Complex river network flood prediction method based on data dynamic cleaning and adaptive recurrent neural network

Abstract

The invention discloses a complex river network flood prediction method based on data dynamic cleaning and an adaptive recurrent neural network. The method comprises the following steps: S1, collecting flood monitoring data in a research area; s2, abnormal value detection and correction are carried out on the data through a speed constraint dynamic cleaning method based on flood classification; s3, preprocessing the cleaned data by using a sliding window mechanism and a data normalization technology, and setting a plurality of prediction period windows for direct multi-step prediction; s4, constructing and optimizing an adaptive recurrent neural network model, wherein the model comprises a gating circulation unit layer and a bottleneck layer; s5, optimizing the hyper-parameters through grid search and an adaptive momentum estimation optimizer, and finally determining model configuration; s6, performing complete training and testing based on the determined model configuration; and S7, flood prediction is carried out through the trained model. The model is combined with actual monitoring data, intelligent river network flood forecasting in different forecasting periods is achieved, and the flood forecasting precision and stability are remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the fields of flood control, disaster reduction and intelligent water safety, and particularly relates to a complex river network flood prediction method based on data dynamic cleaning and adaptive recurrent neural network. Background Art

[0002] Flood disasters in China occur highly frequently, causing huge impacts and losses. With the rapid development of social economy, basin floods and their associated secondary disasters may pose a greater threat to the flood control safety of the basin. Conducting rapid and accurate basin flood prediction and improving the scientific, professional, intelligent and refined level of basin flood disaster prevention are key issues of concern in basin flood disaster prevention.

[0003] Currently, basin flood prediction methods are mainly divided into two types: hydrological and hydrodynamic methods and data-driven methods. Hydrological and hydrodynamic methods usually use mathematical models to simulate the hydrological cycle process based on factors such as the terrain, soil type, and vegetation cover of the basin to predict future flood conditions. Data-driven methods establish flood prediction models through a large amount of observational data and machine learning algorithms, reducing the prior assumptions about basin characteristics and being able to better adapt to the flood prediction needs of different regions and time periods. The main problems are that hydrological and hydrodynamic methods may face challenges such as model parameterization and uncertainty propagation when dealing with complex basin systems; the flood flow process in complex river networks has characteristics such as strong nonlinearity, non-stationarity, and dynamic changes in temporal distribution, and the existing river network flood prediction methods have limited ability to learn complex spatio-temporal patterns from multivariate time series and need to effectively model them; for river network flood prediction, the input of the model is often only the historical monitoring data of the river network, and the quality of the input data greatly affects the performance of the model.

[0004] Aiming at the problem of complex river network flood prediction, the present invention proposes a data cleaning method based on flood classification speed constraint, constructs an adaptive learning recurrent neural network deep learning prediction model, conducts complex river network flood prediction with different lead times, avoids the influence of abnormal monitoring values, realizes rapid and intelligent prediction of complex river networks, and improves the accuracy of prediction results. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to solve the problem of rapid calculation of the non-linear variation of the flood routing process and water depth distribution in complex river networks, and provide technical support for flood disaster prevention in the basin. To this end, the present invention provides a complex river network flood prediction method based on data dynamic cleaning and adaptive recurrent neural network, proposes a data cleaning method based on flood level velocity constraint, constructs an adaptive recurrent neural network deep learning prediction model, conducts complex river network flood prediction with different lead times, avoids the influence of abnormal monitoring values, realizes rapid and automatic prediction of complex river networks, and improves the accuracy of prediction results. Among them, a complex river network refers to a geographical structure with multiple intersecting and interconnected rivers, channels, and water systems in a basin. Such river networks usually have complex hydrodynamic characteristics and hydrological cycle laws.

