Basin water level prediction system and method for complex terrain

Abstract

The application relates to the technical field of water level prediction, and particularly discloses a basin water level prediction system and method for complex terrains, wherein the method is characterized in that: a feature screening algorithm is used to screen a preliminary model feature set, and effective features which have a significant influence on the transit of a flood peak and the risk of backflow are reserved; the effective features are scientifically divided into a training set and a test set; the training set is used to enable the model to fully learn the mapping relationship between the buffer state of a buffer confluence unit, the flood discharge, the movement of the flood peak and the change of the confluence water level under complex terrains; and the fusion of hydrological data and topographic image data covers key influence factors in the time and space dimensions on the data level; the water flow lagging effect is captured through an upstream water level lagging feature set on the feature level; the discharge position is accurately positioned in combination with the discharge channel and the buffer state of the buffer confluence unit; and the transit path of the flood peak and the peak water level are simulated, so that the accurate simulation of the transit of the flood peak in the complex terrain and the reliable prediction of the backflow risk can be realized.

Description

Technical Field

[0001] This invention relates to the field of water level prediction technology, and in particular to a watershed water level prediction system and method for complex terrain. Background Technology

[0002] With the rapid development of artificial intelligence technology and the gradual coverage of watershed rainfall and water level monitoring systems, machine learning for river water level prediction has been applied to flood prevention and early warning systems for rivers at risk of flooding. For example, Chinese patent application CN115713164A discloses a method for predicting downstream water levels in a watershed. Based on upstream water level characteristics, the upstream water level characteristics are gradually shifted forward by several time units while the downstream water level data remains unchanged, constructing upstream water level lag characteristics. Then, the correlation coefficient between the constructed upstream water level lag characteristics and the downstream water level data is calculated, and the upstream water level lag characteristics with correlation coefficients higher than a set value are selected to form the upstream water level lag feature set. The upstream water level lag feature set, together with rainfall characteristics, flood discharge characteristics, upstream water level characteristics, and downstream water level data, forms a preliminary model feature set. The preliminary model feature set is then subjected to lasso feature filtering to obtain the filtered model feature set. The filtered model feature set is used to train and test the water level prediction model to obtain a trained water level prediction model. The trained water level prediction model is then used to predict the downstream water level of the watershed.

[0003] In the aforementioned technical solutions, watershed water level prediction mainly revolves around basic data such as upstream and downstream water levels, rainfall, and flood discharge. Prediction effectiveness is improved by constructing water level lag characteristics or employing deep learning models. However, in the complex scenario of a mountainous watershed where two rivers converge, the rapid flood discharge and confluence of the mountainous terrain, coupled with the initial slow-then-rapid water level changes caused by terraced buffering, make it difficult to accurately capture the nonlinear characteristics of the water level lag effect. Meanwhile, flood discharge in mountainous terrain is characterized by sudden onset and short duration. Data from fixed measuring points suffers from transmission delays, making it impossible to capture the dynamic changes at the flood discharge outlets in real time. This results in delayed identification of the flood discharge location and untimely calculation of the discharge volume. The buffering effect of terraces also affects the discharge volume. Specifically, the overall topography of mountainous terrain forms several terrace-like hierarchical structures. When the terraces are not saturated, the water flow is buffered, and the discharge volume increases slowly. Once saturated, the water flow converges directly, and the discharge volume increases sharply. The linear lag characteristics constructed by the upstream water level shift are difficult to adapt to this nonlinear abrupt change, and it is even more difficult to distinguish the difference in discharge volume between the buffer period and the saturation period, resulting in low accuracy in flood discharge volume prediction.

[0004] In addition, flood peak simulation requires accurate acquisition of three key parameters: peak discharge volume of the two rivers, peak time, and distribution of discharge locations. Due to the rapid rise and fall of flood discharge in mountainous areas, it is necessary to accurately capture the critical point where the discharge volume increases from a slow, buffered increase to a sudden surge to saturation. The above-mentioned schemes do not adapt to the abrupt changes in terraced fields, resulting in the predicted peak discharge volume being too low or too high, and the peak time being too early or too late, leading to prediction deviations between the peak and peak time. Furthermore, the flood peak superposition effect of the topographic parameters in the confluence area of ​​the two rivers is directly related to the discharge location, and the distribution of discharge locations directly affects the water flow confluence path. However, the above-mentioned schemes only model the upstream and downstream unidirectional relationship, ignoring the differentiated impact of different discharge locations on the confluence flow, making it difficult for the confluence path simulation to reflect the true flood peak superposition situation.

[0005] Therefore, there is an urgent need for a watershed water level prediction system and method that can accurately simulate the passage of flood peaks and predict the risk of backflow. Summary of the Invention

[0006] This invention provides a watershed water level prediction system and method for complex terrain, which can accurately simulate the passage of flood peaks and predict the risk of backflow.

[0007] To solve the above-mentioned technical problems, this application provides the following technical solution: A watershed water level prediction method for complex terrain includes the following steps: Step 1: Acquire hydrological data and topographic image data within the watershed; Step 2: Preprocess the hydrological data and topographic image data, unify the acquisition time precision and sort them; analyze the topographic image data to extract relevant data of the buffer confluence unit, flood discharge channel and topographic parameters of the confluence area. The relevant data of the buffer confluence unit includes buffer land, buffer capacity, hierarchical structure and topographic slope; analyze the hydrological data to extract rainfall per unit time, rainfall time, upstream water level data, downstream water level data and upstream flow data; construct a water volume growth model based on the relevant data of the buffer confluence unit, calculate the real-time cumulative water volume according to the rainfall per unit time and rainfall time, and then compare the real-time cumulative water volume with the buffer land capacity to obtain the buffer status of the buffer confluence unit. Step 3: Using upstream water level data as the upstream water level feature, gradually shift the time unit backward, and use downstream water level data as the downstream water level feature while keeping it unchanged, construct multiple upstream water level lag features with different lag levels. Calculate the correlation coefficient between the upstream water level lag features and the downstream water level features, and select upstream water level lag features with correlation coefficients higher than a preset threshold to form an upstream water level lag feature set. Combine the hydrological data, the upstream water level lag feature set, and related data from the buffer confluence unit, the flood discharge channel, the topographic parameters of the confluence area, and the real-time cumulative water volume to form a preliminary model feature set. Step 4: Establish a feature selection algorithm. Input the preliminary model feature set into the feature selection algorithm for fitting training, and retain the effective features with non-zero feature coefficients. Divide the effective features into training set and test set, and obtain the water level prediction model by training the training set and validating it with the test set. Step 5: Determine the flood discharge location based on the water level prediction model combined with the buffer status of the flood discharge channel and the buffer confluence unit; integrate upstream flow data, flood discharge location and topographic parameters of the confluence area to simulate the passage of the flood peak and output the passage time, peak water level and backflow risk level.

