A spring flow rate prediction method and system based on deep learning

Abstract

The application discloses a spring flow prediction method and system based on deep learning, relates to the technical field of karst hydrology prediction, and comprises the following steps: time correlation is performed on the change of a supply end and the response of a spring mouth in a real-time monitoring correlation table of a spring area, a supply response process is identified, and a water chemical lag relationship is constructed; based on the water chemical lag relationship, the supply path and exchange balance state corresponding to each supply response process are determined, and the supply path state constraint table is formed by combining the non-response evidence to exclude the supply paths which are not established; the control condition of the deep learning prediction channel is determined according to the supply path state constraint table, each supply response channel is adjusted in combination with the supply path state and the exchange direction, and the multi-period prediction sequence of the spring flow in the karst geological direction is generated; and the continuity and stability of the multi-period prediction sequence of the spring flow in the karst geological direction are improved.

Description

Technical Field

[0001] This invention relates to the field of karst hydrological prediction technology, and in particular to a method and system for predicting spring flow based on deep learning. Background Technology

[0002] Karst spring flow is influenced by rainfall infiltration, groundwater level changes, river seepage, and aquifer storage and release processes, exhibiting significant time lags and phases. Current spring flow prediction methods typically collect meteorological, hydrological, and springhead monitoring information, employing recurrent neural networks, convolutional neural networks, and combined prediction methods to extract time-series features, and predicting subsequent spring flow changes by analyzing the correlation between historical spring flow and recharge factors. Some studies also combine conductivity, water temperature, and turbidity analysis to examine the recharge sources and springhead response processes.

[0003] Conventional methods still primarily use hydrochemical information as parallel inputs, making it difficult to identify recharge path switching and the state of old and new water exchange by utilizing the hysteresis direction and recovery order between spring flow and hydrochemical changes. At the same time, different recharge stages are usually processed continuously by a unified prediction structure. When the deviation is concentrated in a certain response stage, it is difficult to locate the corresponding recharge response channel and implement directional correction, which easily weakens the stability of multi-time period predictions. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a deep learning-based method for predicting spring flow to address the problems of insufficient identification of supply path switching and difficulty in directional correction of response stage deviations.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a deep learning-based method for predicting spring flow, comprising: Collect monitoring data of the spring recharge area and spring mouth, establish spatial correlation range, and uniformly match the monitoring data of the recharge area and spring mouth to generate a real-time monitoring correlation table of the spring area; Time correlation was performed between changes in the recharge end and the response at the spring mouth in the real-time monitoring correlation table of the spring area to identify the recharge response process and construct the hydrochemical lag relationship. Based on the water chemistry lag relationship, the recharge path and exchange equilibrium state corresponding to each recharge response process are determined, and the invalid recharge path is excluded by combining the evidence of no response, thus forming a recharge path state constraint table. Based on the supply path state constraint table, the control conditions of the deep learning prediction channel are determined. The supply response channels are adjusted in combination with the supply path state and exchange direction to generate a multi-period prediction sequence of spring flow in karst geological direction. By comparing the differences between the predicted and measured processes at each response stage, the deviation of the multi-period prediction sequence of spring flow in the karst geological direction is verified, and the corresponding replenishment response channels are corrected in a targeted manner to generate a dynamic prediction list of spring flow in the karst geological direction.

[0007] As a preferred embodiment of the deep learning-based spring flow prediction method of the present invention, the monitoring data of the spring recharge area includes rainfall, soil moisture content, groundwater level and river seepage change data. The spring monitoring data includes spring flow rate, conductivity, water temperature, and turbidity.

[0008] As a preferred embodiment of the deep learning-based spring flow prediction method of the present invention, the steps of establishing a spatial correlation range and uniformly matching the monitoring data of the recharge area with the monitoring data of the spring outlet to generate a real-time monitoring correlation table of the spring area are as follows: Based on the monitoring data of the spring recharge area, the spatial correspondence between each spring recharge area and the spring mouth is determined, and the spatial correlation range is established; Within the spatial correlation range, the monitoring data of the spring area and the spring mouth are uniformly matched according to the collection time, monitoring location and corresponding recharge area to generate a real-time monitoring correlation table of the spring area.

[0009] As a preferred embodiment of the deep learning-based spring flow prediction method of the present invention, the steps of temporally correlating the changes in the recharge end and the spring outlet response in the real-time monitoring correlation table of the spring area, identifying the recharge response process, and constructing a hydrochemical hysteresis relationship are as follows: Extract the starting point of supply end change, the starting point of spring mouth response, and the recovery node from the real-time monitoring association table of the spring area, and match the corresponding starting point of spring mouth response and recovery node along the time axis based on the starting point of supply end change to establish the time correspondence relationship within the same supply event. The hydrochemical lag value is calculated based on the hydrochemical linkage trajectory of spring flow. The hydrochemical lag value is combined with the sign, absolute value and recovery order of the hydrochemical lag value to construct a hydrochemical lag relationship that characterizes the switching of replenishment path and the exchange state of old and new water. The changes in spring flow rate during each replenishment response process are time-series paired with changes in conductivity, water temperature and turbidity to form a spring flow-water-chemical linkage trajectory. Based on the rotation direction, closure degree, and recovery sequence of the spring flow hydrochemical linkage trajectory, a hydrochemical lag relationship characterizing the switching of replenishment paths and the exchange state of old and new water is constructed.

[0010] As a preferred embodiment of the deep learning-based spring flow prediction method of the present invention, the steps are as follows: The hydrochemical lag relationship corresponding to each recharge response process is matched with the response criteria of different recharge paths, and the recharge path corresponding to each recharge response process is determined based on the matching status. Based on the sequential changes in the spring flow decay process and the water chemical recovery process, it is determined whether the water volume is released from the substrate to the pipeline or reinjected from the pipeline to the substrate, and the situation where the changes tend to be consistent is identified as the exchange state.

[0011] As a preferred embodiment of the deep learning-based spring flow prediction method of the present invention, the step of combining non-response evidence to exclude invalid supply paths and forming a supply path state constraint table is as follows: For monitoring segments where changes have occurred at the recharge end but the flow rate and water chemistry at the spring outlet have not shown continuous changes within the allowable response period, correlation verification is performed to generate evidence of no response. The evidence of no response is compared with the identified supply paths to exclude supply paths that lack actual response support, while retaining the valid supply paths and exchange states to form a supply path state constraint table.

