A Deep Learning-Based Method and System for Determining the Causes of Traffic Split Ratio Deviation

Abstract

This invention discloses a method and system for determining the cause of diversion ratio deviation based on deep learning, relating to the field of smart irrigation canal technology. The method includes: hourly comparing the target allocation ratio of branch canal diversion nodes with the measured allocation ratio of each branch canal within the current scheduling cycle; constructing an event graph centered on the diversion nodes; using graph-based deep learning for message passing and local aggregation on direct paths to obtain the influence characterization parameters of the upstream and downstream sides of each branch canal; identifying the dominant side of hydraulic boundary changes and outputting the determination result of the dominant side upstream or downstream; encapsulating the causal fingerprint verification result into a trial order and binding it with an event number, executing adjustment and verification actions according to the dominant side; performing consistency verification between the evidence package and the causal fingerprint, determining the evidence support status, and outputting the cause of the diversion ratio deviation. This invention can significantly shorten the positioning time, avoid large-scale trial and error and linked oscillations, reduce the risk of exceeding limits, and improve the downstream plan fulfillment rate and operational stability.

Description

Technical Field

[0001] This invention relates to the field of smart irrigation canal technology, and more specifically, to a method and system for determining the causes of deviations in diversion ratios based on deep learning. Background Technology

[0002] With the advancement of irrigation district modernization and water network coordination, parallel branch canal water distribution based on target allocation ratios has become a common practice. Diversion nodes need to stably distribute upstream water to each branch canal proportionally under multi-source disturbances to ensure downstream water use planning and the safety of control sections.

[0003] However, in actual operation, the target ratio often deviates continuously from the measured ratio, affecting execution. Common sources fall into two main categories: first, changes in the hydraulic boundary caused by sudden changes in upstream water flow or downstream backwater rise; second, inadequate execution due to valve command transmission or actuator malfunctions. These two types of factors can overlap in time and mask each other in signal transmission, making it difficult to determine the root cause on-site in a timely manner.

[0004] Existing technologies largely rely on the following approaches: 1) Offline analysis or digital twin verification based on global hydraulic simulation, but model parameters require frequent calibration, boundary conditions are difficult to obtain in real time, computation is intensive, and rapid on-site location is difficult; 2) Rule- or threshold-driven closed-loop water distribution and alarms are triggered only by single-point flow / water level exceeding limits or proportional errors, lacking causal decomposition of upstream drivers and downstream backtracking, easily misjudging as valve problems or blindly increasing control gain, causing oscillations; 3) Traditional data-driven anomaly detection is mostly black-box classification, with weak topology dependence and insensitivity to asynchronous and lag-dependent measurements, making it difficult to provide executable conclusions that can be verified by a single small action. Especially in complex operating conditions such as multi-branch canal coupling, sensor drift, communication lag, and gate wear, existing methods often require large-scale trial and error or multi-point linkage, increasing the risk of exceeding limits and communication costs.

[0005] Therefore, there is an urgent need for a method that can quickly and practically differentiate hydraulic boundary changes and valve malfunctions when the allocation ratio continues to deviate and has already affected downstream operations, without resorting to large-scale handling and trial and error. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a method and system for determining the cause of deviation in the diversion ratio based on deep learning. It has the advantages of significantly shortening the positioning time, avoiding large-scale trial and error and linkage oscillation, reducing the risk of exceeding limits, and improving the downstream plan fulfillment rate and operational stability. This solves the problem that traditional technologies require large-scale trial and error or multi-point linkage, which increases the risk of exceeding limits and communication costs.

[0007] To achieve the aforementioned advantages of significantly shortening positioning time, avoiding large-scale trial and error and linked oscillations, reducing the risk of exceeding limits, and improving the downstream plan fulfillment rate and operational stability, the specific technical solution adopted by this invention is as follows:

[0008] Firstly, a method for determining the cause of traffic splitting ratio deviation based on deep learning is provided, including:

[0009] S1. Compare the target allocation ratio of the branch canal diversion node with the measured allocation ratio of each branch canal in the current scheduling cycle on an hourly basis, filter out the branch canals that do not meet the preset judgment rules, and generate event numbers.

[0010] S2. Construct an event graph centered on the diversion node, including directly related upstream water inflow nodes, branch canal outflow nodes, downstream key section water level nodes and gate control points, establish upstream and downstream information flow paths formed by directed relationships, and export node list and path list.

[0011] S3. Using the node list, path list, target allocation ratio and measured allocation ratio as input, graph-based deep learning is used to perform message passing and local aggregation on the direct path to obtain the influence characterization parameters of the upstream and downstream sides of each branch canal, form the causal fingerprint and calculate the lateral score.

[0012] S4. Verify the causal fingerprint of each branch canal, and combine the difference between the lateral score and the score threshold to identify the dominant side of the hydraulic boundary change, and output the determination result of the dominant side of the upstream or downstream.

[0013] S5. Encapsulate the causal fingerprint verification result into a probe order and bind it to the event number, execute the adjustment and verification actions according to the dominant side; and monitor the regression status of the measured allocation ratio in real time to generate an evidence package;

[0014] S6. Perform consistency verification between the evidence package and the causal fingerprint, determine the evidence support status, output the reasons for the deviation of the diversion ratio, and archive the trigger judgment threshold and constraint rules by event number.

[0015] Furthermore, the target allocation ratio of the branch canal diversion nodes within the current scheduling cycle is compared hourly with the measured allocation ratio of each branch canal. Branch canals that do not meet the preset judgment rules are filtered out, and event numbers are generated, including:

[0016] S11. Using the current sampling time as an index, and based on the joint admission of the preset deviation threshold and the shortest duration, compare the target allocation ratio of the diversion node with the measured allocation ratio of each branch canal. If at least one branch canal in the set of mandatory branch canals deviates in the same direction for more than the shortest duration within the current scheduling cycle, and the absolute value of the deviation is greater than the preset deviation threshold, then the continuous deviation is determined to be valid. Simultaneously record the start time, deviation direction marker, and maximum absolute deviation to generate candidate events.

[0017] S12. The planned water supply satisfaction and satisfaction threshold of the branch canal, the water level and upper limit of the key section, and the flow rate and flow red line of the control section are compared in turn. When any of the three comparison quantities exceeds the limit in a continuous period, it is determined that the candidate event has affected the downstream execution and a unique event number is generated.

[0018] Furthermore, an event graph is constructed centered on the diversion nodes, incorporating directly related upstream inflow nodes, tributary outflow nodes, downstream key section water level nodes, and gate control points. This establishes directed upstream and downstream information flow paths, and derives a node list and a path list, including:

[0019] S21. Take the diversion node corresponding to the event number as the core diversion node and set the event modeling boundary to construct the event graph; add the upstream inflow node set, the downstream key section water level node of the parallel branch canal, the outflow node of each branch canal, the gate set and control point set related to the distribution to the vertex set of the event graph in sequence, and use the principle of proximity and direct access to converge the boundary of the event graph.

[0020] S22. Perform directed modeling of the node relationships in the event graph. Set the upstream inflow node set pointing to the central diversion node as the driving direction, set the downstream water level nodes of each branch canal pointing to the central diversion node as the push direction, and set the central diversion node pointing to the outflow nodes of each branch canal as the distribution output. Connect the influence edges of the gate set and control point set that are directly coupled with the distribution output to the direct relationship of the central diversion node or the branch canal diversion nodes. Use the business rule base to filter the directed edge set of the event graph and retain only the edges that directly cause the diversion ratio to deviate.

[0021] S23. On the event graph, set the upstream and downstream information flow paths and corresponding candidate causes. All direct paths from the upstream inflow node set to each branch canal outflow node through the central diversion node are merged into the upstream driving path set and aggregated into the upstream information flow. Direct paths from the central diversion node to each branch canal outflow node are merged into the downstream backtracking path set and aggregated into the downstream information flow. Direct priority and single convergence point constraints are introduced to avoid the upstream and downstream information flow paths crossing irrelevant branches.

[0022] S24. Establish a unique reference interface at the distribution node, set the central distribution node as the parallel recording position of the upstream and downstream information flow paths, and synchronously mark the direction and magnitude of the upstream and downstream information flows at the parallel recording position; bind the event number using a fixed-length reference observation window, and export the node list and path list from the event graph.

[0023] Furthermore, using the node list, path list, target allocation ratio, and measured allocation ratio as input, graph-based deep learning is employed to perform message passing and local aggregation on direct paths, obtaining the influence characterization parameters of the upstream and downstream sides of each branch canal, forming a causal fingerprint, and calculating lateral scores, including:

[0024] S31. Extract the content belonging to the upstream and downstream information flows from the node list and path list. Using the target allocation ratio and actual allocation ratio of the observation window corresponding to the event number as input, set the central sub-flow node within the analysis scope and set the central sub-flow node as the unique reference position; and use the direct priority strategy to limit the path list to direct paths.

[0025] S32. Perform message passing and local aggregation on the upstream driving subgraph according to the direct path in the path list, gather the driving effects from the upstream integrated node set at the central diversion node, and generate upstream driving impact characterization parameters for each branch canal.

[0026] S33. Perform message passing and local aggregation on the downstream driving subgraph according to the direct path in the path list, and gather the back push effect from the downstream integrated node set at the central diversion node to generate downstream back push impact characterization parameters for each branch canal.

