A Smart Prediction System for Groundwater Storage Variables Based on Multi-Source Data Fusion and Parameter Optimization

Abstract

This invention discloses an intelligent prediction system for groundwater storage variables based on multi-source data fusion and parameter optimization. The invention relates to the field of groundwater storage change prediction technology. The system includes a data acquisition module connected to an IoT hydrological sensor, a satellite remote sensing data interface, and a meteorological data interface. This intelligent prediction system for groundwater storage variables based on multi-source data fusion and parameter optimization solves the prediction lag problem caused by data update delays, inconsistent data spatiotemporal scales, and parameter rigidity in groundwater storage change assessment systems by constructing a dual-loop collaborative architecture of a real-time assimilation loop and a parameter evolution loop. Utilizing high-frequency assimilation and fusion of multi-source real-time data ensures that the model input is synchronized with the actual environmental state; enabling the prediction system to continuously track the transient processes of the groundwater system, effectively improving the timeliness and reliability of groundwater storage change predictions under scenarios such as extreme rainfall, extreme drought, and high-intensity mining.

Description

Technical Field

[0001] This invention relates to the field of groundwater prediction technology, specifically to an intelligent prediction system for groundwater storage variables based on multi-source data fusion and parameter optimization. Background Technology

[0002] Groundwater storage refers to the volume of gravity water accumulated in aquifers during the long-term circulation of groundwater. Predicting changes in groundwater storage is crucial for water resource management and scientific development. Existing prediction and evaluation systems typically employ multi-factor coupling methods, comprehensively considering influencing factors such as hydrogeology, meteorological conditions, and human activities to construct predictive models. However, these systems face significant bottlenecks in practical applications: relying on periodically collected static data and preset fixed parameters, the models struggle to respond in real-time to dynamic changes in groundwater storage. Key factors such as rainfall, groundwater extraction, and evaporation fluctuate constantly, but traditional models are outdated, and the spatiotemporal scales of the multi-source data required by the models differ, resulting in a time lag between the input data and the actual situation. Furthermore, core parameters such as infiltration coefficient, permeability coefficient, specific yield, and release coefficient have shifted under environmental changes, and the system lacks an automated mechanism for real-time calibration, causing the model's state to diverge further from the actual system. This lack of dynamic response capability results in significant time lag in predictions, especially in scenarios involving extreme precipitation, extreme drought, or sudden changes in mining intensity. Lagging information cannot support accurate emergency decision-making and may even trigger secondary risks. While existing technologies continuously optimize factor selection and coupling algorithms, they have not fundamentally solved the problem of the rigidity of model data and parameters in the time dimension. The current technical challenge is to achieve highly timely dynamic data assimilation and adaptive updating of model parameters to eliminate the time lag in prediction results. Summary of the Invention

[0003] To achieve the above objectives, the present invention provides the following technical solution: an intelligent prediction system for groundwater storage variables based on multi-source data fusion and parameter optimization, comprising:

[0004] The data acquisition module connects to IoT hydrological sensors, satellite remote sensing data interfaces, and meteorological data interfaces;

[0005] The real-time assimilation loop is connected to the data acquisition module for high-frequency fusion of ground observation data and multi-source satellite remote sensing data to solve the problems of inconsistent spatiotemporal scales and time delays in data. The multi-source observation data includes GRACE gravity satellite data, precipitation and evapotranspiration remote sensing data, meteorological data, surface hydrological data, aquifer data, and groundwater resource development and utilization data at different spatiotemporal scales.

[0006] The parameter evolution loop is deeply coupled with the real-time assimilation loop, and its input end receives the assimilation data and prediction error data output by the real-time assimilation loop.

[0007] The dual verification module has its inputs connected to the outputs of the real-time assimilation loop and the parameter evolution loop, respectively.

[0008] A cross-cycle feedback bus provides bidirectional connection between the real-time assimilation loop and the parameter evolution loop.

[0009] The output of the parameter evolution loop is fed back to the parameter input of the real-time assimilation loop, and the verification result of the dual verification module triggers the iterative optimization of the real-time assimilation loop or the parameter evolution loop.

[0010] Preferably, the real-time assimilation loop includes:

[0011] The dynamic weight allocation unit adjusts the coupling weights of hydrological factors, meteorological factors, and human mining activity factors in real time based on sensor accuracy level, data timestamps, and environmental change indicators.

[0012] The Kalman filter algorithm unit performs assimilation calculations at different time scales on the weighted multi-source data and outputs predicted groundwater storage values.

[0013] Preferably, the environmental abrupt change indicators include the rainfall intensity change rate, the continuous drought index, and the pumping rate abrupt change threshold.

[0014] Preferably, the parameter evolution loop includes:

[0015] The parameter dynamic modeling unit defines the infiltration coefficient, permeability coefficient, specific yield, release coefficient and water-bearing parameters as functions of spatiotemporal coordinates, and associates them with the pumping accumulation function and drought index function.

[0016] The error tracing unit identifies parameters to be optimized whose contribution rate exceeds a preset threshold based on the prediction error at different time scales of the real-time assimilation loop.

[0017] The hybrid optimization unit adopts a fusion framework of genetic algorithm and particle swarm optimization algorithm to update the parameters to be optimized under the constraint of the groundwater continuity equation.

[0018] Preferably, the hybrid optimization unit sets physical constraint boundaries, including aquifer boundaries, aquifer boundary and thickness limits, infiltration coefficient, permeability coefficient, specific yield, release coefficient and other hydrogeological parameters within reasonable ranges.

[0019] Preferably, the dual verification module includes:

[0020] The physical conservation verification unit verifies the prediction results through the groundwater mass conservation equation. If the mass balance error exceeds the threshold, it sends a data verification command to the real-time assimilation loop.

[0021] The historical scenario verification unit compares the dynamic trend similarity between the real-time prediction curve and the historical curve of the same scenario. If the similarity is lower than the preset standard, it sends a deep optimization instruction to the parameter evolution loop.

[0022] Preferably, the dynamic trend similarity is calculated using a dynamic time warping algorithm.

[0023] Preferably, the cross-cycle feedback bus performs:

[0024] Real-time assimilation of loop parameters, evolutionary loop transport of highly sensitive factor identifiers and prediction error distribution;

[0025] Parameter evolution loop to real-time assimilation loop feedback parameter mapping relationship table.

[0026] Preferably, the error tracing unit works in conjunction with the physical conservation verification unit: when the physical conservation verification fails, the preset threshold of the error tracing unit is automatically adjusted.

[0027] Preferably, the final predicted groundwater storage value should simultaneously meet the following requirements:

[0028] Groundwater mass conservation verification through physical conservation verification unit;

[0029] Verification is performed by validating the dynamic trend similarity of the historical scenario verification unit.

[0030] This invention provides an intelligent prediction system for groundwater storage variables based on multi-source data fusion and parameter optimization. It has the following beneficial effects:

[0031] This loosely porous groundwater storage change prediction system, based on the fusion of data assimilation and parameter optimization, solves the evaluation and prediction lag problem caused by data update delays and parameter rigidity in groundwater storage change prediction systems by constructing a dual-loop collaborative architecture of a real-time assimilation loop and a parameter evolution loop, combined with dynamic weight allocation and environmental mutation response mechanisms. Utilizing high-frequency assimilation and fusion of multi-source data at different spatiotemporal scales ensures that the model input is synchronized with the real environmental state; parameter drift tracking trees and geological constraint optimization enable dynamic adaptive updates of core parameters such as infiltration coefficient, permeability coefficient, specific yield, release coefficient, and water abundance, allowing the prediction system to continuously track the transient processes of the groundwater system and effectively improve the timeliness and reliability of groundwater storage change predictions under scenarios such as heavy rainfall, extreme drought, and high-intensity mining.

