A Dynamic Early Warning Method and System for Soil Erosion in Hydropower Stations Based on Work Area Sensitive Quantitative Coupling
By constructing a four-dimensional sensitive factor weighted model and fusing multi-source data, and dynamically adjusting the early warning threshold, the problem of quantification and dynamic updating of water and soil erosion prevention zones for hydropower stations was solved, and accurate monitoring and early warning of water and soil erosion risks were achieved.
Patent Information
- Application Number
- CN202611204740.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-10
- Publication Date
- 2026-11-17
Smart Images

Figure CN123117768A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of intelligent monitoring and dynamic early warning technology for soil erosion in hydropower stations, specifically involving a dynamic early warning method and system for soil erosion in hydropower stations based on sensitive quantitative coupling of work areas. Background Technology
[0002] Current standards in the hydropower industry, such as the "Technical Specification for Soil and Water Conservation in Water Conservancy and Hydropower Projects," only divide soil and water conservation zones based on the land use functions of the hydropower station, spoil heap, material yard, and construction roads. These zones are only used for statically arranging soil and water retention and vegetation restoration projects, and have the following shortcomings:
[0003] The division of prevention and control zones relies solely on qualitative assessments of site functions, lacking a unified and reproducible quantitative evaluation standard for soil erosion risk. Risk differences between steep slopes and gentle slopes within the same functional area cannot be precisely differentiated. The boundaries and risk levels of prevention and control zones are fixed and cannot be dynamically updated based on monthly excavation disturbances or flood season rainfall, failing to adapt to continuously changing soil erosion conditions during the construction period. All monitoring and early warning assessments use unified parameters across the entire region, failing to couple regional risk differences to data fusion and threshold determination processes. Highly sensitive grids are prone to missing warnings even with minor disturbances, while routine construction on gentle slopes frequently triggers false alarms. Summary of the Invention
[0004] To address the shortcomings of existing technologies, this invention provides a dynamic early warning method and system for soil erosion in hydropower stations based on work area sensitive quantitative coupling, which can effectively solve the above-mentioned problems.
[0005] The technical solution adopted in this invention is as follows:
[0006] In a first aspect, the present invention provides a dynamic early warning method for soil erosion in hydropower stations based on work area-sensitive quantitative coupling, comprising the following steps:
[0007] Step S1: Identify typical monitoring bodies in the hydropower station basin, including slopes and stockpiles; the stockpiles include spoil heaps and material yards; deploy multi-source IoT monitoring equipment, including AI vision cameras, rainfall sensors, and displacement sensors, in each monitoring body; use oblique photography drones to collect aerial remote sensing images of the hydropower station basin, and acquire multi-source IoT monitoring data collected by the multi-source IoT monitoring equipment for each monitoring body, including cumulative slope displacement, stockpile outline, and rainfall;
[0008] Step S2: The U-Net model is used to classify and identify the aerial remote sensing images of the hydropower station basin, and the type of each grid is identified, including construction disturbance area, slope area, spoil disposal area, material yard area and vegetation features. Combined with DEM data and the multi-source IoT monitoring data, the water and soil loss monitoring feature set of each grid is obtained through spatiotemporal registration, including topographic slope, cumulative slope displacement monitoring value, grid period cumulative rainfall, measured value of construction disturbance area, measured value of vegetation coverage and pile accumulation volume.
[0009] Step S3: Construct a set of sensitive factors including terrain slope, vegetation soil-fixing capacity, soil and rock stability, and monthly construction disturbance frequency; set a fixed weighting weight for each sensitive factor; determine the sensitivity coefficient of each sensitive factor for each grid based on the soil erosion monitoring feature set and grid attribute data for each grid; fuse each sensitivity coefficient according to the fixed weighting weight to obtain the comprehensive sensitivity coefficient of each grid; and determine the soil erosion risk level of each grid by combining it with a standardized risk classification system; wherein, the soil erosion risk level includes high risk level, medium risk level, and low risk level.
[0010] Based on the soil and water loss risk level of each grid, a global soil and water loss risk layer with a comprehensive sensitivity coefficient is generated. Each layer consists of grids with the same soil and water loss risk level.
[0011] Step S4: Construct a dynamic weighted fusion mechanism for the time series timeliness coefficient, spatial imaging accuracy coefficient, and comprehensive sensitivity coefficient to correct each type of soil erosion monitoring feature of each grid, and obtain the corrected value of each type of soil erosion monitoring feature of each grid to construct the fusion feature vector of each grid.
[0012] Step S5: Construct a set of soil erosion risk indicators, including construction disturbance area risk, vegetation degradation risk, slope slippage risk, erosion risk of the embankment, and rainfall-induced erosion risk; determine the weight of each soil erosion risk indicator based on the analytic hierarchy process; and introduce a comprehensive sensitivity coefficient to construct a comprehensive risk assessment model that weights and integrates each soil erosion risk indicator.
[0013] Based on the fused feature vector of each grid, the quantitative value of each soil and water loss risk index of each grid is calculated, and then the comprehensive soil and water loss risk value of each grid is obtained by using a comprehensive risk assessment model.
[0014] Step S6: Introduce a comprehensive sensitivity coefficient, substitute it into the composite fitness function, and use a genetic algorithm to iteratively solve the dynamic early warning threshold corresponding to each soil and water loss risk level.
[0015] Step S7: Set up a random forest sub-model corresponding to each soil erosion risk level. The random forest sub-model outputs graded early warning information based on the comprehensive soil erosion risk value of each grid and its dynamic early warning threshold.
[0016] Step S8: Based on the on-site verification results, classify and identify the early warning deviations of the graded early warning information and conduct source analysis, and then adjust the local parameters accordingly.
[0017] Secondly, the present invention provides a dynamic early warning system for soil erosion in hydropower stations based on work area sensitive quantitative coupling, which is used to implement the aforementioned dynamic early warning method for soil erosion in hydropower stations based on work area sensitive quantitative coupling.
[0018] The present invention provides a dynamic early warning method and system for soil erosion in hydropower stations based on work area-sensitive quantitative coupling, which has the following advantages:
[0019] This invention provides a dynamic early warning method and system for soil erosion in hydropower stations based on sensitive quantitative coupling of work areas. The invention constructs a weighted model of four-dimensional sensitive factors, including terrain slope, vegetation soil stabilization capacity, soil and rock stability, and monthly construction disturbance frequency, solves the grid-level comprehensive sensitivity coefficient, and uses the comprehensive sensitivity coefficient as a unified correction factor for the entire domain. It fully embeds a full early warning link, including multi-source data fusion, sub-indicator quantification, comprehensive risk assessment, dynamic early warning threshold iteration, and self-optimization of zonal parameters. The entire set of quantitative calculation rules and process closed-loop system are specifically designed for dynamic disturbance scenarios during the hydropower construction period, and are engineering reproducible and highly implementable. Attached Figure Description
[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0021] Figure 1 The overall flowchart of the dynamic early warning method for soil erosion in hydropower stations based on work area sensitive quantitative coupling provided by the present invention;
[0022] Figure 2 This is a schematic diagram illustrating the classification of the four-dimensional sensing system in the present invention.
[0023] Figure 3 This is a schematic diagram comparing the traditional global threshold with the partitioned dynamic threshold curve of this invention. Detailed Implementation
[0024] To make the technical problems solved, the technical solutions, and the beneficial effects of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and are not intended to limit the invention.
[0025] Current technologies for intelligent early warning of soil erosion in hydropower stations suffer from three core technical deficiencies that are difficult to address simultaneously, thus limiting the accuracy and long-term applicability of engineering early warning systems:
[0026] First, there is no quantitative standard for adapting water and soil risk zoning to different working conditions. Current water conservancy and hydropower specifications only qualitatively divide water and soil conservation zones based on the functions of sites such as hubs, spoil heaps, and material yards. The zoning boundaries are fixed and there are no multi-factor quantitative calculation rules. This makes it impossible to distinguish the differences in erosion endowment between steep slopes and gentle slopes within the same site, and it cannot support differentiated dynamic early warning calculations.