[0012] As a preferred embodiment of the deep learning-based spring flow prediction method of the present invention, wherein the... Read the supply path status and exchange direction from the supply path status constraint table, determine the activation conditions, input range and transmission direction of each deep learning prediction channel and use them as control conditions, and activate the pipeline fast response channel, fracture transition response channel and matrix slow release channel that match each supply response process according to the control conditions. Based on the activation status of each supply response channel, the short-term changes, medium-term transmission changes, and long-term release changes in the corresponding supply response process are written into the corresponding supply response channel to generate sub-channel supply response data. The sub-channel replenishment response data is input into the corresponding deep learning prediction channel. Combined with the exchange direction constraint on the change transmission direction and duration of each replenishment response channel, a sub-channel spring flow prediction sequence is generated. The spring flow prediction sequence is time-aligned, effective prediction segments are selected based on the activation status of each supply response channel, and the connection relationship between different channels in overlapping time periods is processed to generate a multi-time prediction sequence of spring flow in karst geological direction.

[0013] As a preferred embodiment of the deep learning-based spring flow prediction method of the present invention, the step of verifying the deviation of the multi-time period prediction sequence of spring flow in the karst geological direction by comparing the differences in changes between the prediction process and the measured process at each response stage is as follows: The multi-period predicted sequence of spring flow in karst geological direction was time-aligned with the corresponding measured process, and the deviation verification section was divided according to the onset, enhancement, attenuation and recovery process of the replenishment response; Compare the predicted changes and measured changes within each deviation verification section in terms of occurrence time, direction of change, and duration, and generate a response phase deviation record; The deviation records of the response phase are matched with the prediction time period and control conditions corresponding to each deep learning prediction channel to determine the supplementary response channel that caused the deviation.

[0014] As a preferred embodiment of the deep learning-based spring flow prediction method of the present invention, the directional correction corresponds to the replenishment response channel, generating a dynamic prediction list of spring flow in the karst geological direction, the steps are as follows: Based on the response stage and direction of the deviation, the response timing and change transmission relationship of the corresponding supply response channel are corrected in a targeted manner, and a supply response channel correction record is generated. The multi-period prediction sequence of spring flow in the karst geological direction is updated by using the replenishment response channel correction record, and the changes in spring flow and channel correction status in each prediction period are sorted out to generate a dynamic prediction list of spring flow in the karst geological direction.

[0015] Secondly, the present invention provides a deep learning-based spring flow prediction system, comprising: The spring area monitoring and association module collects monitoring data from the spring replenishment area and spring mouth, establishes a spatial association range, and uniformly matches the monitoring data from the replenishment area and spring mouth to generate a real-time monitoring association table for the spring area. The hysteresis relationship construction module correlates the changes at the recharge end and the response at the spring mouth in the real-time monitoring correlation table of the spring area over time, identifies the recharge response process and constructs the hydrochemical hysteresis relationship. The path status determination module, based on the water chemical lag relationship, determines the recharge path and exchange equilibrium state corresponding to each recharge response process, and excludes invalid recharge paths by combining evidence of no response, thus forming a recharge path status constraint table. The prediction channel control module determines the control conditions of the deep learning prediction channel based on the supply path state constraint table, and adjusts each supply response channel in combination with the supply path state and exchange direction to generate a multi-period prediction sequence of spring flow in the karst geological direction. The dynamic deviation correction module verifies the deviation of the multi-period prediction sequence of spring flow in the karst geological direction by comparing the differences between the predicted process and the measured process at each response stage, and makes directional corrections to the corresponding replenishment response channels, generating a dynamic prediction list of spring flow in the karst geological direction.

[0016] The beneficial effects of this invention are as follows: By constructing a hydrochemical lag relationship, the rotation direction, degree of closure, and recovery sequence in the hydrochemical linkage trajectory of spring flow are transformed into the state of replenishment path switching and exchange of old and new water, providing karst hydrological basis for replenishment path determination and verification of no-response evidence; by controlling the pipeline rapid response channel, fracture transition response channel, and matrix slow release channel according to the replenishment path state constraint table, and by combining the response stage deviation record to directionally correct the corresponding replenishment response channel, each response stage can be predicted and updated in a targeted manner, thereby improving the continuity and stability of the multi-period prediction sequence of spring flow in karst geological direction. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. 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.

[0018] Figure 1 This is a flowchart of a deep learning-based method for predicting spring flow.

[0019] Figure 2 This is a schematic diagram of a deep learning-based spring flow prediction system.

[0020] Figure 3 The flowchart for generating the association table for real-time monitoring of the spring area.

[0021] Figure 4 A flowchart for generating channel control and sub-channel supply response data for deep learning prediction. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0025] Reference Figures 1-4 This is one embodiment of the present invention, which provides a deep learning-based method for predicting spring flow, comprising the following steps: S1. Collect monitoring data of the spring recharge area and spring mouth, establish spatial correlation range, and uniformly match the monitoring data of the recharge area and spring mouth to generate a real-time monitoring correlation table of the spring area.

[0026] The monitoring data for the spring recharge area includes rainfall, soil moisture content, groundwater level, and changes in river seepage.

[0027] The spring monitoring data includes spring flow, conductivity, water temperature and turbidity.

[0028] Based on the monitoring data of the spring recharge area, the spatial correspondence between each spring recharge area and the spring mouth is determined, and the spatial correlation range is established.

[0029] Furthermore, the monitoring locations and types of spring recharge areas are read from the monitoring data of the spring recharge areas. Each monitoring location is mapped to a centralized recharge area, a fissure recharge area, and a decentralized recharge area. The spatial correspondence is determined by combining the spatial distance between each spring recharge area and the spring mouth, the topographic flow direction, and the groundwater connection direction. The corresponding monitoring locations are then connected along the water transfer range from the recharge end to the spring mouth to establish a spatial correlation range.

[0030] It should be noted that the concentrated recharge area refers to the spatial range in which rainfall and surface water enter the karst aquifer through sinkholes, underground river inlets, and surface river seepage sections. The fissure recharge area refers to the spatial range in which rainfall seeps into the karst aquifer along rock fissures and fault zones and is transmitted to springs. The dispersed recharge area refers to the spatial range in which rainfall slowly infiltrates through the soil layer and carbonate rock surface layer, replenishing the karst aquifer in a dispersed manner. The spring recharge area refers to the concentrated recharge area, fissure recharge area, and dispersed recharge area defined based on the distribution of carbonate rocks, fault zones, sinkholes, underground rivers, surface river seepage sections, and tracer connectivity.

