A mountain river daily sediment concentration prediction method based on multi-scale decomposition

Abstract

The application discloses a mountainous river daily sediment concentration prediction method based on multi-scale decomposition, and belongs to the technical field of hydrological prediction, and comprises the following steps: obtaining daily sediment concentration sequences and daily influence variable sequences of a plurality of hydrological stations in a target river basin; adopting multivariate empirical mode decomposition to synchronously decompose each sequence to obtain a plurality of intrinsic mode components and residual terms corresponding to each sequence; calculating variance contribution percentages of each intrinsic mode component to the original sequence thereof, removing tail intrinsic mode components and residual terms according to the variance contribution percentages, and accumulating and reconstructing the remaining intrinsic mode components to obtain a denoising reconstructed sequence; obtaining basin static characteristics corresponding to each hydrological station; taking the denoising reconstructed sequence and the basin static characteristics as input characteristics, taking the daily sediment concentration sequence as a prediction target, and adopting a random forest algorithm to construct a daily sediment concentration prediction model. The application can effectively improve the prediction accuracy of the daily sediment concentration of the mountainous river.

Description

Technical Field

[0001] This invention belongs to the field of hydrological prediction technology, and in particular relates to a method for predicting daily sediment content in mountain rivers based on multi-scale decomposition. Background Technology

[0002] Accurate forecasting of hydrological time series is crucial for water resource management, flood control and disaster reduction, and ecological environmental protection. Empirical Mode Decomposition (EMD) and its improved algorithms, as adaptive signal processing methods, can decompose nonlinear and non-stationary time series into several Intrinsic Mode Functions (IMFs) arranged from high to low frequency, along with residual terms, thereby effectively revealing the multi-scale oscillation characteristics and long-term trends of the series. In recent years, combining EMD-like decomposition methods with machine learning models has become an important research direction in the field of time series forecasting. For example, existing techniques use Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) to decompose water quality index series, and then use Long Short-Term Memory (LSTM) neural networks and Autoregressive Integrated Moving Average (ARIMA) models to predict each IMF component. Finally, the prediction results are superimposed to obtain the final predicted value, effectively improving the accuracy of water quality forecasting. Furthermore, some technologies combine EMD with Random Forest (RF) for salinity prediction in tidal river sections. By extracting different frequency components from time-series data through EMD decomposition and combining this with a Random Forest model, the technology can effectively handle the multi-frequency components and lag time issues of complex signals. Similarly, in the field of runoff prediction, there are technical solutions that combine CEEMDAN with Random Forest for medium- and long-term runoff forecasting. These technical solutions demonstrate that the combined modeling strategy of "decomposition-prediction-reconstruction" has good applicability and prediction accuracy in handling nonlinear and non-stationary time series.

[0003] However, the existing technologies still have the following shortcomings. First, existing methods mostly use univariate EMD or its improved algorithms (such as CEEMDAN, EEMD, etc.) to decompose a single target variable, making it difficult to achieve synchronous decomposition of multiple influencing variables within a unified scale framework. This fails to effectively maintain the comparability of different variables at similar time scales, limiting the accuracy of identifying scale dependencies among multiple variables. Second, existing technologies typically use all IMF components obtained from the decomposition for subsequent modeling, lacking an effective screening mechanism based on the contribution of each component to the original sequence information. This fails to remove tail terms and residual terms with limited information, resulting in information redundancy and noise interference in the model input. Third, the input variables of existing prediction models are limited to dynamic time series (such as runoff, tidal level, wind speed, etc.), and have not yet introduced static background features such as watershed topography, slope, land cover, and river network structure to characterize the spatial heterogeneity of different watersheds, making it difficult to achieve unified modeling across multiple stations in different watersheds. Fourth, most existing studies focus on independent modeling and prediction for single stations, failing to construct a unified prediction framework covering multiple stations, thus limiting the model's generalization ability under different watershed conditions. In summary, for the specific technical field of predicting daily sediment content in mountainous rivers, there is an urgent need for an improved solution that can overcome the above-mentioned shortcomings. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention proposes a method for predicting daily sediment concentration in mountainous rivers based on multi-scale decomposition, thereby resolving the issues present in the existing technologies.

[0005] To achieve the above objectives, this invention provides a method for predicting daily sediment concentration in mountainous rivers based on multi-scale decomposition, comprising: S1: Obtain the daily sediment concentration at multiple hydrological stations within the target watershed. The sequence and daily influencing variable sequence, wherein the daily influencing variable sequence includes daily flow sequence, daily rainfall sequence, daily temperature sequence, daily potential evapotranspiration sequence and daily normalized vegetation index sequence; S2: Multivariate empirical mode decomposition is used to simultaneously decompose the daily sediment concentration sequence and each of the daily influencing variable sequences to obtain several intrinsic mode components and residual terms corresponding to each sequence; S3: Calculate the percentage contribution of each intrinsic mode component to the variance of its original sequence, remove the tail intrinsic mode components and the residual terms according to the percentage contribution of variance, and accumulate and reconstruct the remaining intrinsic mode components to obtain the denoised reconstructed sequence. S4: Obtain the watershed static characteristics corresponding to each of the hydrological stations; S5: Using the denoised reconstructed sequence and the watershed static features as input features, and the daily sediment concentration sequence as the prediction target, a daily sediment concentration prediction model is constructed using the random forest algorithm.