[0006] The present invention adopts the following technical solutions:

[0007] A complex river network flood prediction method based on data dynamic cleaning and adaptive recurrent neural network, comprising the following steps:

[0008] Step S1: Collect flood monitoring data within a preset time range in the study area, including hourly flow and water level monitoring data of upstream and downstream hydrological stations and water level stations;

[0009] Step S2: Through an automatic dynamic cleaning method for flow and water level data dynamically determined based on the velocity constraint of flood level, according to the change rate in the flow and water level monitoring data, conduct anomaly detection and correction on the flood monitoring data;

[0010] Step S3: Use the sliding window mechanism and data normalization to preprocess the cleaned data, and set multiple lead time windows for direct multi-step prediction based on the preprocessed data;

[0011] Step S4: Establish a deep learning prediction model based on an adaptive recurrent neural network. The model includes at least one gated recurrent unit layer and a bottleneck layer, configure an appropriate number of neurons, and initialize the model parameters using a method for uniformly initializing the neural network parameters, and determine the hyperparameters of the model, including the number of gated recurrent unit layers, the number of bottleneck layers, the learning rate, and the batch size;

[0012] Step S5: By selecting evaluation indicators, including root mean square error, mean absolute error, and Nash efficiency coefficient, combined with the grid search method, and using the adaptive momentum estimation optimizer to train the model and optimize the hyperparameters to determine the optimal hyperparameters;

[0013] Step S6: Based on the optimal hyperparameters determined in Step S5, conduct complete training of the model and save the trained model parameters;

[0014] Step S7: Utilizing the saved model parameters and combining the actual monitoring data of each hydrological station upstream and downstream of the river network, automatic flood prediction of the river network for different forecast periods is carried out.

[0015] Furthermore, step S2 includes the following steps:

[0016] Step S201: Analyze the key features of the flow and water level monitoring data during the flood process, including peak flow, duration, and rise and fall speed, and use a clustering algorithm to divide the flood process into different levels or categories;

[0017] Step S202: Analyze the rate limits of change in the flow and water level monitoring data for each level of flood process, and formulate speed change constraints applicable to different levels of floods based on these rate limits;

[0018] Step S203: Based on the established speed change constraints, the flow and water level monitoring data of each level of flood are processed separately, including time series outlier detection and data repair.

[0019] Furthermore, in step S2, the time series outlier detection and data repair specifically include time series outlier detection and time series outlier data repair, wherein:

[0020] Time series outlier detection: Set the time series X = [x1, x2, ..., x n ], where each x i Indicates the value of the i-th data point, and sets the interval length to WD, and sets the speed constraint S = (S min , S max ), where the minimum speed S min and maximum speed S max Used to limit the rate of change between data points; for any time interval t j —t i Satisfy 0 <t j —t i <WD的两个数据点x i and x j , requiring sequence X to satisfy the following speed constraints:

[0021]

[0022] Time series outlier data repair: For data points x that do not meet the above conditions i , without changing the time point t in the time series X i In the case of data points x in the sequence iRepair is performed so that the repaired sequence X' satisfies the above speed constraint S; during the repair process, the principle of minimum modification is followed, and the repair distance Δ(x, x') is expressed as the difference between the original time series X and the cleaned time series X', and its calculation formula is:

[0023]

[0024] Furthermore, in step S3, the multiple look-ahead windows for direct multi-step prediction are set as follows: assume the time series X = [x1, x2, …, x n , where each x i represents the value of the i-th data point, the length of the sliding window is L, the moving step size is 1, and L ≤ n. Different look-ahead windows include different numbers of consecutive data points.

[0025] Furthermore, step S4 includes the following steps:

[0026] Step S401: Construct an adaptive recurrent neural network model architecture, including at least one gated recurrent unit layer and at least one bottleneck layer, and determine the number of neurons according to the dimensional and temporal characteristics of the historical flood discharge and water level monitoring data in the study area;

[0027] Step S402: Initialize the weight parameters of the adaptive recurrent neural network model using a method for uniformly initializing neural network parameters;

[0028] Step S403: Configure the hyperparameters for model training, including the number of gated recurrent unit layers, the number of bottleneck layers, the learning rate, and the batch size, to ensure that these initial settings adapt to the training requirements of the model.