[0027] S34. The upstream driving influence characterization parameters and the downstream retrospective influence characterization parameters are cross-compared in the dual-channel comparison module at the central diversion node. A direct path is adopted and the admission criteria are set with a rule base. Channel admission marks are constructed, and the deviation direction and deviation magnitude of each branch channel are determined to generate channel interpretation marks. Lateral scores are obtained by weighting and primacy factor aggregation.

[0028] S35. Assemble the control group results into a causal fingerprint, converge the upstream side as the upstream influence intensity, converge the downstream side as the downstream influence intensity, and collect the directional consistency marker set and amplitude interpretable marker set for each branch canal, and simultaneously generate the initial judgment indication of the dominant side.

[0029] Furthermore, the upstream integrated node set includes: the combination of upstream water inflow nodes, the gate set, and the control point set; the downstream integrated node set includes: the downstream key section water level nodes, the gate set, and the control point set.

[0030] Both upstream and downstream driving influence characterization parameters include channel direction markings, magnitude of action, and relative order of action.

[0031] Furthermore, the formula for calculating the lateral score is as follows:

[0032] ;

[0033] In the formula, , Explanation markers for upstream and downstream channels; , Access is granted for direct routes between upstream and downstream areas; Marking for deviation from direction; The maximum absolute deviation; , Marking the upstream and downstream channel directions; , The amplitude of the channel effect between upstream and downstream; , The relative order of upstream and downstream; Weight of branch canals; The amplitude threshold; It is a very small amount; This is a collection of essential branch canals; k Branch canal serial number; , Lateral scores are given for upstream and downstream areas; As the dominant side; down Downstream; exec Verification of the chain to be executed; This is the boundary of the difference.

[0034] Furthermore, the causal fingerprints of each branch canal are verified, and the difference between the lateral score and the score threshold is compared to identify the dominant side of the hydraulic boundary change. The results of determining the dominant side of the upstream or downstream are output, including:

[0035] S41. The upstream side of the causal fingerprint and the upstream driving influence characterization parameters, channel interpretation markers, lateral scores, deviation direction markers and maximum absolute deviations are checked unilaterally within the observation window; the channel direction markers and deviation direction markers of each branch canal are judged for directional consistency. After the judgment is passed, the amplitude of the branch canal's effect and the maximum absolute deviation are judged for amplitude sufficiency; the proportion of the branch canals that pass both the directional consistency and amplitude sufficiency judgments in the set of mandatory branch canals and the lateral scores are used as a gating, and are judged against the proportion threshold and the score threshold respectively. When both the proportion threshold and the score threshold are satisfied, the hydraulic boundary change is determined to be upstream-dominated, a unique conclusion is output, and the event number is bound to the priority action side;

[0036] S42. The downstream side of the causal fingerprint and the downstream driving influence characterization parameters, channel interpretation markers, lateral scores, deviation direction markers and maximum absolute deviations are checked unilaterally in the observation window; the channel direction markers and deviation direction markers of each branch canal are judged for directional consistency. After the judgment is passed, the amplitude of the branch canal's effect and the maximum absolute deviation are judged for amplitude sufficiency; the proportion of the branch canals that pass both the directional consistency and amplitude sufficiency judgments in the set of mandatory branch canals and the lateral scores are used as a gating, and are judged against the proportion threshold and the score threshold respectively. When both the proportion threshold and the score threshold are satisfied, the hydraulic boundary change is determined to be backwater dominant, a unique conclusion is output, and the branch canal set and the priority action side are recorded.

[0037] S43. Compare the instruction sequence, gate position sequence, and corresponding cross-sectional flow sequence in the observation window with the same gate for causal analysis. If the instruction changes and the gate position sequence is not in place within the tolerance, or if the gate position sequence is in place but the direction and magnitude of the increase or decrease in the cross-sectional flow sequence do not meet the threshold and the hysteresis exceeds the threshold, it is determined that the valve is not executed properly. If the evidence is insufficient, the side with the lateral score closest to the deviation from the requirement is taken as the temporary conclusion, and the dominant side temporary conclusion is output as the only conclusion, and the reversal condition is recorded simultaneously.

[0038] Furthermore, the causal fingerprint verification result is encapsulated into a probe command and bound to the event number, and adjustment and verification actions are executed according to the dominant side; and the regression status of the measured allocation ratio is monitored in real time, generating an evidence package including:

[0039] S51. Encapsulate the unique conclusion, trial side, target regression direction, allowed action amplitude, rate limit, observation window, and limit constraint set into a trial command and bind it to the event number; the diversion node agent and the gate agent receive the trial command and check the action object, amplitude rate, and observation window item by item; after the check is passed, use the diversion node agent and the gate agent to establish a one-time session and lock the session attributes that do not cross-site linkage and do not handle calculations in the middle platform, generate a single-step action template and return field placeholders; if any check fails, reject the session and return the rejection reason and maintain the current state of the device;

[0040] S52. Use changes in hydraulic boundaries as the basis for action. When the upstream is the dominant side, select the set of gates related to the incoming water to perform legal adjustments. When the downstream is the dominant side, select the set of gates related to the backwater to perform legal adjustments. In the observation window, make a coherent judgment on whether the measured distribution ratio of each branch canal continuously regresses along the target regression direction relative to the target distribution ratio. Record the action time, distribution ratio response trajectory, and whether the information of exceeding the limit and prohibition of action is triggered item by item to form executable evidence.

[0041] S53. Using the judgment of incomplete execution as the basis for action, perform a gate position verification action on the relevant gate set. Perform a gate position verification action on the relevant gate set, and use the same observation window to compare whether the corresponding cross-sectional flow sequence produces the same increase or decrease in gate position sequence. Record the sequential relationship and amplitude matching of the command sequence, gate position sequence and cross-sectional flow sequence item by item to form execution chain evidence.

[0042] S54. Combine the executable evidence and the execution chain evidence into an evidence package. Perform consistency verification based on the event number and cause fingerprint. If the evidence supports the original judgment of the unique conclusion, output the final unique cause and generate an action summary and response summary. If the evidence does not support the original judgment of the unique conclusion, switch the unique conclusion to the opposite cause according to the flip condition and terminate the trial.

[0043] Furthermore, the evidence package is checked for consistency with the causal fingerprint to determine the evidence support status, output the reasons for deviations in the diversion ratio, and archive the trigger judgment thresholds and constraint rules by event number, including:

[0044] S61. Encapsulate the unique cause and probing evidence into an output package, using fixed fields and fixed pointers, and bind it to the central distribution node; synchronously push and record the receipt on the distribution node interface; publish the output package with consistent tags for station control and scheduling, and give it a key probing timestamp and observation interval limit;

[0045] S62. Archive the triggering conditions, judgment criteria, and trial evidence of the current event number to generate an archive package. Record the judgment process data in a structured manner and establish a multi-level index using the event number as the primary key to associate and store the archive package with the output package.

[0046] S63. Use the archived results to adjust the triggering and judgment thresholds, and record each adjustment as a rule change, mark the effective time and applicable diversion node, and associate it with the event number range.

[0047] Secondly, a deep learning-based system for determining the cause of traffic splitting ratio deviation is also provided. This system includes:

[0048] The event coding generation module is used to compare the target allocation ratio of the branch canal diversion nodes with the measured allocation ratio of each branch canal in the current scheduling cycle on an hourly basis, filter out branch canals that do not meet the preset judgment rules, and generate event numbers.

[0049] The list export module is used to construct an event graph centered on the diversion node, including directly related upstream water inflow nodes, branch canal outflow nodes, downstream key section water level nodes and gate control points, to establish a directed relationship between upstream and downstream information flow paths, and to export the node list and path list.

[0050] The causal fingerprint generation module is used to take the node list, path list, target allocation ratio and measured allocation ratio as input, and use graph-based deep learning to perform message passing and local aggregation on the direct path to obtain the influence characterization parameters of the upstream and downstream sides of each branch canal, form the causal fingerprint and calculate the lateral score.

[0051] The dominant determination and evaluation module is used to verify the causal fingerprint of each branch canal, and by combining the difference between the lateral score and the score threshold, it identifies the dominant side of the hydraulic boundary change and outputs the determination result of the dominant side of the upstream or downstream.

[0052] The evidence encapsulation module is used to encapsulate the causal fingerprint verification results into a probe order and bind it to the event number, execute adjustment and verification actions according to the dominant side, and monitor the regression status of the measured allocation ratio in real time to generate an evidence package.

[0053] The archiving and verification module is used to verify the consistency between the evidence package and the causal fingerprint, determine the evidence support status, output the reasons for the deviation of the diversion ratio, and archive the trigger judgment threshold and constraint rules by event number.

[0054] Compared with existing technologies, this invention provides a method and system for determining the cause of traffic splitting ratio deviation based on deep learning, which has the following beneficial effects:

[0055] (1) This invention constructs a minimum event graph at the diversion node and extracts two types of information flows, namely upstream drive and downstream pushback, by continuously deviating and affecting the standard dual-gated triggering events, and generates a causal fingerprint. Then, with cross-comparison with consistent direction and sufficient amplitude and light verification, it quickly determines that the deviation mainly comes from changes in hydraulic boundary or valve execution failure, and outputs a unique conclusion. Combined with a small trial with minimum step size, speed limit and safety barrier constraint, it can be verified within a short observation window, which significantly shortens the positioning time, avoids large-scale trial and error and linkage oscillation, reduces the risk of exceeding the limit, and improves the downstream plan fulfillment rate and operational stability.