[0008] The basic principles and beneficial effects of this scheme are as follows: In this scheme, considering the nonlinear hydrological characteristics of the buffer confluence unit's topographic buffer-saturation, relevant data (including buffer area range and actual capacity) of the buffer confluence unit are extracted through UAV topographic image data analysis. Based on the constructed water growth model, combined with the pre-processed unit-time rainfall and rainfall time, the real-time cumulative water volume is dynamically calculated. Then, by comparing the real-time cumulative water volume with the threshold of the buffer area capacity, the buffer state (saturated / unsaturated) of the buffer confluence unit is clarified. This scheme can accurately adapt to the unique pattern of slow water growth during the buffer period and rapid water confluence after saturation in complex terrain, avoiding the prediction deviation of water volume changes caused by the difficulty in depicting the terrain. This provides core support for the accurate calculation of the subsequent flood peak formation time and peak water volume. That is, the short flood discharge time and high intensity in mountainous terrain are due to the concentrated confluence after the buffer confluence unit is saturated. By accurately judging the buffer state of the buffer confluence unit, the timing of flood discharge and the maximum flood discharge volume can be predicted in advance, laying a key foundation for flood peak simulation.

[0009] Meanwhile, the essence of backflow is the "two-way backwater effect," meaning that water from the higher-level side of the confluence flows back into the lower-level side. Predicting backflow requires two conditions: first, accurately calculating the total water level after the superposition of the two river flood peaks; and second, determining whether the water level exceeds the critical value of the confluence's topography. Deviations in flood discharge prediction directly lead to inaccurate calculations of the total water level after the superposition of the two river flood peaks. Underestimating the flood discharge will result in a misjudgment that the total water level has not reached the backflow critical value; overestimating it may cause unnecessary panic. Furthermore, whether backflow occurs depends not only on the total water level but also on the topographic parameters of the confluence area. In the model building phase, a feature selection algorithm was first used to filter the preliminary model feature set, which integrates hydrological data, upstream water level lag feature sets, buffer confluence unit data, flood discharge channels, and topographic parameters of the confluence area. Effective features that significantly impact flood peak passage and backflow risk were retained. These effective features were then scientifically divided into training and testing sets. The training set allowed the model to fully learn the mapping relationship between buffer state of the buffer confluence unit, flood discharge, flood peak movement, and confluence water level changes under complex terrain. The testing set validated and optimized the model parameters, improving the model's generalization ability and prediction accuracy, and avoiding simulation distortion caused by feature redundancy or insufficient training. Furthermore, at the data level, the fusion of hydrological data and topographic image data covers key influencing factors in both time and space dimensions. At the feature level, the lag effect of water flow is captured through the upstream water level lag feature set. In the prediction stage, based on a trained water level prediction model, the flood discharge location is accurately located by combining the buffer status of the flood discharge channel and the buffer confluence unit. Then, the upstream flow data and the topographic parameters of the confluence area are integrated to simulate the passage path and peak water level of the flood peak. Finally, the risk of backflow is predicted by judging whether the peak water level exceeds the critical value of the confluence area topography. It can adapt to the hydrological and topographic characteristics of complex terrain, thereby achieving accurate simulation of the passage of flood peaks in complex terrain and reliable prediction of backflow risk.

[0010] Furthermore, in step 2, the construction of a water volume growth model based on relevant data of the buffer confluence unit includes: constructing a phased water volume growth model based on the buffer land, buffer capacity, hierarchical structure, and terrain slope of the buffer confluence unit; the phased water volume growth model sets two calculation modes: a buffer unsaturation stage and a buffer saturation stage. In the buffer unsaturation stage, a linear growth formula is used to calculate the real-time cumulative water volume, and in the buffer saturation stage, a nonlinear confluence formula is used to calculate the real-time cumulative water volume.

[0011] The beneficial effects are: it can accurately adapt to the phased hydrological characteristics of buffer-saturation of buffer confluence units: a phased model is constructed based on core parameters such as buffer capacity and hierarchical structure, and the water volume of different stages is calculated by combining linear and nonlinear formulas, thereby reducing the prediction deviation of water volume and capturing the dynamic changes of water volume in buffer confluence units under complex terrain.

[0012] Furthermore, in step 2, the calculation of real-time cumulative water volume based on rainfall per unit time and rainfall duration includes: obtaining rainfall per unit time and rainfall duration, and using a linear cumulative formula during the unsaturated buffer phase. The real-time cumulative water volume is calculated using a nonlinear confluence formula during the buffer saturation stage. Calculate the real-time cumulative water volume; where, This is the buffer coefficient. Rainfall per unit time For the duration of rainfall, To maximize the buffer zone capacity, This is the confluence coefficient.

[0013] The beneficial effects are as follows: the linear formula is adapted to the slow water storage characteristics of the unsaturated stage, the nonlinear formula is adapted to the rapid confluence characteristics of the saturated stage, and the combination of quantitative parameters of rainfall and rainfall time can provide high-precision data support for the judgment of buffer status, improve the accuracy of subsequent flood peak superposition total water level calculation, and enhance the reliability of backflow risk prediction under complex terrain.

[0014] Furthermore, in step 3, constructing multiple upstream water level lag features with different lag levels includes: progressively shifting the upstream water level data backward by one time unit to generate multiple sets of upstream water level data with different time offsets; using each set of upstream water level data with different time offsets as an upstream water level lag feature of a lag level, forming a set of upstream water level lag features containing a preset number of different lag levels; and keeping the downstream water level data unchanged from its original time series, using it as a benchmark feature to correspond and match with the upstream water level lag features of each lag level.