[0006] Preferably, multivariate empirical mode decomposition is used to simultaneously decompose the daily sediment concentration sequence and each of the daily influencing variable sequences, specifically including: The daily sediment concentration sequence, daily flow rate sequence, daily rainfall sequence, daily temperature sequence, daily potential evapotranspiration sequence, and daily normalized vegetation index sequence are constructed into a multivariate time series. A set of direction vectors is selected on an n-1 dimensional sphere, where n is the number of variables in the multivariate time series. The multivariate time series is projected along each of the direction vectors to obtain the envelope in each direction; Calculate the mean of the envelope in each direction, and obtain the intrinsic mode components corresponding to each variable through filtering and iteration.

[0007] Preferably, calculating the percentage contribution of each intrinsic mode component to the variance of its respective original sequence specifically includes: Calculate the variance of each intrinsic modal component; The variance contribution percentage is obtained by calculating the percentage of variance of each intrinsic modal component relative to the total variance of the corresponding original sequence.

[0008] Preferably, removing the tail term intrinsic mode components and the residual term according to the variance contribution percentage specifically includes: Sort the intrinsic mode components corresponding to each variable from high to low frequency; The last sorted intrinsic mode component and the residual term are used as the objects to be removed; The sum of the percentage contribution of the removed tail term intrinsic modal components and the residual term to the variance of the original sequence does not exceed 9.6%.

[0009] Preferably, the static characteristics of the watershed include average slope, slope percentage, catchment area per unit width, elevation characteristics, elevation difference, canyon depth, area-elevation integral, land cover type percentage, catchment area, watershed length, morphological indicators, topographic relief ratio, river network length, river network density, and peak ground acceleration.

[0010] Preferably, the step of constructing a daily sediment concentration prediction model using the random forest algorithm includes: The denoised reconstruction sequence and the watershed static features are used as input variables of the random forest model, and the daily sediment concentration sequence is used as output variable of the random forest model. Construct several decision trees, each of which splits its nodes based on a random subset of the input variables; The mean of the prediction results of each decision tree is used as the final prediction output of the random forest model.

[0011] Preferably, the hyperparameters of the random forest model are globally optimized using a Bayesian optimization algorithm, and the hyperparameters include the number of decision trees, the maximum depth, and the minimum number of leaf node samples.

[0012] Preferably, the Bayesian optimization algorithm guides sampling by constructing a surrogate model of the objective function, using the prediction accuracy of the random forest model on the validation set as the optimization objective, and iteratively updates the surrogate model until the termination condition is met.

[0013] Preferably, the input features further include derived features of the daily influencing variable sequence, the derived features including previous cumulative rainfall, previous cumulative flow, rainfall lag term, and flow lag term.

[0014] Preferably, the daily sediment concentration prediction model is a unified model covering the multiple hydrological stations, and the watershed static characteristics of each hydrological station are used to distinguish the spatial heterogeneity of different sub-watersheds.

[0015] Compared with the prior art, the present invention has the following advantages and technical effects: This invention utilizes the technical features of obtaining daily data from multiple hydrological stations within the target watershed in step S1, obtaining the watershed static characteristics corresponding to each hydrological station in step S4, and constructing a daily sediment concentration prediction model in step S5. This enables the prediction model to cover multiple hydrological stations, integrates the dynamic change information and static background information of each station into the same modeling framework, realizes the extraction of common patterns and the adaptation of local characteristics across stations, and solves the problem of spatial heterogeneity characterization in unified modeling of multiple stations.

[0016] The present invention employs a multivariate empirical mode decomposition technique in step S2 to simultaneously decompose the daily sediment concentration sequence and the daily influencing variable sequences. This technique enables the simultaneous decomposition of multiple variables into several intrinsic mode components and residual terms within a unified scale framework, making different variables comparable at similar time scales. This allows for the accurate identification of the scale dependence between each influencing variable and sediment concentration, providing structurally clear multi-scale feature information for subsequent modeling.

[0017] The present invention calculates the percentage contribution of each intrinsic mode component to the variance of its original sequence in step S3, removes tail intrinsic mode components and residual terms based on the percentage contribution, and then accumulates the remaining intrinsic mode components to reconstruct the denoised reconstructed sequence. This invention has the technical characteristics of quantitatively assessing the information contribution of each scale component to the original sequence, effectively filtering out redundant information and noise interference contained in tail components and residual terms that have limited information contribution to the original sequence, so that the denoised reconstructed sequence can more accurately reflect the effective change information of daily sediment content and its influencing variables, and improve the quality of the input signal.