[0056] (2) Adopt the nearest and direct convergence boundary, delete detour or weak causal edges to ensure that the information flow does not cross irrelevant branches; perform single-round message passing and local aggregation on the direct subgraph, and the direction, magnitude and relative order of each branch can be output by deep learning models such as graph neural networks, and the uncertainty is given; complete dual-channel comparison and lateral scoring at the unique comparison interface of the diversion node, which is more accurate, less expensive and more interpretable than traditional threshold alarm or global simulation, and is more robust to asynchronous measurement, lag and local drift, and supports weighting according to branch weight and rules to suppress noise and occasional interference.

[0057] (3) In terms of engineering, the method is decoupled from the station control system: whitelist and no-movement constraints, veto, safety barrier, step size and rate limit, non-middle platform handling, single session lock, to ensure that the trial is controllable; evidence packets are generated and sent back for actions, gate positions, flow and allocation response, and the conclusion is automatically reversed if the evidence is inconsistent; archive all elements of the event and drive the threshold to slightly adapt, support cross-station reuse and online iteration, reduce maintenance and communication costs, improve on-site decision-making efficiency, compliance and reliability, and achieve low-risk, traceable and closed-loop cause determination. Attached Figure Description

[0058] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0059] Figure 1 This is a flowchart of a method for determining the cause of traffic splitting ratio deviation based on deep learning according to an embodiment of the present invention;

[0060] Figure 2 This is a logical flowchart of a deep learning-based method for determining the cause of traffic splitting ratio deviation according to an embodiment of the present invention.

[0061] Figure 3 This is a system principle block diagram of a deep learning-based system for determining the cause of traffic deviation according to an embodiment of the present invention.

[0062] In the picture:

[0063] 1. Event coding generation module; 2. List export module; 3. Causal fingerprint generation module; 4. Leading judgment and evaluation module; 5. Evidence encapsulation module; 6. Archive verification module. Detailed Implementation

[0064] To further illustrate the various embodiments, the present invention provides accompanying drawings, which are part of the disclosure of the present invention. These drawings are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these drawings, those skilled in the art should be able to understand other possible implementation methods and the advantages of the present invention. The components in the drawings are not drawn to scale, and similar component symbols are generally used to represent similar components.

[0065] According to an embodiment of the present invention, a method and system for determining the cause of traffic splitting ratio deviation based on deep learning are provided.

[0066] According to one embodiment of the present invention, the invention will now be further described in conjunction with the accompanying drawings and specific embodiments, such as... Figure 1-Figure 2 As shown, the method for determining the cause of traffic splitting ratio deviation based on deep learning according to an embodiment of the present invention includes:

[0067] S1. Compare the target allocation ratio of the branch canal diversion node with the measured allocation ratio of each branch canal in the current scheduling cycle on an hourly basis, filter out the branch canals that do not meet the preset judgment rules, and generate event numbers.

[0068] Specifically, the target allocation ratio within the current scheduling cycle is compared hourly with the measured allocation ratio of each branch canal. A combined admission criteria of a preset deviation threshold and a minimum duration is adopted. When a continuous deviation in the same direction and exceeding the threshold occurs in the set of mandatory branch canals, a continuous deviation is determined to be established, and the start time, deviation direction, and maximum absolute deviation are recorded. Subsequently, the planned water supply satisfaction, key section water level, and control section flow are compared with their thresholds and minimum impact durations, respectively. If any continuous non-compliance is not met, it is determined that execution has been affected, and an event number is generated.

[0069] In the description of this invention, the method of comparing the target allocation ratio of the branch canal diversion nodes with the measured allocation ratio of each branch canal in the current scheduling cycle on an hourly basis, filtering out branch canals that do not meet the preset judgment rules, and generating event numbers includes:

[0070] S11. Using the current sampling time as an index, and based on the joint admission criteria of a preset deviation threshold and the shortest duration, compare the target allocation ratio of the diversion node with the measured allocation ratio of each branch canal. If at least one branch canal in the set of mandatory branch canals experiences a deviation in the same direction exceeding the shortest duration within the current scheduling cycle, and the absolute value of the deviation is greater than the preset deviation threshold, then a continuous deviation is determined to be valid. Simultaneously record the start time, deviation direction marker, and maximum absolute deviation to generate candidate events.

[0071] Specifically, the target allocation ratio of the parallel branch canal diversion nodes within the current scheduling cycle is compared hourly with the measured allocation ratio of each branch canal; based on time... t (Indicating the sampling time) serves as the index for branch canals. k (Indicates the branch canal number included in the judgment) (Indicating the target allocation ratio) and r k ( t (representing the measured allocation ratio) are compared; a deviation threshold is used. i r (Indicates the absolute value of the deviation threshold) and the shortest duration t min Joint admission (representing the minimum duration of continuous determination); when at least one branch canal in the mandatory set B (representing the set of branch canals to be covered) is within the time period t min The internal direction of deviation continues to appear, i.e., deviation direction markers. sk ∈{+1, -1} (representing the deviation direction) does not flip during this period, and the absolute value of the deviation is continuously greater than 1. i r When the condition is met, a continuous deviation is determined; simultaneously, the start time is recorded. t 0 (indicates the moment when the threshold is first met), deviation direction marker s k The number of records representing the amplitude a k (Indicates the maximum absolute deviation within this segment), generating candidate events that are not included in the causal analysis.

[0072] S12. The planned water supply satisfaction and satisfaction threshold of the branch canal, the water level and upper limit of the key section, and the flow rate and flow red line of the control section are compared in turn. When any of the three comparison quantities exceeds the limit in a continuous period, it is determined that the candidate event has affected the downstream execution and a unique event number is generated.

[0073] Specifically, the impact of the candidate events from the previous step on downstream execution will be determined; for branch canals... k Planned water supply satisfaction M k ( t (representing the planned completion rate) and threshold (Indicating minimum acceptable satisfaction) is used as a reference for key sections. j water level h j ( t (This indicates the real-time water level at the cross-section) and its upper limit. (This indicates the upper limit of the water level control at this section) Compare this to the control section. m Traffic q m ( t (This represents the real-time flow rate at that cross-section) and the red line (Indicates the upper limit of the flow rate at this cross-section) for comparison; the intensity of influence and the shortest duration of influence are used. t imp The hard threshold strategy (representing the minimum duration of continuous influence) applies as long as any of the above three types of quantities is within a continuous time period. t imp If any condition is not met or exceeds the limit, it is determined that execution has been affected. At the same moment this determination is made, the cause determination process is immediately initiated, and a unique event number ID (representing the unique identifier of this event) is generated for reference in subsequent steps.

[0074] S2. Construct an event graph centered on the diversion node, including directly related upstream water inflow nodes, branch canal outflow nodes, downstream key section water level nodes and gate control points, establish upstream and downstream information flow paths formed by directed relationships, and export node list and path list.

[0075] Specifically, an event graph is constructed centered on the diversion node, including only the upstream inflow nodes directly related to water distribution, the outflow nodes of each branch canal, the water level nodes of key cross-sections of each branch canal downstream, and the gates and control points related to distribution. Based on the principles of proximity and direct access and the business rule base, detour or weak causal edges are deleted, and a directed relationship is established to form an upstream driving information flow and a downstream pushback information flow. A unique reference interface is set at the diversion node and a reference window of fixed length is bound to export the node list and path list.

[0076] In the description of this invention, an event graph is constructed centered on the diversion node, incorporating directly related upstream inflow nodes, tributary outflow nodes, downstream key section water level nodes, and gate control points. A directed relationship is established to create upstream and downstream information flow paths, and a node list and path list are derived, including:

[0077] S21. Take the diversion node corresponding to the event number as the core diversion node and set the event modeling boundary to construct the event graph. Add the upstream inflow node set, the downstream key section water level nodes of the parallel branch canals, the outflow nodes of each branch canal, the gate set and control point set related to the allocation to the vertex set of the event graph in sequence, and converge the boundary of the event graph using the principle of proximity and direct access.

[0078] Specifically, the event modeling boundary is determined by using the branch node triggered by S1 as the center, and an event graph G=(V,E) is constructed. The branch node v is then used as the boundary. c V is incorporated as the sole center; the upstream inflow node set U and the downstream key section water level node H of the parallel branch canals are included. k Outflow nodes Q of each branch canal k The gate set A and control point set C related to the allocation are added to V in sequence; only objects that are directly related to water distribution and can affect the allocation result within the observation range of this event are included in V; the boundary is converged by the principle of proximity and direct access, and objects with weak correlation at a distance are eliminated; the event number ID is bound to G to maintain context consistency and prepare the minimum structure for cause determination.

[0079] S22. Perform directed modeling of the node relationships in the event graph. Set the upstream inflow node set pointing towards the central diversion node as the driving direction, set the downstream water level nodes of each branch canal pointing towards the central diversion node as the backflow direction, and set the central diversion node pointing towards the outflow nodes of each branch canal as the distribution output. Connect the influence edges of the gate set and control point set directly coupled to the distribution output to the direct access relationships of the central diversion node or the branch canal diversion nodes. Use a business rule base to filter the directed edge set of the event graph, retaining only edges that directly cause deviations in the diversion ratio.

