River channel on-way water level staging combination forecasting method based on multi-model coupling correction

Abstract

The invention discloses a watercourse on-way water level staging combination forecasting method based on multi-model coupling correction, and the method comprises the steps: building a hydrologic influence factor set of a hydrologic control station water level flow relation based on a plurality of influence factors of the hydrologic control station water level flow relation; key influence factors influencing the water level and flow relation are recognized through the feature selection technology, and the key influence factors are coupled with the water flow influence factors to jointly form a water level forecasting feature factor set; based on various machine learning methods, constructing a water level prediction model library reflecting a complex nonlinear relationship between the water level of the hydrological control station and each prediction characteristic factor; and performing coupling correction on various machine learning prediction results in different scheduling periods, and determining an optimal parameter combination so as to form a whole-process prediction system. The method can improve the watercourse water level prediction precision according to different runoff plume and dryness characteristics of the watercourse and the downstream jacking influence, has the advantages of simple calculation, wide application range and high prediction precision, and can meet the scheduling implementation requirements.

Description

Technical Field

[0001] The present invention belongs to the field of hydrological forecasting, and in particular relates to a method for periodic combined forecasting of water levels along a river channel based on multi-model coupling correction. Background Art

[0002] River water level forecasting is of great practical significance for timely issuing flood and dry season warnings, scientifically and accurately dispatching water projects, and improving the basin's flood control and safety capabilities and the comprehensive utilization of water resources. For large rivers, the complex hydraulic connections between the main and tributary rivers in the basin, coupled with the fuzzy rainfall and runoff processes in uncontrolled intervals, make it difficult to construct a complete hydrodynamic model based on hydrometeorological conditions, topographic and geomorphological characteristics, and mechanisms. As a result, traditional hydrodynamic model-based water level and flow forecasts have low accuracy and computational efficiency that cannot meet real-time dispatch requirements. The completion and commissioning of reservoirs and long-term observations at major control hydrological stations have accumulated a large amount of reservoir operation data, main and tributary water inflow data, and water level and flow data from major control hydrological stations, making data-driven water level forecasting along river channels possible.

[0003] Currently, research on river water levels primarily focuses on analyzing the changing patterns of the relationship between water level and flow during low and high water periods at key hydrological stations and after the completion and commissioning of reservoirs. Limited research exists on water level forecasting. In 2003, Xie Zuotao et al. used a BP artificial neural network to establish a water level forecast model for the Luoshan Station on the Yangtze River. In 2024, Wang Zhenghua et al. used a support vector machine approach to construct a water level forecast model for the Luoshan Station. The selection of characteristic factors in these studies was primarily based on empirical selection, which is highly subjective. Furthermore, using only a single machine learning method to predict water levels throughout the year cannot fully reflect the patterns and characteristics of water level changes during different scheduling periods, resulting in poor adaptability. Therefore, it is necessary to propose a combined, multi-model coupled correction-based forecasting method for river water levels during different phases to address these issues. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a combined forecasting method for water levels along a river channel in stages based on multi-model coupling correction, aiming to overcome the defect that a single machine learning method cannot accurately characterize this complex nonlinear relationship, further improve the accuracy of water level prediction, and can be used to predict the water levels of major control stations along the river channel. The prediction model is simple, has high prediction accuracy, is easy to implement, and takes into account the changing laws of hydrological characteristics in different periods.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is: A method for combined forecasting of water levels along a river channel based on multi-model coupling correction, comprising the following steps: S1: Collect and organize hydrological data of the hydrological control station, its upstream main and tributary reservoirs, and its downstream adjacent hydrological control stations; SS2, based on the water level-discharge relationship of the hydrological control station, constructs the set of hydrological influencing factors of the water level-discharge relationship of the hydrological control station; S3, using a variety of feature selection techniques to identify the key factors affecting the relationship between water level and flow, and considering the propagation time of water from upstream main and tributary rivers and the duration of the impact, the coupled water volume influencing factors together form the water level forecast feature factor set; S4, from the perspective of reservoir operation period and inflow characteristics, the water level forecast characteristic factor set is divided into three subsets: drawdown period, flood season, and water storage period; S5, based on a variety of machine learning methods, constructs a water level prediction model library that reflects the complex nonlinear relationship between the water level of the hydrological control station and various forecast characteristic factors in different scheduling periods; S6, couple and calibrate multiple machine learning prediction results in different scheduling periods, establish a water level combination forecast model, and use intelligent optimization algorithms to optimize the optimal parameter combination of each model in the hydrological forecast model library and the water level period combination forecast model during the training process; S7, using a variety of model evaluation indicators, evaluates and optimizes different water level prediction models to obtain the optimal water level prediction value.