[0029] Furthermore, step S5 includes the following steps:

[0030] Step S501: Select evaluation metrics, including root mean square error, mean absolute error, and Nash efficiency coefficient, where the root mean square error and mean absolute error metrics are used to indicate the prediction error, and the Nash efficiency coefficient is used to indicate the prediction accuracy;

[0031] Step S502: Combine the time series similarity quantification and time series distribution matching algorithm techniques to process the time series data, and use the grid search method to optimize the configured hyperparameters using the adaptive momentum estimation optimizer;

[0032] Step S503: Adjust the hyperparameters according to the optimization results, including the number of gated recurrent unit layers, the number of bottleneck layers, the learning rate, and the batch size, to establish a time series prediction model with optimal performance.

[0033] Compared with the prior art, the advantages of the present invention are:

[0034] 1. The data cleaning method based on the flood classification speed constraint can effectively handle single-point anomalies and consecutive outliers. According to the flow change limitations existing in the flood evolution process, the flood process is divided into different levels using a clustering algorithm, the different characteristics and severity of the flood process are analyzed, the speed constraints corresponding to different levels of floods are formulated, and the speed constraint values are determined, effectively cleaning the single-point anomaly data that is too large or too small in the original values, avoiding the obvious order-of-magnitude differences between the anomaly data and the normal data, and improving the accuracy and robustness in anomaly detection.

[0035] 2. The prediction results of traditional time series analysis methods may not be ideal, and they can only predict non-linear and complex flood changes to a limited extent; the methods based on deep learning usually obtain better prediction results than traditional time series analysis methods, effectively improving the intelligence of spatio-temporal feature extraction and prediction accuracy, having better adaptability of the model, reducing the risk of the algorithm falling into local optima, solving the problem of rapid calculation of the non-linear changes in the flood evolution process and water depth distribution in complex river networks, and providing technical support for basin flood disaster prevention. Description of the Drawings

[0036] Figure 1 It is a schematic flow chart of the complex river network flood prediction method based on the present invention.

[0037] Figure 2 It is a schematic diagram of the input-output structure of the direct multi-step prediction model of the present invention.

[0038] Figure 3 It is a comparison chart of the flood forecast effect evaluation indexes in a typical area of the present invention. Detailed Embodiments

[0039] To make the objectives, technical solutions and advantages of the embodiments of the present invention clearer, the accompanying drawings in the embodiments of the present invention will be combined below, and several key terms involved in the present application will be explained first, and then the technical solutions in the embodiments of the present invention will be described clearly and completely. It should be understood that the embodiments described below are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0040] A complex river network flood prediction method based on data dynamic cleaning and adaptive recurrent neural network includes the following steps:

[0041] Step S1: Collect flood monitoring data within a preset time range in the research area, including hourly flow and water level monitoring data of each upstream and downstream hydrological station and water level station;

[0042] Step S2: An automatic dynamic cleaning method determined by velocity constraints based on flood grading, according to the change rates in the flow rate and water level monitoring data, performs anomaly detection and correction on the flood monitoring data;

[0043] Step S3: Using a sliding window mechanism and data normalization, preprocess the cleaned data, and set multiple look-ahead windows for direct multi-step prediction based on the preprocessed data;

[0044] Step S4: Establish a deep learning prediction model based on an adaptive recurrent neural network. The model includes at least one gated recurrent unit layer and a bottleneck layer, configure an appropriate number of neurons, and initialize the model parameters using a method for uniformly initializing the neural network parameters. Determine the hyperparameters of the model, including the number of gated recurrent unit layers, the number of bottleneck layers, the learning rate, and the batch size;

[0045] Step S5: By selecting evaluation metrics, including root mean square error, mean absolute error, and Nash efficiency coefficient, combined with a grid search method, and using an adaptive momentum estimation optimizer to train the model and optimize the hyperparameters to determine the optimal hyperparameters;

[0046] Step S6: Based on the optimal hyperparameters determined in Step S5, perform a complete training of the model and save the trained model parameters;

[0047] Step S7: Using the saved model parameters, combined with the actual monitoring data of each hydrological station in the upper and lower reaches of the river network, carry out automatic prediction of river network floods with different look-ahead periods.

[0048] The following further elaborates on the preferred embodiments of the present invention.