[0080] Specifically, a directed model is performed on the node relationships in the event graph; the upstream water inflow node set U is directed to the branching node v.c Define the driving direction and set the downstream water level node H of each branch canal. k Point to v c Define the pushback direction and split the flow node v c Pointing to the outflow node Q of each branch canal k Define the distribution output; for the set of gates A and the set of control points C directly coupled to the distribution, connect their influence edges only to v. c Or Q k The direct relationship is established; the business rule base R is used to filter the directed edge set E, retaining only the edges that can directly cause the allocation ratio to deviate, and deleting detour or weak causal edges to eliminate unnecessary coupling.

[0081] S23. Define upstream and downstream information flow paths and corresponding candidate causes on the event graph. Merge all direct paths from the upstream inflow node set to the outflow nodes of each branch canal via the central diversion node into an upstream driving path set, which is then aggregated into an upstream information flow. Merge direct paths from the central diversion node to the outflow nodes of each branch canal into a downstream backtracking path set, which is then aggregated into a downstream information flow. Introduce direct-path priority and single-convergence-point constraints to prevent upstream and downstream information flow paths from crossing irrelevant branches.

[0082] Specifically, two types of information flows are defined on the event graph to correspond to candidate causes; the upstream water inflow node set U will flow through the branching node v. c Effect on the outflow nodes Q of each branch canal k All direct paths are merged into the upstream driving path set P. up And aggregate it into upstream information flow I up This is used to explain the proportion shift caused by changes in inflow water; it will be derived from each H k via v c Affecting Q k All direct paths are merged into a downstream backtracking path set P. down And aggregate it into downstream information flow I down This is used to explain the proportional deflection caused by changes in the backwater; direct priority and single convergence point constraints are adopted to ensure that the two types of information flows do not cross irrelevant branches.

[0083] S24. Establish a unique reference interface at the distribution node, set the central distribution node as the parallel recording position for upstream and downstream information flow paths, and synchronously mark the direction and magnitude of the upstream and downstream information flows at the parallel recording position. Use a fixed-length reference observation window to bind event numbers, and export the node list and path list from the event graph.

[0084] Specifically, a unique reference interface is established at the splitting node; VC is set as the parallel recording location for the two types of information streams, and I is only processed at this location. up with I downThe direction and magnitude of action are synchronously marked; a fixed-length comparison window W is used to bind the event ID, which carries the local time period and context required for subsequent comparison; a node list L is exported from the event graph. node With path list L path L path Contains only P up and P down The direct path in the list is used to eliminate detours and weakly related paths; the above list is used as the sole input for the next step of causal fingerprint extraction.

[0085] S3. Using the node list, path list, target allocation ratio and measured allocation ratio as input, graph-based deep learning is used to perform message passing and local aggregation on the direct path to obtain the influence characterization parameters of the upstream and downstream sides of each branch canal, form the causal fingerprint and calculate the lateral score.

[0086] Specifically, using the node list and path list exported from S2, as well as the target ratio and measured ratio within the window, as input, graph-based deep learning message passing and local aggregation are used only on the direct path to obtain the channel direction markings, influence magnitudes, and relative order of each branch channel on the upstream and downstream sides. At the diversion node, the deviation direction and maximum absolute deviation of the two sides' representations and the records in S1 are cross-referenced. Explanation marks are generated according to the consistency of direction and the sufficiency of magnitude, and converged into lateral scores. If the difference between the scores on both sides exceeds the difference boundary, the dominant side is determined; otherwise, proceed to S4. A causal fingerprint containing the influence intensity of both sides and the set of explanation marks is formed.

[0087] In the description of this invention, taking a node list, a path list, a target allocation ratio, and a measured allocation ratio as inputs, graph-based deep learning is used to perform message passing and local aggregation on direct paths to obtain the influence characterization parameters of the upstream and downstream sides of each branch canal, forming a causal fingerprint and calculating lateral scores, including:

[0088] S31. Extract the content belonging to the upstream and downstream information flows from the node list and path list. Using the target allocation ratio and actual allocation ratio of the observation window corresponding to the event number as input, set the central sub-flow node within the analysis scope and set the central sub-flow node as the unique reference position. And use the direct priority strategy to limit the path list to direct paths.

[0089] Specifically, the node list L exported from S2 node With path list L path The middle belongs to the upstream information flow I up With downstream information flow I down The portion, along with the target proportion within window W corresponding to the event ID. Compared with the measured ratio r k (t) is used as input; the analysis scope revolves only around the branch node v.c Configure, v c As the sole reference location; L is limited using a direct-access priority strategy. path For a direct path, for {U, H k Q k The associated quantities of A and C within window W are referenced but their boundaries are not extended; the start and end points of the window are bound to the event ID, and the business objectives of the two types of information flows are marked to explain the source of this ratio deviation.

[0090] S32. Perform message passing and local aggregation on the upstream driving subgraph according to the direct path in the path list, and gather the driving effects from the upstream integrated node set at the central diversion node to generate upstream driving impact characterization parameters for each branch canal.

[0091] Specifically, the upstream driving subgraph is divided into L... path Direct path P in up Perform a message passing and local aggregation; at the splitter node v c By aggregating the effects from {U, A, C}, a characterization of the upstream driving effects for each branch canal is generated. Among them, the channel direction mark This indicates whether the allocation ratio is increased or decreased, and the magnitude of the effect. Indicates the strength of the explainable effects, in relative order. Indicates the chronological order of events; The service is limited to answering whether changes in upstream water flow dominate the current deviation.

[0092] S33. Perform message passing and local aggregation on the downstream driving subgraph according to the direct path in the path list, and gather the back push effect from the downstream integrated node set at the central diversion node to generate downstream back push impact characterization parameters for each branch canal.

[0093] Specifically, the downstream backtracking subgraph is arranged according to L path Direct path P in down Perform a message passing and local aggregation; at the splitter node v c Collected from {H k The backtracking effect of A and C generates a characterization of the downstream backtracking impact for each branch canal. Among them, the channel direction mark This indicates whether the distribution ratio is suppressed or amplified, and the magnitude of the effect. Indicates the strength of the explainable effects, in relative order. Indicates the chronological order of events; The function is limited to answering whether changes in the backwater level dominate the current deviation.

[0094] S34. The upstream driving influence characterization parameters and the downstream retrospective influence characterization parameters are cross-compared at the central diversion node using a dual-channel comparison module. A direct path is adopted, and access is set using a rule base. Channel access markers are constructed, and the deviation direction and deviation magnitude of each branch channel are determined to generate channel interpretation markers. Lateral scores are calculated by weighting and converging sequential factors.

[0095] Specifically, will and At the split node VC, a cross-comparison is performed in the dual-channel comparison module; using the direct path P. up With P down Based on the rule base R, establish access permissions and construct channel access markers. and For each branch canal, s k Indicates the actual deviation from the direction, expressed in terms of a. k This indicates a deviation from the upper limit of the amplitude, respectively, and , and , Perform direction consistency and amplitude sufficiency checks to generate explanation markers. , Using weight w k and sequential factors , Convergence lateral score S up , S down Select the dominant side D based on the difference boundary η; if it does not exceed the boundary, output the execution chain for verification: ;

[0096] In the formula, , Explanation markers for upstream and downstream channels; , Access is granted for direct routes between upstream and downstream areas; Marking for deviation from direction; The maximum absolute deviation; , Marking the upstream and downstream channel directions; , The amplitude of the channel effect between upstream and downstream; , The relative order of upstream and downstream; Weight of branch canals; The amplitude threshold; It is a very small amount; This is a collection of essential branch canals; k Branch canal serial number; , Lateral scores are given for upstream and downstream areas; As the dominant side; down Downstream; exec Verification of the chain to be executed; This is the boundary of the difference.

[0097] The dual-channel comparison module refers to the comparison interface at the split node, and its function is to represent the upstream driving characteristics generated by the graph network within the same time window. Downstream backtesting characterization Synchronous mapping to actual deviation sequence (s k a k This allows us to determine which side can adequately explain the current deviation.

[0098] First, for each branch canal k, direct paths are filtered according to the rule base. If the channel access flag is set... Then the comparison direction is consistent. or And test the amplitude threshold or When all three conditions are met, an explanation marker is set respectively:

[0099] ;

[0100] ;

[0101] With the weight of the branch canal w k The convergent lateral score is calculated using a confidence reduction factor of 1 / (1+ r up / down Suppressing high uncertainty contributions, we obtain:

[0102] ;

[0103] ;

[0104] like Then D =argmax( S up , S down The process determines the dominant side and outputs "up" or "down"; if the difference is insufficient, it returns "exec" to trigger subsequent execution chain checks. In this way, the direction and magnitude sufficiency of upstream and downstream information flows can be quantitatively compared on the same logical plane, maintaining the consistency of topological causality and avoiding dependence on the global hydraulic model.

[0105] In the description of this invention, the upstream integrated node set includes: the combination of upstream water inflow nodes, the gate set, and the control point set; the downstream integrated node set includes: the downstream key section water level nodes, the gate set, and the control point set.

[0106] Both upstream and downstream driving influence characterization parameters include channel direction markings, magnitude of action, and relative order of action.