[0015] The beneficial effects are as follows: by using this method, the time unit, step size and number of lag levels can be clearly defined, making the feature construction more operable and avoiding insufficient feature effectiveness caused by parameter ambiguity; multi-level lag features comprehensively cover short- and medium-term water level lag effects, accurately match with downstream benchmark features, improve the pertinence of upstream water level lag feature sets, and enhance the fitting accuracy of water level prediction models and the accuracy of flood peak passage simulation.

[0016] Furthermore, in step 3, the calculation of the correlation coefficient between the upstream water level lag characteristics and the downstream water level characteristics includes: using the Pearson correlation coefficient algorithm to calculate the linear correlation between the upstream water level lag characteristics and the downstream water level characteristics for each lag level, and obtaining the corresponding correlation coefficient value.

[0017] The beneficial effects are as follows: the Pearson correlation coefficient can quantify the degree of linear correlation between upstream and downstream water levels, eliminate redundant features with weak correlation, retain core features that have a significant impact on downstream water levels, optimize the quality of the initial model feature set, and reduce the interference of invalid features on subsequent model training.

[0018] Further, in step 4, the establishment of the feature selection algorithm involves inputting the preliminary model feature set into the feature selection algorithm for fitting training, retaining valid features with non-zero feature coefficients, including: selecting a sparse feature selection algorithm model as the feature selection algorithm, setting the value range of the regularization coefficient; inputting the preliminary model feature set into the sparse feature selection algorithm model after standardization, compressing the feature coefficients through L1 regularization constraints; iteratively training until convergence, selecting features with non-zero feature coefficients to form a valid feature set.

[0019] The beneficial effects are as follows: clarifying the type and core parameters of feature selection algorithms improves the accuracy and repeatability of feature selection; the sparse feature selection algorithm combined with L1 regularization constraints can efficiently eliminate redundant features, balancing the rigor and effectiveness of feature selection; the effective feature set after selection focuses on core influencing factors, reducing the computational load of subsequent model training and improving the fitting efficiency and prediction accuracy of the water level prediction model.

[0020] Furthermore, in step 4, the step of dividing the effective features into a training set and a test set, and obtaining the water level prediction model by training the training set and validating the test set, includes: dividing the effective features into a training set and a test set according to a preset ratio; initializing the water level prediction model using a machine learning architecture; fitting the model parameters with the training set data; iteratively optimizing the model loss function until convergence using the gradient descent method; inputting the test set data into the trained model; verifying the model prediction error; and determining the final water level prediction model when the error is lower than a preset threshold.

[0021] The beneficial effects are as follows: dividing the model according to the preset ratio takes into account both the sufficiency of model training and the reliability of validation; the iterative optimization of the gradient descent method can reduce the model fitting error; the error verification of the test set can ensure the generalization ability of the model, avoid overfitting or underfitting the model, improve the adaptability of the water level prediction model to the hydrological laws of complex terrain watersheds, and provide more accurate model support for flood peak simulation and backflow risk prediction.

[0022] Furthermore, in step 5, the determination of the flood discharge location based on the water level prediction model and the buffer status of the flood discharge channel and the buffer confluence unit includes: determining whether the buffer status of the buffer confluence unit is saturated based on the predicted water level output by the water level prediction model; if it is saturated, the flood discharge channel with the closest spatial distance to the buffer confluence unit and the largest flood discharge cross-sectional area is selected as the flood discharge location; if it is not saturated, the flood discharge channel corresponding to the predicted water level exceeding the preset safety threshold is selected as the flood discharge location.

[0023] The beneficial effects are as follows: by combining the dual indicators of the saturation state of the buffer confluence unit and the predicted water level, the decision-making bias caused by relying on a single indicator is avoided. Under saturation conditions, selecting the nearest large-section channel can quickly divert accumulated water, while under unsaturation conditions, selecting the channel according to the threshold can prevent risks in advance. This improves the scientific nature and timeliness of flood discharge decisions, provides an accurate spatial starting point for subsequent flood peak simulation, and ensures the reliability of backflow risk prediction.

[0024] Furthermore, in step 5, the process of integrating upstream flow data, flood discharge location, and topographic parameters of the confluence area to simulate the passage of the flood peak and output the passage time, peak water level, and backflow risk level includes: based on a fluid dynamics confluence model, inputting upstream flow data, flood discharge capacity parameters of the flood discharge location, and topographic parameters of the confluence area; simulating the propagation path and water level change process of the flood peak from the flood discharge location to the confluence, calculating the passage time and peak water level; comparing the peak water level with the critical water level of the confluence topography, and combining the water flow backwater effect judgment rules to classify and output three backflow risk levels: high, medium, and low; based on the classified risk levels, generating a fitted virtual image of the confluence area of ​​the basin, which presents a heat map of the flood peak propagation path, peak water level distribution, and backflow risk level in a three-dimensional visualization form, and simultaneously marking the flood discharge location, critical water level, and risk warning signs.

[0025] The beneficial effects are as follows: by clarifying the core model for flood peak passage simulation and the criteria for determining backflow risk, the accuracy and practicality of prediction are improved; by combining the fluid dynamics model with multi-dimensional key parameters, the propagation law of flood peak under complex terrain can be accurately replicated, avoiding deviations in passage time and peak water level caused by simplified simulation; based on the classification of terrain critical water level and backwater effect, the limitation of simply outputting water level values ​​without being able to determine backflow risk is avoided, providing a quantitative basis for flood control decision-making and strengthening the scientific nature of flood control early warning in complex terrain basins. Attached Figure Description

[0026] Figure 1 This is a flowchart illustrating an embodiment of a watershed water level prediction method for complex terrain.

[0027] Figure 2 This is a topographical diagram of a buffer confluence unit. Detailed Implementation

[0028] The following detailed description illustrates the specific implementation method: This invention provides a watershed water level prediction method for complex terrain, as shown in the appendix. Figure 1 As shown, this embodiment uses a mountainous watershed as the application scenario, including a multi-level terraced buffer confluence unit, three river confluences, and five flood discharge channels.

