

[0056] In an optional embodiment, for each candidate propagation hidden state, the hidden state is either preserved or masked, and the states of the receiving node and downstream nodes are calculated sequentially, specifically including:

[0057] The graph structure of the graph neural network is determined by the pipeline network topology. Pipeline network nodes and connecting edges are respectively used as graph nodes and graph edges. The hidden states of nodes are used as graph node states, and candidate hidden states are propagated along the graph edges.

[0058] It should be noted that the graph structure is established in a one-to-one correspondence with the pipeline topology: each inspection well or drainage node is used as a graph node, and each pipe segment is used as a graph edge. The direction of the graph edge is taken as the preset drainage direction, so the graph is a directed graph. The graph node state is initialized to the corresponding node hidden state. The propagation along the graph edge is that the hidden state of the sending node and the feature of the previous hop edge are used to generate candidate propagation hidden states through the edge message generation function and then sent to the receiving node. The graph neural network adopts the message passing paradigm. Its three sub-processes of message generation, message aggregation and node update are all conventional components in this field: message generation can be achieved by concatenating the hidden state of the sending node and the edge feature and then passing it through a fully connected layer; message aggregation uses summation or attention weighting for each message received by a receiving node; node update uses a gated recurrent unit to fuse the aggregated message with the current hidden state of the receiving node.

[0059] For any candidate propagation hidden state, each subsequent hop edge connecting its propagation source and the receiving node, along with the corresponding downstream node, is combined into a two-hop propagation unit.

[0060] It should be noted that the two-hop propagation unit is the smallest structure used in this embodiment to carry out a complete propagation determination. It means that it is a two-hop path formed by connecting a previous hop edge (from the sending node to the receiving node) and a subsequent hop edge (from the receiving node to the downstream node). The reason for using "each subsequent hop edge to be combined separately" is that a receiving node may drain water to multiple downstream pipe segments at the same time. When the receiving node has multiple subsequent hop edges, the same candidate propagation hidden state needs to form a two-hop propagation unit for each subsequent hop edge, so that whether the upstream change can continue to be transmitted in a certain downstream direction can be determined separately.

[0061] Using the hidden state of the same receiving node as the trial calculation benchmark, keeping the graph neural network parameters and the candidate propagation hidden states of other propagation sources unchanged, the candidate propagation hidden states are retained and masked respectively to obtain the first receiving trial state and the second receiving trial state, and the difference is used to obtain the inbound incremental response.

[0062] It should be noted that the baseline for the trial calculation refers to the hidden state of the receiving node that is shared by both forward inferences and is not superimposed with the influence of the candidate propagation hidden states to be determined. Keeping this baseline, network parameters, and other upstream branch candidates unchanged is to ensure that the difference between the two results is attributed only to the candidate propagation hidden state itself. Retention means incorporating the candidate propagation hidden state into the aggregate input of the receiving node to obtain the first receiving trial state, while masking means setting it to zero from the aggregate input to obtain the second receiving trial state. The difference between the two is the inbound incremental response. Masking is achieved by setting the contribution of the message to zero at the message aggregation point, rather than deleting the edge from the network, thus not changing the graph structure and the aggregation method of other messages.

[0063] The first and second receiving trial states are transmitted along the next hop edge in each two-hop propagation unit. The difference between the two sets of downstream node states is used to obtain the corresponding follow-up incremental response. The propagation source, the next hop edge, the inbound incremental response, the follow-up incremental response and the corresponding time are associated to form a two-hop response record.

[0064] It should be noted that the acquisition method of the continuation incremental response is as follows: the received trial state under the two conditions of retention and shielding is used as the input of the downstream node, and the downstream node state is updated by passing one more hop along the next hop edge of the two-hop propagation unit. The difference between the two sets of downstream node states is calculated. Thus, the inbound incremental response describes the immediate impact of the candidate propagation hidden state on the receiving node, and the continuation incremental response describes whether it can continue to affect the downstream node after passing through the receiving node. The two constitute a pair of influence quantities of "the same state across two hops". The two-hop response record is bound to the inbound incremental response, the continuation incremental response and the timestamp at that moment with the propagation source tuple and the next hop edge number as the key. It is stored in rolling order over time to form a unified data structure for subsequent connection reference and memory processing. When there are multiple next hop edges, a two-hop response record is formed for each next hop edge.

[0065] Through the above technical solution, this embodiment applies retention and masking to a single candidate propagation hidden state under the same trial benchmark and fixed other inputs, and continues the propagation in two-hop propagation units, organizing a single propagation decision into a two-hop response record that can be indexed by the next hop edge.

[0066] In an optional embodiment, after forming the propagation continuation conditions, a two-hop continuation reference formation process is further included, specifically comprising:

[0067] Arrange the two-hop response records that continuously meet the propagation continuation conditions in the same two-hop propagation unit according to the occurrence sequence, extract the inbound incremental response and the continuation incremental response to form inbound response segments and continuation response segments of equal length, move the continuation response segment hourly within the preset time search range, calculate the difference in change amplitude between the two response segments at each moving position, and determine the movement amount corresponding to the moving position with the smallest difference in change amplitude as the time offset.

[0068] It should be noted that the inbound response segment and the follow-up response segment are equal-length subsequences extracted from the inbound incremental response and follow-up incremental response sequences, respectively. Their length is taken as the number of sampling points covered by a typical propagation delay window (typically 10 sampling points, corresponding to 10 minutes). Hourly shift refers to translating the follow-up response segment within a preset time search range with a step size of one sampling period. The preset time search range is from 0 to one times the segment length (typically 0 to 10 sampling periods). The difference in amplitude is the sum (or mean square) of the absolute values ​​of the differences in amplitude at each moment after the two segments are aligned. The direction of the shift is to minimize this difference in amplitude, that is, to find the alignment position that best matches the shape of the follow-up response segment and the inbound response segment. The number of translation sampling points corresponding to this position is the time offset. If the difference in magnitude of change is higher than the preset reference difference threshold throughout the entire search range, it is determined that there is no effective alignment and the two-hop response record will not be included in the subsequent reference update.