[0006] Preferably, in step S1, the hydrological data includes the current hydrological control station series water level Z c and flow Q c , upstream main and tributary reservoir long series discharge flow Q d As well as the long series of water level data of the downstream hydrological control station; the downstream hydrological control station refers to the hydrological control station that is closest to the current hydrological control station and is not affected by the backwater support effect.

[0007] Preferably, in step S2, a set of hydrological influencing factors of the relationship between water level and flow at the hydrological control station is constructed. D , specifically including: There are many factors that affect the hydrological changes at the hydrological control station, including the changes in the flow section caused by scouring and silting of the river channel, changes in the diversion and sediment separation conditions of the upstream estuary, changes in the downstream water flow caused by the upstream reservoir, the flood fluctuation caused by the small downstream tributary water flow and the large main stream water flow, the downstream backwater support, the flood water composition, interval time, rising water level, flood peak shape, etc. Since some of these factors are difficult to express quantitatively, in order to simplify the modeling process, this paper mainly focuses on the hydrological factors directly related to the relationship between water level and flow, including the current water level Zc of the hydrological control station, water level fluctuation, and the water level of the water level. , the water level Zd of the hydrological control station adjacent to the downstream and not affected by the jacking effect, and the water level fluctuation , the water level difference between the current hydrological control station and the downstream hydrological control station .

[0008] Preferably, in step S3, a plurality of feature selection techniques are used to identify key factors affecting the relationship between water level and flow, and the propagation time of water from upstream main and tributary rivers and the duration of the impact are taken into consideration. The coupled water flow influencing factors together constitute the water level forecast characteristic factor set including: The Gini coefficient method based on the random forest model and the feature importance scoring method based on the LightGBM model are used as feature selection methods to screen the pre-selected feature factors, and then the factors with the highest scores are merged to obtain the final feature factor set. This avoids the randomness and redundancy of a single feature selection technology.

[0009] Preferably, the Gini coefficient method of the random forest model and the feature importance scoring method based on the LightGBM model are used to respectively evaluate the hydrological influencing factor set of the water level and flow relationship of the hydrological control station constructed in step S2. D Perform importance scoring, sort the influencing factors by score size, and finally obtain the key influencing factors by combining the top-ranked influencing factors under the two scoring methods. D 1; At the same time, the flow rate of the hydrological control station itself and the water volume related factors of the upstream reservoir discharge and the main and tributary water flow are integrated to form the water level forecast characteristic factor set. D 2.

[0010] Furthermore, during the drawdown period, reservoirs replenish water downstream, raising the downstream water level. During the flood season, reservoirs prevent downstream water levels from exceeding warning and protection levels by intercepting floodwaters. During the impoundment period, reservoirs intercept incoming water to raise the water level. This process of storing water during high seasons and replenishing water during low seasons causes downstream water levels to exhibit different characteristics at different times. Therefore, based on the perspective of reservoir operation periods and inflow characteristics, the water level forecast characteristic factor set is divided into three subsets: the drawdown period, the flood season, and the impoundment period.

[0011] Preferably, in step S5, the multiple machine learning methods used include but are not limited to extreme gradient boosting decision tree XGBoost, lightweight gradient boosting decision tree LightGBM and support vector machine regression model SVM.