[0049] First, execute Step S1 to collect flood monitoring data within a preset time range in the study area, including hourly flow rate and water level monitoring data of each hydrological station and water level station in the upper and lower reaches. The collection and recording of these data are jointly completed by sensors, communication networks, and data warehouses. During the data transmission process, due to interference from the surrounding environment, problems such as unstable transmission of sensor-collected data, network latency, or packet loss may occur, resulting in abnormal flood monitoring data and thus affecting the recording results in the data warehouse.

[0050] Abnormal flood monitoring data can be caused by various factors, including data acquisition sensor failures, communication network failures, or emergencies. However, the actual monitoring data usually has a certain limit on the rate of change, which can be used as a basis for judging outliers. In a time series, the "jump" of data values is usually restricted, and this restriction is the speed constraint. Therefore, in step S2, the speed change constraint condition is used to detect outliers and clean data in the time series. Consider the sequence X = [x1, x2, …, x n , where each x i represents the value of the i-th data point. The speed constraint S = (S min , S max ) with an interval length of WD, that is, the minimum speed S min and the maximum speed S max . The so-called interval size WD refers to a time period with a length of WD. If the timestamps of two data points satisfy 0 < t j - t i < WD, then these two data points are said to be in one interval. It is said that the sequence X satisfies the speed constraint s, which is formally expressed as x| = S. For any two data points x i , x j in one interval of this sequence, the following formula must be satisfied:

[0051]

[0052] A repair (or cleaning) x i of the time series X is to change the value corresponding to the data point without changing the timestamps of the time series, that is, t′ i = t i , x| = S. According to the principle of minimum modification in the field of data cleaning, the so-called modification distance refers to the difference between the original time series x and the cleaned time series x′ i :

[0053]

[0054] The present invention also proposes a method for dynamically determining the speed constraint based on flood grading. When the characteristics of runoff generation and concentration in the study area are relatively complex, it is often difficult to accurately detect and correct flood data using a data cleaning algorithm with a single parameter. To improve the accuracy of data cleaning, the present invention classifies the rainfall-runoff data so that floods with similar data characteristics share a set of correction parameters. The flood process is divided into different levels through a clustering algorithm, the characteristics and severity of the flood process are analyzed, and the speed constraint conditions corresponding to floods at different levels are formulated. In this embodiment, the number of clusters of the clustering algorithm can be set to 4 or 5, the maximum number of iterations can be set to 300, and the tolerance for the change in the centroid position can be set to 1e-4.

[0055] Based on the clustering results, flood processes can be divided into different levels or categories, representing varying degrees of flood severity. Based on the characteristics of the watershed and the magnitude of rainfall and runoff, flood processes can be divided into four or five levels. By analyzing the characteristics and changing trends of each level, corresponding velocity constraints are further formulated to reflect the laws and characteristics of flood processes at each level.

[0056] Next, execute step S3. Considering that the intelligent prediction model of complex river network floods with different forecast periods can provide better early warning for basin flood defense, the present invention adopts a direct multi-step prediction method to directly generate multi-step prediction values through the model, avoiding the error transmission problem and thus improving the prediction accuracy. This embodiment performs direct multi-step prediction of flood flow for different forecast periods. The direct multi-step prediction is implemented in the model through a sliding window mechanism. Suppose a flood flow time series is X = [x1, x2, ..., x n ], the model input step is L, that is, the current window length of the sliding window is L; the moving step of the sliding window is 1, and the forecast period is 1h, 2h, 3h, 6h, 12h, 24h. When the forecast period is a, the input and output structure of the direct multi-step forecast model for the current time t to predict a hours backward is as follows Figure 2 shown.

[0057] Furthermore, to improve model training and evaluate its performance and generalization, this paper employs mean normalization and a reasonable dataset partitioning strategy. The original data is mean normalized to constrain the data range to [0, 1]. The normalized data is then partitioned into training, validation, and test sets in a ratio of 6:3:1, respectively.

[0058] Next, steps S4 and S5 of model building and hyperparameter optimization are performed.