[0029] Step 1: Acquire hydrological data and topographic image data within the watershed. Hydrological data for the past three years was obtained by deploying hydrological monitoring stations within the basin, including rainfall per unit time, rainfall duration, upstream water level data, downstream water level data, and upstream flow data. High-definition remote sensing equipment was used to carry out full-coverage aerial photography of the basin using drones to obtain topographic image data. The aerial photography range covered the entire basin and the surrounding 5km buffer zone, ensuring complete capture of the topographic features of the buffer confluence units, flood discharge channels, and confluence areas.

[0030] Step 2: Preprocess the hydrological data and topographic image data, unify the acquisition time precision and sort them; analyze the topographic image data to extract relevant data of the buffer confluence unit, flood discharge channel and confluence area topographic parameters. The relevant data of the buffer confluence unit includes buffer land, buffer capacity, hierarchical structure and topographic slope; analyze the hydrological data to extract rainfall per unit time, rainfall time, upstream water level data, downstream water level data and upstream flow data; construct a water volume growth model based on the relevant data of the buffer confluence unit, calculate the real-time cumulative water volume according to the rainfall per unit time and rainfall time, and then compare the real-time cumulative water volume with the buffer land capacity to obtain the buffer status of the buffer confluence unit.

[0031] First, the hydrological and topographic image data were preprocessed. The acquisition time precision of all hydrological data was standardized to 10 minutes per acquisition, and the data was sorted in ascending order by timestamp. Abnormal data caused by equipment malfunctions, such as data showing sudden water level changes exceeding reasonable ranges, were removed. ArcGIS software was used to analyze the topographic image data, extracting relevant data on buffer confluence units, flood discharge channel parameters, and topographic parameters of the confluence area. The buffer confluence units are terraced fields within the watershed, and a buffer capacity was assumed. =5000m 3 The terrain consists of eight steps, with each step having a height difference of 0.8-1.2 meters. The slope of the terrain ranges from 15° to 22°. (See attached image.) Figure 2 As shown; the flood discharge channel parameters include the cross-sectional area, length, and spatial coordinates of 5 channels; the topographic parameters of the confluence area include the elevation, slope, and cross-sectional shape of 3 confluence points. Analysis of the hydrological data yielded: the rainfall per unit time P ranged from 5-30 mm / 10 min, the rainfall time t was the duration of continuous rainfall (cumulative in 10-minute increments), the upstream water level ranged from 1.2-3.8 m, the downstream water level ranged from 0.8-2.9 m, and the upstream flow rate ranged from 15-80 m³ / min. 3 / s.

[0032] In this embodiment, buffer capacity This represents the maximum water storage capacity of the buffer runoff collection unit (terraced fields), determined through a combination of topographic surveying and hydrological calculations: First, aerial topographic image data from drones was used, combined with ArcGIS software, to analyze the planar area, hierarchical structure, and effective water storage depth of each terrace; second, based on the calculation standards for runoff interception in water conservancy projects, the value was calculated using the formula... Calculate the total capacity, where For the first The horizontal projected area of ​​the terraced fields. The effective water retention height of this level of terraced field embankment is taken as 1.0m, the midpoint between 0.8-1.2m. Surveying shows that the total horizontal projected area of ​​the 8-level terraced fields is approximately 5000㎡. Based on an effective water retention height of 1.0m, the theoretical maximum water storage capacity is 5000m³. 3 Therefore, it is assumed The core basis of this assumption is: first, the actual topographical survey data of the terraced fields ensures the accuracy and reliability of the area parameters; second, the conventional ridge height of the terraced fields in the southwestern mountainous area is 0.8-1.2m, which conforms to the actual working conditions of local agriculture and water conservancy projects; and third, a safety redundancy of about 10% is reserved to avoid the ridges overflowing under extreme rainfall and ensure the safety of the buffer state determination.

[0033] A phased water volume growth model is constructed based on relevant data from the buffer confluence unit, setting two calculation modes: an unsaturated buffer stage and a saturated buffer stage. The unsaturated buffer stage uses a linear growth formula to calculate the real-time cumulative water volume, while the saturated buffer stage uses a nonlinear confluence formula. The specific calculation formulas are as follows: Buffer unsaturation stage:

[0034] Buffer saturation phase:

[0035] In the formula, k is the buffer coefficient, which is determined according to the hierarchical structure of the buffer confluence unit and the terrain slope. In this embodiment, k=0.75; P is the rainfall per unit time (unit: mm / 10min); t is the rainfall duration (unit: 10 minutes, i.e., t=1 means 10 minutes of rainfall). The maximum capacity of the buffer zone is denoted as 'a', where 'a' is the confluence coefficient, determined based on the cross-sectional area and slope of the flood discharge channel. A larger cross-sectional area allows for a greater maximum flow rate per unit time, resulting in faster water flow and a higher 'a' value. Conversely, a smaller cross-sectional area weakens the confluence capacity, leading to a lower 'a' value. A steeper slope in the buffer zone increases the gravity-driven flow, resulting in faster water flow and a higher 'a' value. Conversely, a gentler slope increases flow resistance, slows the flow, and lowers 'a' value. A positive correlation function between 'a' and the cross-sectional area and slope is constructed. The function parameters are determined using the least squares method, combined with historical hydrological data, ultimately yielding the specific value of 'a'. In this embodiment, the selected flood discharge channel cross-sectional area is 12 m². 2 It belongs to the medium to large size, and the corresponding terrain slope is 18°, which is a medium slope. After fitting calculation, the confluence coefficient a=0.08 is determined.