[0012] Furthermore, both XGBoost and LightGBM are ensemble methods based on gradient boosting decision trees. The latter uses a histogram-based decision tree algorithm instead of the sorting-based decision tree algorithm used by the former. Each has its own characteristics. The XGBoost model can be chosen for small and medium-sized datasets with high accuracy requirements, while the LightGBM model can be chosen for large-scale or high-dimensional sparse datasets. SVM is a support vector machine regression algorithm whose goal is to find a hyperplane or decision boundary that makes the predicted value as close to the true value as possible while allowing a certain error range and having high robustness.

[0013] Preferably, based on multiple machine learning methods, the specific process of constructing a water level prediction model library that reflects the complex nonlinear relationship between the water level of the hydrological control station and various forecast characteristic factors in different scheduling periods is as follows: Water level forecast characteristic factor set by data normalization D 2. Preprocess the decision water level sequence data; The water level forecast characteristic factor set D 2. Divide the training set and test set into decision water level sequence; XGBoost, LightGBM, and SVM were used to train the water level prediction model on the training set, and the model hyperparameters were optimized during the training process; Use model evaluation indicators to test and evaluate the model on the test set to test the model's generalization ability to the data; The optimal water level prediction model based on XGBoost, LightGBM and SVM was obtained.

[0014] Preferably, in step S6, coupling correction is performed on multiple machine learning prediction results in different scheduling periods to establish a water level stage combination forecast model. The intelligent optimization algorithm is used to optimize the optimal parameter combination of the models in the hydrological forecast model library and the water level stage combination forecast model during the training process, including: The water level stage combination forecast model is constructed as follows: ; Where, is the predicted value of the water level combination prediction model; 、 、 They are the predicted values ​​of LightGBM, XGBoost and SVM prediction models respectively. 、 、 are the corresponding weight values, , They are drawdown period, flood season and water storage period; An intelligent optimization algorithm is used to optimize the weight coefficient of the phased combination prediction model. The intelligent optimization algorithm is one of the particle swarm algorithm, genetic algorithm, simulated annealing algorithm, tabu search algorithm or ant colony algorithm.

[0015] Preferably, in step S1, the specific process of optimizing the weight coefficient of the phased combination prediction model using the particle swarm optimization algorithm as the intelligent optimization algorithm is as follows: a. The optimal water level prediction models based on XGBoost, LightGBM and SVM respectively obtain three groups of water level prediction values, namely 、 、 ; b. Initialize the particle swarm, where each particle represents a weighted combination; c. Calculate the fitness value of each particle, that is ; d. Update the global optimal particles and the local optimal particles; e. Update the speed and position of the particles according to the speed and position update formula of the particle swarm algorithm; f. Determine whether the maximum number of iterations or the set termination condition has been reached; g. If the termination condition is met, the process ends; otherwise, return to step c to continue the optimization.

[0016] Preferably, in step S7, multiple model evaluation indicators are used to evaluate and optimize different water level prediction models, and obtaining the optimal water level prediction value includes: Model evaluation indicators include the coefficient of determination R 2 , mean absolute error MAE , mean absolute percentage error MAPE , root mean square error RMSE ,in R 2 The larger the value, the better the model fitting effect, and the smaller the value of other indicators, the better the prediction effect of the model. The calculation formula of each indicator is as follows: ; ; ; ; Where, is the measured value, is the predicted value, n is the number of samples, is the measured sample mean.

[0017] The beneficial effects of the present invention are as follows: 1. Based on the many influencing factors of the water level and flow relationship at the hydrological control station, the present invention constructs a set of hydrological influencing factors of the water level and flow relationship at the hydrological control station. On this basis, feature selection technology is used to identify the key influencing factors affecting the water level and flow relationship, and the propagation time of water from upstream main and tributary rivers and the duration of the impact are taken into account. The coupled water volume influencing factors together form a water level forecast feature factor set. Through multiple feature selection technologies, the accuracy of the basic data is effectively improved, the coupling relationship between the data is fully considered, and the model is prevented from falling into the local optimum.