[0031] Within the spatial correlation range, the monitoring data of the spring area and the spring mouth are uniformly matched according to the collection time, monitoring location and corresponding recharge area to generate a real-time monitoring correlation table of the spring area.

[0032] Furthermore, within the spatial correlation range, based on the acquisition time of the spring recharge area monitoring data, spring mouth monitoring data with acquisition time differences not exceeding the set synchronization interval are classified into the same time slice. The corresponding spring recharge area is determined according to the boundary range of the monitoring location coordinates falling into the centralized recharge area, fissure recharge area, and decentralized recharge area. Then, based on the connection relationship between the spring recharge area and the spring mouth recorded in the spatial correlation range and the tracer arrival time, the spring recharge area monitoring data in the time slice is paired with the spring mouth monitoring data in the corresponding arrival time. Records that simultaneously meet the requirements of acquisition time correspondence, monitoring location affiliation, and spatial connectivity relationship are written into the same correlation item to generate a spring real-time monitoring correlation table.

[0033] S2. Perform time correlation between the changes in the recharge end and the response at the spring mouth in the real-time monitoring correlation table of the spring area, identify the recharge response process, and construct the hydrochemical lag relationship.

[0034] Extract the starting point of supply end change, the starting point of spring mouth response, and the recovery node from the real-time monitoring association table of the spring area. Then, based on the starting point of supply end change, match the corresponding starting point of spring mouth response and the recovery node along the time axis to establish the time correspondence within the same supply event.

[0035] Furthermore, monitoring data of the spring recharge area and spring outlet were retrieved from the real-time monitoring correlation table for each collection period. Using the range of changes over a continuous and stable period as a reference range, the moment when rainfall, soil moisture content, groundwater level, and river seepage first continuously deviated from the reference range was determined as the starting point of the recharge end change. The moment when the spring outlet flow and at least one hydrochemical monitoring item first showed a continuous corresponding change after the starting point of the recharge end change was determined as the spring outlet response starting point. The moment when the spring outlet flow and hydrochemical monitoring items returned to the reference range after the response change and remained stable was determined as the recovery point. Multiple nodes; taking the starting point of each supply end change as the benchmark, find the earliest spring response starting point that meets the continuous change condition along the time axis within the allowable response period corresponding to the spatial association range, and continue to track from the spring response starting point to the corresponding recovery node. When the new supply end change starting point is earlier than the current recovery node, the new supply end change starting point is merged into the same supply event; associate the matched supply end change starting point, spring response starting point and recovery node according to the event identifier, and record the time interval from the supply end to the spring response and from the spring response to the recovery node to establish the time correspondence within the same supply event.

[0036] The hydrochemical lag value is calculated based on the hydrochemical linkage trajectory of spring flow. The positive and negative signs, absolute values ​​and recovery order of the hydrochemical lag value are combined to construct a hydrochemical lag relationship that characterizes the switching of replenishment paths and the exchange state of old and new water.

[0037] The calculation of water chemical lag value is expressed as follows: ; in, Indicates the first The corresponding water chemical hysteresis values ​​for the water chemical parameters, This indicates the total number of monitoring points involved in the same replenishment response process. This indicates the serial number of the monitoring points arranged according to the collection time. Indicates the first Normalized spring flow rate at each collection point Indicates the first The first data collection time Normalized water chemical parameters, Indicates the first [number] in the same supply response process The first data collection time Normalized water chemical parameters, This represents the normalized spring flow rate at the last data collection moment during the same replenishment response process. This represents the normalized spring flow rate at the first data collection moment during the same replenishment response process. This indicates the last acquisition time in the same supply response process. Normalized water chemical parameters, This indicates the first acquisition time in the same supply response process. Normalized water chemical parameters, Indicates the first [number] in the same supply response process Normalized spring flow rate at each collection point.

[0038] Furthermore, spring flow rate, conductivity, water temperature, and turbidity during the same recharge response process are arranged according to the collection time and normalized within the corresponding recharge response process. For each hydrochemical parameter, the normalized spring flow rate and normalized hydrochemical parameter at adjacent collection times are cross-multiplied and the difference is calculated. The cross-difference values ​​of all adjacent collection times are accumulated, and then the closed cross-difference value between the last collection time and the first collection time is added in. The cumulative value is halved to obtain the corresponding hydrochemical hysteresis value. The rotation direction of the spring flow-hydrochemical linkage trajectory is determined based on the sign of the hydrochemical hysteresis value, and the degree of hysteresis between the spring flow rate change and the hydrochemical change is determined based on the absolute value of the hydrochemical hysteresis value. The recharge response process is divided into a rising phase, a peak phase, a decay phase, and a recovery phase, and the spring flow rate changes in each phase are compared. Based on the sequence of hydrochemical changes, the following classification is used: when spring flow changes first and hydrochemical changes lag behind, the state driven by the previously stored water is marked; when spring flow and hydrochemical changes are synchronized, the state of rapid arrival of new water supply is marked; when spring flow begins to decline while hydrochemical changes continue to shift, the state of continuous mixing of old and new water is marked; when hydrochemical recovery occurs earlier than spring flow recovery, the state of rapid water supply path exiting and transitioning to fracture water supply path or matrix release is marked. Thus, by connecting the changes in water supply path and the state of exchange between old and new water according to the direction, degree, and recovery sequence of each stage, a hydrochemical lag relationship is constructed. The hydrochemical lag relationship reveals the switching of different water supply paths and the state of exchange between old and new water by characterizing the response direction, degree, and recovery sequence between spring flow changes and hydrochemical changes, providing a basis with karst geological significance for the selection and constraint of deep learning prediction channels.

[0039] The changes in spring flow rate during each replenishment response process are time-series paired with changes in conductivity, water temperature, and turbidity to form a spring flow-water-chemical linkage trajectory.

[0040] Furthermore, using the spring flow acquisition time during the same replenishment response process as a benchmark, within the allowable synchronization interval, the conductivity, water temperature, and turbidity monitoring values ​​with the closest time are found respectively. The spring flow changes and corresponding water chemical changes are written into the same acquisition time point, and the paired records of each acquisition time point are continuously arranged according to the acquisition time. The spring flow changes and water chemical changes corresponding to adjacent acquisition time points are connected to form a spring flow-water chemical linkage trajectory. The spring flow-water chemical linkage trajectory refers to the change trajectory formed by temporally pairing and continuously connecting the spring flow changes and corresponding conductivity, water temperature, and turbidity changes at each acquisition time during the same replenishment response process.