[0027] Second, risk assessment adopts a unified quantitative standard across the entire area. The existing monitoring indicator assessment formula does not adaptively correct for the soil and water loss risk of the plot. Under the same disturbance conditions, high and steep slopes are prone to missed alarms, while flat material yards frequently generate false alarms, resulting in low early warning accuracy.
[0028] Third, the early warning threshold and model iteration lack scenario-coupled constraints. Conventional optimization algorithms only take minimizing the early warning error as the single objective, without considering regional sensitivity levels or monthly construction disturbance changes. As a result, the early warning accuracy continues to decline after long-term system operation.
[0029] To address the gaps in existing technologies, this invention provides a dynamic early warning method and system for soil erosion in hydropower stations based on sensitive quantitative coupling of work areas. This invention constructs a weighted model of four-dimensional sensitive factors, including terrain slope, vegetation soil stabilization capacity, soil and rock stability, and monthly construction disturbance frequency, to solve for the grid-level comprehensive sensitivity coefficient. The comprehensive sensitivity coefficient is then used as a unified correction factor across the entire domain. This model is fully embedded in a comprehensive early warning link that integrates multi-source data fusion, sub-indicator quantification, comprehensive risk assessment, dynamic early warning threshold iteration, and self-optimization of zonal parameters. The entire set of quantitative calculation rules and closed-loop process system is specifically designed for dynamic disturbance scenarios during the hydropower construction period, making it engineering reproducible and highly implementable.
[0030] This invention distinguishes itself from static and qualitative site-specific soil and water conservation zoning. It proposes a risk grading system based on a four-dimensional weighted comprehensive sensitivity coefficient: Utilizing a five-view oblique photography multi-view stereo matching technology (one vertical and four inclined), a 1m×1m regular DEM grid is generated across the entire area through aerial triangulation, point cloud filtering, and grid interpolation. Using a single DEM grid as the smallest independent evaluation unit, fixed weights are applied to four sensitive factors—topographic slope, vegetation soil-fixing capacity, soil and rock stability, and monthly construction disturbance frequency—for normalized weighted calculation, resulting in a grid-by-grid comprehensive sensitivity coefficient S. Based on the S value range, three levels of soil erosion risk—high, medium, and low—are defined. Based on monthly updated construction disturbance data, S is recalculated across the entire area to achieve dynamic updates of risk level zoning.
[0031] The technical solution of this invention revolves around "full-domain coupling of four-dimensional grid sensitivity coefficients". The comprehensive sensitivity coefficient S of the grid is used as a unified correction factor and embedded hierarchically into the entire early warning process. First, the temporal and spatial dimensions are coupled to achieve adaptive feature fusion of multi-source data. Then, based on S, five quantitative formulas for soil erosion risk indicators are constructed, along with a comprehensive risk assessment model with secondary coupling. Subsequently, a composite fitness function that integrates multiple scenario factors is constructed to iteratively generate dynamic early warning thresholds according to risk levels. Finally, early warning deviations are distinguished according to risk levels to achieve closed-loop correction of parameters for each risk level partition.
[0032] This invention innovatively divides into two core coupled systems: a four-dimensional sensitivity coefficient quantization grid level partitioning and spatiotemporal three-dimensional adaptive feature fusion system, and a sensitivity coefficient coupled sub-item index quantization and composite fitness partitioning dynamic early warning threshold iteration system; and a supporting auxiliary optimization mechanism, namely a supporting auxiliary mechanism for partitioned closed-loop parameter self-correction.
[0033] Please see Figures 1-3 The present invention provides a dynamic early warning method for soil erosion in hydropower stations based on work area-sensitive quantitative coupling, comprising the following steps:
[0034] Step S1: Identify typical monitoring bodies in the hydropower station basin, including slopes and stockpiles; the stockpiles include spoil heaps and material yards; deploy multi-source IoT monitoring equipment, including AI vision cameras, rainfall sensors, and displacement sensors, in each monitoring body; use oblique photography drones to collect aerial remote sensing images of the hydropower station basin, and acquire multi-source IoT monitoring data collected by the multi-source IoT monitoring equipment for each monitoring body, including cumulative slope displacement, stockpile outline, and rainfall;
[0035] As one example, the acquisition cycle for the oblique photography drone to collect aerial remote sensing images of the hydropower station basin is 7 days; the acquisition cycle for the AI vision camera and displacement sensing terminal is 15 minutes; and the acquisition cycle for the rainfall sensing terminal is hours.
[0036] Step S2: The U-Net model is used to classify and identify the aerial remote sensing images of the hydropower station basin, and the type of each grid is identified, including construction disturbance area, slope area, spoil disposal area, material yard area and vegetation features. Combined with DEM data and the multi-source IoT monitoring data, the water and soil loss monitoring feature set of each grid is obtained through spatiotemporal registration, including topographic slope, cumulative slope displacement monitoring value, grid period cumulative rainfall, measured value of construction disturbance area, measured value of vegetation coverage and pile accumulation volume.
[0037] Step S3: Quantize the raster partitions using four-dimensional sensitivity coefficients to form a four-dimensional weighted raster risk partitioning model, which is distinct from water conservation functional prevention and control partitioning.
[0038] Construct a set of sensitive factors including terrain slope, vegetation soil stabilization capacity, soil and rock stability, and monthly construction disturbance frequency; set fixed weights for each sensitive factor, for example, the fixed weights for each sensitive factor are: terrain slope 0.3, vegetation soil stabilization capacity 0.25, soil and rock stability 0.25, and monthly construction disturbance frequency 0.2, with a total weight of 1.
[0039] The fixed weighting of each sensitive factor has the following advantages: Although the slope of the hydropower station material yard, the hardened platform within the site, and the gentle open space is very small (<8°), the soil and rock are stable, and the vegetation is intact, the monthly construction disturbance factor will inevitably be greater than 0 as long as there is construction activity. Even if the other three factors are all close to 0, the construction disturbance weight of 0.2 contributes the value alone, and the minimum value of the comprehensive sensitivity coefficient is naturally locked at 0.2. That is, any hydropower construction site has a minimum risk of soil erosion and disturbance, and there is no grid with a risk of 0.
[0040] Based on the soil and water loss monitoring feature set and grid attribute data of each grid, the sensitivity coefficient of each sensitive factor of each grid is determined. The sensitivity coefficient is the sensitivity coefficient normalized to the interval of 0 to 1. The sensitivity coefficients are fused according to the fixed weighting weight to obtain the comprehensive sensitivity coefficient S of each grid (1m).
[0041] By combining a standardized risk classification system, the soil erosion risk level of each grid is determined; wherein, the soil erosion risk level includes high risk level, medium risk level and low risk level;
[0042] Specifically, to avoid regional imbalance, this invention pre-defines a three-level standardized risk classification system: high risk level, S∈[0.8,1.0] and terrain slope ≥25°; medium risk level, S∈[0.5,0.8), mostly conventional waste dumps; low risk level, S∈[0.2,0.5), basically flat material yards; where S is the comprehensive sensitivity coefficient of the grid.
[0043] Based on the soil and water loss risk level of each grid, a global soil and water loss risk layer with a comprehensive sensitivity coefficient is generated. Each layer consists of grids with the same soil and water loss risk level.
[0044] Using the above method, the entire grid is divided into three risk zones: high-risk zone, medium-risk zone, and low-risk zone.
[0045] This partition uses continuous DEM grids as the calculation unit. The grid boundary is uniquely determined by the projected coordinates and grid size, and is not constrained by artificially delineated spoil heaps and slope vector functional partitions. Multiple levels of risk grids can be subdivided within a single site. The monthly construction excavation disturbance data is imported to refresh the entire domain S, realizing dynamic iteration of risk partitions. The output S serves as a unified correction variable for the entire process. Existing static qualitative prevention and control partitions do not have this quantitative coupling function.