[0059] In step S4, a deep learning prediction model based on an adaptive recurrent neural network is first established. The model includes at least one gated recurrent unit layer and a bottleneck layer, and an appropriate number of neurons are configured to optimize the expressive power of the model. The gated recurrent unit layer can effectively capture the long-term and short-term dependencies in complex time series data, while the bottleneck layer helps prevent overfitting problems and improves the generalization ability of the model by reducing the number of parameters of the model. At the same time, the weight parameters of the model are initialized using a method for uniformly initializing neural network parameters to ensure stability and convergence during training. On this basis, the key hyperparameters of the model are determined, including the number of gated recurrent unit layers, the number of bottleneck layers, the learning rate, and the batch size, to ensure that the model has good initial conditions and lay the foundation for subsequent training and optimization.

[0060] In step S5, the prediction effect of the model is further improved through model training and hyperparameter optimization. The specific operation includes selecting evaluation indicators, focusing on using root mean square error, mean absolute error and Nash efficiency coefficient to measure the prediction performance of the model. The root mean square error and mean absolute error are used to evaluate the prediction error, among which the root mean square error is more sensitive to larger errors, while the mean absolute error gives equal weight to all errors; the Nash efficiency coefficient is used to evaluate the prediction accuracy of the model. The closer the value is to 1, the stronger the prediction ability of the model. The calculation formulas of the root mean square error, mean absolute error and Nash efficiency coefficient are shown in formulas 5 to 7, where RMSE represents the root mean square error, MAE represents the mean absolute error, and NSE represents the Nash efficiency coefficient:

[0061]

[0062] Where: n is the number of experimental predictions; is the predicted value of the i-th model; Y i is the i-th true value; is the average of the true values.

[0063] In order to optimize the prediction accuracy of the adaptive recurrent neural network model for different time periods, we first introduce the time series similarity quantification technology to divide the time series data into the most dissimilar sequences, maximize the use of the shared knowledge of the time series, make the distribution of each period as diverse as possible, and enhance the generalization ability of the model, as shown in Formula 8:

[0064]

[0065] Where: d is the distance metric, is the similarity metric function, d(·,·) is the distance metric function, KT is the number of time series segments, n is the length of the time series, KT0 is a hyperparameter to avoid over-segmentation, is the i-th time series data, Δ1 and Δ2 are predefined parameters.

[0066] This process helps to accurately quantify the distribution information of each time period and provide more accurate input features for subsequent model training. Then, TDM technology is applied. TDM proposes a method similar to domain generalization to learn the optimal model parameters θ ＊ , and dynamically measure the importance of each time state in the RNN unit, using α ＊ Indicates that the learning process of this method is shown in Formulas 9 to 11:

[0067]

[0068] Where: is the TDM loss function used for prediction; Represents a period The k-th label segment; l(·,·) is the MAE loss function; θ is the learning model parameter; is the learning model; is the distribution difference between any two segments in the KT time series; is the hidden state with feature dimension at the t-th moment in the i-th period, is and the distribution importance between them at the t-th moment.

[0069] The time series distribution matching algorithm effectively combines the dynamic change characteristics of the time series distribution, and through the recursive neural network architecture, finally optimizes the performance of the time series prediction model. Subsequently, combined with the grid search method, different hyperparameter combinations are systematically explored, and the model is trained using the adaptive momentum estimation optimizer to maximize the objective function and ensure the minimum error between the actual flood flow and the predicted flood flow. The finally determined hyperparameter configuration will be used to improve the prediction performance of the model and form an optimal time series prediction model.

[0070] To verify the effectiveness of the adaptive recursive neural network model, several benchmark models are also established and tested, including the support vector machine regression model, the artificial neural network model, and the long short-term memory model. These benchmark models will be tested in combination with different data cleaning methods, specifically including not using data cleaning methods, using historical average data cleaning methods, and using dynamic data cleaning methods based on flood grading. By applying different data cleaning methods, the influence of data preprocessing on the model prediction effect can be verified for each benchmark model.