[0107] S35. Assemble the control group results into a causal fingerprint, converge the upstream side as the upstream influence intensity, converge the downstream side as the downstream influence intensity, and collect the directional consistency marker set and amplitude interpretable marker set for each branch canal, and simultaneously generate the initial judgment indication of the dominant side.

[0108] Specifically, the control results are assembled into a causal fingerprint Ψ; the upstream side is converged into an influence intensity. S up The intensity of the impact on the downstream convergence. S down And collect the set of directional consistency markers for each branch canal. B dir ={ } and amplitude interpretable label set B amp ={ }, and at the same time give the initial judgment instructions of the dominant side. D ; Use Ψ as the sole input for subsequent lightweight verification, and define the priority action side, constraint range and comparison window W of the upstream or downstream to drive a verifiable direction for a minimum amplitude action.

[0109] S4. Verify the causal fingerprint to determine the dominant position of the upstream and downstream sides. If the evidence is insufficient, output a temporary dominant conclusion and set the reversal condition.

[0110] Specifically, a lightweight verification of causal fingerprints is performed within the same window: when the upstream side's pass count ratio and lateral score both exceed a preset threshold, the hydraulic boundary change is determined to be upstream-dominant; otherwise, when the downstream side meets the same threshold, it is determined to be backwater-dominant; if neither side meets the threshold, a causal comparison is performed on the command sequence, gate position sequence, and corresponding cross-sectional flow sequence of the same gate. If the command change is not in place, is invalid, or has excessive delay, the valve is determined to be not executed properly; if the evidence is insufficient, the side closer to the deviation from the requirement is output as a temporary conclusion and a flip condition is set, while the trial side, the set of tributaries and gates involved, the minimum step size, and the upper limit of the rate are given.

[0111] In the description of this invention, the causal fingerprint is verified to determine the dominant position of the upstream and downstream sides. When the evidence is insufficient, a temporary dominant conclusion is output and a reversal condition is set, including:

[0112] S41. Perform a one-sided verification of the upstream-driven influence characterization parameters, channel interpretation markers, lateral scores, deviation direction markers, and maximum absolute deviations in the causal fingerprint within the observation window. Perform a directional consistency judgment on the channel direction markers and deviation direction markers of each branch canal. After passing the judgment, perform an amplitude sufficiency judgment on the amplitude of the branch canal's influence and the maximum absolute deviation. Use the proportion and lateral score of the branch canals that pass both the directional consistency and amplitude sufficiency judgments in the mandatory branch canal set as a gating mechanism, and compare them with the proportion threshold and score threshold respectively. When both the proportion threshold and score threshold are satisfied, determine that the hydraulic boundary change is upstream-dominated, output a unique conclusion, and bind the event number to the priority action side.

[0113] Specifically, the upstream side of the S3 genetic fingerprint Ψ , , Recorded with S1 s k , a k Perform a one-sided check within window W; for each k ∈B first and s k A consistency judgment is made regarding direction; those who pass are then... and a k To determine the sufficiency of the amplitude, the following methods are used: i amp As an admission criterion; the proportion of those who pass the count to B will be compared with the lateral score. S up Combined as a gating mechanism, setting a proportional threshold. c up With score threshold s up Perform a judgment; when both conditions are met, determine the hydraulic boundary change (upstream dominant), output a unique conclusion, and bind the event ID to the priority action side, record the adjustable side, minimum step size and rate limit of the branch canal set and gate set A involved, only serve S5 once and do not enter the disposal calculation.

[0114] S42. Perform a one-sided verification of the downstream-side and downstream driving influence characterization parameters, channel interpretation markers, lateral scores, deviation direction markers, and maximum absolute deviations of the causal fingerprint within the observation window. Perform a directional consistency judgment on the channel direction markers and deviation direction markers of each branch canal. After passing the judgment, perform an amplitude sufficiency judgment on the amplitude of the branch canal's effect and the maximum absolute deviation. Use the proportion and lateral score of the branch canals that pass both the directional consistency and amplitude sufficiency judgments in the mandatory branch canal set as a gating mechanism, and compare them with the proportion threshold and score threshold respectively. When both the proportion threshold and score threshold are satisfied, determine that the hydraulic boundary change is dominated by the backwater, output a unique conclusion, and record the branch canal set and the priority action side.

[0115] Specifically, if S41 does not produce a conclusion, the downstream side of Ψ will be... , , and s k , a k Perform the same orientation consistency and amplitude sufficiency check within the same window W; for each Will and s k A directional comparison is performed, and those who pass are then... relatively a k according to i amp Amplitude determination will be made; this will be achieved through the percentage of counts and lateral scores. S down As a gate, set up c down and s down The system makes a judgment; when both conditions are met, it determines the change in hydraulic boundary (backwater dominant), outputs a unique conclusion, and records the branch canal set and the priority action side involved. The conclusion is limited to supporting only one minimum amplitude trial of S5 and does not enter the treatment calculation and complex solution.

[0116] S43. Compare the command sequence, gate position sequence, and corresponding cross-sectional flow sequence in the observation window with the same gate for causal analysis. If a command change occurs and the gate position sequence is not within tolerance, or if the gate position sequence is in place but the direction and magnitude of the increase or decrease in the cross-sectional flow sequence do not meet the threshold, and the hysteresis exceeds the threshold, the valve is determined to be not executing properly. If the evidence is insufficient, the side with the lateral score closest to the deviation from the requirement is taken as a temporary conclusion, and the dominant side temporary conclusion is output as the only conclusion, while the reversal condition is recorded simultaneously.

[0117] Specifically, if both S41 and S42 fail, the instruction sequence within window W will be... u i (t ), gate sequence z i ( t ) and corresponding cross-sectional flow sequence f i ( t For the same gate i Perform causal comparison; use rule-based judgment: when an instruction change occurs... z i ( t Not in tolerance d z Internally in place, or z i ( t It is in place and f i ( t The direction and magnitude of the increase or decrease did not reach the threshold. i f And the delay exceeds d τ If the valve is not properly executed, it is determined that the valve is not performing properly. If the evidence is insufficient, the side with the lateral score that is closer to the side that deviates from the requirement is taken as a temporary conclusion. The unique conclusion is output to support the S5 trial, and the flip condition ΔD (including direction regression and minimum observable increase constraint) is recorded simultaneously without expanding the event boundary and calculation range.

[0118] S5. Encapsulate the causal fingerprint verification result into a probe command and bind it to the event number, then execute the adjustment and verification actions according to the leading side. Monitor the regression status of the measured allocation ratio in real time and generate an evidence package.

[0119] Specifically, a one-time test interface and safety barrier are established at the station control layer. The S4 conclusion is encapsulated as a test order and bound to the event number. After passing the checks of whitelist, amplitude and rate not exceeding the limit, observation window coverage of the shortest perceptible lag, and over-limit constraints, a small legal adjustment or gate position check action is performed on the dominant side. Within the observation window, it is determined whether the allocation ratio continues to regress along the target direction or whether the cross-sectional flow and gate position change are causally consistent, forming an evidence package and transmitting it back.

[0120] In the description of this invention, the causal fingerprint verification result is encapsulated as a probe command and bound to an event number, and an adjustment and verification action is executed according to the dominant side. The regression status of the measured allocation ratio is monitored in real time, and an evidence package is generated including:

[0121] S51. Encapsulate the unique conclusion, trial side, target regression direction, allowed action amplitude, rate ceiling, observation window, and limit constraint set into a trial command and bind it to the event number. The shunt node agent and gate agent receive the trial command and verify the action object, amplitude, rate, and observation window item by item. If the verification passes, establish a one-time session using the shunt node agent and gate agent, locking the session attributes that prevent cross-site linkage and middleware processing calculations, and generate a single-step action template and return field placeholders. If any verification fails, reject the session, return the rejection reason, and maintain the current device state.

[0122] Specifically, the existing control unit at the station control layer will be used as the intelligent agent carrier, with the addition of a one-time test interface and a safety barrier; the unique decision D, test side, and target return direction output by S4 will be... Permissible range of motion Θ act upper limit of speed ν max The observation window W and the set of out-of-bounds constraints Ξ are encapsulated into a trial command ΩID and bound to the event number ID; the ΩID is received by the distribution node agent and the gate agent to determine whether the action object belongs to the whitelist. Whether the amplitude and rate do not exceed Θ act With ν max 1. Check each item to see if the observation window covers the shortest perceptible lag; 2. After the check passes, establish a one-time session and lock the session attribute ΥID that does not involve cross-site linkage or middleware processing calculation, and generate a single-step action template ΛID and a return field placeholder ΠID; 3. If any check fails, use a veto mechanism to return the reason for rejection and keep the current state of the device unchanged.

[0123] S52. The change in hydraulic boundary conditions is used as the basis for action. When the upstream side is dominant, the set of gates related to the incoming water is selected for legal adjustment; when the downstream side is dominant, the set of gates related to the backwater is selected for legal adjustment. Within the observation window, the consistency of the measured distribution ratio of each branch canal relative to the target distribution ratio is assessed to ensure continuous regression along the target regression direction. The action time, distribution ratio response trajectory, and whether any exceedances or prohibitions are triggered are recorded item by item to form executable evidence.