[0032] This system for predicting changes in loose porous groundwater storage, based on the fusion of data assimilation and parameter optimization, ensures the scientific validity and robustness of the prediction results through a dual verification mechanism and deep embedding of geological rules. A physical conservation sieve intercepts physically infeasible solutions using differentiated tolerance rules based on aquifer type; a historical scenario simulator accurately identifies model behavior anomalies through segmented weighted analysis of geological feature-driven curves; and a four-level conflict decision engine initiates targeted responses based on geological priority principles. Ultimately, this forms a closed-loop control chain encompassing data assimilation, parameter evolution, verification interception, and conflict self-healing, providing highly reliable dynamic prediction support for water resource management. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the module interaction of an intelligent prediction system for groundwater storage variables based on multi-source data fusion and parameter optimization according to the present invention.

[0034] Figure 2 This is a schematic diagram of the method flow of an intelligent prediction system for groundwater storage variables based on multi-source data fusion and parameter optimization according to the present invention. Detailed Implementation

[0035] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0036] Please see Figure 1 and Figure 2 This invention provides a technical solution: an intelligent prediction system for groundwater storage variables based on multi-source data fusion and parameter optimization, comprising:

[0037] The data acquisition module connects to IoT hydrological sensors, satellite remote sensing data interfaces, and meteorological data interfaces;

[0038] The real-time assimilation loop is connected to the data acquisition module for high-frequency fusion of multi-source observation data to solve the problem of inconsistent spatiotemporal scales of data. The multi-source observation data includes GRACE gravity satellite data, precipitation and evapotranspiration remote sensing data, meteorological data, surface hydrological data, aquifer data, groundwater level, water quality and development and utilization data at different spatiotemporal scales.

[0039] The parameter evolution loop is nested and deeply coupled to the real-time assimilation loop, and its input receives the prediction error data output by the real-time assimilation loop.

[0040] The dual verification module has its inputs connected to the outputs of the real-time assimilation loop and the parameter evolution loop, respectively.

[0041] A cross-cycle feedback bus provides bidirectional connection between the real-time assimilation loop and the parameter evolution loop.

[0042] The output of the parameter evolution loop is fed back to the parameter input of the real-time assimilation loop, and the verification result of the dual verification module triggers the iterative optimization of the real-time assimilation loop or the parameter evolution loop.

[0043] It should be further explained that during the specific implementation process, upon system startup, the data acquisition module continuously receives groundwater level data, water quality data, and water resource development and utilization data uploaded by IoT hydrological sensors. Simultaneously, it accesses GRACE gravity water storage change data, precipitation, evapotranspiration, and soil moisture content data retrieved from satellite remote sensing, and real-time rainfall forecasts from meteorological departments. The real-time assimilation loop operates at a ten-day frequency: the dynamic weight allocation unit first assesses the reliability of each data source; when the sensor accuracy level is lower than a preset standard, the contribution value of that factor is automatically reduced. Simultaneously, it detects environmental abrupt change indicators; if precipitation intensity, drought index exceed critical thresholds, or pumping rate suddenly becomes abnormal, the coupling weights of meteorological and human activity factors are immediately increased. The Kalman filter algorithm unit, after dynamic weight adjustment of the multi-source data input set, generates ten-day predicted values ​​for groundwater storage changes.

[0044] The parameter evolution loop is deeply coupled with the real-time assimilation loop. The parameter dynamic modeling unit defines parameters such as infiltration coefficient, permeability coefficient, specific yield, release coefficient, and water-bearing capacity as continuous functions of spatial coordinates and time, with their rate of change correlated with the derivative of the pumping accumulation function. When the prediction error of the real-time assimilation loop output exceeds the tolerance limit, the error tracing unit automatically analyzes the contribution distribution of each parameter to the error and locks the two key parameters with the highest contribution rates, such as infiltration coefficient, permeability coefficient, or specific yield. The hybrid optimization unit initiates a genetic-particle hybrid algorithm to iteratively fine-tune the locked parameters under the hard constraint that the difference between the left and right terms of the groundwater continuity equation does not exceed the groundwater mass balance.

[0045] A cross-cycle feedback bus facilitates collaboration between the two loops: the real-time assimilation loop transmits highly sensitive factor identifiers, such as coordinates of sudden increases in pumping rates, to the parameter evolution loop; the parameter evolution loop, in turn, feeds back the optimized parameter mapping table to the parameter input of the real-time assimilation loop. Dual verification modules synchronously monitor the entire process: the physical conservation verification unit calculates the difference between recharge and discharge in real time; if the deviation from the change in groundwater storage exceeds the allowable relative equilibrium difference threshold, it immediately sends a data verification command to the real-time assimilation loop; the historical scenario verification unit retrieves similar scenario curves from the historical database, and when the dynamic time warping distance between the real-time prediction curve and the historical similar scenario curve exceeds the similarity threshold, it forces the parameter evolution loop to activate deep optimization mode. The final prediction result is only output after simultaneously passing both physical conservation verification and historical trend matching verification.

[0046] The dynamic response mechanism is as follows: when the monitored values ​​of rainfall intensity, drought index, or groundwater extraction change rate increase beyond the critical slope within three consecutive time windows, the system automatically adjusts the weight of meteorological factors or human activity factors to the highest level and triggers the high-frequency data assimilation mode; only when the parameter contribution rate ranks in the top three for two consecutive rounds and the error has not converged is it determined to be a parameter to be optimized, thus avoiding invalid calculations.

[0047] When the quality balance check fails, the system traces back to the starting time point of the error mutation, clears all assimilation data thereafter, and reinitializes the model; when the dynamic trend similarity check fails, the parameter evolution loop switches to reinforcement learning mode and uses the historical best parameter combination as the initial population of the hybrid algorithm.

[0048] The high-sensitivity factor identifiers received by the parameter evolution loop are used to construct the spatial priority region of the parameter dynamic modeling unit; the real-time assimilation loop, based on the feedback parameter mapping table, prioritizes the correction of the factor weights of the high-sensitivity region in the next assimilation cycle.

[0049] The data acquisition module only performs standardized data transmission, the Kalman filter algorithm adopts a conventional implementation process, and the final prediction result output interface is a general data protocol.

[0050] The real-time assimilation loop includes:

[0051] The dynamic weight allocation unit adjusts the coupling weights of hydrological factors, meteorological factors, and human mining activity factors in real time based on sensor accuracy level, data timestamps, and environmental change indicators.

[0052] The Kalman filter algorithm unit performs ten-day assimilation calculations on the weighted multi-source data and outputs predicted values ​​of groundwater storage changes.

[0053] It should be further explained that, in the specific implementation process, after the real-time assimilation loop is activated, the dynamic weight allocation unit continuously monitors the status of three types of factors, including hydrological factors, meteorological factors, and human activity factors. Specifically: hydrological factors are matched with pre-set signal levels based on the accuracy certificate number of the IoT sensors; if the device calibration expires or the signal is interrupted, the weight of this factor automatically drops to the lowest level; meteorological factors are judged based on data timestamps for timeliness, with the weight of meteorological data not updated for more than one hour decreasing linearly, while the weight of real-time radar data in areas under red rainstorm warnings is increased to 1.5 times the baseline value; human activity factors are dynamically adjusted in conjunction with environmental abrupt change indicators, including: when the monitored value of the pumping rate abrupt change threshold increases by more than three times the standard deviation of the historical average within one hour, the weight of the human activity factor is immediately locked at the maximum value and an early warning is triggered. Hydrological factors include surface water runoff and irrigation volume, groundwater level, and water quality; meteorological factors include rainfall and evaporation; and human activity factors include pumping rate, groundwater extraction, and irrigation volume.