[0036] Confluence coefficient and the cross-sectional area of ​​the flood discharge channel Topographic slope Positive correlation, model building ,in, The parameters are to be fitted. First, 10 sets of historical hydrological data from the past 3 years were collected for the watershed, each set containing... and actual measurement The least squares method is used to solve for the optimal parameters, with the objective of minimizing the sum of squared errors between the model-calculated values ​​and the measured values. In this embodiment, the fitted result is: Substitute the operating conditions, Calculated Finally determined Buffer coefficient It is a core parameter for quantifying the ability of a buffer and runoff collection unit (terraced field group) to intercept rainfall and delay runoff. Its value is determined by two core factors: the more terraced levels and the higher the effective interception height of each ridge, the greater the total water interception and the more significant the runoff delay effect. The larger the value, the fewer the levels and the lower the interception height. The smaller the value, the gentler the terrain slope, the longer the water stays within the terraced fields, and the higher the water retention efficiency. The larger the value, the steeper the slope, the faster the gravity-driven water flow downstream, reducing the interception efficiency. The smaller the value.

[0037] The calculation logic of "baseline coefficient × slope correction coefficient" is adopted: based on the number of terrace levels and the interception height of the field ridges, the buffer confluence unit in this embodiment is an 8-level terraced field with a height difference of 0.8-1.2m per level, and the interception level is sufficient, so the baseline coefficient is used. A slope of 15°-22° is considered a medium slope, with a corresponding correction factor. 4; Final Calculation The values ​​are consistent with those in the example. If the number of terrace levels is reduced to 5 and the slope increases to 25°-30°, the baseline coefficient is reduced to 0.6, and the correction coefficient is taken as 0.9, resulting in... If the terraced fields increase to 10 levels and the slope decreases to 10°-15°: the baseline coefficient is taken as 0.85, the correction coefficient is taken as 0.95, and the final... .

[0038] If, during a continuous rainfall event, the rainfall per unit time is P = 20 mm / 10 min, and the duration of continuous rainfall is t = 12 time units (i.e., 120 minutes), then first calculate the maximum cumulative water volume during the unsaturated buffer phase (i.e., the buffer capacity Q). max =5000m 3 The critical rainfall duration corresponding to this is approximately 333.3 time units (i.e., 3333 minutes). Since the duration of this rainfall is t=12 < t0, it is in the unsaturated buffer stage, and the real-time cumulative water volume is Q1=180m³. 3 ; Compare real-time cumulative water volume of 180m 3 With a buffer zone capacity of 5000m 3 The buffer bus unit is determined to be in an unsaturated state.

[0039] In this embodiment, the buffer state of the buffer confluence unit can also be accurately identified and switched between unsaturated and saturated states by comparing the real-time accumulated water volume with the buffer capacity and combining the abnormal changes in the flood discharge growth rate. First, the real-time accumulated water volume is calculated based on the phased water volume growth model. and buffer capacity Perform numerical comparison: when At this time, it is determined to be in an unsaturated state. At this time, the terraced fields still have sufficient water storage capacity, all rainfall is intercepted and stored, and the drainage volume of the flood discharge channel remains at a low level. When the flood discharge enters the saturation warning zone, further verification through the flood discharge growth rate is required. Secondly, an indicator of the flood discharge growth rate should be introduced.

[0040] Where η is the flood discharge growth rate. This is the current discharge volume. This represents the discharge volume at the previous moment. If the rainfall per unit time... No increase. However, the growth rate of flood discharge A sudden increase, such as over two consecutive monitoring periods. This indicates that the terraced fields have no water storage capacity, and rainfall is being discharged directly through the flood discharge channels. At this point, the field is determined to be in a saturated state, and the state transition is complete. This approach, which combines water volume thresholds and abnormal flow characteristics, avoids misjudgments based on a single threshold and can accurately capture the transition points between buffer states.

[0041] In this embodiment, the basis for determining the unsaturated state as Q(t) < 0.9Qmax is as follows: 0.9 times the buffer capacity is a 10% safety water storage redundancy reserve; the buffer confluence unit is a terraced water retention facility, which cannot fully utilize its buffer capacity due to soil infiltration, seepage from field ridges, and water surface evaporation. If the saturation threshold is set to... Even a small increase in rainfall can instantly overflow, leaving no buffer or room for adjustment. (0.9) A safety margin is reserved to accommodate actual water storage losses. When the accumulated water volume is below 90% of the total capacity, there are still sufficient water storage gaps within the terraced levels. All rainfall can be intercepted and retained by the terraced levels, and the flood discharge channels will only experience minimal seepage and basic drainage, preventing large-scale overflow. The water flow does not exhibit a significant confluence and acceleration, conforming to the physical laws of linear water storage in the unsaturated stage. (0.9) As a stratification threshold, the regular unsaturated interval and the saturation warning interval are divided; Then, a second verification using the flood discharge growth rate is performed to distinguish between two operating conditions: "nearly full but still capable of holding back water" and "completely saturated and overflowing," avoiding misjudgments based on a single water volume threshold and improving the accuracy of state identification. Based on years of terraced field water storage monitoring data in this watershed, the water volume is below 0.9... At that time, the flow rate of the flood discharge channel remained at a low level, with no sudden increase in flow rate, which was completely consistent with the hydrological performance under unsaturated conditions.

[0042] Step 3: Using upstream water level data as the upstream water level feature, progressively shift the time unit backward, using downstream water level data as the downstream water level feature while keeping it unchanged, constructing multiple upstream water level lag features with different lag levels. Calculate the correlation coefficient between the upstream and downstream water level lag features, and select upstream water level lag features with correlation coefficients higher than a preset threshold to form the upstream water level lag feature set. Combine hydrological data, the upstream water level lag feature set, and related data from buffer confluence units, flood discharge channels, topographic parameters of the confluence area, and real-time cumulative water volume to form a preliminary model feature set. Using upstream water level data as the upstream water level feature, preset... The time unit is 10 minutes, consistent with the data acquisition frequency. The upstream water level data is progressively shifted backward by one time unit, generating 8 sets of upstream water level data with different time offsets, i.e., lag levels 1-8, corresponding to lags of 10 minutes, 20 minutes...80 minutes, respectively. Each set of upstream water level data with different time offsets is used as an upstream water level lag feature for a lag level, forming a set of upstream water level lag features containing 8 different lag levels. The downstream water level data remains unchanged in its original time series and is used as a baseline feature to be matched one-to-one with the upstream water level lag features of each lag level according to timestamps.

