An automated simulation method for river basin flood processes based on time series analysis

Abstract

The present invention relates to the technical field of time series data processing, and specifically to a method for automated deduction of basin flood processes based on time series analysis, comprising: obtaining historical hydrological data of the basin, applying adaptive wavelet transform to process it, and extracting a medium-frequency wavelet feature sequence and a transform residual sequence as data features; establishing a first calculation model to generate a first predicted water level sequence; obtaining an error sequence based on the first predicted water level sequence and the actual water level sequence; constructing a second calculation model using data features and rainfall data, and fitting the error sequence using the second calculation model to obtain a predicted error sequence; performing weighted fusion on the first predicted water level sequence, the predicted error sequence, and the rainfall sequence to generate a second predicted water level sequence; combining the second predicted water level sequence with a hydrodynamic numerical calculation model to obtain predicted data through flow field calculation and analysis. The present invention achieves high-precision automated deduction through automated processing and hybrid modeling.

Description

Technical Field

[0001] The present invention relates to the technical field of time series data processing, and in particular to a method for automatically deducing a basin flood process based on time series analysis. Background Art

[0002] Effective computational analysis and quantitative modeling of time-series data for river basin flood processes are fundamental to understanding the dynamics of complex hydrological systems. However, the complexity of actual hydrological series, along with their frequent presence of noise and non-stationary characteristics, poses challenges to the design of data processing algorithms (e.g., handling non-stationarity and normal transformation of data) and the accuracy of model fitting.

[0003] In the existing technology, traditional physical models are complex to calculate, difficult to determine parameters and are often sensitive to parameters; commonly used time series calculation methods use linear statistical models that are difficult to effectively process complex hydrological digital signals or machine learning calculation models whose algorithmic processes are difficult to explain and rely on large amounts of data.

[0004] At present, there are still obvious technical challenges in automatically building an integrated computing process that effectively couples different computing modules with efficient timing modeling and distribution estimation modules to achieve high-precision quantitative modeling.

[0005] Therefore, an automatic simulation method of basin flood process based on time series analysis is proposed. Summary of the Invention

[0006] The purpose of the present invention is to provide a method for automated deduction of basin flood processes based on time series analysis, which realizes high-precision automated deduction through intelligent processing and hybrid modeling. First, the historical hydrological data of the basin is obtained, processed by applying adaptive wavelet transform, and the intermediate frequency wavelet feature sequence and the transformation residual sequence are extracted as data features; a first calculation model is established to generate a first predicted water level sequence; based on the first predicted water level sequence and the actual water level sequence, an error sequence is obtained; a second calculation model is constructed using the data features and rainfall data, and the error sequence is fitted using the second calculation model to obtain a predicted error sequence; the first predicted water level sequence, the predicted error sequence and the rainfall sequence are weightedly fused to generate a second predicted water level sequence; the second predicted water level sequence is combined with a hydrodynamic numerical calculation model, and predicted data is obtained through flow field calculation and analysis.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] A method for automatically predicting basin flood processes based on time series analysis includes:

[0009] Obtain historical hydrological data of the basin, and select part of the historical hydrological data of the basin as a training set;

[0010] Adaptive wavelet transform is applied to the training set to automatically determine the transform parameters and perform denoising operations to obtain a denoised training set, and the intermediate frequency wavelet feature sequence and transform residual sequence are extracted as data features;

[0011] Based on the denoising training set, a first calculation model is established to generate a first predicted water level sequence; based on the first predicted water level sequence and the actual water level sequence, an error sequence is obtained;

[0012] A second calculation model is constructed using the data features and rainfall data, and the error sequence is fitted using the second calculation model to obtain a predicted error sequence; the first predicted water level sequence, the predicted error sequence and the rainfall data are weightedly fused to generate a second predicted water level sequence; the second predicted water level sequence is used as input, combined with a hydrodynamic numerical calculation model, to obtain flood prediction data through flow field calculation and analysis.