[0054] The Kalman filter algorithm unit receives weighted factors and performs assimilation operations: it extrapolates the current distribution of storage changes based on the previous day's hydrogeological model; it fuses the weighted multi-source observation data, prioritizing the use of high-weight factors to correct prediction biases; if environmental abrupt change indicators remain active, it reduces the iteration step size of the model error covariance matrix to improve sensitivity. The assimilation results output predicted groundwater storage changes every ten days. When the weighted allocation unit triggers the highest-level warning, the assimilation frequency automatically switches to once per hour.

[0055] The equipment fault tolerance mechanism includes: if GRACE gravity satellite data and precipitation-evapotranspiration remote sensing data are interrupted, the gradient replacement scheme will be automatically activated: the first choice is to access the measured values ​​of precipitation and ground evaporation from nearby meteorological stations; the second choice is to call the historical evaporation pattern database for the same date; if there is no replacement data, the meteorological factor weights will be forcibly reset to zero and the confidence level of the prediction results will be downgraded.

[0056] The ensemble Kalman filter employs a domain-based parallel computing strategy, dividing the study area into several subdomains according to hydrogeological units. Each subdomain is assimilated independently and then concatenated using boundary flow equations. When an environmental abrupt change index in a subdomain exceeds the standard, assimilation is initiated only for that region, while the remaining regions continue to be processed on a ten-day basis.

[0057] Emergency intervention for covariance: When updating the model error covariance matrix, if it is detected that the weight allocation unit increases the weight of the same factor three times in a row, the covariance component corresponding to that factor will be automatically compressed to 1 / 10 of its original value, forcing the model to respond quickly to changes in the dominant factor.

[0058] The state transition equation of the Kalman filter is based on the standard groundwater movement theory; the output prediction format is compatible with conventional water resource management platforms.

[0059] Overcoming the limitations of fixed weights or manually set rules, a triple-control weight system based on equipment status, data timeliness, and environmental mutations is established. This system enables autonomous and intelligent adjustment of factor contribution values ​​and dynamic reconstruction of factor weight structures in the event of data anomalies, ensuring prediction reliability. Domain-specific parallel computing addresses the bottleneck of high-frequency computation over large areas, improving efficiency. Emergency covariance intervention directly correlates with weight change trends, ensuring that model sensitivity is strictly matched to the intensity of environmental mutations.

[0060] Environmental abrupt change indicators include the rate of change in rainfall intensity, the rate of change in evaporation, and the threshold for abrupt changes in pumping rate. It should be further explained that, in the specific implementation process, the system scans these three types of environmental abrupt change indicators in real time and executes differentiated responses: the monitoring of the rate of change in rainfall intensity and the rate of change in evaporation uses a sliding time window calculation. When the difference between rainfall or evaporation in adjacent one-hour intervals exceeds one-fifth of the historical extreme value for the same period, it is determined to be a Level 1 abrupt change. At this time, the weight of meteorological factors is automatically increased to the upper limit value, while the covariance update step size of the compressed ensemble Kalman filter is reduced to one-tenth of the normal value. If the abrupt change intensity increases continuously for three time windows, an emergency calibration command for geological parameters is sent to the parameter evolution loop.

[0061] The threshold for sudden changes in pumping rate is determined by incorporating spatial distribution characteristics: In densely pumped industrial areas, when the pumping rate of a single well increases by more than three standard deviations from the average of the previous hour, and more than five wells within a three-kilometer radius exhibit synchronous anomalies, the highest weight level of the human activity factor is activated. In agricultural irrigation areas, the total pumping volume of the area is monitored; if the daily pumping volume accounts for more than a preset empirical coefficient, the irrigation behavior prediction model is immediately corrected.

[0062] During the linkage control of mutation indicators, when two or more mutation indicators are activated simultaneously, the system arbitrates according to priority, as follows:

[0063] Abrupt changes in pumping rate take precedence over abrupt changes in rainfall and evaporation.

[0064] If the duration of sudden changes in rainfall or evaporation exceeds the critical duration, it will be upgraded to the highest level.

[0065] During arbitration, the real-time assimilation loop enters buffer mode, using the parameter set from the previous effective cycle to maintain the prediction.

[0066] During the adaptive threshold adjustment period, the threshold for determining abrupt changes is not a fixed value: the baseline value of the rate of change of rainfall intensity is automatically lowered during the rainy season and raised during the dry season; the baseline value of the standard deviation of pumping rate fluctuates dynamically with the seasonal water use pattern.

[0067] The spatial cascade response of pumping mutation includes: generating a first-level response circle at the coordinates of the mutation well point: with a radius of 1 km, and human activity factors are weighted and topped within the circle; a second-level response circle: with a radius of 5 km, and the weights decay linearly with distance; when the mutation well group forms a spatial cluster, the response circles are automatically merged, and the weights at the boundary are taken as the maximum superposition value.

[0068] The parameter evolution cycle includes:

[0069] The parameter dynamic modeling unit defines parameters such as infiltration coefficient, permeability coefficient, specific yield, release coefficient, and water abundance as functions of spatiotemporal coordinates, and associates them with the pumping accumulation function and drought index function.

[0070] The error tracing unit identifies parameters to be optimized whose contribution rate exceeds a preset threshold based on the prediction error of the real-time assimilation loop every ten days.

[0071] The hybrid optimization unit adopts a fusion framework of genetic algorithm and particle swarm optimization algorithm to update the parameters to be optimized under the constraint of the groundwater continuity equation.

[0072] It should be further explained that, in the specific implementation process, the parameter dynamic modeling unit defines the infiltration coefficient as a continuous function of three-dimensional spatial coordinates and vadose zone lithology and groundwater level depth, dynamically corrected by changes in vadose zone lithology and groundwater level depth; it defines permeability coefficient and water-bearing capacity as continuous functions of three-dimensional spatial coordinates and aquifer lithology, with their rate of change dynamically corrected by pumping volume and water level change functions. When the total pumping volume in a certain area exceeds the quarterly warning line, the permeability coefficient and water-bearing capacity functions of that area are automatically adjusted by an added adjustment factor; the specific yield is associated with aquifer lithology and drought index function, and an exponential decay term is introduced when the meteorological drought level rises to severe. All parameter functions have preset reasonable range boundaries, such as the permeability coefficient not being lower than the empirical minimum value for clay layers.

[0073] The error tracing unit is activated once every ten days, immediately after the ten-day forecast is completed: First, the difference between the predicted reserves and the actual monitored reserves in the real-time assimilation loop is calculated. If the maximum regional error exceeds one-tenth of the aquifer storage, parameter contribution rate analysis is performed. The perturbation method is used to fine-tune each parameter value sequentially and the prediction error is recalculated. Only when adjusting a parameter individually reduces the total error by more than a preset critical value is it included in the parameter set to be optimized. Parameters locked in all three consecutive rounds of analysis are marked as high-sensitivity factors.

[0074] When the hybrid optimization unit starts, a genetic algorithm generates an initial parameter population: each individual represents a spatial distribution pattern of infiltration coefficient, permeability coefficient, and specific yield. Particle swarm optimization then intervenes, iterating using the difference between the left and right terms of the groundwater continuity equation as the fitness function. Key constraints include:

[0075] The change in reserves during a single iteration must not exceed the maximum regulation capacity of the aquifer;

[0076] The parameter update range is limited by the geological parameter database;

[0077] During the optimization process, the status of the physical conservation verification unit is monitored in real time. If an alarm is triggered, the system immediately reverts to the previous valid parameter group. After the optimization results are verified by the parameter drift tracking tree unit, they are updated to the real-time assimilation loop.