[0018] 2. Based on multiple machine learning methods, the present invention constructs a water level prediction model library that reflects the complex nonlinear relationship between the water level of the hydrological control station and various forecast characteristic factors; finally, the multiple machine learning prediction results are coupled and corrected in different scheduling periods, and a new water level combination forecasting method is proposed to determine the optimal parameter combination, thereby forming a full-process forecasting system; it can improve the river water level prediction accuracy according to the different runoff characteristics of the river and the influence of downstream support, and has the advantages of simple calculation, wide adaptability, and high forecast accuracy. It can meet the implementation of scheduling needs and effectively overcome the defect that a single machine learning method cannot accurately characterize this complex nonlinear relationship. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a schematic diagram of a flow chart of the present invention; Figure 2 This is a diagram of the LightGBM model prediction results in an embodiment of the present invention; Figure 3 This is a graph showing the prediction results of the SVM model in an embodiment of the present invention; Figure 4 This is a graph of XGBoost model prediction results in an embodiment of the present invention; Figure 5 This is a diagram of the prediction results of the combined prediction model in an embodiment of the present invention. DETAILED DESCRIPTION

[0020] Example 1: like Figure 1 As shown, a combined forecasting method for river channel water level based on multi-model coupling correction includes the following steps: S1: Collect and organize hydrological data of the hydrological control station, its upstream main and tributary reservoirs, and its downstream adjacent hydrological control stations; SS2, based on the water level-discharge relationship of the hydrological control station, constructs the set of hydrological influencing factors of the water level-discharge relationship of the hydrological control station; S3, using a variety of feature selection techniques to identify the key factors affecting the relationship between water level and flow, and considering the propagation time of water from upstream main and tributary rivers and the duration of the impact, the coupled water volume influencing factors together form the water level forecast feature factor set; S4, from the perspective of reservoir operation period and inflow characteristics, the water level forecast characteristic factor set is divided into three subsets: drawdown period, flood season, and water storage period; S5, based on a variety of machine learning methods, constructs a water level prediction model library that reflects the complex nonlinear relationship between the water level of the hydrological control station and various forecast characteristic factors in different scheduling periods; S6, couple and calibrate multiple machine learning prediction results in different scheduling periods, establish a water level combination forecast model, and use intelligent optimization algorithms to optimize the optimal parameter combination of each model in the hydrological forecast model library and the water level period combination forecast model during the training process; S7, using a variety of model evaluation indicators, evaluates and optimizes different water level prediction models to obtain the optimal water level prediction value.

[0021] Preferably, in step S1, the hydrological data includes the current hydrological control station series water level Z c and flow Q c , upstream main and tributary reservoir long series discharge flow Q d As well as the long series of water level data of the downstream hydrological control station; the downstream hydrological control station refers to the hydrological control station that is closest to the current hydrological control station and is not affected by the backwater support effect.

[0022] Preferably, in step S2, a set of hydrological influencing factors of the relationship between water level and flow at the hydrological control station is constructed. D , specifically including: There are many factors that affect the hydrological changes at the hydrological control station, including the changes in the flow section caused by scouring and silting of the river channel, changes in the diversion and sediment separation conditions of the upstream estuary, changes in the downstream water flow caused by the upstream reservoir, the flood fluctuation caused by the small downstream tributary water flow and the large main stream water flow, the downstream backwater support, the flood water composition, interval time, rising water level, flood peak shape, etc. Since some of these factors are difficult to express quantitatively, in order to simplify the modeling process, this paper mainly focuses on the hydrological factors directly related to the relationship between water level and flow, including the current water level Zc of the hydrological control station, water level fluctuation, and the water level of the water level. , the water level Zd of the hydrological control station adjacent to the downstream and not affected by the jacking effect, and the water level fluctuation , the water level difference between the current hydrological control station and the downstream hydrological control station .

[0023] Preferably, in step S3, a plurality of feature selection techniques are used to identify key factors affecting the relationship between water level and flow, and the propagation time of water from upstream main and tributary rivers and the duration of the impact are taken into consideration. The coupled water flow influencing factors together constitute the water level forecast characteristic factor set including: The Gini coefficient method based on the random forest model and the feature importance scoring method based on the LightGBM model are used as feature selection methods to screen the pre-selected feature factors, and then the factors with the highest scores are merged to obtain the final feature factor set. This avoids the randomness and redundancy of a single feature selection technology.