[0069] Align the inbound response segment and the continuation response segment according to the time offset, calculate the difference in the change amplitude at the corresponding time, and form a jump response difference sequence.

[0070] It should be noted that the jump response difference sequence is a sequence composed of the amplitude difference between the two segments at each time step after alignment with the above time offset. It represents the point-by-point residual of the ingress and follow-up response under the best alignment. The point-by-point residual is calculated at each time step according to the best alignment position.

[0071] For example, the hop response difference sequence of the 1st hop... The elements are:

[0072] ;

[0073] in, The first step of the jump response difference sequence One element; For the inbound response segment The amplitude of each sampling point; The amplitude of the corresponding sampling point after the translation time offset of the continuation response segment; The sampling point number within the segment; since the inbound and continuation responses originate from the influence of the same candidate propagation hidden state across two adjacent hops, the two have the same dimensions and are both the hidden state update amplitude.

[0074] The continuously obtained hop response difference sequence is recursively averaged according to the corresponding sampling position to form a reference difference sequence, and the reference delay range is updated according to the minimum and maximum values ​​of the corresponding time offset. The reference difference sequence and the reference delay range are combined to form a two-hop successive reference state.

[0075] It should be noted that the reference difference sequence is a sequence obtained by recursively averaging multiple consecutive hop response difference sequences at the same sampling position, representing the typical residual shape of the two-hop propagation unit when it is normally connectable. The recursive averaging is implemented with exponential weighting, and the weight of the new sample is typically 0.1 to 0.3, so that the reference difference sequence reflects the recent shape without sudden changes due to a single fluctuation. The reference delay range is determined by taking the minimum and maximum values ​​of the time offsets corresponding to each hop response difference sequence participating in the averaging. The two-hop continuity reference state, that is, the combination of the reference difference sequence and the reference delay range, is saved as the current "normal continuity template" of the two-hop propagation unit and is continuously updated with new samples.

[0076] For the two consecutive hop response records obtained subsequently, the inbound incremental response and the follow-up incremental response are extracted to form inbound response segments and follow-up response segments of equal length. The follow-up response segments are moved hourly within a preset time search range. The time offset is determined based on the difference in the change amplitude of the two response segments at each moving position. The two response segments are then aligned according to the time offset to form a hop response difference sequence. The same segment extraction, hourly movement alignment, and hop response difference sequence calculation method are used for the subsequent two hop response records to be judged as when establishing the reference to ensure that the quantity to be judged and the reference quantity are comparable under the same caliber.

[0077] The difference between the jump response difference sequence and the reference difference sequence is calculated. When the direction of change of the inbound incremental response and the follow-up incremental response is consistent, the difference is not higher than the preset reference difference threshold, and the time offset falls within the reference delay range, it is determined that the response requirements in the propagation continuation conditions are met.

[0078] It should be noted that the difference between the jump response difference sequence to be judged and the reference difference sequence is taken as the norm of the point-by-point difference between the two (typically the mean absolute difference). The judgment uses three parallel conditions: first, directional consistency, i.e., the dominant change direction of the ingress incremental response and the subsequent incremental response within the segment is the same; second, the difference does not exceed a preset reference difference threshold. Third, the current time offset falls within the reference delay range; all three conditions must be met simultaneously to determine if the response requirements in the propagation continuation conditions are met; if any one condition is not met, the response requirements are not met, and hydraulically limited evidence will be used to determine the fracture; one way to determine the preset reference difference threshold is to take the statistical upper bound of the deviation of each jump response difference sequence relative to the reference difference sequence during the reference establishment period (such as the mean plus two standard deviations), and the typical value is that about 95% of the normal continuation samples fall within the threshold;

[0079] It should be noted that the reference itself may become inapplicable as the drainage process evolves. The response pattern on which the reference is based is formed under a certain drainage condition. When the condition changes significantly, judging new samples with the old reference will lead to systematic misjudgment. Therefore, in the next embodiment, a switching and reconstruction mechanism for the reference as the hydraulic state changes is set to ensure that the reference used for judgment matches the current condition.

[0080] For example, the following provides a complete numerical implementation process for determining the formation and response requirements of a two-hop succession reference. Taking a two-hop propagation unit consisting of node A as the sending node, node B as the receiving node, and node C as the downstream node in the branch of this embodiment as an example, the segment length is taken as 6 sampling points.

[0081] During the reference establishment phase, in a two-hop response record that consecutively satisfies the condition, the inbound response segment is [0,2,5,7,6,4], and the continuation response segment is [0,0,1,3,6,7] (units are relative values ​​of hidden state update amplitudes). The continuation response segment is moved hourly within a search range of 0 to 5 sampling periods: the two segments are most closely matched in shape and have the smallest sum of amplitude differences when shifted by 2 periods; therefore, the time offset is... =2.

[0082] according to After aligning with 2, the jump response difference sequence is calculated, resulting in the jump response difference sequence = [|0-1|,|2-3|,|5-6|,|7-7|] = [1,1,1,0] (effective overlapping segment after alignment). The reference difference sequence is obtained by weighting the jump response difference sequences of three consecutive records of the same type according to their position indices. The time offsets of the three records are 2, 2, and 3 respectively, so the reference delay range is [2,3]; the two-hop successive reference state is {reference difference sequence [1,1,1,0], reference delay range [2,3]}.

[0083] During the determination phase, a two-hop response record to be determined is processed in the same way to obtain a time offset of 2 and a hop response difference sequence of [1,2,1,1]. The average absolute difference between this sequence and the reference difference sequence is (0+1+0+1) / 4 = 0.5. Let a preset reference difference threshold be set. =1.0, then 0.5≤1.0 holds true, and the time offset 2 falls within [2,3]. Furthermore, both the entry and continuation transmissions are ascending and descending in the same direction. All three conditions are met simultaneously, and the record is determined to meet the response requirements in the propagation continuation conditions. If the jump response difference sequence of another record to be determined is [3,4,3,2] and the mean absolute difference =2.5>1.0, then the response requirements are not met, and the process is transferred to the hydraulically constrained evidence determination.