[0046] Step S4, Spatiotemporal 3D Adaptive Feature Fusion System:
[0047] A dynamic weighted fusion mechanism is constructed to integrate the time-series timeliness coefficient, spatial imaging accuracy coefficient, and comprehensive sensitivity coefficient. This mechanism corrects the soil erosion monitoring characteristics of each grid cell for each type, resulting in corrected values for each type of soil erosion monitoring characteristics of each grid cell, which are then used to construct the fused feature vector for each grid cell.
[0048] For example, the time series timeliness coefficient is assigned in the following way: the data update interval and the degree of missing historical data of the corresponding data source are used to assign segmented quantitative values. The more frequent the data updates and the lower the missing rate, the higher the time series coefficient value.
[0049] The spatial imaging accuracy coefficient is assigned in a graded manner based on the stability of the on-site camera deployment and the degree of image interference including mountain shadows and fog. The less imaging interference and the more stable the camera deployment, the higher the spatial imaging accuracy coefficient value. The comprehensive sensitivity coefficient is directly called from the comprehensive sensitivity coefficient S pre-calculated for the current grid.
[0050] After normalizing the time series timeliness coefficient, spatial imaging accuracy coefficient, and comprehensive sensitivity coefficient, the correction coefficients for each soil and water loss monitoring feature are obtained by multiplying them. For the two high-risk features of slope displacement and spoil accumulation, the feature proportion is automatically amplified by relying on three-dimensional coupling weights, while weakening low-correlation redundant information such as temperature and humidity and environmental clutter, so as to avoid the weak early erosion features being submerged by noise.
[0051] Based on this, a spatiotemporal weighted interpolation algorithm was used to complete the temporal registration of 7-day cycle low-frequency oblique aerial imagery and 10-minute high-frequency IoT sensor data, filling the time gap between the two types of heterogeneous data. At the same time, the raster comprehensive sensitivity coefficient S was embedded into the DS evidence theory combination rule, and the confidence of different monitoring evidence was corrected according to the raster risk level, and the conflict of multi-source data evidence was resolved. Finally, the fusion feature vector carrying the raster comprehensive sensitivity coefficient was output, providing high-quality hierarchical feature input for subsequent sub-indicator quantification and comprehensive risk assessment.
[0052] Step S5: Construct a set of soil erosion risk indicators, including construction disturbance area risk, vegetation degradation risk, slope slippage risk, erosion risk of the embankment, and rainfall-induced erosion risk; determine the weight of each soil erosion risk indicator based on the analytic hierarchy process; and introduce a comprehensive sensitivity coefficient to construct a comprehensive risk assessment model that weights and integrates each soil erosion risk indicator.
[0053] Based on the fused feature vector of each grid, the quantitative value of each soil and water loss risk index of each grid is calculated, and then the comprehensive soil and water loss risk value of each grid is obtained by using a comprehensive risk assessment model.
[0054] As one implementation method, a comprehensive sensitivity coefficient is introduced into the quantitative value of each soil erosion risk index for each grid, and its calculation method is as follows:
[0055] Construction disturbance area risk quantification value :
[0056] ;
[0057] in: This is the corrected measured value of the construction disturbance area; The baseline value for the unified disturbance area of the watershed is pre-calibrated based on water and soil specifications and historical working conditions; S is the comprehensive sensitivity coefficient of the grid. Characterizes the extent to which construction disturbance exceeds the baseline; This is a correction term, also known as a correction coefficient. To achieve differentiated management by zone, the larger the grid S (steep slope), the smaller the correction coefficient; the smaller the grid S (flat material yard), the larger the correction coefficient. For example, if the value of S is in the range of [0.2, 1.0], the correction coefficient for the correction term will range from 0.5 to 0.9.
[0058] Quantitative value of vegetation degradation risk :
[0059] ;
[0060] in: This is the corrected measured value of vegetation cover; The baseline value for native vegetation cover in the watershed is used. The comprehensive sensitivity coefficient S is coupled and multiplied with the vegetation degradation rate. Even a small amount of damage to highly sensitive slopes can raise the risk value, thus strictly controlling damage to soil-stabilizing vegetation. On flat sites, only large-scale clearing is considered a risk factor. The scheme adds a coupling constraint of S to differentiate the soil and water conservation value of vegetation in different plots.
[0061] Quantitative value of slope slip risk :
[0062] ;
[0063] in: The corrected cumulative displacement monitoring value of the slope; It serves as the benchmark value for safe slope displacement. Based on the zonal scaling logic, multiplying the formula by the comprehensive sensitivity coefficient can directly amplify the displacement risk. Millimeter-level deformation on steep slopes can trigger an early warning, while only large deformation on flat sites is considered a potential hazard, thus accurately capturing early signs of slope slippage.
[0064] Quantitative value of scavenging risk :
[0065] ;
[0066] in: This is the corrected stack volume; This represents the maximum allowable stacking volume for a single grid cell. As a correction term, the risk value under the same stockpile volume in the steep slope spoil disposal area can be simultaneously amplified to prioritize the control of erosion hazards in loose stockpiles. For example, the value range of S is [0.2, 1.0], and the correction coefficient of the correction term ranges from 1.1 to 1.5.
[0067] Rainfall-induced erosion risk quantification value :
[0068] ;
[0069] in: This is the cumulative rainfall over the corrected grid period; This is the critical rainfall threshold for a watershed storm. Based on this quantitative formula, erosion risk can be determined for highly sensitive slopes and spoil heaps with light rain, while low-sensitivity material yards only trigger warnings with extreme rainstorms. This aligns with the disaster-causing mechanisms of rainfall at different sites, and addresses the lack of a dimensionless regional adjustment mechanism for current rainfall risk assessment.
[0070] The comprehensive risk assessment model is as follows:
[0071] ;
[0072] in: The quantitative value of the i-th category of soil and water loss risk index ( , , , , ); The fixed weights are assigned to the i-th type of soil erosion risk indicators; This is a comprehensive risk value for soil and water loss.
[0073] In this embodiment, the fixed weights for each soil erosion risk indicator are as follows: , , , , The fixed weights are 0.30, 0.15, 0.25, 0.20, and 0.10 respectively, with a total weight of 1.
[0074] In this embodiment, the fixed weights of the sub-indicators are determined based on the Analytic Hierarchy Process (AHP). Nine experts with senior engineer titles in hydropower soil and water conservation and slope stability monitoring were selected to conduct a pairwise importance questionnaire survey on a scale of 1–9. The scoring results were summarized to construct a comprehensive judgment matrix, and the initial weights were calculated and normalized using the square root method. After consistency testing, the consistency ratio CR = 0.067 < 0.1, meeting the requirement of logical self-consistency in the evaluation. The weight assignments are aligned with the soil erosion disaster mechanism during the hydropower station construction period: construction excavation disturbance is a continuous wound across the entire area, receiving the highest weight; slope displacement directly induces landslide disasters, receiving the second highest weight; spoil heaps pose a debris flow hazard, receiving the third highest weight; vegetation degradation is a long-term chronic erosion, receiving a medium weight; rainfall is only an external condition triggering erosion, receiving the lowest weight. The entire set of weights is specifically customized for hydropower construction disturbance scenarios and is significantly different from the weights in general models of soil erosion in natural watersheds.
[0075] As one implementation method, if a fixed warning threshold is used, then the risk classification rule can be: P∈[0,0.4) is general risk; P∈[0.4,0.7) is relatively serious risk; P∈[0.7,1.0] is severe risk.
[0076] It should be noted that this invention does not rely on a single indicator to directly determine the warning level. Instead, it uses the risk quantification values of five indicators as intermediate sub-item quantification results, uniformly substituting them into the comprehensive risk model formula P, and then coupling them again with the grid comprehensive sensitivity coefficient S for secondary weighting correction. As can be seen from the aforementioned quantification formulas of the five sub-items and the comprehensive risk assessment model, the grid comprehensive sensitivity coefficient S has a two-level hierarchical coupling correction effect on the calculation of soil erosion risk, forming a two-layer control logic of primary adjustment for sub-item differentiation and secondary weighting amplification of comprehensive risk, thereby achieving differentiated early warning and control of high, medium, and low risk grid levels.