[0024] Preferably, the Gini coefficient method of the random forest model and the feature importance scoring method based on the LightGBM model are used to respectively evaluate the hydrological influencing factor set of the water level and flow relationship of the hydrological control station constructed in step S2.D Perform importance scoring, sort the influencing factors by score size, and finally obtain the key influencing factors by combining the top-ranked influencing factors under the two scoring methods. D 1; At the same time, the flow rate of the hydrological control station itself and the water volume related factors of the upstream reservoir discharge and the main and tributary water flow are integrated to form the water level forecast characteristic factor set. D 2.

[0025] Furthermore, during the drawdown period, reservoirs replenish water downstream, raising the downstream water level. During the flood season, reservoirs prevent downstream water levels from exceeding warning and protection levels by intercepting floodwaters. During the impoundment period, reservoirs intercept incoming water to raise the water level. This process of storing water during high seasons and replenishing water during low seasons causes downstream water levels to exhibit different characteristics at different times. Therefore, based on the perspective of reservoir operation periods and inflow characteristics, the water level forecast characteristic factor set is divided into three subsets: the drawdown period, the flood season, and the impoundment period.

[0026] Preferably, in step S5, the multiple machine learning methods used include but are not limited to extreme gradient boosting decision tree XGBoost, lightweight gradient boosting decision tree LightGBM and support vector machine regression model SVM.

[0027] Furthermore, both XGBoost and LightGBM are ensemble methods based on gradient boosting decision trees. The latter uses a histogram-based decision tree algorithm instead of the sorting-based decision tree algorithm used by the former. Each has its own characteristics. The XGBoost model can be chosen for small and medium-sized datasets with high accuracy requirements, while the LightGBM model can be chosen for large-scale or high-dimensional sparse datasets. SVM is a support vector machine regression algorithm whose goal is to find a hyperplane or decision boundary that makes the predicted value as close to the true value as possible while allowing a certain error range and having high robustness.

[0028] Preferably, based on multiple machine learning methods, the specific process of constructing a water level prediction model library that reflects the complex nonlinear relationship between the water level of the hydrological control station and various forecast characteristic factors in different scheduling periods is as follows: Water level forecast characteristic factor set by data normalization D 2. Preprocess the decision water level sequence data; The water level forecast characteristic factor set D 2. Divide the training set and test set into decision water level sequence; XGBoost, LightGBM, and SVM were used to train the water level prediction model on the training set, and the model hyperparameters were optimized during the training process; Use model evaluation indicators to test and evaluate the model on the test set to test the model's generalization ability to the data; The optimal water level prediction model based on XGBoost, LightGBM and SVM was obtained.

[0029] Preferably, in step S6, coupling correction is performed on multiple machine learning prediction results in different scheduling periods to establish a water level stage combination forecast model. The intelligent optimization algorithm is used to optimize the optimal parameter combination of the models in the hydrological forecast model library and the water level stage combination forecast model during the training process, including: The water level stage combination forecast model is constructed as follows: ; Where, is the predicted value of the water level combination prediction model; 、 、 They are the predicted values ​​of LightGBM, XGBoost and SVM prediction models respectively. 、 、 are the corresponding weight values, , They are drawdown period, flood season and water storage period; An intelligent optimization algorithm is used to optimize the weight coefficient of the phased combination prediction model. The intelligent optimization algorithm is one of the particle swarm algorithm, genetic algorithm, simulated annealing algorithm, tabu search algorithm or ant colony algorithm.