[0084] Through the above technical solution, this embodiment constructs a reference difference sequence and a reference delay range by recursively constructing samples that continuously meet the continuation conditions, and compares them with the hop response difference sequence after time offset alignment. This implements the response requirements in the propagation continuation conditions as a dual comparability criterion of morphological residual and delay window. The difference is that the determination of "whether to continue" is based on the adaptive template of the response morphology of the same candidate hop across two hops, rather than a fixed threshold or single-point comparison, so that the determination is robust to the normal drift of propagation delay.

[0085] In an optional embodiment, after combining the reference difference sequence and the reference delay range to form a two-hop successive reference state, the system further includes a successive reference switching process under hydraulic state transition, specifically including:

[0086] A reference hydraulic state is formed based on the range of head difference variation, full pipe state, and overflow margin variation of the receiving node during the formation of the two-step succession reference state, and is associated with and saved with the two-step succession reference state.

[0087] It should be noted that the reference hydraulic state describes the "under what drainage conditions" a certain two-hop successive reference state is formed. Its fields are: the range of change of the head difference of the next hop side during the establishment period (taking the minimum and maximum values ​​during this period), the full pipe status (full pipe or non-full pipe) during the establishment period, and the range of change of the overflow margin of the receiving node during the establishment period. The reference hydraulic state and the corresponding two-hop successive reference state are stored in pairs under the name of the two-hop propagation unit, so that the same two-hop propagation unit can store multiple sets of "(reference hydraulic state, two-hop successive reference state)", each corresponding to a typical drainage condition.

[0088] Obtain the range of head difference variation, full pipe status, and overflow margin variation of the receiving node for the current time period to form the current hydraulic state; compare the current hydraulic state with the reference hydraulic state, and determine the hydraulic matching condition as the one where the full pipe status is consistent and the range of head difference variation and overflow margin variation overlap.

[0089] It should be noted that the current hydraulic state is obtained by statistically analyzing the same fields as the reference hydraulic state within the nearest time period of the current determination. The time period length is consistent with the segment length when the reference was established to ensure comparability. The hydraulic matching conditions adopt "consistent full pipe state" as a necessary item, and "overlapping head difference variation range" and "overlapping overflow margin variation range" as parallel items: the full pipe state must be either both full pipe or both non-full pipe, because this state indicates whether the drainage has entered the pressurized state, which is a qualitative boundary of the operating condition; the variation range of head difference and overflow margin is considered to be matched in this dimension as long as there is a non-empty overlap with the reference interval. The use of interval overlap rather than point value equality is to allow normal fluctuations of the operating condition within a certain bandwidth.

[0090] When the hydraulic matching condition is met, the two-hop connection reference state is updated with the two-hop response record that meets the propagation connection condition. When the hydraulic matching condition is not met, the current hydraulic state is compared with each reference hydraulic state that has been saved in the same two-hop propagation unit. The reference hydraulic state that meets the hydraulic matching condition is selected, and the corresponding two-hop connection reference state is used for the current propagation connection determination. When no reference hydraulic state is selected, a new two-hop connection reference state is formed according to the two-hop response record under the current hydraulic state, following the two-hop connection reference formation process, and is associated with and saved with the current hydraulic state.

[0091] The order and triggering of the three paths are as follows: First, when the current operating condition matches the currently used reference, the reference is updated using the currently satisfied continuity records in the aforementioned recursive averaging manner, so that it continues to be refined as the operating condition evolves normally; Second, when the current operating condition and the currently used reference do not match, a match is first searched in other reference hydraulic states already saved in the two-hop propagation unit. If a match is found, the corresponding two-hop continuity reference state is switched for judgment to avoid misjudgment based on a mismatched reference; Third, if none of the saved references match (indicating the entry into a new operating condition that has not yet been seen), no old reference is used. Instead, the same process of establishing the reference is followed to re-collect continuously satisfied continuity records under the current operating condition, form a new two-hop continuity reference state, and save it in association with the current hydraulic state; The three paths are triggered exclusively in the order of "update if matched, switch if mismatched, and rebuild if no switchable one", ensuring that only one reference is used for judgment at any given time, and the convergence direction is to ensure that the reference used for judgment always corresponds to the current drainage operating condition.

[0092] Additionally, it should be noted that during the reconstruction period, the two-step propagation unit does not have a mature reference under the current operating conditions. At this time, the unit will not output the continuity judgment conclusion of the response requirements, and will only participate in the fracture judgment based on hydraulically limited evidence. The response requirement judgment will be resumed after the number of participating samples of the new reference reaches the preset minimum number (typically 3 sets of records that continuously meet the continuity requirements) to avoid misjudgment based on immature references.

[0093] Through the above technical solution, this embodiment binds the reference hydraulic state at the time of formation to each group of two-hop connection reference states, and uses full pipe consistency and interval overlap as hydraulic matching conditions to exclusively select among the three paths of "update-switch-rebuild", so that the reference used for connection determination can dynamically adapt as the drainage conditions change; the difference is that the applicability determination of the reference is based on the matching and switching of the operating condition dimension, rather than fixing a single reference or only replacing it with a new one based on time, thereby avoiding systematic misjudgment of new operating condition samples by old operating condition references; the output of the current valid two-hop connection reference state is directly called as the current benchmark for the response requirement determination in the propagation connection conditions.

[0094] In an optional embodiment, the process of writing back the unconnected hidden state to the sending node according to the propagation source, recursively merging it with the historical unconnected hidden states from the same source into an unconnected state memory, and generating subsequent candidate propagation hidden states from the same source together with the hidden state of the sending node, also includes a branch unconnected state memory differentiation process, specifically including:

[0095] Associate the unconnected hidden state with the two-hop propagation unit that generated the unconnected hidden state, and group the unconnected hidden state and the historical unconnected hidden state according to the next hop edge within the same propagation source;

[0096] It should be noted that each discontinuous hidden state originates from a specific two-hop propagation unit, corresponding to a unique preceding edge (propagation source) and a following edge. It is associated with this two-hop propagation unit, that is, it is recorded which downstream direction it occurred in. Grouping by following edge within the same propagation source is because the same upstream branch may propagate through different downstream directions, and the restriction situation of each downstream direction is different. If it is not grouped and is accumulated in general, the restriction of one direction will be mistakenly added to other directions. The following edge number is used as the key for grouping.