[0041] Based on the rotation direction, closure degree, and recovery sequence of the spring flow hydrochemical linkage trajectory, a hydrochemical lag relationship characterizing the switching of replenishment paths and the exchange state of old and new water is constructed.

[0042] Furthermore, the spring flow-water-chemical linkage trajectory is divided into continuous trajectory segments according to the spring flow increase, peak value, decay, and recovery process. The rotation direction of the spring flow-water-chemical linkage trajectory is determined based on the sequential relationship between the changes in spring flow and the changes in conductivity, water temperature, and turbidity in each trajectory segment. The degree of closure is determined based on the proximity of the trajectory endpoint to the starting point and the range enclosed by the trajectory. The order in which spring flow, conductivity, water temperature, and turbidity recover to their stable ranges before the replenishment response is recorded along the time axis. When the spring flow increases rapidly and the water chemical changes intensify synchronously, the rapid pipeline replenishment path is marked as entering the state. When the change in spring flow slows down and the water chemical changes continue to be delayed, the fracture replenishment path is marked as continuing. When the spring flow rate enters a slow decline and the hydrochemical changes gradually recover, the matrix release state is marked. Based on whether the hydrochemical changes occur earlier than, simultaneously with, or later than the spring flow rate changes, the states of priority arrival of new water supply, synchronous mixing of new and old water, or continuous release of previously stored water are determined respectively. The replenishment path state corresponding to each trajectory segment is associated with the exchange state of new and old water in the order of occurrence to construct a hydrochemical lag relationship characterizing the replenishment path switching and the exchange state of new and old water. The hydrochemical lag relationship reveals the replenishment path switching and the exchange state of new and old water by characterizing the rotation direction, closure degree, and recovery order between spring flow rate and hydrochemical changes, providing karst hydrological basis for deep learning prediction channel control.

[0043] It should be noted that the rotation direction refers to whether the linked trajectory of spring flow and hydrochemical changes, formed by connecting them along the collection time, is clockwise or counterclockwise. The degree of closure refers to the proximity of the endpoint and starting point of the linked trajectory and the completeness of the trajectory's enclosed area. The recovery sequence refers to the order in which spring flow, conductivity, water temperature, and turbidity recover to their pre-replenishment stable ranges during the end of the recharge response.

[0044] S3. Based on the water chemistry lag relationship, determine the replenishment path and exchange equilibrium state corresponding to each replenishment response process, and combine the evidence of no response to exclude the invalid replenishment path, forming a replenishment path state constraint table.

[0045] The hydrochemical hysteresis relationship corresponding to each recharge response process is matched with the response criteria of different recharge paths, and the recharge path corresponding to each recharge response process is determined based on the matching status.

[0046] Furthermore, the response criteria for different replenishment paths include: a rapid pipeline replenishment path when spring flow and water chemical changes occur nearly synchronously and the spring flow-water chemical linkage trajectory closes rapidly; a fracture transition replenishment path when spring flow changes precede water chemical changes and there is a continuous interval between their recovery times; and a matrix delayed release path when spring flow slowly decreases and water chemical changes continue to lag and recover with a delay. The response interval, water chemical hysteresis value, degree of closure, and recovery sequence of each replenishment response process are compared with the response criteria for different replenishment paths in turn. Continuous response stages that meet the corresponding criteria are marked as the corresponding replenishment paths. The marks of different stages in the same replenishment response process are connected in the order of occurrence to determine the replenishment path and the replenishment path switching order corresponding to each replenishment response process.

[0047] Based on the sequential changes in the spring flow decay process and the water chemical recovery process, it is determined whether the water volume is released from the substrate to the pipeline or reinjected from the pipeline to the substrate, and the situation where the changes tend to be consistent is determined as the exchange equilibrium state.

[0048] Furthermore, the starting point of spring flow attenuation is compared with the moment when conductivity, water temperature, and turbidity begin to recover to the stable range before the recharge response. When the spring flow remains higher than the level before the recharge response and slowly attenuates, while the water chemical monitoring data has been continuously recovering to the stable range before the recharge response, it is determined that the water volume is released from the substrate to the pipeline. When the spring flow has significantly attenuated, while the water chemical monitoring data still maintains the offset state during the recharge response and recovers with a delay, it is determined that the water volume is reinjected from the pipeline to the substrate. When the spring flow attenuation and water chemical recovery maintain the same direction of change during the continuous collection period, the difference in recovery progress remains within the allowable range, and there is no longer a stable sequential offset, it is determined to be an exchange equilibrium state.

[0049] It should be noted that the spring flow attenuation process refers to the continuous change in spring flow rate after reaching the peak of the recharge response, gradually decreasing and approaching the stable level before the recharge response. The hydrochemical recovery process refers to the continuous change in conductivity, water temperature, and turbidity after the recharge response, gradually returning to the stable range before the recharge response.

[0050] For monitoring segments where changes have occurred at the recharge end but the flow rate and hydrochemistry at the spring outlet have not shown continuous changes within the allowable response period, correlation verification is performed to generate evidence of no response.

[0051] Furthermore, based on the starting point of the change at the recharge end, the flow rate, conductivity, water temperature, and turbidity of the spring are continuously read within the allowable response period corresponding to the spatial correlation range. Monitoring segments in which the flow rate at the spring does not continuously deviate from the stable range before the recharge response and the conductivity, water temperature, and turbidity do not show continuous changes in the same direction are marked as candidate non-response segments. Then, it is checked whether the change at the recharge end meets the duration requirement and the influence of monitoring interruption and abnormal jumps is excluded. The recharge end change record, the corresponding spring monitoring segment, and the allowable response period are correlated to generate evidence of non-response.

[0052] It should be noted that, based on the starting point of the recharge end change, the flow rate, conductivity, water temperature, and turbidity of the spring are continuously read within the allowable response period corresponding to the spatial correlation range. Monitoring segments in which the changes in rainfall, soil moisture content, groundwater level, and river seepage after the starting point of the recharge end change continuously exceed the stable range before the recharge response and last for no less than three collection periods are identified as valid recharge end change segments. Monitoring segments in which the time interval between adjacent collection times exceeds the set collection interval and is accompanied by missing data are marked as monitoring interruption segments. If the change amplitude of a single collection time point is significantly higher than that of the consecutive collection time points before and after, the change at the next collection time point is immediately restored to the original value. Monitoring segments showing a trend and no corresponding changes at adjacent monitoring locations are marked as anomalous jump segments, and monitoring interruption segments and anomalous jump segments are excluded. Within the retained allowable response period, monitoring segments where the spring flow rate does not continuously deviate from the stable range before the recharge response and where conductivity, water temperature, and turbidity do not show continuous unidirectional changes are correlated with the corresponding effective recharge end change segments and spatial correlation range to generate no-response evidence. No-response evidence confirms that an effective change has occurred at the recharge end, but the spring flow rate and water chemistry do not show a corresponding response, excluding recharge paths lacking actual transmission support, and avoiding incorrect learning of ineffective recharge relationships by deep learning predictions.