[0124] Specifically, the determination of hydraulic boundary changes is used as the basis for action, and the set of gates related to the incoming water is selected when the upstream is dominant. Implement minor, legal adjustments, selecting the backwater-related gate set when the downstream is dominant. Perform small, legal adjustments; use the minimum step size Δz and not exceed Θ. act Execution speed not exceeding ν max And strictly constrained by Ξ; within the observation window W, the measured proportion rk(t) of each branch canal relative to the target proportion is... Whether along The direction is continuously regressed to make a coherent judgment; the action time t is used.act The allocation ratio response trajectory and whether it touches Ξ are recorded item by item to form executable evidence, without proceeding to further processing calculations.

[0125] S53. Using the judgment of incomplete execution as the basis for action, perform a gate position verification action on the relevant gate set. In the same observation window, compare whether the corresponding cross-sectional flow sequence produces the same increase or decrease in gate position sequence. Record the sequential relationship and amplitude matching of the command sequence, gate position sequence and cross-sectional flow sequence item by item to form execution chain evidence.

[0126] Specifically, the decision of incomplete execution will be used as the basis for action, and the relevant gate set will be... Perform a gate position check operation; using a small opening / closing or single step method, the gate position sequence is required to be... zi ( t Reaching the instruction sequence u i ( t And remain stable for a short time; within the same observation window W, for the corresponding cross-sectional flow sequence f i ( t Does it produce the same as z i ( t A causal comparison was made between consistent increases and decreases; u i ( t ), z i ( t ), f i ( t The order and magnitude of each item are recorded one by one, and any obvious delays or blockages are noted to form evidence of the execution chain, which is used to confirm whether the execution chain has been restored.

[0127] S54. Combine the executable evidence and the execution chain evidence into an evidence package. Perform consistency verification based on the event number and causal fingerprint. If the evidence supports the original judgment of a unique conclusion, output the final unique cause and generate an action summary and a response summary. If the evidence does not support the original judgment of a unique conclusion, switch the unique conclusion to the opposite cause according to the flip condition and terminate the trial.

[0128] Specifically, the evidence package formed by S52 or S53 is sent back to the judgment process, where the triage node comparison module performs consistency verification based on ΩID and causal fingerprint Ψ. If the evidence supports the original judgment, the conclusion is immediately confirmed and the final unique cause JID is output, while an action summary ΣID is generated. act With response summary ΣID respIf the evidence does not support the conclusion, the conclusion will be switched to another cause and the trial will be terminated according to the preset flip condition ΔD. At the same time, the trigger condition, action details and response summary will be archived as RID with the event number ID for subsequent fine-tuning of triggering and judgment thresholds.

[0129] S6. Perform consistency verification between the evidence package and the causal fingerprint, determine the evidence support status, output the reasons for the deviation of the diversion ratio, and archive the trigger judgment threshold and constraint rules by event number.

[0130] Specifically, based on the S5 evidence, consistency is checked with the causal fingerprint in the diversion node comparison module. If the evidence supports the original judgment, the final unique cause, action summary, and response summary are output. If the evidence does not support the original judgment, the cause is switched to another cause according to the preset flip condition and the trial is terminated. At the same time, the triggering conditions, action and response trajectories, as well as the over-limit and prohibited action constraints, whitelist and rule base version are archived by event number for subsequent fine-tuning of triggering and judgment thresholds.

[0131] In the description of this invention, the consistency of the evidence package and the causal fingerprint is checked to determine the evidence support status, the cause of the deviation in the diversion ratio is output, and the trigger judgment threshold and constraint rules are archived by event number, including:

[0132] S61. Encapsulate the sole cause and probing evidence into an output packet, using fixed fields and fixed pointers, and bind it to the central distribution node. Simultaneously push and record the response on the distribution node interface. Publish the output packet with a tag consistent with station control and scheduling, and assign a key probing timestamp and observation interval boundary.

[0133] Specifically, the final unique cause JID and response summary ΣID formed by S4 and S5 are... resp Encapsulated as an output packet OID; fields include cause category, dominant side D, tributary set B, trial action object and step size Δz, observation window W, and response summary ΣID. resp Confirm or flip mark b cfm Information on exceeding limits and prohibiting movement (Ξ), and event ID; using fixed fields and explicit pointers, bound to the traffic distribution node (v). c The label is only for selecting the direction of measures and for on-site communication, and does not contain disposal instructions; it is simultaneously pushed to the diversion node interface and the receipt is recorded; the OID is issued with the same label as the station control and dispatch, including the whitelist A of controlled objects. white The explanation of the reason for the movement restriction includes a key timestamp t for the probe. act Clearly define the boundaries of the observation interval, the feedback channels, and the handling methods for timeouts to ensure that the shift leader can select the communication and follow-up direction accordingly.

[0134] S62. Archive the triggering conditions, judgment criteria, and trial evidence for the current event number to generate an archive package. Record the judgment process data in a structured manner and establish a multi-level index using the event number as the primary key to associate and store the archive package with the output package.

[0135] Specifically, the triggering conditions, judgment criteria, and probing evidence for this event will be archived; the triggering time t0, the duration of the deviation, and the direction of the deviation will be recorded. s k Impact determination, causal fingerprint Ψ key fields, lightweight verification conclusion Z, trial action timeline t act Gate position and flow response z i ( t ), f i ( t ), allocation ratio regression trajectory r k ( t )relatively Changes are recorded in a structured manner; the event ID is used as the primary key, and the records are categorized by station and distribution node. c 1. Establish multi-level indexes for branch canals k and solidify the currently effective over-limit and prohibited movement constraints Ξ. 2. Whitelist A white Associate archive packages with OIDs for storage, ensuring they can be retrieved, reviewed, and compared later.

[0136] S63. Use the archived results to adjust the triggering and judgment thresholds, and record each adjustment as a rule change, mark the effective time and applicable diversion node, and associate it with the event number range.

[0137] Specifically, the triggering and judgment thresholds are fine-tuned manually using archived results. When similar events self-recover multiple times and probing becomes unnecessary, the deviation threshold is increased. i r Or extend the minimum duration t min When multiple instances are dominated by a single boundary and can be verified with a small step size Δz, maintain the current Δz and the observation window W; when there are multiple false trigger risks, tighten the action whitelist A. white With the constraint of exceeding the limit Ξ; record each fine-tuning as a rule change U, and mark the effective time and applicable diversion node v. c It is also associated with the event ID range, so that subsequent events enter the judgment process according to the new threshold.

[0138] According to another embodiment of the invention, such as Figure 3 As shown, a deep learning-based system for determining the cause of traffic splitting ratio deviation is also provided. This system includes:

[0139] Event coding generation module 1 is used to compare the target allocation ratio of the branch canal diversion node with the measured allocation ratio of each branch canal in the current scheduling cycle on an hourly basis, filter out branch canals that do not meet the preset judgment rules, and generate event numbers.

[0140] The list export module 2 is used to construct an event graph centered on the diversion node, including directly related upstream water inflow nodes, branch canal outflow nodes, downstream key section water level nodes and gate control points, to establish upstream and downstream information flow paths formed by directed relationships, and to export the node list and path list.

[0141] The causal fingerprint generation module 3 is used to take the node list, path list, target allocation ratio and measured allocation ratio as input, and use graph-based deep learning to perform message passing and local aggregation on the direct path to obtain the influence characterization parameters of the upstream and downstream sides of each branch canal, form the causal fingerprint and calculate the lateral score.

[0142] The dominant determination and evaluation module 4 is used to verify the causal fingerprint and determine the dominant position of the upstream and downstream sides. When the evidence is insufficient, a temporary dominant conclusion is output and a reversal condition is set.

[0143] Evidence encapsulation module 5 is used to encapsulate the causal fingerprint verification result into a probe order and bind it to the event number, and execute adjustment and verification actions according to the leading side. It also monitors the regression status of the measured allocation ratio in real time and generates an evidence package.

[0144] The archiving verification module 6 is used to verify the consistency between the evidence package and the causal fingerprint, determine the evidence support status, output the reasons for the deviation of the diversion ratio, and archive the trigger judgment threshold and constraint rules by event number.

[0145] To facilitate understanding of the above technical solutions of the present invention, the working principle or operation method of the present invention in actual process will be described in detail below.

[0146] During the 2023-2024 irrigation season, a water resources dispatch center conducted a special rectification campaign to address long-term water distribution deviations in the three parallel branch canals at the main canal. The upstream water diversion at this node was affected by the coupling effect of torrential rains in the mountainous area and the opening and closing of sluice gates, while the downstream area was adjacent to a backup urban water source. Previously, relying on manual inspections and threshold alarms, the average time for cause determination exceeded 6 hours, and more than 30% of hydraulic boundary fluctuations were mistakenly attributed to valve jamming, leading to unnecessary on-site maintenance and dispatch oscillations. The dispatch center collaborated with the Water Network Technology Research Institute to embed the method of this invention into a SCADA hybrid platform, performing offline training on historical data from June 2022 to December 2023, and launching real-time inference on March 1, 2024.

[0147] The sensing side integrates 12 flow meters, 8 water level gauges, and 3 electric gate position encoders. The historical window spans 12 months and includes 241 reviewed deviation events, 67% of which originated from sudden increases in upstream water flow or rises in downstream backwater levels, while the remaining 33% were due to inadequate execution. The event graph converges to 27 nodes and 44 edges based on two direct paths: upstream water flow—diversion node—tribute outflow, and downstream water level of the tributary—diversion node. After removing weak causal edges using a rule base, each inference requires only a single round of message passing. The graph network employs two layers of GraphSAGE with gated attention. Node features include five sequences: {Q, W, P, Cmd, Pos}. The loss function simultaneously minimizes the direction consistency judgment error and the amplitude interpretation residual. Offline validation shows that the dominant side's judgment accuracy A... acc =93.8%, F1 score 0.914, a significant improvement compared to the traditional Kalman-PID model's 78.5%.