[0018] The present invention obtains the watershed static features corresponding to each hydrological station in step S4, and constructs a random forest model in step S5 by using the denoised reconstruction sequence and watershed static features as input features. This allows the introduction of background constraint information reflecting the inherent attributes of each watershed, such as topography, slope, cover, and river network structure, on the basis of dynamic time series information. This enables the model to characterize the spatial heterogeneity of different watersheds, thereby achieving unified modeling of multiple hydrological stations and effectively improving the generalization performance of the model under different watershed conditions. Attached Figure Description

[0019] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart of a method for predicting daily sediment concentration in mountainous rivers based on multi-scale decomposition, according to an embodiment of the present invention. Figure 2 This is a schematic diagram illustrating the temporal variation trends of sediment content, flow rate, precipitation, potential evapotranspiration, average temperature, and NDVI series in an embodiment of the present invention. Figure 3 This is a schematic diagram of the intrinsic mode function of daily sediment concentration and influencing factors based on MEMD decomposition in an embodiment of the present invention. Figure 4 This is a schematic diagram illustrating the percentage (%) of the IMF (Influence Factor) of daily sediment concentration and its influencing variables relative to the variance of the original data in an embodiment of the present invention. Figure 5 This is a schematic diagram comparing the performance metrics of the model test set in an embodiment of the present invention; Figure 6 This is a schematic diagram illustrating the simulated effect of the sand content test set in an embodiment of the present invention. Detailed Implementation

[0020] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0021] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0022] Example 1 like Figure 1 As shown, this embodiment provides a method for predicting daily sediment concentration in mountainous rivers based on multi-scale decomposition, including: S1: Obtain the daily sediment concentration sequence and daily influencing variable sequence of multiple hydrological stations within the target watershed. The daily influencing variable sequence includes the daily flow sequence, daily rainfall sequence, daily temperature sequence, daily potential evapotranspiration sequence, and daily normalized vegetation index sequence. The study area of ​​this embodiment is the lower reaches of the Jinsha River, specifically the section from the confluence of the Yalong River to Yibin, with a total length of approximately 768 km and an area of ​​86,000 square kilometers. The basin has a complex and diverse topography, with deep valleys and steep slopes. The average river gradient is approximately 0.93‰, and gravity erosion events such as landslides and collapses occur frequently along the river. The basin has active sediment sources and intense erosion and sediment production, making it one of the most severely affected areas of soil erosion in the upper reaches of the Yangtze River. Influenced by the monsoon climate, rainfall distribution throughout the year is extremely uneven, with rainfall from May to October accounting for over 80% of the total annual rainfall. Sediment production and transport processes are easily affected by geological structures and extreme events, resulting in steep rises and falls, sharp and slender mountain-shaped floods. The water-sediment relationship exhibits strong nonlinearity and pulse-like characteristics. The basin contains numerous typical mountain rivers, such as the Meigu River and the Heishui River, most of which have a catchment area of ​​over 10... 2 ~10 4 Between these points, there are small to medium-sized mountain rivers.

[0023] The lower reaches of the Jinsha River exhibit three key characteristics: ① a dense distribution of cascade reservoirs, highlighting the need for sediment regulation and reservoir operation management; ② a typical mountainous river sediment-producing area with complex and highly representative water-sediment processes; and ③ a wide distribution of basins with limited or no data, necessitating urgent sediment forecasting. Therefore, this study not only has academic value in terms of sediment dynamics and data-driven modeling methods but also provides engineering references for sediment regulation in cascade reservoirs and water-sediment management in mountainous rivers.

[0024] The collected water and sediment observation data were all obtained from the *Yangtze River Basin Hydrological Yearbook*. Data was unavailable in a wide area of ​​the lower Jinsha River basin, and some stations had short time series and incomplete monitoring. Therefore, considering the inconsistent observation periods and varying degrees of data loss among stations, severely missing periods were removed, and daily data from 2011 to 2016, shared by all stations, were selected as the unified modeling period. Rainfall (P) and temperature (T) data were obtained from the meteorological gridded product released by the National Earth System Science Data Center. This dataset was obtained by interpolation from the original meteorological station data, with a spatial resolution of 1 km × 1 km. The Normalized Difference Vegetation Index (NDVI) remote sensing dataset also came from this platform, with a spatial resolution of 5 km × 5 km. Potential evapotranspiration (PE) data were obtained from the National Tibetan Plateau Science Data Center.