[0053] The evidence of no response is compared with the identified supply paths to exclude supply paths that lack actual response support, while retaining the valid supply paths and exchange states to form a supply path state constraint table.

[0054] Furthermore, the non-response evidence is matched one by one with the identified replenishment paths according to the replenishment response process identifier, spatial correlation range, and allowable response period. It is verified whether the effective replenishment end change position in the non-response evidence is consistent with the starting point of the replenishment path, and whether the corresponding spring flow and hydrochemical monitoring segments cover the expected response period of the replenishment path. When the effective replenishment end change corresponding to the same replenishment path has been established, but the spring flow and hydrochemical have not shown continuous response, and there are no other response segments that can support the establishment of the replenishment path in the same period, the corresponding replenishment path is marked as lacking actual response support and is excluded. When the spring flow change, hydrochemical change, and exchange state corresponding to the replenishment path have continuous response records, and the non-response evidence does not cover the effective response period of the corresponding replenishment path, the corresponding replenishment path and exchange state are retained, and the excluded state, retained state, and corresponding exchange state are linked and organized to form a replenishment path state constraint table.

[0055] S4. Based on the supply path state constraint table, determine the control conditions of the deep learning prediction channel, and adjust each supply response channel in combination with the supply path state and exchange direction to generate a multi-period prediction sequence of spring flow in karst geological direction.

[0056] Read the supply path status and exchange direction from the supply path status constraint table, determine the activation conditions, input range and transmission direction of each deep learning prediction channel and use them as control conditions, and activate the pipeline fast response channel, fracture transition response channel and matrix slow release channel that match each supply response process according to the control conditions.

[0057] Furthermore, the replenishment path status, exchange direction, retention status, and effective response period corresponding to each replenishment response process are read from the replenishment path status constraint table. The pipeline rapid replenishment path is assigned to the pipeline rapid response channel, the fracture transition replenishment path to the fracture transition response channel, and the matrix lag release path to the matrix slow release channel. The condition for activation is that the replenishment path is in a retained state and the corresponding exchange direction is consistent with the replenishment response process. The input range is determined by the changes in the replenishment end and the spring response related to the corresponding replenishment path within the effective response period. The matrix release to the pipeline is set as the matrix slow release channel transmitting towards the direction of increasing spring flow, and the pipeline to matrix recharge is set as the pipeline rapid response channel transmitting towards the direction of decreasing spring flow. The exchange balance state is set to maintain the current transmission direction. Based on the activation conditions, the pipeline rapid response channel, fracture transition response channel, and matrix slow release channel matching each replenishment response process are opened. The pipeline rapid response channel is used to handle short-term rapid changes caused by concentrated replenishment and to predict the rapid rise and peak response of the spring flow. The fracture transition response channel is used to handle the medium-term delayed changes caused by fracture propagation, predicting the transition of spring flow from a rapid response to a slow release. The matrix slow release channel is used to handle the long-term changes caused by the continuous release of water from the matrix storage, predicting the spring flow decay phase and continuous outflow during the dry season.

[0058] Based on the activation status of each supply response channel, the short-term changes, medium-term transmission changes, and long-term release changes in the corresponding supply response process are written into the corresponding supply response channel to generate sub-channel supply response data.

[0059] Furthermore, the activation status and corresponding input range of each replenishment response channel are read, and monitoring data of the spring replenishment area, spring flow changes, and hydrochemical changes within the effective response period are extracted from the replenishment response process. The continuous monitoring content from the start of the replenishment end change to the rapid rise of the spring flow is written into the pipeline rapid response channel according to the acquisition time. The continuous monitoring content from the end of the rapid rise of the spring flow to the continuous transmission of hydrochemical changes is written into the fracture transition response channel. The continuous monitoring content from the spring flow entering the attenuation stage to the recovery node is written into the matrix slow release channel. The replenishment response process, acquisition time, exchange direction, and effective response period are marked for the continuous monitoring content written into each replenishment response channel. The monitoring content corresponding to the unactivated replenishment response channel is excluded to generate sub-channel replenishment response data.

[0060] It should be noted that short-term changes refer to the rapid increase in spring flow rate within a short period after the initial change at the recharge end, approaching its peak value. Medium-term transmission changes refer to the process where, after the rapid increase in spring flow rate ends, the recharge influence continues to propagate along the fissures, driving further changes in spring flow rate and hydrochemistry. Long-term release changes refer to the process where, after the spring flow rate enters the decay phase, the matrix reservoir continuously releases water into the fissures and karst conduits, maintaining the outflow from the spring.

[0061] The sub-channel replenishment response data is input into the corresponding deep learning prediction channel. By combining the exchange direction constraint with the change transmission direction and duration of each replenishment response channel, a sub-channel spring flow prediction sequence is generated.

[0062] Furthermore, the sub-channel replenishment response data are input into the pipeline fast response channel, the fracture transition response channel, and the matrix slow release channel according to the acquisition time. The pipeline fast response channel extracts short-term surge and peak arrival features, the fracture transition response channel extracts delayed transmission and continuous change features, and the matrix slow release channel extracts attenuation continuation and slow recovery features, and predicts the spring flow changes for the corresponding time periods. When the exchange direction is matrix release to the pipeline, the positive change in spring flow maintained in the matrix slow release channel is retained, and the pipeline fast response channel is limited to further enhance. When the exchange direction is pipeline reinjection to the matrix, the increase in spring flow in the pipeline fast response channel is limited, and the transmission period of the fracture transition response channel is extended. When the exchange direction is in an exchange equilibrium state, the current change direction of each replenishment response channel is maintained. At the same time, the effective response period corresponding to each replenishment response channel is used as the boundary, and the continuous prediction values ​​within the boundary are extracted, while the change transmission outside the boundary is terminated, generating a sub-channel spring flow prediction sequence.