[0078] Dynamic parameter modeling, including: pumping cumulative decay mechanism, including:

[0079] Industrial Zone Response: When the monthly pumping volume exceeds the historical peak, functions such as permeability coefficient, specific yield, and release coefficient will have a negative exponential correction term added within a one-kilometer radius of the pumping point; if excessive pumping continues for three months, the attenuation range will expand to three kilometers and be superimposed with a linear attenuation layer.

[0080] Agricultural zone response: When the cumulative pumping volume during the irrigation season exceeds the estimated value of the crop water requirement model, functions such as permeability coefficient, water supply, and water release coefficient are embedded with a step-by-step attenuation factor in the irrigation zone; during the fallow period, the attenuation is automatically removed and the natural recovery model is activated.

[0081] Drought index linkage rules: When the meteorological drought level reaches the orange alert level, the shallow aquifer permeability coefficient and water yield function are reduced by 0.7 times; when the alert level is raised to red, the reduction coefficient is reduced to 0.5 and the cross-aquifer recharge correction term is triggered; after the alert is lifted, the gradual recovery algorithm is started, and the original value is restored to one-tenth every day until it returns to normal.

[0082] During the hybrid optimization constraint period, the geological history constraint database is applied, including: if the parameter update value exceeds the fluctuation range of the measured value within the local ten years, it needs to be backtested through the historical scenario verification unit: if the match is successful, the special geological event marker is unlocked; if the match fails, it is forced to revert to the historical average.

[0083] The real-time conservation monitoring interruption mechanism is as follows: when the physical conservation verification unit issues an alarm during the optimization process, the current iteration is immediately paused; the spatiotemporal overlap between the alarm parameters and the optimization parameters is compared; when the overlap area exceeds 60% of the optimization domain, the current population is cleared and reinitialized; when the overlap is less than 30%, only the parameter individuals in the alarm area are removed.

[0084] Among them, the basic parameter monitoring adopts conventional groundwater level measurement equipment, the drought level data comes from the public interface of the meteorological department, and the initial population generation of the genetic algorithm uses a standard random method.

[0085] Overcoming the limitations of static parameter tables, a dynamic response chain is established, encompassing the decline in pumping accumulation, permeability coefficient, specific yield, and release coefficient, as well as the decline in drought index and specific yield, enabling the parameter model to adapt to environmental conditions. A historical constraint database comparison mechanism is used to transform geological historical data into hard boundaries for parameter optimization. Physical conservation verification directly intervenes in the optimization process, forming a closed-loop control chain of calculation, verification, interruption, and repair.

[0086] The hybrid optimization unit sets physical constraint boundaries, including aquifer boundaries, thickness limits, and reasonable ranges for hydrogeological parameters such as permeability coefficient, specific yield, and release coefficient.

[0087] It should be further explained that, in the specific implementation process, when the hybrid optimization unit performs parameter updates, physical constraint boundaries are embedded as insurmountable hard conditions throughout the entire iteration process: the aquifer boundary is determined based on the geological exploration report, the aquifer thickness limit value adopts a three-level threshold control, the base value is set based on the geological exploration report, the safety threshold is 0.8 times the base value, and the critical threshold is 0.6 times the base value. When the optimization candidate solution causes the calculated aquifer thickness in a certain area to be lower than the critical threshold, multi-source data verification is automatically triggered: the latest geophysical inversion data is called first for verification, and if geophysical data is missing, the most unfavorable value of the historical dry season is used for comparison. Only when the verification confirms the lower limit of the allowable thickness of the current geological structure, the candidate solution is retained for the next iteration; otherwise, the optimization process is immediately aborted and reverted to the previous generation parameter set.

[0088] Dynamic calibration is implemented to constrain the reasonable range of hydrogeological parameters such as infiltration coefficient, permeability coefficient, water yield, and water release coefficient: conventional sedimentary strata are matched with preset ranges according to lithology type, such as the permeability coefficient of sand and gravel should not exceed the daily per 100-meter level; real-time inversion verification is added in areas with strong human activity. If the sudden change in the unit drawdown and inflow of the pumping well group points to parameter anomalies, the optimization is immediately frozen and the field exploration command is initiated.

[0089] Adaptive adjustment of constraint boundaries under special geological scenarios, including the following:

[0090] Extreme rainfall period: When the daily rainfall exceeds the once-in-a-century value, the lower limit constraint on aquifer thickness is temporarily lifted;

[0091] During the dynamic management of constraints, the three-level circuit breaker mechanism for aquifer thickness includes:

[0092] Warning level, i.e. when the thickness reaches the safety threshold: send geological uncertainty markers to the real-time assimilation ring, start the adjacent area thickness compensation algorithm, and balance the reserve distribution within a five-kilometer radius.

[0093] Critical level, i.e. when the thickness exceeds the critical threshold: Forcefully terminate the current parameter optimization and call the historical scenario verification unit to backtrack similar events within the past ten years;

[0094] When no matching cases are found, a manual intervention protocol is initiated.

[0095] Anti-fraud verification of parameters such as permeability coefficient, specific yield, and release coefficient: When the optimization candidate solution causes the single increase of parameters such as permeability coefficient, specific yield, and release coefficient to exceed the historical maximum jump value: if there is no reasonable cause, it is judged as an optimization fraud solution and automatically eliminated; three consecutive fraud solutions trigger the parameter evolution loop self-checking program.

[0096] It affects the maintenance of regular constraints outside the domain.

[0097] The basic geological exploration data comes from publicly available databases of geological departments.

[0098] By overcoming the limitations of fixed boundaries, a three-level threshold-multi-source verification-conditional circuit breaker intelligent control chain is established to address the risk of the optimization process deviating from geological reality. An anti-fraud verification mechanism is used to identify and intercept physically infeasible parameter jumps, ensuring the scientific validity of the optimization. Real-time linkage between torrential rain, persistent drought, and constraint boundaries enables the system to adapt to environmental changes.

[0099] The two-factor authentication module includes:

[0100] The physical conservation verification unit verifies the prediction results through the groundwater mass conservation equation. If the mass balance error exceeds the threshold, it sends a data verification command to the real-time assimilation loop.

[0101] The historical scenario verification unit compares the dynamic trend similarity between the real-time prediction curve and the historical curve of the same scenario. If the similarity is lower than the preset standard, it sends a deep optimization instruction to the parameter evolution loop.

[0102] It should be further explained that, in the specific implementation process, the physical conservation verification unit executes the groundwater mass conservation equation once every ten days, synchronized with the prediction cycle: it calculates the difference between regional recharge and discharge in real time, comparing changes in groundwater storage; where regional recharge includes rainfall infiltration, irrigation infiltration, surface water infiltration, and lateral runoff, and discharge includes pumping consumption, evaporation, and downstream discharge. If the deviation exceeds one-tenth of the current aquifer storage, a Level III response is immediately initiated.

[0103] Level 1 deviation, i.e., tolerance of 1-1.5 times: send a local data verification command to the real-time assimilation loop, clear the assimilation data of the most recent hour and reinitialize the model boundary conditions;

[0104] Secondary deviation, i.e., tolerance 1.5-2 times: freeze the optimization process of parameter evolution loop, and call geophysical equipment to perform real-time scanning of the anomaly area;

[0105] Level 3 deviation, i.e., tolerance of more than 2 times: the entire system enters a safe mode, and the current forecast is covered by the historical average parameter set of the same season.

[0106] The historical scenario verification unit extracts historical scenarios from the geological archive that match the current meteorological conditions and mining intensity, with a matching degree of over 80%. A dynamic time warping algorithm is used to quantify the morphological similarity between the real-time prediction curve and the historical curve. When the similarity is below a preset threshold: if there is an active environmental mutation event in the real-time assimilation loop, a re-examination is delayed by 1 hour; if there is no mutation event, a deep optimization instruction is immediately sent to the parameter evolution loop, activating the reinforcement learning module and restarting optimization with the historically optimal parameter combination as the initial population; if two consecutive re-examinations fail to meet the standard, a forced switch to the backup prediction model is initiated, and a geological verification work order is generated.