[0025] Considering that sediment transport in mountainous rivers is controlled not only by dynamic conditions such as rainfall and runoff, but also by underlying surface conditions such as topography, land cover, and river network structure, this paper further constructs a static characteristic system for the watershed. All indicators are represented by watershed average values ​​to reflect the spatial heterogeneity and long-term background constraints of the watershed. Topographic scale and river system indicators are extracted using a 30m×30m resolution ASTER GDEM; land cover data are derived from a 1km resolution land use remote sensing monitoring dataset released by the Resource and Environmental Science and Data Center, and integrated according to the primary land use classification system of the Second National Land Survey. This embodiment selects data from tributary hydrological stations, as detailed in Table 1.

[0026] Table 1 S2: Multivariate empirical mode decomposition is used to simultaneously decompose the daily sediment concentration sequence and each of the daily influencing variable sequences to obtain several intrinsic mode components and residual terms corresponding to each sequence; Furthermore, multivariate empirical mode decomposition is used to simultaneously decompose the daily sediment concentration sequence and the daily influencing variable sequences. Specifically, this includes: constructing a multivariate time series from the daily sediment concentration sequence, the daily flow rate sequence, the daily rainfall sequence, the daily temperature sequence, the daily potential evapotranspiration sequence, and the daily normalized vegetation index sequence; selecting a set of direction vectors on an n-1 dimensional sphere, where n is the number of variables in the multivariate time series; projecting the multivariate time series along each direction vector to obtain the envelope in each direction; calculating the mean of the envelope in each direction; and obtaining the intrinsic mode components corresponding to each variable through iterative filtering.

[0027] Specifically, Multivariate Empirical Mode Decomposition (MEMD) can simultaneously decompose multiple variables into several intrinsic mode functions (IMFs) arranged from high to low frequency, along with residual terms, thereby identifying local oscillation characteristics and long-term variation components of multivariate time series within a unified scale framework. Unlike traditional methods that pre-set basis functions, MEMD has strong data adaptability, better handling daily hydrological series with complex fluctuations and obvious local characteristics, and enhancing the comparability of different variables at similar time scales. After variable decomposition, the instantaneous frequencies are calculated using Hilbert transform, and further transformations are performed to obtain the spatial representation scale unique to IMFs. Assumptions Represent a n Meta-time series, decomposed into MEMD k IMFs at various scales: (1) In the formula: express Thek One IMF, express n Yuan Residual, Order For along the angle Given a direction vector, where Defined in ( n -1) on the sphere.

[0028] S3: Calculate the percentage contribution of each intrinsic mode component to the variance of its original sequence, remove the tail intrinsic mode components and the residual terms according to the percentage contribution of variance, and accumulate and reconstruct the remaining intrinsic mode components to obtain the denoised reconstructed sequence. Further, the percentage contribution of each intrinsic modal component to the variance of its respective original sequence is calculated, specifically including: calculating the variance value of each intrinsic modal component; calculating the percentage of the variance value of each intrinsic modal component to the total variance value of the corresponding original sequence, to obtain the variance contribution percentage.

[0029] Further, the process of removing the tail-end intrinsic modal components and the residual terms based on the variance contribution percentage specifically includes: sorting the intrinsic modal components corresponding to each variable in descending order of frequency; selecting the last-ranked intrinsic modal component and the residual term as the removal targets; wherein the sum of the variance contribution percentages of the removed tail-end intrinsic modal components and the residual terms to the original sequence does not exceed 9.6%.

[0030] S4: Obtain the watershed static characteristics corresponding to each of the hydrological stations; Furthermore, the static characteristics of the watershed include average slope, slope percentage, catchment area per unit width, elevation characteristics, elevation difference, canyon depth, area-elevation integral, land cover type percentage, catchment area, watershed length, morphological indicators, topographic relief ratio, river network length, river network density, and peak ground acceleration.

[0031] S5: Using the denoised reconstructed sequence and the watershed static features as input features, and the daily sediment concentration sequence as the prediction target, a daily sediment concentration prediction model is constructed using the random forest algorithm.

[0032] Furthermore, the step of constructing a daily sediment concentration prediction model using the random forest algorithm includes: using the denoised reconstructed sequence and the watershed static features as input variables of the random forest model, and using the daily sediment concentration sequence as the output variable of the random forest model; constructing several decision trees, with each decision tree splitting its nodes based on a random subset of the input variables; and using the mean of the prediction results of each decision tree as the final prediction output of the random forest model.

[0033] Furthermore, the hyperparameters of the random forest model are globally optimized using a Bayesian optimization algorithm. The hyperparameters include the number of decision trees, the maximum depth, and the minimum number of leaf node samples.