[0077] The first level is a single-stage correction for each sub-indicator: each indicator is coupled with an S-coupling constraint to differentiate between different plots. Vegetation degradation, slope displacement, and rainfall indicators are corrected using a linear correction by directly multiplying by S. For the two core anthropogenic disturbances—excavation disturbance and spoil heap—decreasing correction terms are designed separately. and incremental correction terms The two types of reverse-matching correction terms are designed based on different mechanisms of soil and water erosion. Excavation disturbance only damages the surface vegetation, and the original soil still has the ability to resist erosion. A decreasing correction term is used to appropriately reduce the quantitative base of the individual term. The waste is an unconsolidated loose accumulation, which is prone to landslides and debris flows when it rains, and the disaster risk is stronger. An increasing correction term is used to simultaneously raise the quantitative base of the individual term. The difference in risk potential between the two types of disturbances is distinguished in the sub-item calculation stage.
[0078] The second level is the comprehensive risk layer secondary coupling correction: all component quantification results are uniformly input into the calculation formula of the comprehensive soil and water loss risk value P of the grid, and after weighted summation, it is multiplied again by the corresponding grid comprehensive sensitivity coefficient S to complete the full-domain risk calibration. Even if the component quantification values are similar under the same working conditions, the comprehensive soil and water loss risk value P of the high-sensitivity grid will be amplified due to the larger value of S, and the early warning judgment scale will be automatically tightened; the comprehensive soil and water loss risk value of the low-sensitivity grid will be compressed due to the smaller value of S, and the judgment standard will be appropriately relaxed.
[0079] This dual-layer coupling design comprehensively addresses the shortcomings of traditional unified indicators and thresholds that fail to differentiate between the erosion endowments of different land parcels. It distinguishes the inherent hazard potential of excavation and spoil heap disturbances during the sub-item calculation stage, and leverages grid sensitivity during the comprehensive risk rating stage to achieve tiered control by varying the severity of each zone. This simultaneously considers both the degree of indicator exceedance and the inherent erosion endowment of each land parcel, reducing the probability of missed alarms in highly sensitive areas and false alarms in less sensitive areas from a quantitative perspective. This achieves strict control over steep slopes and reduces false alarms in gentler material yards. In contrast, existing comprehensive evaluation models only weight the raw monitoring data without incorporating pre-quantitative zoning results for secondary correction, resulting in insufficient refined grading capabilities.
[0080] Step S6: Introduce a comprehensive sensitivity coefficient, substitute it into the composite fitness function, and use a genetic algorithm to iteratively solve the dynamic early warning threshold corresponding to each soil and water loss risk level.
[0081] As one implementation method, the composite fitness function is:
[0082] ;
[0083] ;
[0084] in: The composite fitness is the iterative fitness of the genetic algorithm. The larger the value, the better the adaptation effect of the partition threshold. This corresponds to the historical early warning error factor of the risk grid; This represents the time-series change rate of construction disturbances for the current month. This represents the total area of construction disturbance caused by the corresponding risk grid in the current month. This represents the total area of construction disturbance during the same period last month; α is the overall sensitivity coefficient of the raster; β is the error correction weight, β is the scene adaptation weight, and γ is the dynamic weight of the construction condition. For example, the fixed configuration parameters are: , , .
[0085] For example, three types of hydropower-specific scenario factors: The historical early warning error factor for the corresponding risk grid over the past 30 days is defined, with a value range of [0,1]; S is the current grid comprehensive sensitivity coefficient (i.e., the land parcel sensitivity level factor), with a value range of [0.2,1.0]. This is the monthly construction disturbance time-series change rate (i.e., the monthly construction dynamic condition factor).
[0086] To prevent extreme working conditions from causing Infinite scaling of values leads to composite fitness Numerical distortion, iterative divergence of genetic algorithms, and the fact that the monthly disturbance increase or decrease in hydropower projects rarely exceeds ±100% will be addressed during project use. The truncation constraint is set to the interval [−0.5, 1.0]. After unifying the numerical range, it matches the scales Err∈[0,1] and S∈[0.2, 1.0]. The weighted operation of the three types of factors is balanced, and no single factor weight will suppress the other two items. A negative value indicates a reduction in construction disturbance. A value of 0 indicates that the disturbance size remains unchanged. A positive value indicates perturbation expansion. The higher the perturbation increase, the higher the fitness function output value, driving the genetic algorithm to iterate and obtain a more stringent partition warning threshold.
[0087] Iteration rule: Using the initial baseline threshold of each risk partition as the initial population, it can iterate for 100 generations. The maximum value is the convergence target, and the dynamic early warning threshold is solved independently for high, medium and low risk units.
[0088] This invention's composite fitness function includes three types of factors specific to hydropower station soil erosion early warning scenarios: a zonal historical early warning error factor, a grid sensitivity level factor representing the inherent erosion endowment of the land parcel, and a monthly construction dynamic condition factor representing the intensity of phased development. These three types of factors are weighted and coupled to construct a specific adaptation target. This overcomes the shortcomings of existing genetic algorithms that rely solely on a single dimension of early warning error for optimization. It simultaneously considers the characteristics of three scenarios: historical early warning accuracy, the natural risk endowment of the land parcel, and real-time construction disturbance conditions. Iterative iteration across risk grids yields a more adaptable dynamic early warning threshold, fundamentally reducing the false alarms and missed alarms caused by a uniform threshold across the entire domain.
[0089] Step S7: Set up a random forest sub-model corresponding to each soil erosion risk level. The random forest sub-model outputs graded early warning information based on the comprehensive soil erosion risk value of each grid and its dynamic early warning threshold.
[0090] Step S8: Based on the on-site verification results, classify and identify the warning deviations of the graded warning information and conduct source analysis, and adjust the local parameters accordingly. For example, when adjusting the local parameters, only the parameters corresponding to the soil and water loss risk level are adjusted, including the dynamic warning threshold, the model parameters of the random forest sub-model, and the weights of the composite fitness function.
[0091] The existing early warning optimization scheme adopts a unified batch update strategy across the entire domain. Regardless of whether there is an early warning deviation in the grid, all fusion weights, early warning thresholds, and model parameters are updated synchronously. This not only causes a large amount of ineffective computing power consumption, but also frequently disturbs the parameters of stable and unbiased work areas. Long-term operation is prone to causing fluctuations in early warning accuracy.
[0092] This invention constructs a tiered digital workflow for handling hidden dangers, and configures differentiated on-site verification time limits for high, medium and low risk work areas; it collects on-site verification ledgers for the entire basin on a monthly basis, and classifies deviations into three categories according to their causes: rainfall-induced false alarms in high-sensitivity areas, pile disturbance-induced false alarms in medium-sensitivity areas, and extreme working condition adaptation deviations in low-sensitivity areas, and establishes independent deviation datasets for each risk unit.
[0093] This mechanism has two core features: targeted regional analysis and a closed-loop process.
[0094] 1. Partition-based targeted optimization
[0095] During the parameter iteration phase, only the three-dimensional fusion weights, index quantification parameters, composite fitness coefficients, and partitioned random forest sub-model neuron parameters of the risk grid pattern that generates early warning deviations are updated. The parameters of stable risk units without deviations remain unchanged to avoid the waste of computing power and accuracy fluctuations caused by indiscriminate updates across the entire domain. At the same time, the corresponding coupling parameters are adjusted according to the deviation type, tightening the sensitive constraints in the case of missed alarms and appropriately relaxing the judgment scale in the case of false alarms, so as to achieve targeted correction of deviation causes.
[0096] 2. Closed-loop iterative link
[0097] A complete closed loop is formed, which includes risk warning push, on-site graded review, deviation classification and collection, local parameter update of each zone, and the next round of basin warning calculation. A round of full-domain parameter self-iteration is completed every month based on the latest review data.
[0098] This mechanism can continuously adapt to dynamic changes in operating conditions such as excavation, surcharge, and rainfall during the construction period of hydropower stations, suppress the decline in the accuracy of the overall early warning in the long term, and maintain the stability of the graded early warning judgment across the entire basin.