[0071] A total of eight test schemes are set, namely: Scheme 1 "Adaptive Recursive Neural Network Model", Scheme 2 "Support Vector Machine Regression Model", Scheme 3 "Artificial Neural Network Model", Scheme 4 "Long Short-Term Memory Model", Scheme 5 "Adaptive Recursive Neural Network Model + Dynamic Data Cleaning Algorithm Based on Flood Grading", Scheme 6 "Support Vector Machine Regression Model + Dynamic Data Cleaning Algorithm Based on Flood Grading", Scheme 7 "Artificial Neural Network Model + Dynamic Data Cleaning Algorithm Based on Flood Grading", Scheme 8 "Long Short-Term Memory Model + Dynamic Data Cleaning Algorithm Based on Flood Grading". The comparison chart of the flood forecasting effect evaluation indicators in the typical area is as Figure 3 shown.

[0072] The same hyperparameter search and optimization strategy was employed across all schemes. Specifically, a grid search approach was used to systematically explore the optimal hyperparameter combination for each model, and an adaptive momentum estimation optimizer was employed for model training and parameter adjustment, ensuring comparable results across different test scenarios. Across all schemes, root mean square error, mean absolute error, and Nash efficiency coefficient were used as unified evaluation metrics to ensure a fair comparative analysis of the prediction performance of different models. By comparing and analyzing the prediction performance of these schemes, the significant advantages of the adaptive recurrent neural network model in flood prediction for complex river networks were ultimately verified.

[0073] Furthermore, for each model's specific experimental process, the ratio of the training set, validation set, and test set was 6:3:1, respectively, to ensure sufficient model training and reasonable model validation. Each model used the same initial conditions and data preprocessing process to reduce variable interference during the experiment. This rigorous experimental design and systematic hyperparameter optimization fully demonstrated the performance differences between the models under different data processing conditions, highlighting the dominant position of the adaptive recurrent neural network model in complex river network flood prediction.

[0074] Finally, steps S6 and S7 were executed. After grid search and hyperparameter optimization, the optimal time series forecasting model was identified and trained. This model underwent rigorous testing to ensure its stability and accuracy in various environments. Subsequently, this model, combined with actual monitoring data, was used to perform intelligent flood forecasting for different forecast periods, providing efficient and reliable technical support for early warning and management of floods in the basin.

[0075] Through the overall method of the present invention, accurate prediction of floods in complex river networks was successfully achieved, which greatly improved the accuracy of prediction and the generalization ability of the model, and has important application value and broad promotion prospects.

[0076] Thus far, the technical solutions of the present invention have been described in conjunction with the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art may make equivalent changes or substitutions to the relevant technical features, and the technical solutions after such changes or substitutions will fall within the scope of protection of the present invention.

Claims

1. A complex river network flood prediction method based on data dynamic cleaning and adaptive recurrent neural network, characterized in that, It includes the following steps: Step S1: Collect flood monitoring data within a preset time range in the flood prediction area, including hourly flow and water level monitoring data of upstream and downstream hydrological stations and water level stations; Step S2: Through an automatic dynamic cleaning method for flow and water level data dynamically determined based on flood classification and velocity constraints, and based on the change rates in the flow and water level monitoring data, conduct anomaly detection and correction on the flood monitoring data; Step S3: Use a sliding window mechanism and data normalization to preprocess the cleaned data, and set multiple look-ahead windows for direct multi-step prediction based on the preprocessed data; Step S4: Establish a deep learning prediction model based on an adaptive recurrent neural network. The deep learning prediction model includes at least one gated recurrent unit layer and a bottleneck layer. Configure the number of neurons, and use a method for uniformly initializing neural network parameters to initialize the model parameters, and determine the hyperparameters of the deep learning prediction model, including the number of gated recurrent unit layers, the number of bottleneck layers, the learning rate, and the batch size; Step S5: Select metrics for evaluating the performance of the deep learning prediction model, including root mean square error, mean absolute error, and Nash efficiency coefficient. Combine with the grid search method, and use the adaptive moment estimation optimizer to train and optimize the hyperparameters of the deep learning prediction model to determine the optimal hyperparameters; Step S6: Based on the optimal hyperparameters determined in Step S5, conduct a complete training of the deep learning prediction model and save the trained model parameters; Step S7: Use the saved model parameters, combined with the actual monitoring data of upstream and downstream hydrological stations in the river network, to carry out automatic prediction of river network floods with different look-ahead periods.