[0063] The deep learning prediction channel consists of an input encoding layer, a causal temporal convolutional layer, a gated recurrent layer, and a spring flow output layer. The input encoding layer converts the sub-channel supply response data into unified temporal features. The causal temporal convolutional layer extracts local changes and delay features using only data prior to the current time. The gated recurrent layer stores the state of supply changes propagating to subsequent time periods. The spring flow output layer maps the temporal state to continuous spring flow prediction values. The pipeline fast response channel uses a short time window and a small convolutional receptive range to focus on retaining the features of sudden increases and peak arrivals in spring flow. The fracture transition response channel uses dilated causal convolution and a longer gated memory to focus on learning the process of delayed propagation of supply effects along the fracture. The matrix slow release channel uses a multi-layer gated recurrent structure and a long-term state retention range to focus on learning the slow decay and continuous release process of spring flow. Thus, the three types of deep learning prediction channels correspond to short-term fast response, medium-term delayed propagation, and long-term slow release, respectively.

[0064] The historical replenishment response process was divided into three paths based on the replenishment path status: rapid pipeline replenishment path, fracture transition replenishment path, and matrix delayed release path. The corresponding short-term changes, medium-term transmission changes, and long-term release changes were then organized into training samples for the rapid pipeline response channel, fracture transition response channel, and slow matrix release channel, respectively. Monitoring data of the spring replenishment area, spring outlet monitoring data, water chemical hysteresis relationships, and exchange directions within a continuous time period were used as inputs. Measured spring flow sequences for subsequent corresponding time periods were used as training labels. Training and validation data were divided according to time sequence, and a sliding time window was used to continuously form the input sequence and training labels. The training samples were sequentially fed into the input encoding layer, causal temporal convolutional layer, gated recurrent layer, and spring flow output layer to learn short-term surges, delayed transmission, and long-term release characteristics, respectively. The deviation between the spring flow prediction sequence output by each deep learning prediction channel and the measured spring flow sequence was calculated. The network parameters were adjusted layer by layer through backpropagation, and training was stopped when the deviation in the validation data remained stable, thus fixing the network parameters of the corresponding deep learning prediction channel.

[0065] It should be noted that the exchange direction refers to the directional state of water flow between the matrix and karst conduits, from the matrix to the conduits, from the conduits back to the matrix, or maintaining an exchange balance. The change transmission direction refers to the direction in which the predicted spring flow changes in each recharge response channel continue in subsequent periods, either increasing, decreasing, or maintaining the current trend.

[0066] The spring flow prediction sequence of each channel is time-aligned, effective prediction segments are selected based on the activation status of each supply response channel, and the connection relationship between different channels in overlapping time periods is processed to generate a multi-time prediction sequence of spring flow in the karst geological direction.

[0067] Furthermore, the spring flow prediction sequences for each channel are uniformly mapped to the same prediction time. Based on the activation status of each supply response channel, prediction segments within the effective response period and whose change direction conforms to the exchange direction constraint are retained, while segments corresponding to inactive channels and extended segments exceeding the effective response period are removed. For prediction segments of different channels within overlapping periods, the dominant segments are determined in the order of rising stage corresponding to the pipeline rapid response channel, transition stage corresponding to the fracture transition response channel, and decay stage corresponding to the matrix slow release channel. The spring flow difference and change direction at adjacent prediction times are compared, and prediction segments with abrupt jumps at the junction are continuously transitioned to ensure that the end of the previous channel and the starting point of the next channel maintain the same change trend, thus generating a multi-period prediction sequence of spring flow in the karst geological direction.

[0068] S5. By comparing the differences between the predicted process and the measured process at each response stage, the deviation of the multi-period prediction sequence of spring flow in the karst geological direction is verified, and the corresponding replenishment response channel is corrected in a targeted manner to generate a dynamic prediction list of spring flow in the karst geological direction.

[0069] The multi-period predicted sequence of spring flow in karst geological directions was time-aligned with the corresponding measured process, and deviation verification sections were divided according to the initiation, enhancement, attenuation and recovery process of the replenishment response.

[0070] Furthermore, the multi-period prediction sequence of spring flow in the karst geological direction is unified with the corresponding measured process to the same acquisition interval and prediction time. Spring flow prediction values ​​with consistent acquisition times are directly paired with measured spring flow values. For records with deviations in acquisition time, the record with the smallest time difference and not exceeding the allowable synchronization interval is selected for pairing. The range of measured spring flow changes during the continuous stable period before the replenishment response is used as the stage division benchmark. The period from the first continuous deviation of measured spring flow from the stable range to the formation of a continuous upward trend is divided into the initial segment. The period from the continuous rise of spring flow to the peak value of this replenishment response is divided into the enhancement segment. The period from the continuous decline of spring flow after exceeding the peak value to the re-entering the stable range boundary is divided into the attenuation segment. The period from the re-entering the stable range boundary and continuously maintaining stability is divided into the recovery segment. The predicted values ​​of the corresponding times in the multi-period prediction sequence of spring flow in the karst geological direction are written into the same segment to form the initial deviation verification segment, enhancement deviation verification segment, attenuation deviation verification segment, and recovery deviation verification segment.

[0071] Compare the predicted changes and measured changes within each deviation verification section in terms of occurrence time, direction of change, and duration, and generate a response phase deviation record.

[0072] Furthermore, within each deviation verification section, based on the moment when the corresponding measured spring flow first begins to change continuously, the time difference between the moment when the predicted spring flow begins to change and the moment when the measured spring flow begins to change is calculated to determine whether the predicted change is early, synchronous, or lagging. Along the continuous prediction time, the direction of increase or decrease of the predicted spring flow value and the measured spring flow value are compared, and continuous periods with the same direction are marked as periods of consistent change, while continuous periods with opposite directions are marked as periods of deviation. The duration of the predicted change and the duration of the measured change are statistically analyzed to determine whether the prediction process has been terminated early, is insufficient, or continues for too long. The corresponding deviation verification section, the time difference of occurrence, the direction of change, and the duration are associated and written into the deviation record of the response stage.

[0073] The deviation records of the response phase are matched with the prediction time period and control conditions corresponding to each deep learning prediction channel to determine the supplementary response channel that caused the deviation.