[0099] As one embodiment, a dynamic early warning method for soil erosion in hydropower stations based on work area-sensitive quantitative coupling is provided, such as... Figure 1 and Figure 3 As shown, Figure 1 This is a flowchart of the overall process of a dynamic early warning method for soil erosion in hydropower stations based on work area sensitive quantitative coupling according to the present invention. From top to bottom, the modules are: multi-source data acquisition and preprocessing module, four-dimensional sensitive quantitative work area zoning module, spatiotemporal three-dimensional sensitive coupling feature fusion module, zone-specific risk assessment module, and zone-targeted closed-loop iterative optimization module. Each module performs calculations in sequence, and the closed-loop iterative module outputs updated parameters and feeds them back to the upstream coupled calculation modules to achieve full-process adaptive optimization. Figure 2 This is a schematic diagram of the four-dimensional sensitive coefficient quantitative zoning classification of the present invention; the base map is the DEM grid site outline of the hydropower station watershed, the dashed box is the boundary of the traditional water and soil conservation functional prevention and control zoning in the industry, and the grid texture distinguishes the high, medium and low water and soil loss risk grids automatically divided by the present invention; the attached figure on the left shows the formula for calculating the sensitivity coefficient S by four-dimensional factor weighting and the grid classification judgment standard, which intuitively reflects the innovation of the present invention in grid-based refined dynamic zoning that is different from the traditional block-based qualitative prevention and control zoning. Figure 3 This diagram illustrates the comparison between traditional global thresholds and the dynamic threshold curves of this invention. The horizontal axis represents the construction sequence, and the vertical axis represents the risk warning threshold. The black horizontal line represents the fixed static threshold across the entire domain in the prior art, while the three curves represent the adaptive dynamic thresholds corresponding to the high, medium, and low risk grids of this invention. The diagram clearly shows that this invention can automatically adjust the judgment scale according to the sensitivity of the work area and the construction sequence, thus solving the problem of prominent false alarms and missed alarms in the traditional unified threshold.
[0100] The specific method is as follows:
[0101] S1. Multi-source data standardization preprocessing
[0102] A five-lens oblique photography system with one vertical and four horizontal lenses was used, with 85% overlap in the flight direction and 80% overlap in the lateral direction, to complete the aerial photography of the entire watershed every 7 days. Based on multi-view image stereo matching, a three-dimensional dense point cloud was constructed, and a 1m×1m full-area DEM raster was generated through vegetation layer filtering, ground point extraction, and natural neighborhood interpolation. Simultaneously, IoT images, rainfall, and slope displacement monitoring data collected every 10 minutes were accessed.
[0103] The aerial downward-looking images were corrected for steep slope distortion, removed mountain shadows, and semantically segmented for ground features. The sensor signals were denoised using a micro-motion feature preservation method to filter out equipment jitter interference. The terrain slope was extracted from the DEM raster in batches, grid by grid. All data were uniformly converted to the watershed Gaussian plane coordinate system, and outliers were removed to form a standardized monitoring dataset.
[0104] S2. Calculate the raster comprehensive sensitivity coefficient using a weighted average of four-dimensional sensitivity factors and classify the area into three risk level zones.
[0105] Using a fixed weighting ratio of 0.3 for terrain slope, 0.25 for vegetation soil stabilization capacity, 0.25 for rock and soil stability, and 0.2 for monthly construction disturbance frequency, the comprehensive sensitivity coefficient S is calculated by normalizing and weighting each 1m DEM raster. Based on the range of S values, the system automatically divides the risk into three risk levels: high risk, medium risk, and low risk, and outputs a zone vector layer with sensitive labels. The monthly construction disturbance area data is updated, and S is recalculated across the entire region to achieve dynamic iterative updates of risk zones.
[0106] S3. Three-dimensional sensitive coupling adaptive multi-source feature fusion
[0107] The timeliness coefficient of the data source, the spatial imaging accuracy coefficient of the monitoring point, and the grid comprehensive sensitivity coefficient S are calculated and normalized to dynamically adjust the fusion weight of various soil and water loss features. The Kalman filter is improved with S as a constraint to complete the time alignment of high and low frequency monitoring data. Based on S, the DS evidence combination rule is modified to resolve the conflict of multi-source monitoring evidence and output the grid fusion feature vector carrying the grid sensitivity label.
[0108] S4. Quantification of Sensitive Coupled Soil Erosion Risk Indicators and Analysis of Zonal Dynamic Early Warning Thresholds
[0109] The comprehensive sensitivity coefficient S of the raster is substituted into the five sets of differentiated soil and water loss risk index quantification formulas to obtain the risk quantification value of each sub-item of a single raster; the sub-item quantification value is input into the partitioned coupled comprehensive risk assessment model to calculate the comprehensive soil and water loss risk value P of the raster and divide it into three levels of early warning; the three-dimensional scene is substituted into the coupled composite fitness function, and the dynamic early warning threshold is solved iteratively by risk unit using a genetic algorithm; the corresponding risk partition random forest sub-model is matched to complete the graded early warning judgment, and the graded early warning information is pushed to multiple terminals simultaneously.
[0110] S5. Partition-based targeted closed-loop parameter self-optimization
[0111] Based on the work process of handling hierarchical hazards, the on-site early warning verification results are entered and collected according to risk zones. Three types of deviation data are collected: missed alarms, false alarms, and working condition adaptation. Only for the risk grid pattern that causes early warning deviation, the three-dimensional fusion weight, the supporting parameters of indicator quantification, and the composite fitness coefficient are updated. A round of self-iterative optimization of the whole domain coupling mechanism is completed every month.
[0112] This embodiment provides a dynamic early warning system for soil erosion in hydropower stations based on work area sensitive quantitative coupling. It is used to implement the dynamic early warning method described in any of the aforementioned embodiments. It uses four-dimensional grid sensitivity coefficient full-domain coupling as the underlying operating logic, adopts a layered modular design, and features software and hardware collaboration. It is engineering-applicable, and each module corresponds one-to-one with the method steps.
[0113] 1. Multi-source data acquisition and preprocessing module: integrates an oblique aerial flight data access unit, a field IoT sensor receiving unit, and a DEM topographic raster import unit; it collects three types of multi-source heterogeneous monitoring data: aerial remote sensing data of hydropower station basins, IoT monitoring data, and DEM topographic data. It has built-in image distortion correction, micro-motion preservation and denoising, and lightweight raster slope extraction operators to complete the unification of multi-source data coordinates and noise filtering, and output a standardized monitoring dataset.
[0114] 2. Four-dimensional Sensitivity System Quantitative Engineering Zone Zoning Module: Construct a four-dimensional quantitative evaluation system for topography, vegetation, soil and rock, and construction disturbance. Use fixed weights to calculate the comprehensive sensitivity coefficient S of the grid-scale work area. Based on the sensitivity coefficient, divide the watershed into three differentiated water and soil loss risk units: high, medium, and low.
[0115] Specifically, it has built-in four-dimensional indicator normalization weighting operator and risk grid classification judgment operator; it automatically calculates the comprehensive sensitivity coefficient S grid by grid, divides three types of risk units according to preset value range, generates and updates the partition vector layer monthly, and provides a unified correction factor for the back-end calculation module.
[0116] 3. Sensitive Coupling Spatiotemporal Adaptive Fusion Module: Using the sensitivity coefficient S as the coupling correction factor, a three-dimensional coupling spatiotemporal adaptive three-level fusion mechanism of time-space-work area sensitivity is constructed to dynamically adjust the fusion weight of each soil and water loss feature and output the fusion feature vector of the partition.
[0117] Specifically, it has a built-in three-dimensional weighted coupling operation unit, a sensitive constraint Kalman time-series registration unit, and an S-corrected DS evidence fusion unit; it receives standardized monitoring data and raster sensitive layers, dynamically calculates feature fusion weights, and outputs partition-specific fusion feature vectors.