[0148] Within three months of its launch, the system captured 27 persistent deviations that were confirmed to affect execution. A single cause was identified on average within 345 seconds. Of these, 19 were determined to be changes in hydraulic boundaries, and 8 were due to valve malfunctions. On-site verification results were consistent across all cases. A typical case occurred on April 17, 2024, at 11:32 AM, when a torrential rain upstream caused a 17% instantaneous increase in water inflow. The measured proportion at the diversion node shifted by +0.12 towards branch canal I. The system triggered the event at a threshold δ=0.05 and a duration τ=180 seconds. The lateral score was calculated as follows. S up =0.83, S down =0.21, difference| S up - S down |=0.62>η=0.2, immediately determine upstream dominance; then initiate Δ Pos A small-scale opening test with a margin of +2% was conducted. Within 10 minutes of observation, the proportion regressed to 0.09 along the target direction. The evidence package met the consistency check, and the conclusion was confirmed. The entire closed loop took 7 minutes, which is 88% shorter than the manual process.

[0149] To evaluate the overall benefits, data from three months before and after the 2024 irrigation season were compared: Option 1 was the present invention, and Option 2 was threshold alarm + manual cause investigation. Statistics show that Option 1 had a deviation cause identification accuracy rate of approximately 92.2% (71 / 77); the average response time T... resp=5.4min; the number of additional gate actions was reduced by 41%; the downstream planned water supply fulfillment rate increased from 87.2% to 95.1%. Scheme 2 had an accuracy of only 68.3% and a response time of 38 min. The results demonstrate that by combining a minimum event graph, dual-channel GNN representation, and a small-scale trial, this invention can quickly, with low risk, and verifiably pinpoint the main causes of diversion ratio deviations in actual irrigation district environments, providing reliable support for precise scheduling.

[0150] In summary, by utilizing the above-mentioned technical solution of this invention, this invention constructs a minimum event graph at the diversion node and extracts two types of information flows—upstream drive and downstream pushback—through continuous deviation and dual-gated triggering events affecting compliance, generating a causal fingerprint. Subsequently, through cross-comparison with consistent direction and sufficient amplitude, and lightweight verification, it quickly determines that the deviation mainly originates from changes in hydraulic boundaries or inadequate valve execution, and outputs a unique conclusion. Combined with a small-scale trial with minimum step size, speed limit, and safety barrier constraints, it can be verified within a short observation window, significantly shortening the positioning time, avoiding large-scale trial and error and linked oscillations, reducing the risk of exceeding limits, and improving the downstream plan fulfillment rate and operational stability. By employing proximity and direct convergence boundaries and removing detours or weakly causal edges, the system ensures that information flow does not cross irrelevant branches. Single-round message passing and local aggregation are performed on the direct subgraph, and deep learning models such as graph neural networks can output the direction, magnitude, and relative order of each branch canal, along with the uncertainty. Dual-channel comparison and lateral scoring are completed at the unique comparison interface of the diversion node, which is more accurate, less expensive, and more interpretable than traditional threshold alarms or global simulations. It is also more robust to asynchronous, delayed, and local drift measurements and supports weighting by branch canal weights and rules to suppress noise and occasional interference. In terms of engineering, the method is decoupled from the station control system: whitelists and no-movement constraints, veto power, safety barriers, step size and rate limits, non-centralized handling, and single session locking ensure that the trial is controllable; evidence packets are generated and transmitted back for actions, gate positions, traffic and allocation responses, and the conclusion is automatically reversed if the evidence is inconsistent; all elements of the archived events are used to drive the threshold to slightly adapt, supporting cross-station reuse and online iteration, reducing maintenance and communication costs, improving on-site decision-making efficiency, compliance and reliability, and achieving low-risk, traceable and closed-loop cause determination.

[0151] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for determining the cause of traffic splitting ratio deviation based on deep learning, characterized in that, include: S1. Compare the target allocation ratio of the branch canal diversion node with the measured allocation ratio of each branch canal in the current scheduling cycle on an hourly basis, filter out the branch canals that do not meet the preset judgment rules, and generate event numbers. S2. Construct an event graph centered on the diversion node, including directly related upstream water inflow nodes, branch canal outflow nodes, downstream key section water level nodes and gate control points, establish upstream and downstream information flow paths formed by directed relationships, and export node list and path list. S3. Using the node list, path list, target allocation ratio and measured allocation ratio as input, graph-based deep learning is used to perform message passing and local aggregation on the direct path to obtain the influence characterization parameters of the upstream and downstream sides of each branch canal, form the causal fingerprint and calculate the lateral score. S4. Verify the causal fingerprint of each branch canal, and combine the difference between the lateral score and the score threshold to identify the dominant side of the hydraulic boundary change, and output the determination result of the dominant side of the upstream or downstream. S5. Encapsulate the causal fingerprint verification result into a probe command and bind it to the event number, and execute the adjustment and verification actions according to the dominant side; It also monitors the regression status of the measured allocation ratio in real time and generates an evidence package; S6. Perform consistency verification between the evidence package and the causal fingerprint, determine the evidence support status, output the reasons for the deviation of the diversion ratio, and archive the trigger judgment threshold and constraint rules by event number.

2. The method for determining the cause of traffic splitting ratio deviation based on deep learning according to claim 1, characterized in that, The step of comparing the target allocation ratio of the branch canal diversion nodes with the measured allocation ratio of each branch canal in the current scheduling cycle hourly, filtering out branch canals that do not meet the preset judgment rules, and generating event numbers includes: S11. Using the current sampling time as an index, and based on the joint admission of the preset deviation threshold and the shortest duration, compare the target allocation ratio of the diversion node with the measured allocation ratio of each branch canal. If at least one branch canal in the set of mandatory branch canals deviates in the same direction for more than the shortest duration within the current scheduling cycle, and the absolute value of the deviation is greater than the preset deviation threshold, then the continuous deviation is determined to be valid. Simultaneously record the start time, deviation direction marker, and maximum absolute deviation to generate candidate events. S12. The planned water supply satisfaction and satisfaction threshold of the branch canal, the water level and upper limit of the key section, and the flow rate and flow red line of the control section are compared in turn. When any of the three comparison quantities exceeds the limit in a continuous period, it is determined that the candidate event has affected the downstream execution and a unique event number is generated.

3. The method for determining the cause of traffic splitting ratio deviation based on deep learning according to claim 1, characterized in that, The event graph, centered on the diversion node, incorporates directly related upstream inflow nodes, tributary outflow nodes, downstream key section water level nodes, and gate control points. It establishes directed upstream and downstream information flow paths and derives a node list and a path list, including: S21. Take the diversion node corresponding to the event number as the core diversion node and set the event modeling boundary to construct the event graph; add the upstream inflow node set, the downstream key section water level node of the parallel branch canal, the outflow node of each branch canal, the gate set and control point set related to the distribution to the vertex set of the event graph in sequence, and use the principle of proximity and direct access to converge the boundary of the event graph. S22. Perform directed modeling of the node relationships in the event graph. Set the upstream inflow node set pointing to the central diversion node as the driving direction, set the downstream water level nodes of each branch canal pointing to the central diversion node as the push direction, and set the central diversion node pointing to the outflow nodes of each branch canal as the distribution output. Connect the influence edges of the gate set and control point set that are directly coupled with the distribution output to the direct relationship of the central diversion node or the branch canal diversion nodes. Use the business rule base to filter the directed edge set of the event graph and retain only the edges that directly cause the diversion ratio to deviate. S23. On the event graph, set the upstream and downstream information flow paths and corresponding candidate causes. All direct paths from the upstream inflow node set to each branch canal outflow node through the central diversion node are merged into the upstream driving path set and aggregated into the upstream information flow. Direct paths from the central diversion node to each branch canal outflow node are merged into the downstream backtracking path set and aggregated into the downstream information flow. Direct priority and single convergence point constraints are introduced to avoid the upstream and downstream information flow paths crossing irrelevant branches. S24. Establish a unique reference interface at the distribution node, set the central distribution node as the parallel recording position of the upstream and downstream information flow paths, and synchronously mark the direction and magnitude of the upstream and downstream information flows at the parallel recording position; bind the event number using a fixed-length reference observation window, and export the node list and path list from the event graph.