[0043] The Pearson correlation coefficient algorithm was used to calculate the linear correlation between the upstream water level lag characteristics and the downstream water level characteristics for each lag level, obtaining the corresponding correlation coefficient values. The Pearson correlation coefficient is used to quantify the linear correlation between two continuous variables. Its core logic is as follows: first, calculate the covariance of the two sets of data, then divide it by the product of the standard deviations of the two sets of data to eliminate the influence of dimensions, and finally obtain the correlation coefficient with a value range of [-1, 1]. The closer the absolute value is to 1, the stronger the linear correlation.

[0044] The calculation formula is as follows:

[0045] In the formula, This represents the upstream water level lag characteristic sequence at a certain lag level. This is a downstream water level characteristic sequence. , , are the means of the two sets of sequences, This represents the number of samples.

[0046] In this embodiment, for the upstream water level sequences with eight lag levels, the upstream and downstream water level sequences are substituted into the formula for calculation: taking lag level 4 as an example, the covariance and standard deviation of the upstream water level data with a 40-minute lag and the corresponding downstream water level data are calculated point by point, finally obtaining a correlation coefficient of 0.75; similarly, the coefficients of the remaining seven lag levels are calculated, and then a threshold of 0.6 is used for filtering, retaining the features of lag levels 3, 4, and 5 to form a feature set. This algorithm can accurately quantify the linear correlation between upstream and downstream water levels, providing a quantitative basis for feature selection. Suppose the calculation results are as follows: lag level 1 correlation coefficient 0.32, lag level 2 0.45, lag level 3 0.68, lag level 4 0.75, lag level 5 0.72, lag level 6 0.58, lag level 7 0.41, and lag level 8 0.29; with a preset correlation coefficient threshold of 0.6, upstream water level lag features with correlation coefficients higher than 0.6 at lag levels 3, 4, and 5 are selected to form the upstream water level lag feature set. The analyzed hydrological data, the upstream water level lag feature set, buffer confluence unit related data, flood discharge channels, topographic parameters of the confluence area, and real-time cumulative water volume are combined to form a preliminary model feature set, which contains features in 23 dimensions.

[0047] Step 4: Establish a feature selection algorithm. Input the preliminary model feature set into the feature selection algorithm for fitting training and retain effective features with non-zero feature coefficients. Divide the effective features into training set and test set, and obtain the water level prediction model by training the training set and verifying it with the test set.

[0048] A sparse feature selection algorithm (specifically the Lasso algorithm) is chosen as the feature selection algorithm. The regularization coefficient λ is set to a range of 0.01-0.1; in this embodiment, λ=0.05 is selected. The initial model feature set is standardized, for example, using Z-score standardization to make the mean of each feature 0 and the standard deviation 1. This standardization is then input into the Lasso algorithm model, and the feature coefficients are compressed using L1 regularization constraints. Iterative training continues until the model loss function value stabilizes, i.e., the model converges. Features with non-zero coefficients are selected to form an effective feature set. For example, if 12 dimensions of effective features are finally selected, including rainfall per unit time, rainfall time, upstream flow data, upstream water level lag features of 3-5 levels, buffer capacity, terrain slope, cross-sectional area of ​​flood discharge channel, and elevation of confluence, etc. The effective features are divided into training and test sets at a preset ratio of 7:3. A random forest machine learning architecture is used to initialize the water level prediction model, with 100 decision trees and a maximum tree depth of 10. The model parameters are fitted with the training set data, and the model loss function is iteratively optimized using gradient descent until convergence. Iteration stops when the number of iterations reaches 500 or the loss function value is less than 0.001. The test set data is input into the trained model to verify the model's prediction error. If the mean absolute error of the test set is 0.08m, which is lower than the preset threshold of 0.1m, the model is determined to be the final water level prediction model.

[0049] In this embodiment, λ is the L1 regularization strength coefficient of the Lasso algorithm. The value controls the feature compression strength. When λ < 0.01, the regularization constraint is very weak, which cannot compress redundant features and is prone to overfitting. When λ > 0.1, the regularization constraint is too strong, and a large number of key hydrological and topographic feature coefficients are forced to zero, resulting in the loss of effective prediction information and underfitting. Based on the feature dimension and sample size of the watershed water level prediction task, 0.01-0.1 is defined as a reasonable adjustment range. Within the interval, a grid search was used to traverse five sets of parameters: 0.01, 0.03, 0.05, 0.07, and 0.1. Lasso models were trained for each set, and the filtering effect was verified. When λ=0.01 and 0.03, redundant features were not completely removed, resulting in poor model generalization. When λ=0.07 and 0.1, some core features such as lagging water levels and buffer capacity were removed, leading to a decrease in prediction accuracy. λ=0.05 balanced the regularization constraint strength, removing only irrelevant noise features and retaining 12 effective features. The mean absolute error on the test set was 0.08m, achieving the optimal result. This balanced model simplification and prediction accuracy; therefore, λ=0.05 was selected for this implementation example.

[0050] Step 5: Based on the water level prediction model, determine the flood discharge location by combining the flood discharge channel and the buffer status of the buffer confluence unit; integrate upstream flow data, flood discharge location and topographic parameters of the confluence area to simulate the passage of the flood peak and output the passage time, peak water level and backflow risk level.

[0051] Based on the predicted water level output by the water level prediction model, if the upstream water level is predicted to rise to 2.8m within the next 2 hours, and considering the buffer status of the confluence unit (unsaturated), the flood discharge channel corresponding to the predicted water level exceeding the preset safety threshold (preset safety threshold is 2.5m) is selected as the flood discharge location. After matching, the predicted water level in the basin area corresponding to flood discharge channel number 2 will reach 2.8m, exceeding the safety threshold; therefore, this channel is determined as the flood discharge location. Based on the fluid dynamics confluence model (such as the MIKE11 one-dimensional hydrodynamic model), upstream flow data (e.g., upstream flow of 65m³) is input. 3 / s), flood discharge capacity parameters of the flood discharge location (e.g., the flood discharge capacity of flood discharge channel No. 2 is 50m). 3 The simulation included the topographic parameters of the confluence area (e.g., elevation of 580m, slope of 12°, rectangular cross-section, and cross-sectional width of 5m at the confluence); the simulation also included the propagation path and water level change of the flood peak from the discharge point to the confluence; the calculation showed that the flood peak would pass in 1.5 hours and the peak water level would be 581.2m; the peak water level of 581.2m was compared with the critical water level of the confluence (e.g., 581.0m based on the elevation and cross-sectional shape of the confluence); and the backflow effect judgment rule was used (when the peak water level is greater than the critical water level, there is a risk of backflow, and the greater the water level difference, the higher the risk level). Since the peak water level was 0.2m higher than the critical water level, the backflow risk level was determined to be medium and output.