[0097] The unconnected hidden states of each group are recursively merged according to their order of occurrence to form the unconnected state memory of the branch corresponding to the next hop edge. The unconnected state memories of each branch together constitute the unconnected state memory of the corresponding propagation source.

[0098] It should be noted that the unconnected hidden states within each group are recursively merged according to their occurrence time to form the memory of the unconnected state of the branch corresponding to the next hop edge. ( For receiving nodes, (For downstream nodes, identify the specific next-hop edge); one implementation of recursive fusion is to weight the historical memory of the branch according to the forgetting factor and then superimpose the unconnected hidden state written back this time. The forgetting factor is typically 0.8 to 0.95, so that the influence of recent limitation dominates while the influence of long-term limitation gradually fades out; the unconnected state memory of the branches of each next-hop edge under a certain propagation source together constitute the unconnected state memory of that propagation source. ,Right now Composed of several The next side is arranged in parallel.

[0099] For each two-hop propagation unit with the same propagation source, the corresponding branch discontinuity state memory is extracted based on the next hop edge. The branch discontinuity state memory and the hidden state of the sending node are input into the candidate propagation hidden state generation process of the corresponding two-hop propagation unit to obtain the subsequent candidate propagation hidden state of the two-hop propagation unit.

[0100] It should be noted that when generating the subsequent candidate propagation hidden state of a two-hop propagation unit, only the unconnected state memory of the branch corresponding to the next hop edge is extracted for generation. The candidate generation of the sending node towards other next hop edges uses the memory of its corresponding branch and is not applied to each other. Thus, the restriction in a certain downstream direction only inhibits or changes the subsequent candidate propagation hidden state in that direction, without affecting the propagation of the sending node towards other unrestricted downstream directions. The "extracting the memory of the corresponding branch according to the next hop edge and then inputting it for generation" means that in the two-hop propagation unit structure, the effect of memory on generation is applied to each downstream direction.

[0101] A key issue to note here is that if the generation is not based on the next hop edge but on a single node-level memory, the restriction of a receiving node in the direction of A will also suppress its normal propagation in the direction of B, resulting in false suppression. In this embodiment, the memory of the branch not connected is extracted according to the next hop edge and applied only to the candidate generation in the corresponding direction, thus avoiding this cross-directional crosstalk.

[0102] Through the above technical solution, this embodiment splits the unconnected state memory into branch unconnected state memory according to the next hop edge, and only uses the memory of the corresponding branch when generating the subsequent candidate propagation hidden state of each two-hop propagation unit. This allows the restriction of downstream drainage on the subsequent propagation of the upstream to be applied in each downstream direction. The difference is that the effect of memory on generation is organized at the granularity of "each next hop edge" rather than uniformly applied at the node level, which avoids the false suppression of other normal directions of the same node by a certain restricted direction.

[0103] In an optional embodiment, after obtaining the subsequent candidate propagation hidden states of the two-hop propagation unit, a branch discontinuation state memory recovery process is further included, specifically including:

[0104] When the subsequent candidate propagation hidden states of the target two-hop propagation unit meet the propagation continuation conditions, keep the hidden state of the sending node and the graph neural network parameters unchanged, retain the memory of the uncontinuation state of the target branch and replace the memory with a zero state respectively, and generate the first branch candidate propagation hidden state and the second branch candidate propagation hidden state.

[0105] It should be noted that the recovery process is only triggered when the subsequent candidate propagation hidden state of the target two-hop propagation unit satisfies the propagation continuation condition again, that is, the downstream direction has been reconnected. In order to measure whether the memory of the current branch's unconnected state still has a substantial effect on the generation, the same method as the counterfactual calculation is adopted: keep the hidden state of the sending node and the network parameters unchanged, once retain the memory of the target branch's unconnected state to participate in the generation of the first branch's candidate propagation hidden state, and once replace the memory with a zero state to participate in the generation of the second branch's candidate propagation hidden state. The difference between the two is the current influence of the memory on the candidate generation.

[0106] Using the same hidden state of the receiving node as the benchmark, while keeping the candidate hidden states of other propagation sources unchanged, the candidate hidden states of the first branch and the candidate hidden states of the second branch are updated at the receiving node and the message is passed to the next hop along the target two-hop propagation unit. The difference between the two sets of receiving node states and the two sets of downstream node states is calculated to obtain the memory inbound impact response and the memory continuation impact response.

[0107] It should be noted that by performing a two-hop trial calculation along the target two-hop propagation unit for each of the two branch candidate propagation hidden states, the difference between the two sets of receiving node states is obtained as the memory inbound impact response, and the difference between the two sets of downstream node states is obtained as the memory continuation impact response. These two parameters describe the indirect impact of "the branch memory being generated via the candidate" on the receiving node and downstream node. They are in the same two-hop trial calculation caliber as the inbound incremental response and continuation incremental response obtained from the trial calculation of the candidate propagation hidden states, so they can be measured using the same set of connection criteria. The memory inbound impact response and memory continuation impact response obtained here are still impact quantities and do not involve segmentation of node states.

[0108] The cumulative number of consecutive recovery times is calculated when the direction consistency, morphological difference, and time offset of the memory ingress impact response and the memory continuation impact response meet the direction consistency, morphological difference, and time offset conditions corresponding to the propagation continuation conditions.

[0109] It should be noted that, based on the current effective two-hop connection reference state, the memory ingress impact response and memory continuation impact response are subject to the same three criteria: directional consistency, the difference between the cross-hop response difference sequence and the reference difference sequence does not exceed a preset reference difference threshold, and the time offset falls within the reference delay range. If all three criteria are met simultaneously, it indicates that the restricted impact carried by the branch memory has become re-acceptable downstream, meaning the direction has indeed been restored, and the number of consecutive restorations will be... The count is incremented once; if any condition is not met, it indicates that the recovery has not yet stabilized, and the number of consecutive recovery counts is reset to zero to ensure that the recovery is continuous and to prevent a single fluctuation from being mistaken for a recovery.