[0074] Furthermore, the deviation verification section, occurrence time difference, change direction status, and duration status in the deviation record of the response phase are read and matched with the prediction period, activation status, input range, transmission direction, and duration of each deep learning prediction channel. Records whose deviation occurrence time falls within the prediction range of the pipeline fast response channel and whose deviation is concentrated in the initial and enhancement sections are matched to the pipeline fast response channel. Records whose deviation occurrence time falls within the prediction range of the fracture transition response channel and whose deviation is concentrated in the later part of the enhancement section and the earlier part of the attenuation section are matched to the fracture transition response channel. Records whose deviation occurrence time falls within the prediction range of the matrix slow release channel and whose deviation is concentrated in the attenuation and recovery sections are matched to the matrix slow release channel. When multiple deep learning prediction channels overlap in the same time period, the proximity of the deviation start time to the activation time of each deep learning prediction channel is compared, and it is checked whether the deviation direction violates the corresponding transmission direction and whether the deviation duration exceeds the corresponding duration. The deep learning prediction channel that simultaneously satisfies the time overlap, direction correspondence, and duration status correspondence is determined as the replenishment response channel that generated the deviation.

[0075] Based on the response stage and direction of the deviation, the response timing and change propagation relationship of the corresponding supply response channel are corrected in a targeted manner, and a supply response channel correction record is generated.

[0076] Furthermore, the system reads the supply response channel and corresponding response stage that caused the deviation. When the predicted change is earlier than the measured change, the activation time of the supply response channel is adjusted backward and the advance transmission period is shortened. When the predicted change is later than the measured change, the activation time of the supply response channel is adjusted forward and the missing transmission period is supplemented. When the direction of the predicted change is opposite to the direction of the measured change, the corresponding supply response channel is restricted from continuing to transmit in the wrong direction according to the exchange direction, and the subsequent predicted values ​​are adjusted to the same increase or decrease direction as the measured change. When the predicted change is consistently insufficient, the effective response period of the corresponding supply response channel is extended. When the predicted change continues for too long, the predicted segment exceeding the termination time of the measured change is truncated. The activation time, transmission direction, duration, and corresponding response stage before and after the correction are associated and recorded to generate a supply response channel correction record.

[0077] It should be noted that the change transmission relationship refers to the correspondence between the spring flow change in the replenishment response channel and the subsequent prediction period according to the predetermined activation time, increase / decrease direction and duration. For example, after the pipeline rapid response channel is activated, it will transmit the rapid increase trend of spring flow to the peak period, the fracture transition response channel will transmit the delayed change trend to the transition period, and the matrix slow release channel will transmit the slow decay trend to the recovery period.

[0078] The multi-period prediction sequence of spring flow in the karst geological direction is updated by using the replenishment response channel correction record, and the changes in spring flow and channel correction status in each prediction period are sorted out to generate a dynamic prediction list of spring flow in the karst geological direction.

[0079] Furthermore, the correction channel, correction response stage, activation time, transmission direction, and duration are read from the replenishment response channel correction record. The corresponding prediction segment is located in the multi-time prediction sequence of spring flow in the karst geological direction. The corrected activation time is written into the corresponding prediction starting point, and the corrected transmission direction is written into the continuous prediction period. The corresponding prediction segment is extended or truncated according to the corrected duration. The updated prediction segment is time-connected with the adjacent uncorrected prediction segment to eliminate spring flow jumps at the connection point and retain the prediction segment without deviation. The updated spring flow change, corresponding replenishment response channel, correction response stage, and channel correction status are recorded sequentially according to the prediction time to generate a dynamic prediction list of spring flow in the karst geological direction.

[0080] This embodiment also provides a deep learning-based spring flow prediction system, including: The spring area monitoring and association module collects monitoring data from the spring replenishment area and spring mouth, establishes a spatial association range, and uniformly matches the monitoring data from the replenishment area and spring mouth to generate a real-time monitoring association table for the spring area. The hysteresis relationship construction module correlates the changes at the recharge end and the response at the spring mouth in the real-time monitoring correlation table of the spring area over time, identifies the recharge response process and constructs the hydrochemical hysteresis relationship. The path status determination module, based on the water chemical lag relationship, determines the recharge path and exchange equilibrium state corresponding to each recharge response process, and excludes invalid recharge paths by combining evidence of no response, thus forming a recharge path status constraint table. The prediction channel control module determines the control conditions of the deep learning prediction channel based on the supply path state constraint table, and adjusts each supply response channel in combination with the supply path state and exchange direction to generate a multi-period prediction sequence of spring flow in the karst geological direction. The dynamic deviation correction module verifies the deviation of the multi-period prediction sequence of spring flow in the karst geological direction by comparing the differences between the predicted process and the measured process at each response stage, and makes directional corrections to the corresponding replenishment response channels, generating a dynamic prediction list of spring flow in the karst geological direction.

[0081] In summary, this invention constructs a hydrochemical lag relationship, transforming the rotation direction, closure degree, and recovery sequence in the hydrochemical linkage trajectory of spring flow into the state of replenishment path switching and exchange of old and new water, providing karst hydrological basis for replenishment path determination and verification of no-response evidence; by controlling the pipeline rapid response channel, fracture transition response channel, and matrix slow release channel according to the replenishment path state constraint table, and by combining the response stage deviation record to directionally correct the corresponding replenishment response channel, each response stage can be predicted and updated in a targeted manner, improving the continuity and stability of the multi-period prediction sequence of spring flow in karst geological directions.

[0082] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A deep learning-based method for predicting spring flow, characterized in that, include: Collect monitoring data of the spring recharge area and spring mouth, establish spatial correlation range, and uniformly match the monitoring data of the recharge area and spring mouth to generate a real-time monitoring correlation table of the spring area; Time correlation was performed between changes in the recharge end and the response at the spring mouth in the real-time monitoring correlation table of the spring area to identify the recharge response process and construct the hydrochemical lag relationship. Based on the water chemistry lag relationship, the recharge path and exchange equilibrium state corresponding to each recharge response process are determined, and the invalid recharge path is excluded by combining the evidence of no response, thus forming a recharge path state constraint table. Based on the supply path state constraint table, the control conditions of the deep learning prediction channel are determined. The supply response channels are adjusted in combination with the supply path state and exchange direction to generate a multi-period prediction sequence of spring flow in karst geological direction. By comparing the differences in changes between the predicted and measured processes at each response stage, the deviation of the multi-period prediction sequence of spring flow in the karst geological direction is verified, and the corresponding replenishment response channels are corrected in a targeted manner to generate a dynamic prediction list of spring flow in the karst geological direction.

2. The deep learning-based spring flow prediction method as described in claim 1, characterized in that: The monitoring data for the spring recharge area includes rainfall, soil moisture content, groundwater level, and river seepage changes. The spring monitoring data includes spring flow rate, conductivity, water temperature, and turbidity.