[0118] 4. Zonal Dynamic Threshold Risk Assessment Module: Based on the sensitivity coefficient S of the work area, a quantification formula for differentiated indicators of each work area is constructed and a zonal coupled comprehensive risk assessment model is established. By integrating the zonal early warning error, grid sensitivity coefficient, and construction time sequence change rate into a three-dimensional scene coupled composite fitness function, a genetic algorithm is used to iterate the zonal-specific dynamic early warning threshold to complete the risk classification assessment.
[0119] Specifically, it has five built-in indicator quantification operators, a partitioned coupled risk assessment operator, a composite fitness genetic iteration operator, and a partitioned random forest sub-model library; it reads fusion features and grid sensitivity coefficients to complete indicator quantification, comprehensive risk rating, partitioned dynamic threshold solution, and outputs hierarchical early warning information.
[0120] 5. Partition-based targeted closed-loop iterative optimization module: Based on risk units, it distinguishes three types of deviations: missed alarms, false alarms, and operating condition adaptation, and implements partition-based targeted closed-loop correction of fusion weights and thresholds.
[0121] Specifically, it connects to the Activiti workflow engine for graded hazard handling, with built-in deviation classification and identification operators and local parameter update engines; it collects early warning and review deviations from each risk unit, and updates the corresponding grid coupling operation parameters in a targeted manner to achieve long-term adaptive operation of the system.
[0122] 6. Multi-terminal visualization display sub-module (with supporting implementation unit): Deploy PC management terminal and mobile mini-program, display watershed risk heat map, dynamic threshold change curve, and hidden danger rectification ledger in layers according to risk unit, control the early warning push delay within 1 second, and persistently store the full process monitoring and iterative data for long-term working condition traceability and review.
[0123] Implementation and comparison results:
[0124] A six-month synchronous comparative monitoring experiment was conducted in the basin during the construction period of a cascade hydropower station in Southwest China. The control group adopted a traditional global fixed weight and unified static threshold early warning scheme, while this invention adopts a four-dimensional sensitive global coupled early warning system for parallel operation.
[0125] (1) Basic working conditions of the project
[0126] The monitoring area includes three typical work areas: steep rock slopes with a gradient of 26° to 38°, large-scale layered spoil heaps, and flat sand and gravel yards; during the flood season, there is concentrated heavy rainfall, and excavation and spoil heap construction are carried out continuously.
[0127] Hardware configuration: A five-lens oblique photography drone with 1 vertical and 4 horizontal lenses is used, with 85% overlap in the flight direction and 80% overlap in the side direction, and a full-area aerial photography is completed every 7 days; AI vision cameras, rainfall and displacement sensors are deployed on slopes, spoil heaps and material yards, with data collection intervals of 10 minutes; the watershed uses a 1m precision DEM grid and a unified Gaussian plane coordinate system.
[0128] Preliminary calibration reference parameters: =50m 2 , =0.75、 =8mm =120m 3 , =45mm; Fixed weights for four-dimensional factors: terrain slope 0.3, vegetation soil stabilization capacity 0.25, soil and rock stability 0.25, monthly construction disturbance frequency 0.2; Fixed coefficient for composite fitness: , , .
[0129] (2) Step S1: Multi-source data acquisition and standardization preprocessing
[0130] Every 7 days, a full-area oblique aerial photography is carried out. Dense point clouds are generated by multi-view image stereo matching. After vegetation layering filtering, ground point extraction, and natural neighborhood interpolation, a full-area 1m×1m DEM raster is generated, and a corrected DOM orthophoto image is output simultaneously.
[0131] The IoT terminal outputs raw data on the pile outline, cumulative slope displacement, and hourly rainfall every 10 minutes.
[0132] Preprocessing stage: DOM images are used to remove steep slope stretching distortion and mountain shadows, and an improved U-Net is used to segment construction disturbances, slopes, waste, and vegetation features; sensor data are filtered using a micro-motion retention filter to remove equipment jitter noise; DEM is used to calculate terrain slope in batches on a grid-by-grid basis; all data are unified with coordinates, extreme outliers are removed, and a standardized dataset is output and stored in a cloud database.
[0133] (3) Step S2: Four-dimensional sensitivity coefficient calculation and raster partitioning example
[0134] A grid was selected on the slope: slope 32°, vegetation cover 0.12, high degree of soil weathering, and large disturbance area in the current month. The normalized values of the four sensitive factors were 0.95, 0.9, 0.92, and 0.85, respectively.
[0135] The overall sensitivity coefficient S = 0.3 × 0.95 + 0.25 × 0.9 + 0.25 × 0.92 + 0.2 × 0.85 = 0.90;
[0136] If the slope satisfies S∈[0.8,1.0] and the slope is ≥25°, it is classified as a high-risk slope grid.
[0137] Select the grid cell in the middle of the spoil heap: S=0.62, which falls within the interval [0.5,0.8), and determine it as a medium-risk grid cell;
[0138] Select a grid cell on a flat platform in the material yard: S=0.31, which falls within the range of [0.2,0.5) and is therefore classified as a low-risk grid cell;
[0139] The total area of construction disturbance in the entire region is imported monthly, S is recalculated grid by grid, and the risk zoning layer of the entire watershed is dynamically updated.
[0140] (4) Step S3: Example of spatiotemporal three-dimensional coupled feature fusion
[0141] Taking the aforementioned highly sensitive slope grid as an example:
[0142] The temporal efficiency coefficient is 0.92 (10 min high-frequency sensing), the spatial imaging accuracy coefficient is 0.95, and the grid integrated sensitivity coefficient S=0.90; after normalization and coupling, the weights of slope displacement and rainfall characteristics are amplified.
[0143] Based on the S-constraint, the Kalman filter is improved to fill the time discontinuity between the 7-day flight low-frequency and 10-minute sensing high-frequency time series; by substituting S=0.90 into the DS evidence correction rule, the confidence of deformation and rainfall evidence is amplified, and the fused feature vector of this raster is output.
[0144] (5) Step S4: Quantification of soil and water loss risk indicators, comprehensive risk, and complete iterative calculation of alarm threshold.
[0145] Substitute the measured data of the highly sensitive slope grid (S=0.90) into the complete set of quantitative formulas:
[0146] ;
[0147] 1) Construction disturbance area =96m 2 , =(96 / 50)×(1−0.9 / 2)=1.92×0.55=1.056;
[0148] 2) Vegetation coverage =0.12, =1−(0.12 / 0.75)×0.9=0.856;
[0149] 3) Cumulative slope displacement =7.2mm, =(7.2 / 8)×0.9=0.81;
[0150] 4) There is no waste accumulation in this grid. Not included in the overall evaluation;
[0151] 5) Cumulative rainfall over a period of time =42mm, =(42 / 45)×0.9=0.84;
[0152] Based on the Analytic Hierarchy Process (AHP), the weights of the quantitative values of construction disturbance area, vegetation cover, slope displacement, spoil accumulation, and rainfall were determined to be 0.30, 0.15, 0.25, 0.20, and 0.10, respectively, with a total weight of 1.
[0153] Substitute the four valid quantitative values into the comprehensive risk model.
[0154] The overall risk value P = (0.30 × 1.056 + 0.15 × 0.856 + 0.25 × 0.81 + 0.20 × 0 + 0.10 × 0.84) × 0.9 = 0.66.
[0155] This was determined to be a relatively high risk. Other grid cells were calculated using the same algorithm.
[0156] Take the early warning error rate of this grid over the past 30 days =0.08, the rate of change of disturbance in the current month =0.45; Substitute into the 3D scene coupled composite fitness function:
[0157] The fixed configuration parameters are: , ,
[0158] = 0.4×(1−0.08)+0.35×0.90+0.25×0.45=0.80;
[0159] Using the initial baseline threshold of each risk partition as the first generation population of the genetic algorithm, the number of iterations is set to 100 generations. With the maximization of Fit as the convergence objective, the dynamic warning threshold of the high-sensitivity grid is obtained after iterative convergence. The comprehensive risk value P=0.66 is input into the random forest sub-model of the matching high-sensitivity partition, which is judged as a relatively heavy risk and a graded warning information is pushed.