[0107] The physical conservation three-level circuit breaker includes a dynamic tolerance mechanism for water supply: the tolerance threshold is automatically raised to 1.2 times that of the dry season during the rainy season to cope with dynamic changes in the unsaturated zone; when the pumping intensity in the human activity area exceeds the design value, the tolerance threshold is lowered to 0.7 times the normal value.

[0108] The replenishment-discharge balance diagnostic tree includes: when the conservation check fails, the system performs cause tracing: priority is given to detecting the discharge end: if the pumping rate change exceeds three times the standard deviation, it is marked as a human cause; then the replenishment end is checked: if the rainfall data is more than 1 hour old and there is no change event, it is judged as data distortion; if there is no clear cause, the emergency tracing module of the parameter evolution loop is activated.

[0109] During the application of dynamic time warping enhancement, the prediction curve is divided into dry season, wet season and sudden event segments according to the inflection point, and the similarity is calculated for each segment; double weight is given to the sudden drop in water level; when the similarity of the sudden drop segment is less than 60% of the overall threshold, deep optimization is directly triggered.

[0110] Deep optimization is triggered by gradients, including: primary optimization: adjusting the genetic mutation probability of the parameter evolution loop to 3 times the normal value; intermediate optimization: resetting the initial position of the particle swarm to the neighborhood of the historical best solution; and ultimate optimization: launching a cross-aquifer coupling model to integrate precipitation-surface water-groundwater joint analysis.

[0111] The basic conservation equation adopts the industry standard water balance formula, historical scenario matching calls public hydrological databases, and the backup prediction model is a simplified version of neural network.

[0112] By breaking through the limitations of fixed tolerance thresholds, a dynamic tolerance system is established, which includes corrections for permeability coefficient, water supply, and water release coefficient, as well as suppression of human activities and fluctuations during the rainy season, to ensure that the verification results conform to geological realities. A three-level circuit breaker response chain is used, consisting of data reset at level one, optimization by freezing at level two, and system rollback at level three, to form a gradient safety barrier.

[0113] Dynamic trend similarity is calculated using the dynamic time warping algorithm. It should be further explained that, in the specific implementation process, when the historical scenario verification unit performs curve comparison, the dynamic time warping algorithm first identifies the key turning points between the real-time predicted curve and the historical curve: it captures the starting point of a sudden drop in water level based on a first-order difference threshold, and verifies the authenticity of the turning point by combining geological event markers; these geological event markers include the start date and stop time of pumping wells, groundwater level, and water quality changes. The curve is divided into three segments: a stable period, a gradual change period, and a sudden change period, with double the calculation weight assigned to the sudden change period. When the similarity of the sudden change period is less than 60% of the overall threshold, the parameter evolution loop depth optimization is directly triggered.

[0114] The segmented comparison rule base is configured based on groundwater level and drawdown cone, including the following:

[0115] The key monitoring areas are the water level and the shape of the drawdown funnel in the pumping-affected zone, requiring that the slope of the drawdown funnel be at least 80% similar.

[0116] If the overall curve similarity meets the standard but there are differences in the mutation period, initiate the causal analysis decision tree, including the following:

[0117] Human activity triggers: Examine whether there are new sources of interference in current human activity factors;

[0118] Causes of data distortion: Verify whether sensor data corresponding to the period of sudden change is missing or abnormal;

[0119] Geological variation triggers: When there are no clear external factors, generate a warning report of suspected geological structural changes.

[0120] Geological features drive segmentation: The criteria for determining a sudden drop in water level in three stages are as follows:

[0121] A normal sudden drop, i.e., the slope is greater than 1 times the historical average, triggers the regular weight calculation and associates the operation logs of pumping wells within a five-kilometer radius.

[0122] Abnormal sudden drop, i.e., slope > twice the historical average: automatically increase weight to three times; forcibly start the water well operation scanning mode with a radius of ten kilometers.

[0123] A catastrophic drop, i.e., a slope greater than three times the historical average: pause curve comparison; directly activate cross-departmental emergency response protocols.

[0124] The specific tolerance rules for aquifers include the following:

[0125] Falling funnel slope tolerance: Tolerance = Basic tolerance × (1 - 0.2 × Pumping well density index);

[0126] During the in-depth diagnosis of the underlying cause, the following are included:

[0127] Novel interference source identification technology: Compare the current and historical human activity factor feature matrices: mark yellow interference sources when the expanded irrigation area exceeds twice the historical area; exempt the area from similarity requirements during abrupt change when a red interference source is located within ten kilometers of the center of the water level funnel; and launch a special impact assessment model.

[0128] Multiple sources of geological variation are confirmed: the curve difference area is spatially superimposed with the following data: underground engineering excavation warning circle; ground subsidence area; when two or more data overlap, a high-risk geological anomaly alarm is generated.

[0129] Among them, the basic inflection point detection adopts the standard difference algorithm, the historical curve library is connected to the national groundwater monitoring engineering database, and the emergency protocol follows the water resources management regulations.

[0130] By overcoming the limitations of homogenized comparison, a segmented verification system based on aquifer structure, characteristic parameters, and tolerance rules is established to solve the problem of misjudgment caused by geological diversity. By utilizing a three-level response mechanism for sudden drops, the curve morphology is directly linked to emergency management.

[0131] The cross-cycle feedback bus includes:

[0132] Real-time assimilation of loop parameters, evolutionary loop transport of highly sensitive factor identifiers and short-term prediction error distribution;

[0133] Parameter evolution loop to real-time assimilation loop feedback parameter mapping relationship table.

[0134] It should be further explained that, in the specific implementation process, a bidirectional channel is established between the real-time assimilation loop and the parameter evolution loop via the cross-cycle feedback bus: During uplink transmission, the real-time assimilation loop packages high-sensitivity factor identifiers and short-term prediction error distributions hourly. High-sensitivity factors are graded and marked according to their spatial influence range. Locally sensitive factors, i.e., influence radius ≤ 1 km, are transmitted using discrete point coordinates; regionally sensitive factors, i.e., influence radius > 5 km, are supplemented with spatial contour vector data. The prediction error distribution is compressed by hydrogeological unit partitioning to avoid the mean masking local abrupt changes.

[0135] During downlink feedback, the parameter mapping table generated by the parameter evolution loop is intelligently simplified, including: only transmitting parameters whose changes in the last two rounds of optimization exceed the historical fluctuation extremes; such as transmitting continuous surface interpolation coefficients for porous aquifers; and automatically attaching parameter reliability labels to areas with strong human activity, including: Level A: drilling test verification, Level B: geophysical inversion, and Level C: algorithm estimation.

[0136] The bus has a built-in geological feature routing engine, including: under normal conditions, data flows through a standard verification channel, and the transmission delay is controlled within 1 hour; during periods of sudden environmental changes, such as a red rainstorm warning: the priority routing protocol is activated to allocate a dedicated channel for meteorological factor-related data; during the conservation verification alarm period: the bus is forced to enter read-only mode, and the downlink feedback of parameters is frozen until physical balance is restored.

[0137] The sensitivity factor spatial clustering algorithm includes a three-level spatial encapsulation strategy, specifically as follows:

[0138] Point-based sensitive sources, i.e., sudden changes in pumping speed in a single well: transmission precise coordinates + influence radius; radius calculation correlates with aquifer permeability coefficient: R = base radius × (1 + 0.5 × K / K) 基准 );

[0139] Linear sensitive zone, namely: river seepage section: extract the buffer zone extending to both sides of the river centerline; the buffer zone width is dynamically adjusted according to the thickness of riverbed sediment.