[0110] The memory retention coefficient is determined based on the number of consecutive recovery cycles, which decreases as the number of consecutive recovery cycles increases. The memory of the unconnected state of the target branch is scaled according to the memory retention coefficient. When the number of consecutive recovery cycles reaches the preset number of recovery cycles, the memory of the unconnected state of the target branch is updated to zero.

[0111] It should be noted that the memory retention coefficient To increase with the number of consecutive recovery times The coefficient, which increases and decreases monotonically, ranges between 0 and 1. One implementation is a tiered approach, such as 0.6 for one consecutive recovery, 0.3 for two, and 0 for three, or an exponential approach that decays with the number of consecutive recoveries. The memory of the unconnected state of the target branch is scaled according to this coefficient, meaning its amplitude gradually shrinks as the recovery continues. When the number of consecutive recoveries reaches the preset number (typically three), the memory of the unconnected state of the target branch is directly updated to zero, completely removing the inhibition of subsequent candidate generation in that downstream direction. The reason for using tiered scaling instead of a one-time zeroing is that downstream recovery may be repetitive. Tiered scaling ensures that the memory retains some influence before a stable recovery is confirmed, avoiding the oscillation of inhibition and release caused by a single suspected recovery being completely released.

[0112] For example, one expression for the decay of the memory retention coefficient with the number of consecutive recoveries is:

[0113] ;

[0114] in, The memory retention factor; This is the attenuation base, typically taken as 0.5; The number of consecutive recovery cycles is used. Since a larger number of consecutive recovery cycles indicates a more certain downstream recovery and a weaker retention of memory for that branch, a base less than 1 is used to make the memory retention coefficient monotonically decrease with the number of consecutive recovery cycles.

[0115] For example, the following provides a complete numerical implementation of the memory recovery process for a branch. Suppose that the branch from node B to node C (the next hop edge BC) was previously unconnected, and the memory amplitude of its unconnected state accumulated to a relative value of 1.0. At a certain moment, the subsequent candidate propagation hidden state in this direction satisfies the continuation again, triggering a recovery trial: the first trial shows that the direction of the memory ingress impact response and the memory continuation impact response is consistent, the average absolute difference between the jump response difference sequence and the current reference difference sequence is 0.4 (lower than the preset reference difference threshold of 1.0), and the time offset 2 falls within the reference delay range [2,3]. When all three conditions are met, the number of consecutive recovery times increases from 0 to 1, the memory retention coefficient is taken as 0.6, and the branch memory scaling is 1.0 × 0.6 = 0.6; at the next moment, it is still satisfied, the number of consecutive recovery times is 2, the coefficient is 0.3, and the memory scaling is 0.6 × (0.3 / 0.6) = 0.3 (i.e., relative to the initial value of 0.3); at the next moment after that, it is still satisfied, the number of consecutive recovery times reaches the preset 3 times, the branch unconnected state memory is updated to zero state, and the generation of the subsequent candidate propagation hidden state from node B to node C is no longer suppressed by this memory. If any of the three items in the second trial calculation are invalid (e.g., the mean absolute difference rises to 1.5 > 1.0), then the number of consecutive recovery attempts is reset to zero, and the memory is maintained at the current amplitude without scaling.

[0116] Through the above technical solution, this embodiment measures the current impact of the branch memory by using two-hop trial calculations from the same source when the downstream direction becomes reconnectable, and scales it monotonically decreasing until it is cleared by the number of consecutive recovery cycles, so that the memory of the branch not being connected is controlled to fade away as the downstream restriction is actually relieved.

[0117] In an optional embodiment, the process of forming a propagation fracture chain by connecting adjacent propagation fracture edges according to the pipeline topology and the order of occurrence also includes a downstream support constraint verification process, specifically including:

[0118] For the next jump edge that has been recorded as a propagation fracture edge, obtain the head difference, full pipe state and downstream node overflow margin determined at both ends along the preset drainage direction. When the head difference changes from a non-negative value to a negative value, and is accompanied by the downstream node full pipe state or downstream node overflow margin continuously decreasing and continuously maintaining a preset restricted time, the propagation fracture edge is determined as a downstream restricted fracture edge; otherwise, it is determined as a non-downstream restricted fracture edge.

[0119] It should be noted that the purpose of this verification is to identify from the recorded propagation break edges that are indeed caused by downstream backwater constraint. A change from a non-negative to a negative head difference indicates that the downstream piezometer head surpasses the upstream head, i.e., a backwater trend has emerged. Accompanying a downstream node at full capacity or a continuously decreasing overflow margin indicates that the downstream is indeed under pressure or experiencing increased water accumulation. These conditions must be maintained for a preset constraint period (typically 3 sampling periods) to confirm the break edge and exclude instantaneous fluctuations. If all three conditions are met, it is recorded as a downstream constrained break edge; otherwise, it is recorded as a non-downstream constrained break edge. This verification only reads existing hydraulic state quantities such as head difference, full capacity, and overflow margin; it does not generate any reverse candidate propagation latent states, does not perform reverse transmission, and does not assume the existence of a real flow along the reverse drainage direction.

[0120] For two-step propagation units that have been selected with reference hydraulic state, when the current hydraulic state of the unit changes from non-full pipe to full pipe relative to the selected reference hydraulic state, or when the range of head difference changes in the negative direction, it is determined that the hydraulic state of the two-step propagation unit has migrated on the restricted side, and the downstream restricted fracture side of the corresponding next step side is marked as hydraulic migration confirmed.

[0121] It should be noted that the hydraulic state migration on the restricted side uses the maintained reference hydraulic state as a reference system: if the current hydraulic state changes from a non-full pipe to a full pipe relative to the selected reference hydraulic state, or if the range of head difference changes in the whole direction, it indicates that the working condition is evolving in the restricted direction; accordingly, the corresponding downstream restricted fracture edge is marked as hydraulic migration confirmation, as strengthening evidence that the fracture is indeed in the process of intensifying restriction.