[0030] Preferably, in step S1, the specific process of optimizing the weight coefficient of the phased combination prediction model using the particle swarm optimization algorithm as the intelligent optimization algorithm is as follows: a. The optimal water level prediction models based on XGBoost, LightGBM and SVM respectively obtain three groups of water level prediction values, namely 、 、 ; b. Initialize the particle swarm, where each particle represents a weighted combination; c. Calculate the fitness value of each particle, that is ; d. Update the global optimal particles and the local optimal particles; e. Update the speed and position of the particles according to the speed and position update formula of the particle swarm algorithm; f. Determine whether the maximum number of iterations or the set termination condition has been reached; g. If the termination condition is met, the process ends; otherwise, return to step c to continue the optimization.

[0031] Preferably, in step S7, multiple model evaluation indicators are used to evaluate and optimize different water level prediction models, and obtaining the optimal water level prediction value includes: Model evaluation indicators include the coefficient of determinationR 2 , mean absolute error MAE , mean absolute percentage error MAPE , root mean square error RMSE ,in R 2 The larger the value, the better the model fitting effect, and the smaller the value of other indicators, the better the prediction effect of the model. The calculation formula of each indicator is as follows: ; ; ; ; Where, is the measured value, is the predicted value, n is the number of samples, is the measured sample mean.

[0032] Example 2: This embodiment discloses the specific process of the Gini coefficient method based on the random forest model in step S3: The random forest model uses the Gini coefficient to evaluate the importance of each feature factor, and uses the binary recursive partitioning technique to generate a simple binary tree as the basis for selecting influencing factors, thus avoiding the subjectivity of feature selection. D The Gini coefficient, and then calculate the characteristic factors d The Gini coefficient.

[0033] The calculation formula of the Gini coefficient of the characteristic factor set is: ; Where, is the Gini coefficient; N is the size of the feature factor set; p i For the i Characteristic factors d i The proportion in the feature factor set.

[0034] Characteristic Factor d The calculation formula of the Gini coefficient is: ; Where: v is the characteristic factor set D The number of subsets to be divided; For the i feature subsets.

[0035] Example 3: This embodiment discloses the specific process of the feature importance scoring method based on the LightGBM model in step S3: There are two ways to evaluate the importance of feature factors in the LightGBM model. One is to compare the total number of times each feature is split in all decisions, and the other is to calculate the information gain obtained by the feature as a split point in all decision trees. The more times or the greater the gain, the more important the feature. f The importance of is the feature of each decision tree in the LightGBM model f The result of superposition of importance, feature f In a decision tree, a node d The importance calculation formula is: ; Where, w d For nodes d The ratio of the amount of data to the total amount of data; For nodes d Information gain after splitting.

[0036] Example 4: This example uses water level forecasting at the S City station in the middle and lower reaches of the Yangtze River as a case study. Considering the new water and sediment conditions after the Three Gorges Reservoir is normally filled to 175 m, the study uses measured water level and flow data from the S City Hydrological Control Station from January 1, 2010, to October 31, 2024, as well as measured water level data from Chenglingji, and measured outflow data from the Three Gorges Reservoir and Qingjiang Gaobazhou. A total of 5,113 data samples from 2010 to 2023 serve as the training set for the prediction model, while 305 data samples from January 1 to October 31, 2024, serve as the test set. Missing data in the original data are filled using linear interpolation.

[0037] The contribution of various factors influencing the water level-discharge relationship at S City Station was assessed using the Gini coefficient method (i.e., contribution degree) of the random forest model and the feature importance scoring method based on the LightGBM model. The evaluation results are shown in Table 1. As can be seen from Table 1, the water level at S City Station has the greatest impact on the discharge at S City Station, followed by the water level difference between S City and the LH hydrological station. Other factors have relatively low contributions to the discharge at S City Station. Therefore, the water level at S City Station and the water level difference between S City and the LH hydrological station are used as the key factors influencing the discharge at S City Station.

[0038] Table 1: Contribution assessment of factors affecting the relationship between water level and flow at S city station;

[0039] The prediction models were trained using a total of 5,113 data samples from 2010 to 2023. The PSO algorithm was used to optimize the key hyperparameters of LightGBM, XGBoost, and SVM. The prediction performance of each prediction model on the test samples is shown in Table 2.