[0034] Specifically, the random forest model algorithm is an ensemble machine learning algorithm based on decision trees. By constructing multiple decision trees during training and integrating their outputs, it can effectively simulate and predict complex nonlinear processes, exhibiting advantages such as high stability, high accuracy, and avoidance of overfitting. Utilizing the feature importance evaluation function in the random forest algorithm, the relative importance of influencing factors is ranked, quantifying the contribution of each potential driving factor to sediment concentration, thereby revealing the sediment concentration driving mechanism. The sediment concentration prediction output of a single decision tree is as follows: (2) In the formula: For the first k Decision trees for samples x The predicted output, For the first k Tree No. m The average sand content of each leaf node For the first k Tree No. m The sample space of leaf nodes, For indicator functions (samples) x belong If the value is 1, then the value is 0.

[0035] The final prediction result of the random forest is the mean of the outputs of all decision trees: (3) In the formula: K For the number of decision trees, This is the final predicted value for sediment content.

[0036] Furthermore, the Bayesian optimization algorithm guides sampling by constructing a surrogate model of the objective function, using the prediction accuracy of the random forest model on the validation set as the optimization objective, and iteratively updates the surrogate model until the termination condition is met.

[0037] Specifically, Bayesian optimization is a global optimization strategy based on a probabilistic model. It guides sampling by constructing a surrogate model of the objective function, thereby efficiently finding the global optimum. Its advantage lies in using historical evaluation information to approach the global optimum with fewer iterations, avoiding local traps while reducing computational costs. In this study, Bayesian optimization is used to automatically search for the optimal hyperparameter combination (number of decision trees, maximum depth, minimum number of leaf node samples, etc.) of a random forest model, ensuring that the RF model achieves optimal performance.

[0038] This embodiment uses the Kling-Gupta efficiency coefficient ( KGE ) and coefficient of determination ( R 2 () is used as a model evaluation index. KGE The coefficients comprehensively consider the correlation, bias, and variability between simulated and observed values, and can comprehensively evaluate the model performance. R ² Reflects the model's ability to explain the variation in the dependent variable. The closer both values ​​are to 1, the better the model's simulation performance. The calculation formula is as follows: (4) (5) In the formula: r The correlation coefficient between observed and predicted values; α It is the ratio of the mean of the observed value to the mean of the predicted value. ; β It is the ratio of the standard deviations of the observed values ​​to the predicted values. ; , These are the measured and predicted values ​​of daily sediment content, respectively. i These are the observations from the first year; N This represents the total number of observations.

[0039] Furthermore, the input features also include derived features of the daily influencing variable sequence, which include previous cumulative rainfall, previous cumulative flow, rainfall lag term, and flow lag term.

[0040] Furthermore, the daily sediment concentration prediction model is a unified model covering the multiple hydrological stations, and the watershed static characteristics of each hydrological station are used to distinguish the spatial heterogeneity of different sub-watersheds.

[0041] Based on the above technical solution, this embodiment employs the following steps for the research area: 1. Temporal variation characteristics of sediment content and its influencing factors; The temporal variation characteristics of daily sediment concentration (S), flow rate (Q), precipitation (P), temperature (T), potential evapotranspiration (PE), and normalized differential vegetation index (NDVI) in the control area of ​​10 hydrological stations on the lower reaches of the Jinsha River are as follows: Figure 2As shown in the table. During the study period, daily sediment concentration generally remained at a low level, with high values ​​concentrated during the flood season and exhibiting an intermittent distribution: sediment concentration often rapidly climbed to a peak in a short period of time and then dropped sharply, forming a peak-type sediment transport process. Combined with the coefficients of variation (Table 2), the coefficient of variation for daily sediment concentration was 360.2%–438.0%, far higher than other variables, fully confirming its significant high-frequency pulse fluctuations and non-stationary characteristics. Discharge and precipitation also showed significant irregular fluctuations on a daily scale, with coefficients of variation, although lower than those for daily sediment concentration, reaching 91.3%–268.9% and 192.6%–266.3%, respectively. Potential evapotranspiration and mean temperature showed parabolic variations throughout the year, exhibiting the same seasonal variation pattern as NDVI, peaking in summer and reaching their lowest point in winter. Overall, the complexity of sediment transport mechanisms in mountain tributaries and the nonlinearity of hydrological processes are very significant. The temporal trends of sediment concentration, discharge, precipitation, potential evapotranspiration, mean temperature, and NDVI series are shown in the table. Figure 2 As shown.

[0042] Table 2 2. The multi-scale relationship between sediment content and its influencing factors; Since the influence of potential influencing variables on sediment concentration at different time scales is not a simple additive process, a simple linear function is insufficient to identify its complex nonlinear correlations. Therefore, using MEMD to identify multi-time-scale results of water and sediment processes is essential for revealing the association between sediment concentration and its influencing variables. This paper decomposes the daily sediment concentration and its potential influencing variables from 10 hydrological stations into 9-11 IMFs and a residual (…) using MEMD. Figure 3 Among them, Doushaguan Station yielded 9 IMFs, Niujie Station and Dashadian Station yielded 10 IMFs, and the remaining stations yielded 11 IMFs each. The differences in the number of IMFs among the stations reflect certain differences in the non-stationarity and complexity of the input signals in each catchment area.