[0140] The parameter mapping table dynamic compression includes a change amplitude threshold linkage mechanism: when the parameter evolution loop is in deep optimization mode, the amplitude threshold is automatically reduced to 50% of the normal value.

[0141] The reliability label decision tree includes A-level labels and C-level labels; where:

[0142] Grade A labels must meet the following requirements: parameter update values ​​are verified by on-site pumping tests; the deviation from the geophysical inversion results is less than the allowable error; and the solution has not been marked as fraudulent for three consecutive rounds of optimization.

[0143] C-level label triggering conditions: optimized candidate solutions exceed the historical fluctuation range; the area is not covered by monitoring wells; geological uncertainty warnings are automatically attached during transmission.

[0144] The basic data transmission uses a standard communication protocol; the spatial contour lines are generated using a general GIS algorithm; and network bandwidth management relies on the QoS function of commercial routers.

[0145] Overcoming the limitations of homogeneous transmission, a three-level encapsulation system (point, line, and surface) plus a dynamic routing system based on geological features is established, thereby improving data transmission efficiency. A reliability tagging system is used to provide downstream modules with a basis for parameter confidence decisions. Deep optimization and threshold linkage prevent the omission of critical parameters.

[0146] The collaboration between the error tracing unit and the physical conservation verification unit includes: when the physical conservation verification fails, the preset threshold of the error tracing unit is automatically lowered. It should be further explained that, in the specific implementation process, the error tracing unit and the physical conservation verification unit establish a real-time collaborative channel: when the physical conservation verification unit detects that the mass balance deviation exceeds the threshold, it immediately sends the geological anomaly coordinates and deviation magnitude to the error tracing unit. Based on this, the error tracing unit dynamically lowers the preset threshold. When the deviation is within the first-level tolerance range, the preset threshold is reduced to four-fifths of its original value; in the second-level tolerance range, it is reduced to three-fifths; and in the third-level tolerance range, threshold filtering is directly disabled, forcing the entire parameter set to participate in optimization.

[0147] The spatial focusing strategy is executed during the collaboration process, including:

[0148] Geological anomaly zone identification: A priority analysis circle with a radius of three kilometers is generated centered on the peak point of the conservation check deviation, and the weight of monitoring well data within the circle is increased to three times;

[0149] Parameter correlation recalibration: The weight of the error contribution rate calculation for hydrogeological parameters within the priority circle is doubled, and the weight of parameters outside the priority circle is halved;

[0150] Cross-period data linkage: Call historical conservatism verification records. If the current abnormal area overlaps with the historical problem area by more than 70%, automatically match the historical optimization scheme as the initial solution.

[0151] Special geological scenarios trigger synergistic enhancements, including the following:

[0152] During the red alert period for heavy rain: the conservation verification threshold will be temporarily relaxed to 1.5 times the normal value to avoid frequent false alarms;

[0153] Pumping beyond design capacity: The preset threshold for error contribution rate is lowered by two levels to strengthen the monitoring of human activity factors.

[0154] The tolerance interval-threshold linkage matrix satisfies the following table:

[0155]

[0156] The matching rules for the historical issues area are as follows:

[0157] Spatial matching: Activated when the overlap rate between the current anomaly area and the historical polygon database is greater than 70%;

[0158] Trigger matching: The dominant factors, such as pumping / rainfall, are of the same type;

[0159] Conditions for reusing the scheme: The historical optimization scheme has been validated for more than three months; there has been no new geological or engineering interference recently.

[0160] Special scenario synergistic enhancement: The weight of the error contribution rate of parameters such as permeability coefficient, water supply, and water release coefficient within the target area is increased to five times; the physical conservation verification is temporarily exempted from the balance requirement outside the target area.

[0161] Emergency response to excessive pumping: When the pumping volume in a region exceeds the design value by 1.5 times: the error tracing threshold is lowered by two levels; the contribution rate of human activity factors is forcibly increased by one level; if the optimization fails to converge after three consecutive rounds, an over-extraction hazard warning report is generated.

[0162] Among them, the basic spatial analysis adopts GIS overlay technology, and the historical scheme database is stored in a standard database format.

[0163] By breaking through the limitations of independent operation modules, a closed-loop response chain of conservation deviation, source tracing threshold, and spatial focus is established, enabling analysis and judgment to be accurately applied to the core area of ​​the problem. The dual matching of space and cause ensures the reliability of solution reuse.

[0164] The final predicted value of groundwater storage change must simultaneously meet the following requirements:

[0165] Groundwater mass conservation verification through physical conservation verification unit;

[0166] Verification is performed by validating the dynamic trend similarity of the historical scenario verification unit.

[0167] It should be further explained that, during the specific implementation process, before the final prediction result is output, the system performs physical conservation verification and historical trend verification in parallel: the physical conservation verification unit dynamically adjusts the tolerance threshold according to the aquifer type. For example, for porous aquifers, a linear tolerance is adopted, meaning that the difference between recharge and discharge does not exceed one-tenth of the storage change. When the verification fails, spatial focusing backtracking is initiated: an influence domain with a radius of three kilometers is generated centered on the deviation peak point, the assimilation data of the most recent hour within the domain is cleared, and the calculation is recalculated.

[0168] The historical trend verification unit performs aquifer differentiation matching: for example, in the pore region, the similarity of the entire curve must be greater than or equal to a preset threshold. If verification fails, it automatically extracts three sets of neighboring parameters from the historical best solution, injects them into a parameter evolution loop, and restarts the optimization process.

[0169] When the double validation results conflict, the four-level decision tree is activated, and the process is as follows:

[0170] Level 1 conflict, i.e., physical verification failed + historical verification passed: physical backtracking is executed first;

[0171] Historical verification results are converted into reference recommendations; prediction results are marked as "geological uncertainty".

[0172] Secondary conflict, i.e., physical verification passes + historical verification fails: freeze the current parameter set; call the best parameter coverage of the same historical scenario; start the reinforcement learning module to trace the cause of the difference.

[0173] Level 3 conflict refers to: minor deviation from double verification: narrowing the analysis time and space scope to the anomaly area; on-site verification using high-precision geophysical equipment; selectively adopting verification conclusions based on the verification results.

[0174] Level 4 conflict, namely: severe failure of dual verification: switch to simplified neural network backup model; generate geological emergency verification list; system enters safe operation and maintenance mode.

[0175] The geological priority principle for Level 1 conflicts includes mandatory priority for physical verification conclusions in the following scenarios: verification deviations exceeding Level 2 tolerance; spatial overlap between the anomaly zone and the pumping over-extraction zone exceeding 60%; and recent occurrences of ground subsidence or collapse events. After priority execution of the backtracking, historical verification units must complete a cause tracing report.

[0176] The emergency response chain for a Level 4 conflict includes the following:

[0177] Simplified model switching: adopts a neural network model validated by historical disaster events; input factors are simplified to three core data: rainfall, pumping, and water level.

[0178] Geological inventory generation: Automatically marks the spatial coordinates of conflict zones and suspected causes; links them to geological monitoring anomaly records from the most recent three months; and pushes them to the priority queue of the local management department.

[0179] Safe operation and maintenance mode: Stop parameter evolution loop optimization, reduce the frequency of assimilation loop in real time, and output system self-test report.

[0180] The basic tolerance threshold is based on industry standards, the historical best solution is extracted using standard database queries, and the neural network model is built on an open-source framework.

[0181] Overcoming the limitations of homogeneous verification, a pore-differentiated tolerance rule was established to ensure that verification results conform to geological realities. Discrete point criticality determination technology was utilized to solve the challenge of heterogeneous verification. A geological priority principle ensured reliability in critical scenarios; a three-tiered emergency response mechanism improved crisis management efficiency.