[0122] When connecting adjacent propagating fracture edges according to the pipeline topology and the order of occurrence, only the downstream restricted fracture edge is included in the upstream connection of the propagating fracture chain, and the non-downstream restricted fracture edge is excluded from the propagating fracture chain.

[0123] It should be noted that the downstream restricted fracture edge is used as the admission condition when connecting the fracture chains: only the downstream restricted fracture edge can participate in the upstream connection of the fracture chain, and the non-downstream restricted fracture edge (such as the one whose water level rises mainly due to local heavy rainfall and is not caused by downstream backflow) is excluded from the chain; thus, the fracture chain depicts the process that is indeed restricted downstream and develops upstream along the pipeline network, avoiding the direction and warning area of ​​isolated fracture pollution chains caused by non-backflow.

[0124] When one of two adjacent downstream restricted fracture edges is not marked as hydraulic migration confirmed and its unconnected state memory no longer undergoes continuous recursive fusion, the upstream connection of the propagation fracture chain is terminated at that downstream restricted fracture edge.

[0125] It should be noted that the termination condition is used to define a clear boundary for the upstream extension of the propagation fracture chain: when the downstream confined fracture edge on the upstream side is neither confirmed by hydraulic migration (the condition at this point does not continue to evolve in the confined direction) nor does its unconnected state memory continue to recursively merge (the confining of this branch is not continuous, and unconnected hidden states are no longer generated and merged at each time step), it indicates that the confined extension no longer has supporting evidence, and therefore the upstream connection is terminated at the downstream confined fracture edge. This termination is based solely on two types of existing evidence: whether the condition continues to deteriorate and whether the confining continues, without inferring the actual hydraulic state further upstream.

[0126] It should be noted that the "unconnected state memory no longer continuously undergoes recursive fusion" used here is the same criterion as "unconnected state memory continuously undergoes recursive fusion" in the early warning area determination. That is, whether the branch continuously generates and merges unconnected hidden states at continuous moments is used as a unified criterion for whether the restriction is continuous, so that the termination boundary of the broken chain and the delineation of the early warning area are coordinated with each other.

[0127] Through the above technical solution, this embodiment identifies downstream confinement based on the recorded propagation fracture edge when the head difference turns negative and is accompanied by increased downstream pressure or water accumulation. Only downstream confined fracture edges are allowed to participate in the series connection, and the termination boundary for upstream extension is based on whether the operating conditions continue to deteriorate and whether the confinement continues. This ensures that the propagation fracture chain excludes isolated fractures not caused by backwater and has a clear upstream termination. The difference lies in establishing the inclusion and termination of the fracture chain on the verification of existing hydraulic state quantities and the duration of confinement, rather than indiscriminately connecting all fracture edges or extending infinitely upstream, thereby improving the accuracy of the early warning area delineation. The verified propagation fracture chain output serves as the direct basis for determining the early warning area and mapping the early warning level.

[0128] See Figure 2 As shown, this scheme proposes a rainstorm flooding early warning system based on dual-propagation hidden state decoupling, used to implement the aforementioned rainstorm flooding early warning method based on dual-propagation hidden state decoupling, including:

[0129] The topology configuration module is used to obtain the node water level and rainfall sequence, pipeline topology and hydraulic properties during rainstorms, and to determine the previous and next hop edges formed by sequentially connecting the sending node, receiving node and downstream node according to the pipeline topology.

[0130] The state encoding module is used to obtain the head difference, full pipe state and overflow margin determined along the preset drainage direction from the node water level and hydraulic properties, encode the node water level and rainfall time sequence as the node hidden state, generate candidate propagation hidden states along the previous hop edge, and mark the propagation source with the sending node and the previous hop edge.

[0131] The incremental calculation module is used to retain and mask the hidden state of each candidate propagation hidden state, and to calculate the state of the receiving node and downstream node in turn. The difference between the two receiving node states is used to obtain the inbound incremental response, and the difference between the two downstream node states is used to obtain the retransmission incremental response. The modules are then matched according to the propagation source.

[0132] The continuity determination module is used to determine the continuity delay range based on the consistency of direction, morphological differences and time offset of the continuously paired inbound incremental response and the subsequent incremental response, and to form the propagation continuity conditions by combining the variation range of head difference and overflow margin and the full pipe state.

[0133] The breakage identification module is used to write the candidate propagation hidden state into the hidden state of the receiving node when the propagation continuation conditions are met. When the propagation continuation conditions are not met, and the head difference of the next hop side continues to decrease, and a full pipe state or the overflow margin of the receiving node continues to decrease, it is recorded as an unconnected hidden state, and the next hop side is recorded as a propagation breakage side.

[0134] The memory write-back module is used to write back the unconnected hidden state to the sending node according to the propagation source, and recursively merge it with the historical unconnected hidden states of the same source into the unconnected state memory, and generate subsequent candidate propagation hidden states of the same source together with the hidden state of the sending node.

[0135] The urban flooding early warning module is used to form a propagation fracture chain by connecting adjacent propagation fracture edges in series according to the pipeline topology and the order of occurrence. The range of continuous nodes on the chain where the unconnected state memory continuously occurs and the overflow margin continuously decreases is determined as the early warning area, and the early warning level is determined according to the overflow margin.

[0136] In another embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above embodiments.

[0137] In one embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps described above.

[0138] In one embodiment, a computer program product or computer program is provided, the computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and executes the computer instructions, causing the computer device to perform the steps described above.

[0139] The technical scope of this invention is not limited to the content described above. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this invention, and all such modifications and variations should fall within the protection scope of this invention.