[0043] As the order of the IMF increases, the frequency oscillations gradually decrease, while the periodic scale gradually increases. Hilbert transform results show that the periodic scales of the high-frequency components IMF1-IMF3 differ relatively little among different stations. The average period for IMF1 is 2.7-4.8 days, for IMF2 it is 6.7-8.3 days, and for IMF3 it is 11.8-15.7 days, indicating certain commonalities in short-term water and sediment responses across different catchment areas. The average periods of the mid-frequency components IMF4-IMF8 show significantly greater differences among different stations, indicating variations in daily sediment concentration and related variables over longer time intervals. For example, the average period for IMF8 at Fengtun station, which has the smallest catchment area, is 169 days, while at Dashadian station, which has the largest catchment area and a stronger reservoir influence, the average period for IMF8 can reach 313 days. The periodic scales of the low-frequency components IMF9-IMF11 further increase, gradually converging at IMF11. The average period of IMF9 is 313 days to 761.2 days; the average period of IMF10 is 548 days to 1096 days; for stations including IMF11, except for Qingkoutang station where the average period of IMF11 is 1035.2 days, the average period of the other stations is 1096 days. These results mainly reflect the long-term trends under the combined influence of regional climate background, intensity of human activities, and evolution of the underlying surface of the watershed. Specifically, the daily sediment concentration at Gaoqiao station generally first increased and then decreased, while Ningnan and Zhaojue stations showed a decreasing trend, and the other stations generally increased. Overall, the water and sediment sequences generally exhibit a significant multi-scale structure, with different variables showing synchronous oscillations at similar scales, especially at short-period scales (IMF1-IMF3).

[0044] The variance contribution percentage is the percentage of the intrinsic mode function (IMF) variance of each variable relative to the corresponding original variance, reflecting the explanatory power of the original sequence. From... Figure 3It can be seen that the main contribution factors (IMFs) of different variables in the study area are not consistent. For daily sediment concentration and precipitation, IMF1-IMF3 make the largest contributions, with the total contribution rate of sediment concentration IMF1-IMF3 reaching 49.1%~109.4% and precipitation reaching 41.4%~109.1%. This indicates that the changes in daily sediment concentration and precipitation are mainly controlled by short-term events. In contrast, the main variance contribution of flow varies significantly among different stations. Stations such as Qingkoutang, Ningnan, Doushaguan, and Dashadian are more concentrated in IMF7-IMF9, while the contribution of other stations is more dispersed. The main contribution components of mean temperature (T), potential evapotranspiration (PE), and NDVI are significantly concentrated in the low-frequency components, with most stations reaching their peak in IMF8 or IMF9. The influence of these three factors on sediment transport is indirect and lagged, and is more prominent on longer time scales, rather than being short-term events dominated by precipitation or runoff. In addition, it can be found that for any given variable, the sum of the variance contributions of each IMF is not strictly equal to 100%. This may be because the relevant processes are not independent at different scales, or because different IMFs of a variable lack complete orthogonality. Taking advantage of this characteristic, MEMD is very suitable for processing complex daily hydrological sequences formed by the superposition of multiple processes. Its decomposition results can not only identify the dominant information of different variables at different time scales, but also provide a basis for subsequent construction of recombinant input sequences for daily sediment concentration simulation.

[0045] 3. MEMD-RF model performance; based on Figure 4 The variance contribution analysis showed that the tail terms (IMF10 ~ IMF12) of each variable contributed less than 5% to the variance of the original sequence on average, far lower than that of the higher-order components (IMF1 ~ IMF9), indicating that they contained very limited information. Therefore, this paper attempts to remove the aforementioned tail terms and the residual terms remaining from MEMD decomposition, retaining only the effective IMF components that reflect the daily-scale response to reduce noise. Based on this, watershed static features are introduced as background feature variables to distinguish different sub-watersheds, and a multi-site RF model is constructed. The hyperparameters of the model are globally optimized using a Bayesian optimization algorithm to achieve unified modeling of multi-watershed data, and the performance evaluation indexes on the corresponding test sets of each site are calculated and output. Simultaneously, to verify the effectiveness of the MEMD reconstruction strategy, a control model is constructed using the original variable sequence without MEMD decomposition as input.

[0046] Figure 5 The results show that the MEMD-reconstructed sequence model outperforms the original sequence-based model on the test set overall. The former's model metrics on the test set... R 2 Reached 0.740, KGE Approximately 0.669, the latter R 2 and KGE The values ​​were only 0.689 and 0.639 respectively, indicating that MEMD decomposition and recombination can improve the quality of input information and enhance the model's fitting ability to a certain extent. Considering the simulation fitting results... Figure 6 The model can capture the fluctuations and seasonal changes in sediment concentration during the flood season relatively well, and the simulation of low-value phases is also relatively stable, indicating that the model can learn the cross-river water and sediment response patterns to a certain extent.