[0182] It should be further explained that, during the specific implementation process, after the system starts up, the data acquisition module continuously receives real-time data streams from IoT hydrological sensors, GRACE gravity satellites, precipitation-evaporation multi-source remote sensing, and meteorological interfaces. The real-time assimilation loop first performs dynamic weight allocation: hydrological factors are assigned a basic confidence level based on the sensor calibration status, and are automatically downweighted if equipment maintenance exceeds its time limit; the weight of meteorological factors fluctuates according to data freshness, and linearly decays if data is not updated within a certain time limit; the weight of human activity factors is linked to environmental abrupt change indicators, and when the monitored value of the pumping rate abrupt change threshold increases by more than three times the standard deviation of the historical average within one hour, the highest weight is immediately locked and an alarm is triggered. The weighted multi-source data is input into a Kalman filter unit to generate predicted values ​​of groundwater storage changes at a ten-day frequency.

[0183] The parameter evolution loop is deeply coupled with the assimilation loop: the parameter dynamic modeling unit defines parameters such as infiltration coefficient, permeability coefficient, specific yield, and release coefficient as spatiotemporally continuous functions, and their rate of change is related to the cumulative pumping volume in the region. When the monthly pumping volume exceeds the warning line, the function automatically adds a decay correction term; the specific yield further responds to the meteorological drought level, and an exponential decay term is introduced during severe drought. The error source tracing unit analyzes and predicts errors, and only marks a parameter as a parameter to be optimized when adjusting a parameter alone can reduce the total error by more than a preset critical value. The hybrid optimization unit initiates a genetic-particle swarm optimization algorithm to iteratively update parameters under the hard constraint that the difference in the groundwater continuity equation does not exceed the groundwater mass balance.

[0184] A cross-cycle feedback bus connects two loops: the uplink transmits spatial distribution information of sensitive factors, with precise coordinates of local sensitive points and vector data of regional sensitive zones layered and encapsulated; the downlink feedback uses a simplified parameter mapping table, transmitting only parameters whose changes exceed historical extremes. A dual verification module monitors in parallel, and a physical conservation verification unit calculates the supply-discharge balance difference. If the deviation exceeds the aquifer's dynamic tolerance, a three-level response is initiated based on the deviation magnitude, including: Level 1 deviation: local data reset; Level 2 deviation: freezing, optimization, and scanning for geological anomalies; Level 3 deviation: enabling historical parameter set coverage. The aquifer's dynamic tolerance includes an automatic upward floating tolerance during the rainy season. The historical scenario verification unit segments the prediction curve into stable, gradual, and abrupt change periods, assigning double weight to abrupt change segments to calculate morphological similarity; loose porous regions require full curve matching.

[0185] Final output preprocessing for verification conflicts: If physical verification fails but historical verification passes, spatial backtracking is prioritized and geological uncertainties are marked; if historical verification fails but physical verification passes, parameters are frozen and the historical optimal solution is called; if there is a slight deviation in both verifications, the analysis scope is narrowed for on-site verification; if both verifications fail severely, a simplified model is switched and a geological emergency list is generated. Geological rules are embedded throughout the system: during periods of sudden rainfall changes, the filtering step size is compressed; when pumping exceeds the scale, the error threshold is lowered.

[0186] When satellite data is interrupted, gradient substitution is initiated: the first choice is to access the measured values ​​of ground meteorological stations, the second choice is to call the historical database of the same model, and when there is no substitution data, the weight is reset to zero and the prediction confidence is downgraded.

[0187] When key monitoring well data is missing, the weighted average of adjacent wells is used as a substitute, with the weights distributed inversely proportional to the distance. If more than 30% of the monitoring points in the area are invalid, the system automatically switches to the remote sensing inversion-dominated mode and marks the data reliability as downgraded.

[0188] When the monthly pumping volume in an industrial zone exceeds the historical peak, functions such as permeability coefficient, specific yield, and release coefficient are adjusted with a negative exponential factor within one kilometer of the pumping point; if the excess exceeds the limit for three consecutive months, the adjustment is expanded to a linear attenuation layer over a three-kilometer radius. When the cumulative quarterly pumping volume in an agricultural irrigation area exceeds the crop water requirement model, functions such as permeability coefficient, specific yield, and release coefficient are adjusted with a stepped attenuation factor, and during the fallow period, a natural recovery algorithm is activated to restore one-tenth of the original value daily.

[0189] When the physical conservation check issues an alarm during optimization, the iteration is immediately paused and the spatial overlap between the alarm region and the optimization domain is compared. If the overlap exceeds 60%, the population is cleared and reinitialized; if it is less than 30%, only individuals with parameters in the alarm region are removed.

[0190] When the deviation exceeds the tolerance limit, the discharge end is checked first. If the pumping rate changes abruptly by more than three times the standard deviation, it is marked as a human-induced cause. The supply end is checked next. If the rainfall data is out of time and there is no sudden change, the data is judged to be distorted. If there is no clear cause, emergency tracing of parameters is initiated.

[0191] After segmenting the curve by inflection point, the segment with a sudden drop in water level is given double the weight. Ordinary sudden drop, i.e., when the slope exceeds the historical average by one time, it is correlated with the surrounding pumping logs; abnormal sudden drop, i.e., when it exceeds twice the historical average, the weight is increased to three times and a scan of wells within a 10-kilometer radius is forcibly initiated; catastrophic sudden drop, i.e., when it exceeds three times the historical average, the comparison is suspended and the emergency protocol is activated directly.

[0192] Point-like sensitive sources have their influence radius calculated based on the aquifer permeability coefficient; linear sensitive zones have their buffer bandwidth adjusted according to the riverbed sediment thickness; and areal sensitive areas have their polygon vertices reconstructed based on monitoring data.

[0193] Updated parameters verified by on-site pumping tests and with geophysical deviations within limits are marked as Grade A; parameters without monitoring well coverage and exhibiting fluctuations exceeding historical levels are marked as Grade C and accompanied by a geological warning.

[0194] If the deviation reaches the second-level tolerance limit or above, or if the abnormal area overlaps with the pumping over-extraction area by more than 60%, or if ground subsidence or collapse has occurred recently, the physical conclusions will be retrospectively applied without conditions.

[0195] Switch to a simplified three-factor model of rainfall, pumping, and water level; automatically link the geological inventory with monitoring anomalies within three months; output a self-inspection report after system downgrade until manual cancellation.

[0196] It should be further explained that a method for predicting loose porous groundwater storage variables based on the fusion of data assimilation and parameter optimization includes the following steps:

[0197] Step S1: Connect IoT hydrological sensors, GRACE gravity satellite, precipitation-evaporation multi-source remote sensing and meteorological real-time data stream; assess sensor calibration status and assign basic confidence to hydrological factors, automatically reducing weight when equipment maintenance exceeds the time limit; detect the freshness of meteorological data, linearly reducing weight if data is not updated within the time limit; monitor pumping rate mutation values, locking the highest weight of human activity factors and triggering an alarm when the rate exceeds three times the standard deviation of historical fluctuation extreme values.

[0198] Step S2: Perform ensemble Kalman filtering on the weighted multi-source data: extrapolate the current groundwater storage distribution based on the hydrogeological model; prioritize the correction of prediction bias by fusing high-weight factors; reduce the error covariance iteration step size when environmental changes are active; output the predicted value of groundwater storage change every ten days, and automatically switch to high-frequency calculation when the weight alarm is triggered.