[0160] (6) Step S5: Practical Specifications for Partition-Targeted Closed-Loop Iteration
[0161] Set time limits for on-site verification at different levels: on-site verification within 2 hours of warning for highly sensitive slopes, within 4 hours of warning for spoil heaps, and within 12 hours of warning for material yards;
[0162] During the 6-month monitoring period, the high slope grid showed 3 instances of missed alarms induced by rainfall, which were included in the high-sensitivity area-specific bias dataset.
[0163] During monthly iterations, only the 3D fusion weights, composite fitness β coefficients, and displacement index quantization parameters of the high-sensitivity rasters are locally fine-tuned, while the parameters of medium- and low-risk rasters remain unchanged, thus completing closed-loop self-optimization.
[0164] The actual test results are as follows:
[0165] First, the fidelity of multi-source data fusion features: the traditional fixed-weight fusion scheme has a fidelity of only about 72%, while the spatiotemporal three-dimensional sensitive coupling fusion scheme of this invention can reach about 93%, and the retention effect of weak soil erosion deformation and pile disturbance features is significantly improved.
[0166] Second, overall early warning accuracy: The traditional static early warning threshold scheme has a long-term accuracy rate of 70% to 72%, while this invention relies on regional dynamic early warning thresholds and monthly closed-loop iteration, and the early warning accuracy rate is maintained at 86% to 89%.
[0167] Third, the optimization effect of false alarms and missed alarms: For highly sensitive grids such as steep slopes, the problem of missed alarms has been greatly eliminated; the number of invalid false alarms caused by conventional excavation and surcharge construction in flat sand and gravel yards has been significantly reduced;
[0168] Fourth, long-term operational stability: Traditional models lack an adaptive iteration mechanism, and their early warning accuracy decays to about 64% after three months of continuous operation; this invention performs closed-loop correction of the zoned targeted parameters monthly, and the early warning accuracy does not significantly decay throughout the entire monitoring cycle, maintaining a stable level of over 86%.
[0169] Fifth, engineering practice value: During the 6-month pilot monitoring period, the system identified a total of 13 potential soil and water loss hazards, such as slope erosion and spoil sliding, and pushed out disposal work orders for all of them and completed on-site verification and rectification. The closed-loop disposal rate of the pilot project reached 100%.
[0170] Extended Explanation: This invention provides a complete quantitative early warning system for four-dimensional work area sensitive full-domain coupling. It has universal adaptability and can be adapted to different types of hydropower projects, such as water diversion, pumped storage, and roller-compacted concrete dams, by simply replacing the target watershed DEM topography, soil and rock properties, and monthly construction disturbance basic input parameters. At the same time, it can meet the needs of two application scenarios: dynamic disturbance monitoring during the construction period and steady-state water and soil risk management during the operation period. It has a wide range of engineering reuse.
[0171] This invention also provides a dynamic early warning system for soil erosion in hydropower stations based on work area-sensitive quantitative coupling, used to implement the dynamic early warning method for soil erosion in hydropower stations based on work area-sensitive quantitative coupling described in any of the foregoing embodiments. The system features a hierarchical and modular design, fully implementing all the aforementioned coupling innovation mechanisms, and simultaneously possessing all the aforementioned beneficial technical effects:
[0172] (1) Construct a standardized and quantitative risk zoning system to break through the limitations of traditional qualitative prevention and control zoning.
[0173] This invention utilizes oblique photogrammetry and stereo matching to generate a 1m×1m continuous DEM grid. It features a unique four-dimensional weighted quantitative model incorporating topography, vegetation, soil and rock, and construction disturbance. This model automatically calculates the comprehensive sensitivity coefficient S for each grid cell and divides the area into high, medium, and low risk zones based on a unified numerical standard. Unlike traditional soil and water conservation zoning that only delineates boundaries and cannot subdivide the area within the site, this invention is not constrained by site vector boundaries. It allows for precise differentiation of high and low risk grids within the same spoil heap or slope. Simultaneously, it dynamically updates risk levels monthly based on construction disturbance data, providing a unified and reproducible quantitative correction benchmark for differentiated early warning throughout the entire process, filling the gap in existing technologies for risk assessment standards in chemical industrial zones.
[0174] (2) Three-dimensional sensitive coupling dynamic fusion to solve the inundation defects of weak water and soil risk characteristics.
[0175] This invention couples the raster sensitivity coefficient with temporal timeliness and spatial accuracy to construct a three-dimensional dynamic weighting mechanism. Based on S-improved Kalman temporal registration and DS evidence fusion logic, it can adaptively amplify the weights of highly sensitive features such as slope displacement and spoil erosion according to the risk level of the plot, while suppressing irrelevant and redundant information such as temperature and humidity. Compared with industry fixed-weight fusion schemes, this invention significantly improves the feature fidelity of multi-source data fusion, and can completely preserve early soil erosion features that are easily obscured, such as small deformations of steep slopes and small amounts of surface spoil, thereby improving the quality of early warning underlying data from the data source.
[0176] (3) The secondary coupled grid sensitivity coefficient, along with the differential correction formula of increasing and decreasing, significantly reduces false alarms and missed alarms caused by the uniform threshold across the entire domain.
[0177] To address the differences in the disaster-causing mechanisms of two types of erosion carriers—excavation disturbance and loose waste—this invention designs… Inhibitory This invention utilizes an enlarged paired correction formula, coupled with sensitive coupling quantification rules for vegetation, displacement, and rainfall, to achieve a tiered judgment system that is "stricter in highly sensitive areas and more lenient in gentler terrain." Traditional technologies use the same set of judgment criteria for all indicators, easily overlooking minor disturbances on steep slopes and frequently generating false alarms during routine excavation at material yards. This invention adaptively adjusts the risk threshold of indicators through sensitivity coefficients, and simultaneously combines this with a secondary coupled comprehensive risk assessment model, significantly improving the accuracy of tiered judgment. The accuracy rate of engineering measurement and early warning is consistently maintained within the range of 86% to 89%.
[0178] (4) A unique multi-factor composite fitness function is used to realize the dynamic threshold adaptive iteration of partitioning.
[0179] Abandoning the well-known approach of genetic algorithms that solely optimize for minimizing prediction error, this invention couples three scenario-specific factors—early warning error, grid sensitivity level, and monthly construction disturbance change rate—to construct a composite fitness function. This function then iterates dynamically on early warning thresholds independently for high, medium, and low-risk grids. Traditional global static thresholds cannot adapt to changes in construction excavation or flood season rainfall, resulting in a significant decline in accuracy after three months of operation. This invention, by independently optimizing thresholds for each region, can adapt to dynamic disturbance changes during the construction period, effectively mitigating the long-term decline in early warning accuracy.
[0180] (5) Partition-based targeted closed-loop self-optimization to achieve long-term adaptive operation of the system.
[0181] This invention differentiates between missed alarms, false alarms, and operational condition adaptation deviations based on risk type, and establishes a tiered verification workflow. It updates the fusion weights, quantization parameters, and iteration coefficients only for grid areas with deviations, without altering parameters in other stable work areas. Unlike existing full-domain synchronous iteration schemes, this significantly reduces invalid parameter adjustments, lowers computing power consumption, completes a round of autonomous optimization monthly, exhibits no significant performance degradation in early warning throughout the entire monitoring cycle, and demonstrates stronger long-term operational stability.
[0182] (6) The complete coupling system has strong versatility and is suitable for the entire life cycle of various types of hydropower projects.
[0183] The entire four-dimensional sensitive coupling algorithm framework is not bound to a single project type. It can be adapted to various types of hydropower stations such as water diversion, pumped storage, and roller compaction dams by simply replacing the basic parameters of watershed DEM, geotechnical engineering, and construction disturbance. At the same time, it takes into account two scenarios: dynamic excavation monitoring during construction and steady-state water and soil management during operation. There is no need to rebuild the underlying logic of integration, analysis, and iteration. The engineering reuse cost is low and the application scope is wide.
[0184] (7) Full-chain quantitative closed loop, with outstanding engineering practicality and control value.