[0047] Focusing on the local site level, the reconstructed sequence model achieved varying degrees of improvement across most sites. R 2 The accuracy improved from 0.26–0.87 in the original sequence model to 0.44–0.94. KGE The performance improved from 0.21~0.72 to 0.37~0.77. The improvement was particularly noticeable at stations where the original sequence model performed weakly, further demonstrating that MEMD has a good denoising and information reconstruction effect on complex fluctuating sequences. Nevertheless, significant differences in model performance remain between different stations. For example, at the Doushaguan station… R 2 The value reached as high as 0.94, while at Takahashi Station R 2 The value was only 0.44. This indicates that while the model can extract common patterns across watersheds, its adaptability to the localized, specific processes of some tributaries remains limited. The reason for this is not insufficient optimization of model parameters, but rather the significant differences in sediment generation and transport mechanisms among different tributaries. For watersheds with relatively stable sediment generation and transport processes and consistent water and sediment responses, the model can more easily establish effective mapping relationships; however, for watersheds with significant regional water and sediment characteristics, relying solely on conventional hydrological and meteorological variables and macroscopic static indicators is insufficient to fully characterize their sediment generation and transport processes.

[0048] In summary, the MEMD recombination strategy primarily enhances the model's efficiency in utilizing multi-scale dynamic information and improves the simulation results for conventional processes. However, its characterization ability is limited for sudden, pulse-like sediment transport processes dominated by short-term, high-intensity extreme events. This explains why the model fits most conventional processes well, but still underestimates some extreme peak processes. The original daily sequences contain high-frequency fluctuations, low-frequency trends, and random noise, making direct modeling susceptible to interference from non-stationarity and information redundancy. MEMD, on the other hand, can separate information at different scales within a multivariate unified scale framework, making it easier for the model to extract effective signals directly related to sediment concentration changes, thereby improving simulation accuracy.

[0049] 4. Feature importance analysis.

[0050] To reveal which input features the model primarily relies on for feature importance analysis in daily sediment concentration simulation, this example presents the results of the top-ranked factors. Rainfall (P) is the most important driving factor in the model, contributing 26.8%; followed by flow (Q), contributing 9.1%. Rainfall is the direct trigger for sediment production, and its spatiotemporal distribution directly affects the soil erosion process. Flow, on the other hand, characterizes the intensity of water scouring and sediment transport capacity, comprehensively reflecting the watershed confluence process, flood fluctuations, and short-term water and sediment response characteristics. The combined contribution of these two factors reaches 35.9%. Furthermore, the cumulative rainfall P from the preceding period, which is closely related to the rainfall and flow processes, is also a significant factor. sum3 P sum7 (Total 8.4%), Rainfall / Flow Lag Item P lag1 Q lag1 The precipitation rate (totaling 9.2%) and mutation rate (totaling 5.0%) also showed high importance. This indicates that daily sediment concentration is not solely controlled by daily rainfall; the cumulative effect and lag characteristics implied by continuous sediment inflow during flood events are also crucial. If rainfall, flow, and their respective derived characteristics are combined, rainfall-related characteristics contribute approximately 50.2%, and flow-related characteristics contribute approximately 19.1%. In summary, the combination of precipitation and flow reflects both the long-term cumulative effect within the watershed and compensates for the spatial or temporal deficiencies of a single data source, collectively determining approximately 69.3% of daily sediment concentration changes. In contrast, the static characteristics of the watershed themselves are not particularly important, accounting for only 9.5% of all single-factor watershed static characteristics. However, their interaction with rainfall and flow demonstrates strong importance, accounting for a cumulative 12.7%. This suggests that watershed static characteristics such as the underlying surface are not strong driving factors in daily-scale modeling, but rather play a background regulatory role by modulating rainfall-runoff response and sediment transport efficiency.

[0051] This embodiment, based on long-term measured water and sediment data from multiple stations, constructs a daily sediment concentration prediction model for random forests using MEMD recombination features and watershed static features, achieving unified modeling across multiple stations, and yields the following main conclusions: (1) The daily sediment concentration of the downstream tributaries of the Jinsha River has significant high-frequency fluctuations and non-stationary characteristics, with a coefficient of variation as high as 360.2%~438%. MEMD results show that the daily sediment concentration sequence has clear multi-scale results, of which IMF1-IMF3 contribute more than 49% cumulatively, indicating that the daily sediment concentration change is mainly controlled by short-period processes.

[0052] (2) The multi-site model constructed based on the dynamic features of MEMD recombination achieved good results on the test set, and the overall performance was good. KGE and R 2 The scores were 0.669 and 0.740 respectively, which are better than the original variable model without MEMD processing. KGE0.639 R 2 :0.689), single-station local fitting R 2 The improvement from 0.26~0.87 to 0.44~0.94 indicates that multi-scale decomposition, noise reduction, and feature recombination help improve the model's ability to extract effective information and its cross-basin generalization performance.