Claims

1. A method for early warning of urban flooding caused by rainstorms based on dual-propagation hidden state decoupling, characterized in that, The method includes: Obtain node water levels and rainfall sequence, pipeline topology and hydraulic properties during the rainstorm, and determine the previous and next hop edges formed by sequentially connecting the sending node, receiving node and downstream node according to the pipeline topology; The head difference, full pipe state and overflow margin determined along the preset drainage direction are obtained from the node water level and hydraulic properties. The node water level and rainfall time sequence are encoded as node hidden states. Candidate propagation hidden states are generated along the previous hop edge, and the propagation source is marked by the sending node and the previous hop edge. For each candidate propagation hidden state, the hidden state is retained and masked respectively. The states of the receiving node and downstream node are calculated in turn. The inbound incremental response is obtained by the difference between the two receiving node states, and the retransmission incremental response is obtained by the difference between the two downstream node states. They are then paired according to the propagation source. The continuity delay range is determined based on the consistency of direction, morphological differences, and time offset of the continuously paired incremental responses at the inlet and the subsequent incremental responses. The propagation continuity conditions are then formed by combining the variation range of head difference and overflow margin with the full pipe condition. When the propagation continuation conditions are met, the candidate propagation hidden state is written into the hidden state of the receiving node. When the propagation continuation conditions are not met, and the head difference of the next hop edge continues to decrease, and a full pipe state or the overflow margin of the receiving node continues to decrease, it is recorded as an uncontinuation hidden state, and the next hop edge is recorded as a propagation break edge. The unconnected hidden state is written back to the sending node according to the propagation source, and recursively merged with the historical unconnected hidden states of the same source into an unconnected state memory, and together with the hidden state of the sending node, it generates subsequent candidate propagation hidden states of the same source. Based on the pipeline topology and the sequence of occurrence of adjacent propagation fracture edges to form a propagation fracture chain, the range of continuous nodes on this chain where unconnected state memories continuously occur and merge recursively, and the overflow margin continuously decreases, is determined as the warning area, and the warning level is determined according to the overflow margin.

2. The method according to claim 1, characterized in that, For each candidate propagation hidden state, the hidden state is either preserved or masked, and the states of the receiving node and downstream nodes are calculated sequentially, specifically including: The graph structure of the graph neural network is determined by the pipeline network topology. Pipeline network nodes and connecting edges are respectively used as graph nodes and graph edges. The hidden states of nodes are used as graph node states, and candidate hidden states are propagated along the graph edges. For any candidate propagation hidden state, each subsequent hop edge connecting its propagation source and the receiving node, along with the corresponding downstream node, is combined into a two-hop propagation unit. Using the hidden state of the same receiving node as the trial calculation benchmark, keeping the graph neural network parameters and the candidate propagation hidden states of other propagation sources unchanged, the candidate propagation hidden states are retained and masked respectively to obtain the first receiving trial state and the second receiving trial state, and the difference is used to obtain the inbound incremental response. The first and second receiving trial states are transmitted along the next hop edge in each two-hop propagation unit. The difference between the two sets of downstream node states is used to obtain the corresponding follow-up incremental response. The propagation source, the next hop edge, the inbound incremental response, the follow-up incremental response and the corresponding time are associated to form a two-hop response record.

3. The method according to claim 2, characterized in that, After the propagation continuation conditions are established, the process for forming a two-hop continuation reference also includes: Arrange the two-hop response records that continuously meet the propagation continuation conditions in the same two-hop propagation unit according to the occurrence sequence, extract the inbound incremental response and the continuation incremental response to form inbound response segments and continuation response segments of equal length, move the continuation response segment hourly within the preset time search range, calculate the difference in change amplitude between the two response segments at each moving position, and determine the movement amount corresponding to the moving position with the smallest difference in change amplitude as the time offset. Align the inbound response segment and the continuation response segment according to the time offset, calculate the difference in the change amplitude at the corresponding time, and form a jump response difference sequence. The continuously obtained hop response difference sequence is recursively averaged according to the corresponding sampling position to form a reference difference sequence, and the reference delay range is updated according to the minimum and maximum values ​​of the corresponding time offset. The reference difference sequence and the reference delay range are combined to form a two-hop successive reference state. For the two consecutive hop response records obtained subsequently, the inbound incremental response and the resume incremental response are extracted to form inbound response segments and resume response segments of equal length. The resume response segments are moved hourly within a preset time search range. The time offset is determined according to the difference in the change amplitude of the two response segments at each moving position. The two response segments are then aligned according to the time offset to form a hop response difference sequence. The difference between the jump response difference sequence and the reference difference sequence is calculated. When the direction of change of the inbound incremental response and the follow-up incremental response is consistent, the difference is not higher than the preset reference difference threshold, and the time offset falls within the reference delay range, it is determined that the response requirements in the propagation continuation conditions are met.

4. The method according to claim 3, characterized in that, After combining the reference difference sequence and the reference delay range to form a two-hop successive reference state, the process also includes a successive reference switching procedure under hydraulic state transition, specifically including: A reference hydraulic state is formed based on the range of head difference variation, full pipe state, and overflow margin variation of the receiving node during the formation of the two-step succession reference state, and is associated with and saved with the two-step succession reference state. Obtain the range of change in head difference of the next jump side, full pipe status, and range of change in overflow margin of the receiving node for the current time period to form the current hydraulic state; By comparing the current hydraulic state with the reference hydraulic state, the hydraulic matching condition is defined as the full pipe state being consistent and the ranges of head difference and overflow margin variation overlapping respectively. When the hydraulic matching condition is met, the two-hop connection reference state is updated with the two-hop response record that currently meets the propagation connection condition. When the hydraulic matching condition is not met, the current hydraulic state is compared with each reference hydraulic state that has been saved in the same two-hop propagation unit. The reference hydraulic state that meets the hydraulic matching condition is selected, and the corresponding two-hop connection reference state is used for the current propagation connection determination. When no reference hydraulic state is selected, the two-step succession reference formation process is followed to form a new two-step succession reference state based on the two-step response record under the current hydraulic state, and the new state is associated with and saved with the current hydraulic state.