[0013] Preferably, the transformation parameters include a wavelet transformation threshold, a wavelet basis function type, and a wavelet decomposition layer number. The specific steps of automatically determining the transformation parameters are:

[0014] quantifying the data discreteness of the highest frequency wavelet detail coefficient of the training set, automatically generating a signal background noise intensity estimate based on the discreteness, and calculating a specific threshold value of the wavelet coefficient based on the noise intensity estimate and a general threshold calculation formula;

[0015] The wavelet basis function type is selected by matching the skewness and sample entropy values of the training set with predefined criteria; the wavelet decomposition level is determined by analyzing the time length distribution of the data and applying the rule of cumulative energy percentage.

[0016] Preferably, the intermediate frequency wavelet feature sequence acquisition process includes: calculating the local energy value of the wavelet detail coefficient of the selected intermediate decomposition level within the sliding window to form an intermediate frequency wavelet feature sequence that changes with time.

[0017] Preferably, the transformation residual sequence includes: the remaining portion after subtracting all detail components of each layer extracted in the wavelet transformation process from the original signal sequence.

[0018] Preferably, the error sequence acquisition process includes: subtracting the first predicted water level sequence from the actual water level observation value to obtain the error sequence.

[0019] Preferably, the second calculation model includes:

[0020] a rainfall encoder configured to receive a rainfall sequence as input, process the rainfall sequence using a first recurrent neural network layer, and generate a first hidden state sequence;

[0021] an intermediate frequency encoder configured to receive the intermediate frequency wavelet feature sequence as input and process it using a second recurrent neural network layer to generate a second hidden state sequence;

[0022] a residual encoder configured to receive the transformed residual sequence as input and process it using a third recurrent neural network layer to generate a third hidden state sequence;

[0023] a feature fusion unit, connected to the output ends of the three input encoding branches, configured to combine the first, second, and third hidden state sequences at corresponding time steps to generate a combined feature sequence;

[0024] A temporal attention module analyzes the combined feature sequence through a temporal attention mechanism to generate a context feature sequence, where the context feature sequence is the result of weighted aggregation of historical information;

[0025] The prediction output unit is a multi-layer perceptron, whose input end is connected to the output end of the temporal attention module and is configured to calculate and output a prediction error sequence based on the context vector sequence.

[0026] Preferably, the weighted fusion of the first predicted water level sequence, the error sequence predicted value and the rainfall data comprises:

[0027] According to the predicted water level sequence, prediction error sequence and rainfall data of the current time step, a set of dynamic weight coefficients is calculated;

[0028] The dynamic weight coefficient is used to weight the first predicted water level sequence, the prediction error sequence and the rainfall data to generate a second predicted water level sequence.

[0029] Preferably, the prediction data obtained through flow field calculation and analysis includes: driving the second predicted water level sequence as the input condition of the hydrodynamic numerical calculation model, and processing the spatiotemporal change results output by the hydrodynamic numerical calculation model to determine the peak arrival time and peak water level at a predetermined location in the basin, and the spatiotemporal change results include spatial distribution data of water level and water depth.

[0030] Preferably, obtaining prediction data through flow field calculation and analysis also includes:

[0031] Based on the spatial distribution data of the maximum water surface elevation during floods output by the hydrodynamic numerical calculation model and combined with the digital elevation model, the submerged water depth distribution map in the study area is determined.

[0032] Compared with the prior art, the present invention has the following beneficial effects:

[0033] 1. By introducing adaptive wavelet transforms, this invention automatically optimizes and determines key signal processing parameters based on the noise and non-stationary characteristics of actual hydrological time series, extracting multi-scale features that better reflect physical processes. This automated processing based on data characteristics overcomes the limitations of existing technologies, which rely on manual parameter selection and are difficult to adapt to data changes. It provides more reliable and targeted data input for subsequent high-precision quantitative modeling.

[0034] 2. This invention effectively couples a basic time series model capable of capturing linear patterns with an error correction model capable of learning complex nonlinear dynamics, driven by previously extracted multi-scale features. A learnable dynamic weighting mechanism is employed to intelligently fuse the outputs of different models based on real-time contextual information. This coupled and adaptive fusion approach enables the model to more comprehensively and accurately simulate complex water level changes, significantly improving the accuracy of time series modeling and achieving the goal of high-precision quantitative modeling.