[0052] In this embodiment, based on the high, medium, and low backflow risk levels, a fitted virtual image of the watershed confluence area is generated. Using watershed topographic mapping data as a base, a three-dimensional visualization scene is constructed. The flood peak propagation path simulated by the hydrodynamic confluence model and the peak water level data at each node are mapped onto the scene. High, medium, and low backflow risk areas are marked with red, yellow, and green heat maps, respectively. Simultaneously, the flood discharge location, critical water level markers at the confluence, and risk warning prompts are overlaid. The fitted virtual image intuitively presents the flood peak evolution process from downstream of the No. 2 flood discharge channel to the confluence, clearly demonstrating the spatial range of the medium-level backflow risk area. This provides visual support for flood control scheduling decisions, facilitating staff to quickly grasp the watershed's hydrological and risk situation.

[0053] This embodiment also provides a watershed water level prediction system for complex terrain, which is used to execute a watershed water level prediction method for complex terrain. The system includes a data acquisition module, a data processing module, a feature construction module, a model construction module, and a prediction output module. The modules are connected through a bus to realize data transmission and interaction.

[0054] The data acquisition module includes a hydrological monitoring unit and a remote sensing aerial photography unit. The hydrological monitoring unit consists of hydrological monitoring stations deployed within the basin. Each monitoring station is equipped with a water level sensor, a flow sensor, and a rainfall sensor to collect real-time data on rainfall per unit time, rainfall duration, upstream water level data, downstream water level data, and upstream flow data, and transmits the collected data to the data processing module. The remote sensing aerial photography unit uses DJI drones equipped with full-frame high-definition cameras to conduct full-coverage aerial photography of the basin, acquiring topographic image data with a resolution of 0.1m, and transmitting the image data to the data processing module.

[0055] The data processing module uses an Intel Core i7-12700H processor and is equipped with ArcGIS 10.8 geographic information processing software and Python 3.9 data processing environment. Its functions include: preprocessing hydrological data and topographic image data transmitted by the data acquisition module, unifying the data acquisition time accuracy and sorting, and removing abnormal data; parsing topographic image data to extract relevant data of buffer confluence units, flood discharge channel parameters and confluence area topographic parameters; parsing hydrological data to extract various hydrological characteristic parameters; constructing a phased water volume growth model based on relevant data of buffer confluence units, calculating real-time cumulative water volume by combining rainfall and rainfall time, and comparing the buffer capacity to obtain the buffer status of the buffer confluence units.

[0056] The feature construction module is developed based on the NumPy and Pandas libraries in the Python 3.9 environment. Its functions include: constructing multiple upstream water level lag features with different lag levels based on upstream water level data and according to preset time units and step sizes; calculating the correlation coefficient between each lag feature and the downstream water level feature using the Pearson correlation coefficient algorithm, and selecting features with correlation coefficients higher than preset thresholds to form the upstream water level lag feature set; and combining hydrological data, lag feature set and other topographic and water volume related data into a preliminary model feature set.

[0057] The model building module is deployed on a server equipped with an NVIDIA GeForce RTX 3060 graphics card and developed using the Scikit-learn machine learning library. Its functions include: establishing a sparse feature selection algorithm model (Lasso algorithm model); inputting the initial model feature set into the model after standardization; filtering out effective features with non-zero feature coefficients through L1 regularization constraints; dividing the training set and test set according to a preset ratio; initializing the water level prediction model using a random forest architecture; fitting the model parameters through the training set; optimizing the model loss function until convergence using gradient descent; verifying the model error through the test set; and determining the final water level prediction model.

[0058] The prediction output module includes a decision-making unit and a visualization unit. The decision-making unit, based on the constructed water level prediction model, determines the flood discharge location by combining flood discharge channel parameters and the buffer status of the buffer confluence unit. It then calls the MIKE11 one-dimensional hydrodynamic model, integrating upstream flow data, flood discharge location parameters, and topographic parameters of the confluence area to simulate the flood peak passage process, calculate the flood peak passage time and peak water level, and compares the results with the critical water level to determine the backflow risk level. The visualization unit uses Tableau visualization software to display the flood peak passage time, peak water level, and backflow risk level in chart format (such as line charts, bar charts, and risk level heat maps), and outputs the results in real time through a display screen. The results can also be exported as an Excel file for easy viewing and archiving by staff. This system, through the collaborative work of its modules, can efficiently and accurately execute the prediction method described above. It is suitable for water level prediction and backflow risk prevention in complex terrain basins such as the southwestern mountainous areas, and has high practicality and reliability.