2. The complex river network flood prediction method based on data dynamic cleaning and adaptive recurrent neural network according to claim 1, wherein, Step S2 includes the following steps: Step S201: Analyze the key features of the flow and water level monitoring data during the flood process, including peak flow, duration, and rising and falling speeds, and use a clustering algorithm to divide the flood process into different levels; Step S202: For each level of flood process, analyze the change rate limits in its flow and water level monitoring data, and respectively formulate velocity change constraints applicable to different levels of floods according to the change rate limits; Step S203: Based on the formulated velocity change constraints, respectively process the flow and water level monitoring data of each level of flood, including time series outlier detection and data repair.

3. The complex river network flood prediction method based on dynamic data cleaning and adaptive recursive neural network according to claim 2 is characterized in that: In Step S2, the time series outlier detection and data repair specifically include time series outlier detection and time series outlier data repair, where: Time series outlier detection: Set the time series X = [x1, x2, ..., x n ], where each x i Indicates that at the i-th time point t i The time series data point value is set, and the time series interval length is set to the period WD, and the speed change constraint S=(S min , S max ), where the minimum speed S min and maximum speed S max Used to limit the speed change between time series data points; for any two data points x in the time series X i and x j The corresponding time point t i and t j , if the time interval t j —t i Satisfy 0 <t j —t i <WD,则要求序列X满足以下速度约束: Time series outlier data repair: For time series data points x that do not meet the above conditions i , without changing the time point t in the time series X i In the case of time series data point x i Repair is performed so that the repaired sequence X' satisfies the above speed constraint S. The repair process follows the principle of minimum modification. The repair distance Δ(x, x') is expressed as the difference between the original time series X and the cleaned time series X', and its calculation formula is:

4. The complex river network flood prediction method based on dynamic data cleaning and adaptive recursive neural network according to claim 1 is characterized in that: In step S3, the multiple look-ahead windows for direct multi-step prediction are set as follows: Given a time series X = [x1, x2, …, x n , where each x i represents the value of the time series data point at the i-th time point t i . The length of the time series sliding window is L, the moving step size is 1, and L ≤ n. Each sliding window includes L consecutive time series data points. Different look-ahead windows are used to achieve multi-step prediction for different future time points by setting time series sliding windows of different lengths.

5. The complex river network flood prediction method based on data dynamic cleaning and adaptive recurrent neural network according to claim 1, characterized in that, Step S4 includes the following steps: Step S401: Construct the deep learning prediction model architecture, including at least one gated recurrent unit layer and at least one bottleneck layer, and determine the number of neurons based on the dimensional and temporal characteristics of the historical flood flow and water level monitoring data in the study area; Step S402: Use a method for uniformly initializing neural network parameters to initialize the weight parameters of the deep learning prediction model; Step S403: Configure the hyperparameters for model training, including the number of gated recurrent unit layers, the number of bottleneck layers, the learning rate, and the batch size, to ensure that the above initial settings adapt to the training requirements of the deep learning prediction model.

6. The complex river network flood prediction method based on dynamic data cleaning and adaptive recursive neural network according to claim 1 is characterized in that: Step S5 includes the following steps: Step S501: Select the metrics for evaluating the performance of the deep learning prediction model, including the root mean square error, the mean absolute error, and the Nash efficiency coefficient, where the root mean square error and the mean absolute error metrics are used to indicate the prediction error, and the Nash efficiency coefficient is used to indicate the prediction accuracy; Step S502: Combine the time series similarity quantification and the time series distribution matching algorithm techniques to process the time series data, and use the grid search method to optimize the hyperparameters of the deep learning prediction model with the adaptive moment estimation optimizer; Step S503: Adjust the hyperparameters according to the optimization results, including the number of gated recurrent unit layers, the number of bottleneck layers, the learning rate, and the batch size, to establish a time series prediction model.

Images