[0035] 3. This invention automatically couples and drives high-precision data-driven time-series modeling results with a physics-based hydrodynamic model, creating an end-to-end integrated computational process from data input to final state estimation. This process not only automatically outputs highly accurate single-point water level simulations, but also automatically calculates and outputs flood peak characteristics at multiple key locations, as well as spatialized inundation range and depth maps, achieving comprehensive, high-precision estimation of flood conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 A schematic flow chart of a method for automatically simulating a river basin flood process based on time series analysis provided by an embodiment of the present invention;

[0037] Figure 2 A flowchart of an adaptive wavelet transform process provided by an embodiment of the present invention;

[0038] Figure 3 This is a flow chart of the second calculation model provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0039] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0040] See also Figures 1 to 3 The present invention provides a method for automatically simulating river basin flood processes based on time series analysis. The technical solution is as follows:

[0041] Example 1

[0042] The current flood peak arrival time and peak water level have large errors compared with the actual ones, and the inundation range cannot be predicted. Therefore, there is an urgent need for a new solution that can automatically deduce and predict the flood arrival time, peak water level and inundation range.

[0043] Figure 1 A schematic flow chart of a method for automatically simulating river basin flood processes based on time series analysis is provided in this embodiment.

[0044] like Figure 1 As shown in FIG, a method for automatically predicting basin flood processes based on time series analysis includes:

[0045] S10. Obtain historical hydrological data records of the watershed, and select part of the historical hydrological data of the watershed as a training set.

[0046] First, the historical hydrological data records of Basin A are obtained from the regional hydrological database and meteorological database. The historical hydrological data records include the record timestamp, the corresponding water level data, and the rainfall data that can reflect the precipitation conditions of the basin during the time period. In this embodiment, the hourly time series data of the outlet section of Basin A for the past five years are collected. The data format is a CSV file containing three columns: time, water level, and basin rainfall data. Then, the acquired data is preprocessed to check the data integrity and format, convert the timestamp into a unified time format, delete duplicate entries, use the IQR method to detect outliers, and treat them as missing values. Then, linear interpolation is used to fill in all missing values. Table 1 shows the historical hydrological data.

[0047] Table 1 Historical hydrological data

[0048]

[0049] Furthermore, after completing the above-mentioned data collection and preprocessing, the hourly, complete and outlier-free hydrological and meteorological time series data of Basin A for the past five years are obtained, and the complete time series data are divided into a training set, a validation set and a test set in strict chronological order and according to a predetermined ratio; in this embodiment, the continuous data segments starting from the starting time point of the complete time series data and occupying the first 90% of the total length in the time span are selected and defined as the training set; the continuous data segments immediately following the end time point of the training set and occupying the middle 6% of the total length in the time span are selected and defined as the validation set; the continuous data segments immediately following the end time point of the validation set and occupying the last 4% of the total length in the time span are selected and defined as the test set.

[0050] S20. Apply adaptive wavelet transform to process the training set; automatically determine transformation parameters and perform denoising operations to generate a denoised training set.

[0051] Figure 2 This is a flowchart of the adaptive wavelet transform processing in this embodiment.

[0052] First, a preliminarily selected wavelet basis function is selected. In this embodiment, db4 is used. Multi-layer wavelet decomposition is performed on the training set time series to obtain the highest-frequency detail coefficient sequence. The data dispersion of the detail coefficient sequence is quantified by calculating the absolute median difference. The standard deviation of the Gaussian white noise in the signal is estimated based on the absolute median difference to estimate the noise intensity. The specific calculation formula is as follows:

[0053]

[0054] in, represents the estimated noise intensity, represents the absolute median difference, represents the highest frequency detail coefficient sequence, Represents a preset constant parameter used to Converted to an estimated value of the standard deviation. In this embodiment, based on the assumption that the noise approximately obeys the Gaussian distribution, the parameter The value of is 0.6745.

[0055] Furthermore, a general threshold calculation logic is used to automatically calculate the threshold based on the estimated noise intensity and the length of historical hydrological data. The specific calculation is as follows:

[0056]

[0057] in, represents the estimated noise intensity, Indicates automatic calculation of the threshold, Indicates the length of historical hydrological data after preprocessing, Represents the logarithm with base e.

[0058] Furthermore, a candidate library containing a plurality of commonly used wavelet basis functions with different characteristics is pre-set. The candidate library includes:

[0059] The Daubechies series (e.g., db4, db8) is characterized by tight support and good vanishing moment, and is suitable for analyzing asymmetric signals and signals containing transient characteristics.