3. The deep learning-based spring flow prediction method as described in claim 2, characterized in that, The steps for establishing the spatial correlation range and uniformly matching the monitoring data of the replenishment area with the monitoring data of the spring mouth to generate a real-time monitoring correlation table for the spring area are as follows: Based on the monitoring data of the spring recharge area, the spatial correspondence between each spring recharge area and the spring mouth is determined, and the spatial correlation range is established; Within the spatial correlation range, the monitoring data of the spring area and the spring mouth are uniformly matched according to the collection time, monitoring location and corresponding recharge area to generate a real-time monitoring correlation table of the spring area.

4. The deep learning-based spring flow prediction method as described in claim 1, characterized in that, The steps for correlating changes at the recharge end with the spring mouth response in the real-time monitoring correlation table of the spring area, identifying the recharge response process, and constructing hydrochemical lag relationships are as follows: Extract the starting point of supply end change, the starting point of spring mouth response, and the recovery node from the real-time monitoring association table of the spring area, and match the corresponding starting point of spring mouth response and recovery node along the time axis based on the starting point of supply end change to establish the time correspondence relationship within the same supply event. The hydrochemical lag value is calculated based on the hydrochemical linkage trajectory of spring flow. The hydrochemical lag value is combined with the sign, absolute value and recovery order of the hydrochemical lag value to construct a hydrochemical lag relationship that characterizes the switching of replenishment path and the exchange state of old and new water. The changes in spring flow rate during each replenishment response process are time-series paired with changes in conductivity, water temperature and turbidity to form a spring flow-water-chemical linkage trajectory. Based on the rotation direction, closure degree, and recovery sequence of the spring flow hydrochemical linkage trajectory, a hydrochemical lag relationship characterizing the switching of replenishment paths and the exchange state of old and new water is constructed.

5. The deep learning-based spring flow prediction method as described in claim 1, characterized in that, The steps are as follows: The hydrochemical lag relationship corresponding to each recharge response process is matched with the response criteria of different recharge paths, and the recharge path corresponding to each recharge response process is determined based on the matching status. Based on the sequential changes in the spring flow decay process and the water chemical recovery process, it is determined whether the water volume is released from the substrate to the pipeline or reinjected from the pipeline to the substrate, and the situation where the changes tend to be consistent is identified as the exchange state.

6. The deep learning-based spring flow prediction method as described in claim 5, characterized in that, The steps for combining non-responding evidence to exclude invalid supply paths and forming a supply path state constraint table are as follows: For monitoring segments where changes have occurred at the recharge end but the flow rate and water chemistry at the spring outlet have not shown continuous changes within the allowable response period, correlation verification is performed to generate evidence of no response. The evidence of no response is compared with the identified supply paths to exclude supply paths that lack actual response support, while retaining the valid supply paths and exchange states to form a supply path state constraint table.

7. The deep learning-based spring flow prediction method as described in claim 1, characterized in that, The Read the supply path status and exchange direction from the supply path status constraint table, determine the activation conditions, input range and transmission direction of each deep learning prediction channel and use them as control conditions, and activate the pipeline fast response channel, fracture transition response channel and matrix slow release channel that match each supply response process according to the control conditions. Based on the activation status of each supply response channel, the short-term changes, medium-term transmission changes, and long-term release changes in the corresponding supply response process are written into the corresponding supply response channel to generate sub-channel supply response data. The sub-channel replenishment response data is input into the corresponding deep learning prediction channel. Combined with the exchange direction constraint on the change propagation direction and duration of each replenishment response channel, a sub-channel spring flow prediction sequence is generated. The spring flow prediction sequence of each channel is time-aligned, effective prediction segments are selected based on the activation status of each supply response channel, and the connection relationship between different channels in overlapping time periods is processed to generate a multi-time prediction sequence of spring flow in the karst geological direction.

8. The deep learning-based spring flow prediction method as described in claim 1, characterized in that, The deviation verification of the multi-time period prediction sequence of spring flow in karst geology is carried out by comparing the differences between the predicted process and the measured process at each response stage. The steps are as follows: The multi-period predicted sequence of spring flow in karst geological direction was time-aligned with the corresponding measured process, and the deviation verification section was divided according to the onset, enhancement, attenuation and recovery process of the replenishment response; Compare the predicted changes and measured changes within each deviation verification section in terms of occurrence time, direction of change, and duration, and generate a response phase deviation record; The deviation records of the response phase are matched with the prediction time period and control conditions corresponding to each deep learning prediction channel to determine the supplementary response channel that caused the deviation.

9. The deep learning-based spring flow prediction method as described in claim 8, characterized in that, The targeted correction corresponds to the replenishment response channel, generating a dynamic prediction list of spring flow in the karst geological direction. The steps are as follows: Based on the response stage and direction of the deviation, the response timing and change propagation relationship of the corresponding supply response channel are corrected in a targeted manner, and a supply response channel correction record is generated. The multi-period prediction sequence of spring flow in the karst geological direction is updated by using the replenishment response channel correction record, and the changes in spring flow and channel correction status in each prediction period are sorted out to generate a dynamic prediction list of spring flow in the karst geological direction.

10. A deep learning-based spring flow prediction system, based on the deep learning-based spring flow prediction method according to any one of claims 1 to 9, characterized in that, include: The spring area monitoring and association module collects monitoring data from the spring replenishment area and spring mouth, establishes a spatial association range, and uniformly matches the monitoring data from the replenishment area and spring mouth to generate a real-time monitoring association table for the spring area. The hysteresis relationship construction module correlates the changes at the recharge end and the response at the spring mouth in the real-time monitoring correlation table of the spring area over time, identifies the recharge response process and constructs the hydrochemical hysteresis relationship. The path status determination module, based on the water chemical lag relationship, determines the recharge path and exchange equilibrium state corresponding to each recharge response process, and excludes invalid recharge paths by combining evidence of no response, thus forming a recharge path status constraint table. The prediction channel control module determines the control conditions of the deep learning prediction channel based on the supply path state constraint table, and adjusts each supply response channel in combination with the supply path state and exchange direction to generate a multi-period prediction sequence of spring flow in the karst geological direction. The dynamic deviation correction module verifies the deviation of the multi-period prediction sequence of spring flow in the karst geological direction by comparing the differences between the predicted process and the measured process at each response stage, and makes directional corrections to the corresponding replenishment response channels, generating a dynamic prediction list of spring flow in the karst geological direction.

Images