[0059] The above are merely embodiments of the present invention. The invention is not limited to the fields covered by these embodiments. Commonly known structures and characteristics in the solutions are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are able to access all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A watershed water level prediction method for complex terrain, characterized in that, Including the following steps: Step 1: Acquire hydrological data and topographic image data within the watershed; Step 2: Preprocess the hydrological data and topographic image data, unify the acquisition time precision and sort them; analyze the topographic image data to extract relevant data of the buffer confluence unit, flood discharge channel and topographic parameters of the confluence area. The relevant data of the buffer confluence unit includes buffer land, buffer capacity, hierarchical structure and topographic slope; analyze the hydrological data to extract rainfall per unit time, rainfall time, upstream water level data, downstream water level data and upstream flow data; construct a water volume growth model based on the relevant data of the buffer confluence unit, calculate the real-time cumulative water volume according to the rainfall per unit time and rainfall time, and then compare the real-time cumulative water volume with the buffer land capacity to obtain the buffer status of the buffer confluence unit. Step 3: Using upstream water level data as the upstream water level feature, gradually shift the time unit backward, and use downstream water level data as the downstream water level feature while keeping it unchanged, construct multiple upstream water level lag features with different lag levels. Calculate the correlation coefficient between the upstream water level lag features and the downstream water level features, and select upstream water level lag features with correlation coefficients higher than a preset threshold to form an upstream water level lag feature set. Combine the hydrological data, the upstream water level lag feature set, and related data from the buffer confluence unit, the flood discharge channel, the topographic parameters of the confluence area, and the real-time cumulative water volume to form a preliminary model feature set. Step 4: Establish a feature selection algorithm. Input the preliminary model feature set into the feature selection algorithm for fitting training, and retain the effective features with non-zero feature coefficients. Divide the effective features into training set and test set, and obtain the water level prediction model by training the training set and validating it with the test set. Step 5: Based on the water level prediction model, determine the flood discharge location by combining the flood discharge channel and the buffer status of the buffer confluence unit; integrate upstream flow data, flood discharge location and topographic parameters of the confluence area to simulate the passage of the flood peak and output the passage time, peak water level and backflow risk level.

2. The watershed water level prediction method for complex terrain according to claim 1, characterized in that, In step 2, the water volume growth model is constructed based on the relevant data of the buffer confluence unit, including: constructing a phased water volume growth model based on the buffer land, buffer capacity, hierarchical structure and terrain slope of the buffer confluence unit; the phased water volume growth model sets two calculation modes: a buffer unsaturation stage and a buffer saturation stage. The buffer unsaturation stage uses a linear growth formula to calculate the real-time cumulative water volume, and the buffer saturation stage uses a nonlinear confluence formula to calculate the real-time cumulative water volume.

3. The watershed water level prediction method for complex terrain according to claim 2, characterized in that, In step 2, the calculation of real-time cumulative water volume based on rainfall per unit time and rainfall duration includes: obtaining rainfall per unit time and rainfall duration; and using a linear cumulative formula during the unsaturated buffer phase. The real-time cumulative water volume is calculated using a nonlinear confluence formula during the buffer saturation stage. Calculate the real-time cumulative water volume; where, This is the buffer coefficient. Rainfall per unit time For the duration of rainfall, To maximize the buffer zone capacity, This is the confluence coefficient.

4. The watershed water level prediction method for complex terrain according to claim 3, characterized in that, In step 3, constructing multiple upstream water level lag features with different lag levels includes: progressively shifting upstream water level data backward by one time unit to generate multiple sets of upstream water level data with different time offsets; using each set of upstream water level data with different time offsets as an upstream water level lag feature of a lag level, forming a set of upstream water level lag features containing a preset number of different lag levels; and keeping the downstream water level data unchanged from its original time series, using it as a baseline feature to correspond and match with the upstream water level lag features of each lag level.

5. The watershed water level prediction method for complex terrain according to claim 4, characterized in that, In step 3, the calculation of the correlation coefficient between the upstream water level lag characteristics and the downstream water level characteristics includes: using the Pearson correlation coefficient algorithm to calculate the linear correlation between the upstream water level lag characteristics and the downstream water level characteristics for each lag level, and obtaining the corresponding correlation coefficient value.

6. The watershed water level prediction method for complex terrain according to claim 5, characterized in that, In step 4, the establishment of the feature selection algorithm involves inputting the preliminary model feature set into the feature selection algorithm for fitting training and retaining effective features with non-zero feature coefficients. This includes: selecting a sparse feature selection algorithm model as the feature selection algorithm and setting the range of regularization coefficients; inputting the preliminary model feature set into the sparse feature selection algorithm model after standardization and compressing the feature coefficients through L1 regularization constraints; iteratively training until convergence and selecting features with non-zero feature coefficients to form an effective feature set.

7. The watershed water level prediction method for complex terrain according to claim 6, characterized in that, In step 4, dividing the effective features into a training set and a test set, and training the model using the training set and validating it using the test set to obtain the water level prediction model, includes: dividing the effective features into a training set and a test set according to a preset ratio; initializing the water level prediction model using a machine learning architecture; fitting the model parameters with the training set data; iteratively optimizing the model loss function until convergence using the gradient descent method; inputting the test set data into the trained model; verifying the model prediction error; and determining the final water level prediction model when the error is lower than a preset threshold.

8. The watershed water level prediction method for complex terrain according to claim 7, characterized in that, In step 5, determining the flood discharge location based on the water level prediction model and the buffer status of the flood discharge channel and the buffer confluence unit includes: determining whether the buffer status of the buffer confluence unit is saturated based on the predicted water level output by the water level prediction model; if it is saturated, the flood discharge channel with the closest spatial distance to the buffer confluence unit and the largest flood discharge cross-sectional area is selected as the flood discharge location; if it is not saturated, the flood discharge channel corresponding to the predicted water level exceeding the preset safety threshold is selected as the flood discharge location.

9. The watershed water level prediction method for complex terrain according to claim 8, characterized in that, In step 5, the process of integrating upstream flow data, flood discharge location, and topographic parameters of the confluence area to simulate the passage of the flood peak and output the passage time, peak water level, and backflow risk level includes: inputting upstream flow data, flood discharge capacity parameters of the flood discharge location, and topographic parameters of the confluence area based on a fluid dynamics confluence model; simulating the propagation path and water level change process of the flood peak from the flood discharge location to the confluence, calculating the passage time and peak water level; comparing the peak water level with the critical water level of the confluence topography, and combining the backflow effect judgment rules to classify and output three backflow risk levels: high, medium, and low; and generating a fitted virtual image of the confluence area based on the risk levels. The fitted virtual image presents a heat map of the flood peak propagation path, peak water level distribution, and backflow risk level in a three-dimensional visualization form, and simultaneously labels the flood discharge location, critical water level, and risk warning signs.

10. A watershed water level prediction system for complex terrain, characterized in that, Used to perform the method according to any one of claims 1-9.

Images