[0060] The Symlets series (e.g., sym4, sym8) is an approximately symmetric version of the Daubechies wavelet, which reduces phase distortion while maintaining good characteristics and is suitable for signals that require symmetry.

[0061] The Coiflets series (e.g., coif2, coif5) has scaling functions and wavelet functions with good vanishing moments and is approximately symmetrical, making it suitable for processing relatively smooth signals.

[0062] Furthermore, the skewness and sample entropy values are calculated for the preprocessed water level time series training set. The skewness of the numerical distribution of the time series is calculated to quantitatively assess its symmetry. A skewness close to zero indicates a more symmetrical data distribution, while a skewness far from zero indicates a left- or right-skewed distribution. The sample entropy value of the time series is calculated. The embedding dimension m and the tolerance threshold r are set when calculating the sample entropy. The tolerance threshold is set based on the standard deviation of the sequence and is used to quantitatively assess its complexity or irregularity. Smaller sample entropy values indicate more regular and self-similar time series patterns, while larger values indicate more complex, random, and irregular sequences. Then, based on the calculated skewness and sample entropy values.

[0063] The most suitable wavelet basis function is automatically determined from the candidate library based on pre-defined criteria and matching rules. A pre-defined symmetry threshold, used to distinguish between high and low symmetry levels, and a pre-defined complexity threshold, used to distinguish between high and low complexity levels, are pre-set. These thresholds can be determined based on prior knowledge or experimental optimization. If the calculated symmetry index indicates good signal symmetry (below the symmetry threshold) and the complexity index indicates a regular signal pattern (below the complexity threshold), the signal characteristics are considered suitable for wavelet basis functions with good symmetry and smooth morphology. In this case, specific wavelet models from the Symlets or Coiflets families are preferentially selected from the candidate library. If the calculated symmetry index indicates significant signal asymmetry (reaching or exceeding the symmetry threshold) or the complexity index indicates a high level of irregularity (reaching or exceeding the complexity threshold), the signal characteristics are considered more suitable for wavelet basis functions that can effectively capture local details or asymmetric structures. In this case, specific wavelet models from the Daubechies family, particularly those with high vanishing moments, are preferentially selected from the candidate library.

[0064] Furthermore, after the wavelet basis function is determined, the decomposition level used in this wavelet transform is automatically determined. This process uses a rule based on the cumulative energy percentage. First, for the preprocessed water level time series of the training set with a length of N, an initial multi-layer wavelet decomposition of sufficient depth is performed using the previously automatically selected wavelet basis function (for example, the number of decomposition levels can be referred to ), the goal is to obtain coefficients at each layer for energy analysis. Then, for each layer of detail coefficient sequence generated by this initial depth decomposition and the final approximate coefficient sequence, the sum of the squares of all coefficient values within the sequence is calculated to obtain the contained energy. The cumulative energy percentage rule is applied to automatically determine the final number of decomposition layers to be adopted, namely, the sum of the energy of all detail coefficients and the energy of the final approximate coefficients. Then, starting from the lowest decomposition layer, layer 1, the energy of the detail coefficients at each layer is accumulated layer by layer, and the percentage of this cumulative energy to the total signal energy is calculated. The first decomposition layer is found where the cumulative energy percentage reaches or exceeds a preset high percentage threshold. This layer is determined as the final decomposition layer used in this wavelet transform. For example, if the cumulative energy percentage is set to 99.5%, and the cumulative energy percentage at the seventh layer is 99.2%, which does not reach the set energy percentage, decomposition continues. When the cumulative energy percentage at the eighth layer is 99.7%, which exceeds the set energy percentage, the number of decomposition layers is automatically determined to be 8.

[0065] By establishing a computational process based on the data's inherent characteristics (such as noise level, symmetry, complexity, energy distribution, and time scale) and clear rules, the selection of key parameters (denoising threshold, wavelet basis function type, and decomposition level) is objective and adaptable to different input data. This not only avoids the subjective arbitrariness and extensive duplication of effort that can be introduced by traditional manual parameter adjustment, ensuring the repeatability of analysis results and significantly improving processing efficiency, but also enhances the effectiveness of wavelet transform processing by selecting parameters that better match the data's inherent characteristics and hydrophysical processes. This potentially provides higher-quality data input for subsequent model training, helping to improve the simulation accuracy of the entire flood simulation system.