[0199] Step S3: Model parameters such as infiltration coefficient, permeability coefficient, water yield, and release coefficient as spatiotemporal functions. When the monthly pumping volume exceeds the warning line, the function is superimposed with a negative exponential decay term within one kilometer of the pumping point. The water yield is further correlated with the meteorological drought level, and an exponential decay term is embedded when there is severe drought. When excessive pumping is detected for three consecutive months, the parameter decay range is expanded to three kilometers and a linear decay layer is superimposed.

[0200] Step S4: Analyze the prediction error every ten days. Only when the adjustment of a single parameter causes the total error to decrease by more than the preset critical value, it is marked as a parameter to be optimized. Physical conservation verification deviation exceeds the dynamic tolerance limit, that is: the tolerance limit is automatically raised during the rainy season, and the response is initiated according to the magnitude: the first-level deviation locally resets the data, the second-level deviation freezes the optimization and scans for geological anomalies, and the third-level deviation enables the historical parameter overlay.

[0201] Step S5: Start the genetic-particle swarm fusion algorithm to iterate parameters; hard constraint that the difference in the groundwater continuity equation does not exceed the groundwater mass balance; if a physical verification alarm occurs during optimization, pause the iteration; if the spatial overlap exceeds 60%, clear the population and restart the optimization.

[0202] Step S6: Uplink transmission of spatial information of sensitive factors: the influence radius of point sensitive sources is calculated according to the aquifer permeability coefficient, the bandwidth of linear sensitive zones is adjusted according to the riverbed sediment thickness, and the polygon vertices of areal sensitive areas are reconstructed according to the monitoring data; downlink feedback of simplified parameter mapping table, and geological warning labels are added to parameters without monitoring well coverage.

[0203] Step S7: Parallel execution:

[0204] Physical conservation verification: Linear tolerance is applied to the loose porous region;

[0205] Historical trend verification: The segmentation curves represent the plateau period, gradual change period, and abrupt change period, with the abrupt change segment assigned double weight;

[0206] In case of conflict, initiate geological priority decision-making: if physical verification fails but historical verification passes, spatial backtracking is performed and geological uncertainty is marked; if historical verification fails but physical verification passes, parameters are frozen and the historical optimal solution is called; if both verifications fail severely, switch to the rainfall-pumping-water level three-factor simplified model and generate a geological emergency list.

[0207] Step S8: Output the final predicted value only when both verifications pass; Level 4 conflict triggers system degradation: stop parameter optimization, real-time assimilation loop frequency reduction, and output a self-check report; the geological inventory is automatically associated with recent monitoring anomaly records and pushed to the management department.

[0208] By constructing a dual-loop collaborative architecture of a real-time assimilation loop and a parameter evolution loop, combined with dynamic weight allocation and environmental mutation response mechanisms, the prediction lag problem caused by data update delays and parameter rigidity in groundwater prediction systems is solved. High-frequency assimilation and fusion of multi-source real-time data ensures that model input is synchronized with the actual environmental state; parameter drift tracking trees and geological constraint optimization enable dynamic adaptive updates of core parameters such as infiltration coefficient, permeability coefficient, specific yield, release coefficient, and water abundance, allowing the prediction system to continuously track the transient processes of the groundwater system and effectively improve the timeliness and reliability of groundwater storage change predictions under scenarios such as extreme rainfall, extreme drought, and high-intensity extraction.

[0209] By employing a dual verification mechanism and deeply embedding geological rules, the scientific rigor and robustness of the prediction results are ensured. The physical conservation sieve intercepts physically infeasible solutions through differentiated tolerance rules based on aquifer type. The historical scenario simulator, driven by geological features, performs segmented weighted analysis of the curves to accurately identify abnormal model behavior. A four-level conflict decision engine initiates targeted responses based on geological priority principles, enabling the system to adapt to complex geological environments. Ultimately, this forms a closed-loop control chain encompassing data assimilation, parameter evolution, verification interception, and conflict self-healing, providing highly reliable dynamic prediction support for water resource management.

[0210] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0211] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A groundwater storage variable intelligent prediction system based on multi-source data fusion and parameter optimization, characterized in that, include: The data acquisition module connects to IoT hydrological sensors, satellite remote sensing data interfaces, and meteorological data interfaces; The real-time assimilation loop is connected to the data acquisition module for high-frequency fusion of multi-source observation data, which includes precipitation-evapotranspiration remote sensing data, GRACE gravity satellite data, meteorological data, surface hydrological data, aquifer data, groundwater level, water quality and development and utilization data at different spatiotemporal scales. The parameter evolution loop is nested and deeply coupled to the real-time assimilation loop, and its input receives the prediction error data output by the real-time assimilation loop. The dual verification module has its inputs connected to the outputs of the real-time assimilation loop and the parameter evolution loop, respectively. A cross-cycle feedback bus provides bidirectional connection between the real-time assimilation loop and the parameter evolution loop. The output of the parameter evolution loop is fed back to the parameter input of the real-time assimilation loop, and the verification result of the dual verification module triggers the iterative optimization of the real-time assimilation loop or the parameter evolution loop. The real-time assimilation loop includes: The dynamic weight allocation unit adjusts the coupling weights of hydrological factors, meteorological factors, and human mining activity factors in real time based on sensor accuracy level, data timestamps, and environmental change indicators. The Kalman filter algorithm unit is used to perform assimilation calculations at different time scales on the weighted multi-source data and output the predicted value of groundwater storage change. The parameter evolution loop includes: The parameter dynamic modeling unit defines the infiltration coefficient, permeability coefficient, specific yield, release coefficient, and water-bearing parameters as functions of spatiotemporal coordinates, and associates them with the pumping accumulation function and the drought index function. The error tracing unit identifies parameters to be optimized whose contribution rate exceeds a preset threshold based on the prediction error at different time scales of the real-time assimilation loop. The hybrid optimization unit adopts a fusion framework of genetic algorithm and particle swarm optimization algorithm to update the parameters to be optimized under the constraint of groundwater continuity equation. The dual verification module includes: The physical conservation verification unit verifies the prediction results through the groundwater mass conservation equation. If the mass balance error exceeds the threshold, it sends a data verification command to the real-time assimilation loop. The historical scenario verification unit compares the dynamic trend similarity between the real-time prediction curve and the historical curve of the same scenario. If the similarity is lower than the preset standard, it sends a deep optimization instruction to the parameter evolution loop. The cross-cycle feedback bus executes: Real-time assimilation of loop parameters, evolutionary loop transport of highly sensitive factor identifiers and short-term prediction error distribution; Parameter evolution loop to real-time assimilation loop feedback parameter mapping table; The error tracing unit works in conjunction with the physical conservation verification unit: when the physical conservation verification fails, the preset threshold of the error tracing unit is automatically adjusted.

2. The intelligent prediction system for groundwater storage variables based on multi-source data fusion and parameter optimization according to claim 1, characterized in that: The environmental abrupt change indicators include the rate of change in rainfall intensity, extreme drought, and the threshold for abrupt changes in pumping rate.

3. The intelligent prediction system for groundwater storage variables based on multi-source data fusion and parameter optimization according to claim 1, characterized in that: The hybrid optimization unit sets physical constraint boundaries, including aquifer boundaries, aquifer thickness limits, and reasonable ranges for hydrogeological parameters such as infiltration coefficient, permeability coefficient, specific yield, and release coefficient.

4. The intelligent prediction system for groundwater storage variables based on multi-source data fusion and parameter optimization according to claim 1, characterized in that: The dynamic trend similarity is calculated using the dynamic time warping algorithm.

5. The intelligent prediction system for groundwater storage variables based on multi-source data fusion and parameter optimization according to claim 1, characterized in that: The final predicted value of groundwater storage change must simultaneously meet the following requirements: Verification was performed using the groundwater mass conservation equation of the physical conservation verification unit; Verification is performed by validating the dynamic trend similarity of the historical scenario verification unit.

Images