[0040] Table 2: Evaluation index values ​​of prediction effect of each prediction model;

[0041] As can be seen from Table 2, the coefficient of determination of each model prediction R 2 The values ​​all reached 0.99, indicating that the model has a good fitting effect; from other evaluation indicators, it can be seen that the combined prediction model RMSE, MAE, MAPE The value of is the lowest compared with other models, indicating that its prediction effect is the best. MAE The value is 0.071m, which is lower than 0.1m, indicating that it has a high prediction accuracy.

[0042] Figure 2-Figure 5 The prediction effect of each prediction model on the test set is intuitively demonstrated; Figure 2-Figure 5 It can be seen that the overall prediction trend of the combined prediction model is basically consistent with the test samples, and it shows better prediction results than other prediction models in high, medium, and low water levels, as well as in the water level rise and fall stages. Table 3 shows the distribution of absolute errors and the range of absolute errors of each model for the water level of S City Station on the test samples.

[0043] Table 3: Statistics of absolute error range of water level prediction at Shashi station;

[0044] As can be seen from Table 3, the accuracy of the combined prediction model in predicting water level absolute error within ±0.1m is 76.72%, which is significantly better than other prediction models and shows high prediction accuracy.

[0045] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the methods and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A combined forecasting method for river water level along the river based on multi-model coupling correction, characterized in that: The following steps are involved: S1: Collect and organize hydrological data of the hydrological control station, its upstream main and tributary reservoirs, and its downstream adjacent hydrological control stations; SS2, based on the water level-discharge relationship of the hydrological control station, constructs the set of hydrological influencing factors of the water level-discharge relationship of the hydrological control station; S3, using a variety of feature selection techniques to identify the key factors affecting the relationship between water level and flow, and considering the propagation time of water from upstream main and tributary rivers and the duration of the impact, the coupled water volume influencing factors together form the water level forecast feature factor set; S4, from the perspective of reservoir operation period and inflow characteristics, the water level forecast characteristic factor set is divided into three subsets: drawdown period, flood season, and water storage period; S5, based on a variety of machine learning methods, constructs a water level prediction model library that reflects the complex nonlinear relationship between the water level of the hydrological control station and various forecast characteristic factors in different scheduling periods; S6, couple and calibrate multiple machine learning prediction results in different scheduling periods, establish a water level combination forecast model, and use intelligent optimization algorithms to optimize the optimal parameter combination of each model in the hydrological forecast model library and the water level period combination forecast model during the training process; S7, using a variety of model evaluation indicators, evaluates and optimizes different water level prediction models to obtain the optimal water level prediction value.

2. The method for combined forecasting of water level along a river based on multi-model coupling correction according to claim 1 is characterized in that: In step S1, the hydrological data includes the current hydrological control station series water level Z c and flow Q c , upstream main and tributary reservoir long series discharge flow Q d As well as the long series of water level data of the downstream hydrological control station; the downstream hydrological control station refers to the hydrological control station that is closest to the current hydrological control station and is not affected by the backwater support effect.

3. The method for combined forecasting of water level along a river based on multi-model coupling correction according to claim 1 is characterized in that: In step S2, a set of hydrological influencing factors of the relationship between water level and flow at the hydrological control station is constructed. D , including the current water level at the hydrological control station Z c , water level fluctuations , the water level of the hydrological control station adjacent to the downstream and not affected by the jacking effect Z d , water level fluctuations , the water level difference between the current hydrological control station and the downstream hydrological control station .

4. The method for combined forecasting of water level along a river based on multi-model coupling correction according to claim 1 is characterized in that: In step S3, a variety of feature selection techniques are used to identify key factors affecting the relationship between water level and flow. The propagation time of water from upstream main and tributary rivers and the duration of the impact are taken into consideration. The coupled water flow influencing factors together form the water level forecast feature factor set, which includes: The Gini coefficient method based on the random forest model and the feature importance scoring method based on the LightGBM model were used as feature selection methods to screen the pre-selected feature factors respectively, and then the factors with the highest scores were merged to obtain the final set of feature factors.