[0066] S30. Extract the intermediate frequency wavelet feature sequence and the transformation residual sequence as data features.

[0067] A complete n-layer discrete wavelet decomposition is performed on the water level sequence of the training set using the automatically determined parameters to obtain the detail coefficient sequence of the n-layer and the approximate coefficient sequence of the n-th layer. The automatically determined denoising threshold is used to perform threshold denoising on the detail coefficient level required for feature extraction to reduce the impact of noise on energy calculation. One or more intermediate decomposition levels that can represent a predefined intermediate frequency range are selected. These selected level detail coefficient sequences are considered to mainly contain information on the dynamic changes of the flood body.

[0068] Furthermore, a sliding window width W (in terms of the number of data points) and a step size S are pre-set. The window width W should be selected to capture the time scale of the local dynamic features of interest. For each selected intermediate frequency detail coefficient sequence, a sliding window with a width of W and a step size of S is used to traverse the intermediate frequency detail coefficient sequence. At each position where the window slides, the sum of the squares of all W coefficients in the window is calculated to obtain the local energy value of the level at that point in time. This process is repeated for each selected intermediate frequency level to generate multiple independent local energy sequences. The energy values of the multiple independent local energy sequences are summed at each identical time point to obtain an intermediate frequency wavelet feature sequence that characterizes the flood dynamic process.

[0069] By calculating the local energy of the signal in the mid-frequency band, the dynamic intensity and change rate of the flood process are effectively captured, which can better highlight the occurrence, development and subsidence stages of flood events than the original water level signal; at the same time, by selecting specific frequency bands and calculating the sum of square energies, it enhances the features and suppresses noise and trend interference, providing more effective and informative input for subsequent machine learning models that rely on this dynamic information for error correction.

[0070] After performing the aforementioned n-layer discrete wavelet decomposition on the training set water level time series, coefficient sequences representing different frequency components are obtained. The n-th layer approximate coefficient sequence is obtained after the maximum number of layers of low-pass filtering and downsampling, and therefore contains information about the lowest-frequency, slowest-changing components in the signal. Next, n zero vectors of the same length as the corresponding level detail coefficient sequence are created. These zero vectors are input into the inverse transform in place of the actual detail coefficients. The approximate coefficient sequence and the zero vectors are subjected to a multi-layer discrete wavelet inverse transform, ensuring that the same wavelet basis functions as those used in the forward transform are used. After performing n-layer inverse wavelet transforms, an output sequence is obtained, the length of which is the same as the preprocessed water level sequence initially input. The sequence obtained through selective reconstruction is the transformation residual sequence of this embodiment. It is a relatively smooth sequence that primarily reflects the long-term trend, seasonal pattern, or slow changes in baseflow in the original hydrological time series. All higher-frequency, faster-changing detail components are filtered out during this process.

[0071] It effectively separates and extracts the base flow or long-term trend component with very slow changes from the original complex signal, providing a smoother input for subsequent models that need to distinguish or process trend information separately. As the coarsest scale component of wavelet multiscale decomposition, it together with the previously extracted intermediate frequency feature sequence constitutes a decomposition representation of the original signal at different time scales, which helps to more comprehensively understand the hydrological process.

[0072] S40. Based on the denoising training set, establish a first calculation model to generate a first predicted water level sequence; based on the first predicted water level sequence and the actual water level sequence, obtain an error sequence.

[0073] First, the stationarity of the series is determined by performing an ADF test. If the test result indicates that the series is nonstationary, it is then subjected to differencing until the differenced series passes the stationarity test. The number of differences performed is the order d. For the differenced stationary series, the autocorrelation function (ACF) and partial autocorrelation function (PACF) plots are calculated and analyzed. Based on the observed truncation and tailing patterns of the ACF and PACF, a preliminary range of candidate values for the autoregressive order p and the moving average order q can be determined. Then, combining the information criterion (AIC) minimization principle, the optimal (p, q) combination is selected within the candidate range through calculation and comparison, ultimately obtaining the optimal order (p, d, q) for the ARIMA model.