[0053] (3) Rainfall series and flow series characteristics are key factors in the simulation of daily sediment concentration, and together they explain about 69.3% of the variation in daily sediment concentration. Although the static characteristics of the watershed contribute relatively little (9%) on their own, they enhance the model’s ability to capture spatial heterogeneity, enabling the model to integrate the characteristics of each station and capture the overall pattern pieced together from various regions in a wider spatial domain.

[0054] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for predicting daily sediment concentration in mountain rivers based on multi-scale decomposition, characterized in that, Includes the following steps: S1: Obtain the daily sediment concentration sequence and daily influencing variable sequence of multiple hydrological stations within the target watershed. The daily influencing variable sequence includes the daily flow sequence, daily rainfall sequence, daily temperature sequence, daily potential evapotranspiration sequence, and daily normalized vegetation index sequence. S2: Multivariate empirical mode decomposition is used to simultaneously decompose the daily sediment concentration sequence and each of the daily influencing variable sequences to obtain several intrinsic mode components and residual terms corresponding to each sequence; S3: Calculate the percentage contribution of each intrinsic mode component to the variance of its original sequence, remove the tail intrinsic mode components and the residual terms according to the percentage contribution of variance, and accumulate and reconstruct the remaining intrinsic mode components to obtain the denoised reconstructed sequence. S4: Obtain the watershed static characteristics corresponding to each of the hydrological stations; S5: Using the denoised reconstructed sequence and the watershed static features as input features, and the daily sediment concentration sequence as the prediction target, a daily sediment concentration prediction model is constructed using the random forest algorithm.

2. The method according to claim 1, characterized in that, The daily sediment concentration sequence and the daily influencing variable sequences are simultaneously decomposed using multivariate empirical mode decomposition, specifically including: The daily sediment concentration sequence, daily flow rate sequence, daily rainfall sequence, daily temperature sequence, daily potential evapotranspiration sequence, and daily normalized vegetation index sequence are constructed into a multivariate time series. A set of direction vectors is selected on an n-1 dimensional sphere, where n is the number of variables in the multivariate time series. The multivariate time series is projected along each of the direction vectors to obtain the envelope in each direction; Calculate the mean of the envelope in each direction, and obtain the intrinsic mode components corresponding to each variable through filtering and iteration.

3. The method according to claim 1, characterized in that, Calculate the percentage contribution of each intrinsic mode component to the variance of its respective original sequence, specifically including: Calculate the variance of each intrinsic modal component; The variance contribution percentage is obtained by calculating the percentage of variance of each intrinsic modal component relative to the total variance of the corresponding original sequence.

4. The method according to claim 1, characterized in that, The intrinsic modal components of the tail terms and the residual terms are removed based on the variance contribution percentage, specifically including: Sort the intrinsic mode components corresponding to each variable from high to low frequency; The last sorted intrinsic mode component and the residual term are used as the objects to be removed. The sum of the percentage contribution of the removed tail term intrinsic modal components and the residual term to the variance of the original sequence does not exceed 9.6%.

5. The method according to claim 1, characterized in that, The static characteristics of the watershed include average slope, slope ratio, catchment area per unit width, elevation characteristics, elevation difference, canyon depth, area-elevation integral, land cover type ratio, catchment area, watershed length, morphological indicators, topographic relief ratio, river network length, river network density, and peak ground acceleration.

6. The method according to claim 1, characterized in that, The method of constructing a daily sediment concentration prediction model using the random forest algorithm includes: The denoised reconstruction sequence and the watershed static features are used as input variables of the random forest model, and the daily sediment concentration sequence is used as output variable of the random forest model. Construct several decision trees, each of which splits its nodes based on a random subset of the input variables; The mean of the prediction results of each decision tree is used as the final prediction output of the random forest model.

7. The method according to claim 6, characterized in that, The hyperparameters of the random forest model are globally optimized using a Bayesian optimization algorithm. These hyperparameters include the number of decision trees, the maximum depth, and the minimum number of leaf node samples.

8. The method according to claim 7, characterized in that, The Bayesian optimization algorithm guides sampling by constructing a surrogate model of the objective function, using the prediction accuracy of the random forest model on the validation set as the optimization objective, and iteratively updates the surrogate model until the termination condition is met.

9. The method according to claim 1, characterized in that, The input features also include derived features of the daily influencing variable sequence, which include previous cumulative rainfall, previous cumulative flow, rainfall lag term, and flow lag term.

10. The method according to claim 1, characterized in that, The daily sediment concentration prediction model is a unified model covering the multiple hydrological stations, and the watershed static characteristics of each hydrological station are used to distinguish the spatial heterogeneity of different sub-watersheds.

Images