5. The method for combined forecasting of water level along a river based on multi-model coupling correction according to claim 4 is characterized in that: The Gini coefficient method of the random forest model and the feature importance scoring method based on the LightGBM model are used to evaluate the hydrological influencing factor set of the water level and flow relationship of the hydrological control station constructed in step S2. D Perform importance scoring, sort the influencing factors by score size, and finally obtain the key influencing factors by combining the top-ranked influencing factors under the two scoring methods. D 1; At the same time, the flow rate of the hydrological control station itself and the water volume related factors of the upstream reservoir discharge and the main and tributary water flow are integrated to form the water level forecast characteristic factor set. D 2.

6. The method for combined forecasting of water level along a river based on multi-model coupling correction according to claim 1 is characterized in that: In step S5, the multiple machine learning methods used include but are not limited to extreme gradient boosting decision tree XGBoost, lightweight gradient boosting decision tree LightGBM and support vector machine regression model SVM.

7. The method for combined forecasting of water level along a river based on multi-model coupling correction according to claim 6 is characterized in that: Based on a variety of machine learning methods, the specific process of constructing a water level prediction model library that reflects the complex nonlinear relationship between the water level of the hydrological control station and various forecast characteristic factors in different scheduling periods is as follows: Water level forecast characteristic factor set by data normalization D 2. Preprocess the decision water level sequence data; The water level forecast characteristic factor set D 2. Divide the training set and test set into decision water level sequence; XGBoost, LightGBM, and SVM were used to train the water level prediction model on the training set, and the model hyperparameters were optimized during the training process; Use model evaluation indicators to test and evaluate the model on the test set to test the model's generalization ability to the data; The optimal water level prediction model based on XGBoost, LightGBM and SVM was obtained.

8. The method for combined forecasting of water level along a river based on multi-model coupling correction according to claim 1 is characterized in that: In step S6, a plurality of machine learning prediction results are coupled and corrected in different scheduling periods to establish a water level stage combination forecast model. The intelligent optimization algorithm is used to optimize the optimal parameter combination of the models in the hydrological prediction model library and the water level stage combination forecast model during the training process, including: The water level stage combination forecast model is constructed as follows: ； Where, is the predicted value of the water level combination prediction model; 、 、 They are the predicted values ​​of LightGBM, XGBoost and SVM prediction models respectively. 、 、 are the corresponding weight values, , They are drawdown period, flood season and water storage period; An intelligent optimization algorithm is used to optimize the weight coefficient of the phased combination prediction model. The intelligent optimization algorithm is one of the particle swarm algorithm, genetic algorithm, simulated annealing algorithm, tabu search algorithm or ant colony algorithm.

9. The method for combined forecasting of water level along a river based on multi-model coupling correction according to claim 8 is characterized in that: In step S1, the specific process of optimizing the weight coefficient of the phased combination prediction model using the particle swarm optimization algorithm as the intelligent optimization algorithm is as follows: a. The optimal water level prediction models based on XGBoost, LightGBM and SVM respectively obtain three groups of water level prediction values, namely 、 、 ; b. Initialize the particle swarm, where each particle represents a weighted combination; c. Calculate the fitness value of each particle, that is ; d. Update the global optimal particles and the local optimal particles; e. Update the speed and position of the particles according to the speed and position update formula of the particle swarm algorithm; f. Determine whether the maximum number of iterations or the set termination condition has been reached; g. If the termination condition is met, the process ends; otherwise, return to step c to continue the optimization.

10. The method for combined forecasting of water level along a river based on multi-model coupling correction according to claim 1, characterized in that: In step S7, various model evaluation indicators are used to evaluate and optimize different water level prediction models, and the optimal water level prediction value is obtained, including: Model evaluation indicators include the coefficient of determination R 2 , mean absolute error MAE , mean absolute percentage error MAPE , root mean square error RMSE ,in R 2 The larger the value, the better the model fitting effect, and the smaller the value of other indicators, the better the prediction effect of the model. The calculation formula of each indicator is as follows: ； ； ； ； Where, is the measured value, is the predicted value, n is the number of samples, is the measured sample mean.

Images