[0185] This invention forms a complete digital process from aerial photography data raster generation, sensitivity coefficient calculation, feature fusion, risk assessment to closed-loop rectification. All calculation rules, parameters, and grading standards are quantifiable and implementable. During pilot operation, it can identify potential soil erosion hazards such as slope erosion and spoil displacement in advance, and is equipped with a graded and time-limited verification mechanism. The closed-loop hazard handling is highly efficient, and it can avoid secondary disasters such as landslides and debris flows in advance, providing complete technical support for the digital management of soil and water conservation in hydropower projects.
[0186] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.
Claims
1. A dynamic early warning method for soil erosion in hydropower stations based on work area-sensitive quantitative coupling, characterized in that, Includes the following steps: Step S1: Identify typical monitoring bodies in the hydropower station basin, including slopes and stockpiles; the stockpiles include spoil heaps and material yards; deploy multi-source IoT monitoring equipment, including AI vision cameras, rainfall sensors, and displacement sensors, in each monitoring body; use oblique photography drones to collect aerial remote sensing images of the hydropower station basin, and acquire multi-source IoT monitoring data collected by the multi-source IoT monitoring equipment for each monitoring body, including cumulative slope displacement, stockpile outline, and rainfall; Step S2: The U-Net model is used to classify and identify the aerial remote sensing images of the hydropower station basin, and the type of each grid is identified, including construction disturbance area, slope area, spoil disposal area, material yard area and vegetation features. Combined with DEM data and the multi-source IoT monitoring data, the water and soil loss monitoring feature set of each grid is obtained through spatiotemporal registration, including topographic slope, cumulative slope displacement monitoring value, grid period cumulative rainfall, measured value of construction disturbance area, measured value of vegetation coverage and pile accumulation volume. Step S3: Construct a set of sensitive factors including terrain slope, vegetation soil stabilization capacity, soil and rock stability, and monthly construction disturbance frequency; Set fixed weights for each sensitive factor, and determine the sensitivity coefficient of each sensitive factor for each grid based on the soil and water loss monitoring feature set and grid attribute data for each grid. The sensitivity coefficients are fused according to the fixed weighting to obtain the comprehensive sensitivity coefficient of each grid. Combined with the standardized risk classification system, the soil erosion risk level of each grid is determined. The soil erosion risk level includes high risk level, medium risk level and low risk level. Based on the soil and water loss risk level of each grid, a global soil and water loss risk layer with a comprehensive sensitivity coefficient is generated. Each layer consists of grids with the same soil and water loss risk level. Step S4: Construct a dynamic weighted fusion mechanism for the time series timeliness coefficient, spatial imaging accuracy coefficient, and comprehensive sensitivity coefficient to correct each type of soil erosion monitoring feature of each grid, and obtain the corrected value of each type of soil erosion monitoring feature of each grid to construct the fusion feature vector of each grid. Step S5: Construct a set of soil erosion risk indicators, including construction disturbance area risk, vegetation degradation risk, slope slippage risk, erosion risk of the embankment, and rainfall-induced erosion risk; determine the weight of each soil erosion risk indicator based on the analytic hierarchy process; and introduce a comprehensive sensitivity coefficient to construct a comprehensive risk assessment model that weights and integrates each soil erosion risk indicator. Based on the fused feature vector of each grid, the quantitative value of each soil and water loss risk index of each grid is calculated, and then the comprehensive soil and water loss risk value of each grid is obtained by using a comprehensive risk assessment model. Step S6: Introduce a comprehensive sensitivity coefficient, substitute it into the composite fitness function, and use a genetic algorithm to iteratively solve the dynamic early warning threshold corresponding to each soil and water loss risk level. Step S7: Set up a random forest sub-model corresponding to each soil erosion risk level. The random forest sub-model outputs graded early warning information based on the comprehensive soil erosion risk value of each grid and its dynamic early warning threshold. Step S8: Based on the on-site verification results, classify and identify the early warning deviations of the graded early warning information and conduct source analysis, and then adjust the local parameters accordingly.
2. The method for dynamic early warning of soil erosion in hydropower stations based on work area sensitive quantitative coupling according to claim 1, characterized in that, The acquisition cycle for oblique photography drones to collect aerial remote sensing images of the hydropower station basin is 7 days; the acquisition cycle for AI vision cameras and displacement sensing terminals is 15 minutes; and the acquisition cycle for rainfall sensing terminals is hours.
3. The method for dynamic early warning of soil erosion in hydropower stations based on work area sensitive quantitative coupling according to claim 1, characterized in that, The fixed weighting of each sensitive factor is as follows: topographic slope is 0.3, vegetation soil stabilization capacity is 0.25, rock and soil stability is 0.25, and monthly construction disturbance frequency is 0.
2. The standardized risk classification system is as follows: high risk level, S∈[0.8,1.0] and terrain slope≥25°; medium risk level, S∈[0.5,0.8); low risk level, S∈[0.2,0.5); where S is the comprehensive sensitivity coefficient of the raster.
4. The method for dynamic early warning of soil erosion in hydropower stations based on work area sensitive quantitative coupling according to claim 1, characterized in that, The time series timeliness coefficient is assigned as follows: it is assigned in segments based on the data update interval of the corresponding data source and the degree of missing historical data. The more frequent the data update and the lower the missing rate, the higher the time series coefficient value. The spatial imaging accuracy coefficient is assigned in the following way: it is graded according to the stability of the on-site camera deployment and the degree of image interference including mountain shadows and fog. The less imaging interference and the more stable the point deployment, the higher the value of the spatial imaging accuracy coefficient.
5. The method for dynamic early warning of soil erosion in hydropower stations based on work area sensitive quantitative coupling according to claim 1, characterized in that, The quantitative values of each soil and water loss risk indicator for each grid are calculated as follows: Construction disturbance area risk quantification value : ; in: This is the corrected measured value of the construction disturbance area; The baseline value for the unified disturbance area of the watershed; S is the comprehensive sensitivity coefficient of the grid; Characterizes the extent to which construction disturbance exceeds the baseline; This is a correction item; Quantitative value of vegetation degradation risk : ; in: This is the corrected measured value of vegetation cover; This serves as the baseline value for native vegetation cover in the watershed. Quantitative value of slope slip risk : ; in: The corrected cumulative displacement monitoring value of the slope; This is the benchmark value for the safe displacement of the slope; Quantitative value of scavenging risk : ; in: This is the corrected stack volume; This represents the maximum allowable stacking volume for a single grid cell. This is a correction item; Rainfall-induced erosion risk quantification value : ; in: This is the cumulative rainfall over the corrected grid period; This is the critical rainfall amount for a rainstorm in the basin.
6. The method for dynamic early warning of soil erosion in hydropower stations based on work area sensitive quantitative coupling according to claim 5, characterized in that, The comprehensive risk assessment model is as follows: ; in: This represents the quantified value of the i-th category of soil and water loss risk indicator; The fixed weights are assigned to the i-th type of soil erosion risk indicators; This is a comprehensive risk value for soil and water loss.
7. The method for dynamic early warning of soil erosion in hydropower stations based on work area sensitive quantitative coupling according to claim 1, characterized in that, The composite fitness function is: ; ; in: For composite fitness; This corresponds to the historical early warning error factor of the risk grid; This represents the time-series change rate of construction disturbances for the current month. This represents the total area of construction disturbance caused by the corresponding risk grid in the current month. This represents the total area of construction disturbance during the same period last month; α is the overall sensitivity coefficient of the raster; β is the error correction weight; β is the scene adaptation weight; and γ is the dynamic weight of the construction condition.
8. The method for dynamic early warning of soil erosion in hydropower stations based on work area sensitive quantitative coupling according to claim 1, characterized in that, When adjusting local parameters using feedback, only the parameters corresponding to the soil erosion risk level are adjusted, including the dynamic early warning threshold, the model parameters of the random forest sub-model, and the weights of each item in the composite fitness function.
9. A dynamic early warning system for soil erosion in hydropower stations based on work area-sensitive quantitative coupling, characterized in that, This method is used to implement the dynamic early warning method for soil erosion in hydropower stations based on work area sensitive quantitative coupling as described in any one of claims 1 to 8.