5. The method according to claim 2, characterized in that, The process of writing back the unconnected hidden state to the sending node according to the propagation source, recursively merging it with the historical unconnected hidden states from the same source into an unconnected state memory, and generating subsequent candidate propagation hidden states from the same source together with the hidden state of the sending node, also includes a branch unconnected state memory differentiation process, specifically including: Associate the unconnected hidden state with the two-hop propagation unit that generated the unconnected hidden state, and group the unconnected hidden state and the historical unconnected hidden state according to the next hop edge within the same propagation source; The unconnected hidden states of each group are recursively merged according to their order of occurrence to form the unconnected state memory of the branch corresponding to the next hop edge. The unconnected state memories of each branch together constitute the unconnected state memory of the corresponding propagation source. For each two-hop propagation unit with the same propagation source, the corresponding branch discontinuity state memory is extracted based on the next hop edge. The branch discontinuity state memory and the hidden state of the sending node are input into the candidate propagation hidden state generation process of the corresponding two-hop propagation unit to obtain the subsequent candidate propagation hidden state of the two-hop propagation unit.

6. The method according to claim 5, characterized in that, After obtaining the subsequent candidate propagation hidden states of the two-hop propagation units, the process also includes a branch unconnected state memory recovery procedure, which specifically includes: When the subsequent candidate propagation hidden states of the target two-hop propagation unit meet the propagation continuation conditions, keep the hidden state of the sending node and the graph neural network parameters unchanged, retain the memory of the uncontinuation state of the target branch and replace the memory with a zero state respectively, and generate the first branch candidate propagation hidden state and the second branch candidate propagation hidden state. Using the same hidden state of the receiving node as the benchmark, while keeping the candidate hidden states of other propagation sources unchanged, the candidate hidden states of the first branch and the candidate hidden states of the second branch are updated at the receiving node and the message is passed to the next hop along the target two-hop propagation unit. The difference between the two sets of receiving node states and the two sets of downstream node states is calculated to obtain the memory inbound impact response and the memory continuation impact response. The cumulative number of consecutive recovery times is calculated when the direction consistency, morphological difference, and time offset of the memory ingress impact response and the memory continuation impact response meet the direction consistency, morphological difference, and time offset conditions corresponding to the propagation continuation conditions. The memory retention coefficient, which decreases with the number of consecutive recoveries, is determined based on the number of consecutive recoveries, and the memory of the unconnected state of the target branch is scaled according to the memory retention coefficient. When the number of consecutive recovery attempts reaches the preset number of recovery attempts, the unconnected state memory of the target branch will be updated to zero.

7. The method according to claim 4, characterized in that, The process of forming a propagation fracture chain by connecting adjacent propagation fracture edges according to the pipeline topology and the order of occurrence also includes a downstream support constraint verification process, which specifically includes: For the next jump edge that has been recorded as a propagation fracture edge, obtain the head difference, full pipe state and downstream node overflow margin determined at both ends along the preset drainage direction. When the head difference changes from a non-negative value to a negative value, and is accompanied by the downstream node full pipe state or downstream node overflow margin continuously decreasing and continuously maintaining a preset restricted time, the propagation fracture edge is determined as a downstream restricted fracture edge; otherwise, it is determined as a non-downstream restricted fracture edge. For two-step propagation units that have been selected with reference hydraulic state, when the current hydraulic state of the unit changes from non-full pipe to full pipe relative to the selected reference hydraulic state, or when the range of head difference changes in the negative direction, it is determined that the hydraulic state of the two-step propagation unit has migrated on the restricted side, and the downstream restricted fracture side of the corresponding next step side is marked as hydraulic migration confirmed. When connecting adjacent propagating fracture edges according to the pipeline topology and the order of occurrence, only the downstream restricted fracture edge is included in the upstream connection of the propagating fracture chain, and the non-downstream restricted fracture edge is excluded from the propagating fracture chain. When one of two adjacent downstream restricted fracture edges is not marked as hydraulic migration confirmed and its unconnected state memory no longer undergoes continuous recursive fusion, the upstream connection of the propagation fracture chain is terminated at that downstream restricted fracture edge.

8. A rainstorm flooding early warning system based on dual-propagation hidden state decoupling, characterized in that, A method for implementing the rainstorm flooding early warning method based on dual-propagation hidden state decoupling as described in any one of claims 1-7 includes: The topology configuration module is used to obtain the node water level and rainfall sequence, pipeline topology and hydraulic properties during rainstorms, and to determine the previous and next hop edges formed by sequentially connecting the sending node, receiving node and downstream node according to the pipeline topology. The state encoding module is used to obtain the head difference, full pipe state and overflow margin determined along the preset drainage direction from the node water level and hydraulic properties, encode the node water level and rainfall time sequence as the node hidden state, generate candidate propagation hidden states along the previous hop edge, and mark the propagation source with the sending node and the previous hop edge. The incremental calculation module is used to retain and mask the hidden state of each candidate propagation hidden state, and to calculate the state of the receiving node and downstream node in turn. The difference between the two receiving node states is used to obtain the inbound incremental response, and the difference between the two downstream node states is used to obtain the retransmission incremental response. The modules are then matched according to the propagation source. The continuity determination module is used to determine the continuity delay range based on the consistency of direction, morphological differences and time offset of the continuously paired inbound incremental response and the subsequent incremental response, and to form the propagation continuity conditions by combining the variation range of head difference and overflow margin and the full pipe state. The breakage identification module is used to write the candidate propagation hidden state into the hidden state of the receiving node when the propagation continuation conditions are met. When the propagation continuation conditions are not met, and the head difference of the next hop side continues to decrease, and a full pipe state or the overflow margin of the receiving node continues to decrease, it is recorded as an unconnected hidden state, and the next hop side is recorded as a propagation breakage side. The memory write-back module is used to write back the unconnected hidden state to the sending node according to the propagation source, and recursively merge it with the historical unconnected hidden states of the same source into the unconnected state memory, and generate subsequent candidate propagation hidden states of the same source together with the hidden state of the sending node. The urban flooding early warning module is used to form a propagation fracture chain by connecting adjacent propagation fracture edges in series according to the pipeline topology and the order of occurrence. The range of continuous nodes on the chain where the unconnected state memory continuously occurs and the overflow margin continuously decreases is determined as the early warning area, and the early warning level is determined according to the overflow margin.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method described in any one of claims 1-7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1-7.

Images