4. The method for determining the cause of traffic splitting ratio deviation based on deep learning according to claim 1, characterized in that, The process involves taking a node list, a path list, a target allocation ratio, and a measured allocation ratio as inputs, and employing graph-based deep learning to perform message passing and local aggregation on direct paths. This yields the influence characterization parameters for the upstream and downstream sides of each branch canal, forming a causal fingerprint and calculating lateral scores, including: S31. Extract the content belonging to the upstream and downstream information flows from the node list and path list. Using the target allocation ratio and actual allocation ratio of the observation window corresponding to the event number as input, set the central sub-flow node within the analysis scope and set the central sub-flow node as the unique reference position; and use the direct priority strategy to limit the path list to direct paths. S32. Perform message passing and local aggregation on the upstream driving subgraph according to the direct path in the path list, gather the driving effects from the upstream integrated node set at the central diversion node, and generate upstream driving impact characterization parameters for each branch canal. S33. Perform message passing and local aggregation on the downstream driving subgraph according to the direct path in the path list, and gather the back push effect from the downstream integrated node set at the central diversion node to generate downstream back push impact characterization parameters for each branch canal. S34. The upstream driving influence characterization parameters and the downstream retrospective influence characterization parameters are cross-compared in the dual-channel comparison module at the central diversion node. A direct path is adopted and the admission criteria are set with a rule base. Channel admission marks are constructed, and the deviation direction and deviation magnitude of each branch channel are determined to generate channel interpretation marks. Lateral scores are obtained by weighting and primacy factor aggregation. S35. Assemble the control group results into a causal fingerprint, converge the upstream side as the upstream influence intensity, converge the downstream side as the downstream influence intensity, and collect the directional consistency marker set and amplitude interpretable marker set for each branch canal, and simultaneously generate the initial judgment indication of the dominant side.

5. The method for determining the cause of traffic splitting ratio deviation based on deep learning according to claim 4, characterized in that, The upstream integrated node set includes: upstream water inflow node combination, gate set and control point set; the downstream integrated node set includes: downstream key section water level node, gate set and control point set. Both upstream and downstream driving influence characterization parameters include channel direction markings, magnitude of action, and relative order of action.

6. The method for determining the cause of traffic splitting ratio deviation based on deep learning according to claim 4, characterized in that, The formula for calculating the lateral score is as follows: ； In the formula, , Explanation markers for upstream and downstream channels; , Access is granted for direct routes between upstream and downstream areas; Marking for deviation from direction; The maximum absolute deviation; , Marking the upstream and downstream channel directions; , The amplitude of the channel effect between upstream and downstream; , The relative order of upstream and downstream; Weight of branch canals; The amplitude threshold; It is a very small amount; This is a collection of essential branch canals; k Branch canal serial number; , Lateral scores are given for upstream and downstream areas; As the dominant side; up Upstream; down Downstream; exec Verification of the chain to be executed; This is the boundary of the difference.

7. The method for determining the cause of traffic splitting ratio deviation based on deep learning according to claim 1, characterized in that, The process of verifying the causal fingerprints of each branch canal, and combining the difference between the lateral score and the score threshold, identifies the dominant side of the hydraulic boundary change, and outputs the determination result of the dominant side of the upstream or downstream, including: S41. The upstream side of the causal fingerprint and the upstream driving influence characterization parameters, channel interpretation markers, lateral scores, deviation direction markers and maximum absolute deviations are checked unilaterally within the observation window; the channel direction markers and deviation direction markers of each branch canal are judged for directional consistency. After the judgment is passed, the amplitude of the branch canal's effect and the maximum absolute deviation are judged for amplitude sufficiency; the proportion of the branch canals that pass both the directional consistency and amplitude sufficiency judgments in the set of mandatory branch canals and the lateral scores are used as a gating, and are judged against the proportion threshold and the score threshold respectively. When both the proportion threshold and the score threshold are satisfied, the hydraulic boundary change is determined to be upstream-dominated, a unique conclusion is output, and the event number is bound to the priority action side; S42. The downstream side of the causal fingerprint and the downstream driving influence characterization parameters, channel interpretation markers, lateral scores, deviation direction markers and maximum absolute deviations are checked unilaterally in the observation window; the channel direction markers and deviation direction markers of each branch canal are judged for directional consistency. After the judgment is passed, the amplitude of the branch canal's effect and the maximum absolute deviation are judged for amplitude sufficiency; the proportion of the branch canals that pass both the directional consistency and amplitude sufficiency judgments in the set of mandatory branch canals and the lateral scores are used as a gating, and are judged against the proportion threshold and the score threshold respectively. When both the proportion threshold and the score threshold are satisfied, the hydraulic boundary change is determined to be backwater dominant, a unique conclusion is output, and the branch canal set and the priority action side are recorded. S43. Compare the instruction sequence, gate position sequence, and corresponding cross-sectional flow sequence in the observation window with the same gate for causal analysis. If the instruction changes and the gate position sequence is not in place within the tolerance, or if the gate position sequence is in place but the direction and magnitude of the increase or decrease in the cross-sectional flow sequence do not meet the threshold and the hysteresis exceeds the threshold, it is determined that the valve is not executed properly. If the evidence is insufficient, the side with the lateral score closest to the deviation from the requirement is taken as the temporary conclusion, and the dominant side temporary conclusion is output as the only conclusion, and the reversal condition is recorded simultaneously.

8. The method for determining the cause of traffic splitting ratio deviation based on deep learning according to claim 1, characterized in that, The result of the causal fingerprint verification is encapsulated into a probe command and bound to the event number, and the adjustment and verification actions are executed according to the dominant side. It also monitors the regression status of the measured allocation ratio in real time and generates an evidence package including: S51. Encapsulate the unique conclusion, trial side, target regression direction, allowed action range, rate limit, observation window and limit constraint set into a trial command and bind it to the event number. The diversion node agent and the gate agent receive the probe command and verify the action object, amplitude rate and observation window item by item. After the verification is passed, a one-time session is established using the diversion node agent and the gate agent and the session attributes of not cross-site linkage and not middle platform processing calculation are locked. A single-step action template and return field placeholders are generated. If any verification fails, the session is rejected and the rejection reason is returned and the current state of the device is maintained. S52. Use changes in hydraulic boundaries as the basis for action. When the upstream is the dominant side, select the set of gates related to the incoming water to perform legal adjustments. When the downstream is the dominant side, select the set of gates related to the backwater to perform legal adjustments. In the observation window, make a coherent judgment on whether the measured distribution ratio of each branch canal continuously regresses along the target regression direction relative to the target distribution ratio. Record the action time, distribution ratio response trajectory, and whether the information of exceeding the limit and prohibition of action is triggered item by item to form executable evidence. S53. Using the judgment of incomplete execution as the basis for action, perform a gate position verification action on the relevant gate set. Perform a gate position verification action on the relevant gate set, and use the same observation window to compare whether the corresponding cross-sectional flow sequence produces the same increase or decrease in gate position sequence. Record the sequential relationship and amplitude matching of the command sequence, gate position sequence and cross-sectional flow sequence item by item to form execution chain evidence. S54. Combine the executable evidence and the execution chain evidence into an evidence package. Perform consistency verification based on the event number and cause fingerprint. If the evidence supports the original judgment of the unique conclusion, output the final unique cause and generate an action summary and response summary. If the evidence does not support the original judgment of the unique conclusion, switch the unique conclusion to the opposite cause according to the flip condition and terminate the trial.

9. The method for determining the cause of traffic splitting ratio deviation based on deep learning according to claim 1, characterized in that, The process of verifying the consistency between the evidence package and the causal fingerprint, determining the evidence support status, outputting the reasons for deviations in the split ratio, and archiving the trigger judgment threshold and constraint rules by event number includes: S61. Encapsulate the unique cause and probing evidence into an output package, using fixed fields and fixed pointers, and bind it to the central distribution node; synchronously push and record the receipt on the distribution node interface; publish the output package with consistent tags for station control and scheduling, and give it a key probing timestamp and observation interval limit; S62. Archive the triggering conditions, judgment criteria, and trial evidence of the current event number to generate an archive package. Record the judgment process data in a structured manner and establish a multi-level index using the event number as the primary key to associate and store the archive package with the output package. S63. Use the archived results to adjust the triggering and judgment thresholds, and record each adjustment as a rule change, mark the effective time and applicable diversion node, and associate it with the event number range.

10. A deep learning-based system for determining the cause of traffic splitting ratio deviation, used to implement the deep learning-based method for determining the cause of traffic splitting ratio deviation as described in any one of claims 1-9, characterized in that, The system includes: The event coding generation module is used to compare the target allocation ratio of the branch canal diversion nodes with the measured allocation ratio of each branch canal in the current scheduling cycle on an hourly basis, filter out branch canals that do not meet the preset judgment rules, and generate event numbers. The list export module is used to construct an event graph centered on the diversion node, including directly related upstream water inflow nodes, branch canal outflow nodes, downstream key section water level nodes and gate control points, to establish a directed relationship between upstream and downstream information flow paths, and to export the node list and path list. The causal fingerprint generation module is used to take the node list, path list, target allocation ratio and measured allocation ratio as input, and use graph-based deep learning to perform message passing and local aggregation on the direct path to obtain the influence characterization parameters of the upstream and downstream sides of each branch canal, form the causal fingerprint and calculate the lateral score. The dominant determination and evaluation module is used to verify the causal fingerprint of each branch canal, and by combining the difference between the lateral score and the score threshold, it identifies the dominant side of the hydraulic boundary change and outputs the determination result of the dominant side of the upstream or downstream. The evidence encapsulation module is used to encapsulate the causal fingerprint verification results into a probe order and bind it to the event number, execute adjustment and verification actions according to the dominant side, and monitor the regression status of the measured allocation ratio in real time to generate an evidence package. The archiving and verification module is used to verify the consistency between the evidence package and the causal fingerprint, determine the evidence support status, output the reasons for the deviation of the diversion ratio, and archive the trigger judgment threshold and constraint rules by event number.

Images