[0074] Furthermore, the training set data is used to fit an ARIMA model of the selected order using maximum likelihood estimation (MLE) to establish a first computational model. The training data is then re-entered into the trained ARIMA model for prediction, resulting in a first predicted water level series. The error series is calculated by subtracting the predicted values of the first predicted water level series from the actual water level observations at each time point in the training set.

[0075] The ARIMA model effectively captures and fits the linear autocorrelation patterns in the denoised hydrological series, providing a reliable and repeatable baseline forecast for the entire complex flood simulation process. Furthermore, the ARIMA model's fitting error sequence is accurately calculated. This error sequence clearly separates the nonlinear dynamics and external factors in the original signal that cannot be explained by the linear model, providing a clear and necessary learning objective and training basis for the subsequent error correction model.

[0076] S50. Use the data features and rainfall data to construct a second calculation model, and use the second calculation model to fit the error sequence to obtain a predicted error sequence.

[0077] Figure 3 This is a flow chart of the second calculation model provided in this embodiment.

[0078] The second calculation model includes:

[0079] a rainfall encoder configured to receive a rainfall sequence as input, process the rainfall sequence using a first recurrent neural network layer, and generate a first hidden state sequence;

[0080] an intermediate frequency encoder configured to receive the intermediate frequency wavelet feature sequence as input and process it using a second recurrent neural network layer to generate a second hidden state sequence;

[0081] a residual encoder configured to receive the transformed residual sequence as input and process it using a third recurrent neural network layer to generate a third hidden state sequence;

[0082] a feature fusion unit connected to the output ends of the three input encoding branches, configured to combine the first, second, and third hidden state sequences at corresponding time steps to generate a combined feature sequence;

[0083] A temporal attention module analyzes the combined feature sequence through a temporal attention mechanism to generate a contextual feature sequence, where the contextual feature sequence is the result of weighted aggregation of historical information;

[0084] The prediction output unit is a multi-layer perceptron, whose input end is connected to the output end of the temporal attention module and is configured to calculate and output a prediction error sequence based on the context vector sequence.

[0085] Specifically, the rainfall sequence collected in step S10, the medium-frequency wavelet feature sequence and the transformation residual sequence obtained in step S30 are used as the input of the second computing model, and the prediction error sequence output by the second computing model and the error sequence are used as the loss to regress the second computing model. Each recurrent neural network layer of the second computing model adopts an LSTM structure, which includes 2 layers and 64 units in each layer. The feature fusion unit combines hidden states through splicing operations. The temporal attention module adopts the Bahdanau attention mechanism, and the time window is 24 time steps. The prediction output unit is an MLP with 2 hidden layers, 32 nodes in each layer, using the ReLU activation function and adding Dropout (ratio 0.2) after the hidden layer, and the output layer is linearly activated. The Adam optimizer is used for training, and the initial learning rate is 10 −3 The batch size is 64, the mean square error is used as the loss function, and the early stopping mechanism is adopted in the training process to monitor the loss on the validation set. When the loss on the validation set no longer improves within the preset number of rounds, the training is stopped to prevent the model from overfitting, and the maximum number of training rounds is set to 300 rounds.

[0086] By effectively integrating multiple key inputs, including external driving forces as well as internal dynamic characteristics and background states previously extracted through adaptive wavelet transforms, a more comprehensive and in-depth understanding of the error sources and patterns of the ARIMA model is achieved. This not only fully leverages the multi-scale information decomposition provided by wavelet analysis, but also effectively captures complex hydrological response relationships, enabling highly accurate and reliable flood simulations.

[0087] S60. Perform weighted fusion on the first predicted water level sequence, the prediction error sequence and the rainfall sequence to generate a second predicted water level sequence.

[0088] A multi-layer perceptron (MLP) is trained to calculate a weight coefficient in real time at each time point based on the model's current operating state and relevant input information. The MLP comprises two hidden layers, each with, for example, 32 nodes. The hidden layer activation function is preferably Reinforced Luminance (ReLU). The number of nodes in the input layer depends on the dimensionality of the input information. The output layer comprises three nodes and uses a Softmax activation function to ensure that the three output weights are non-negative and sum to 1. These correspond to the contribution weights of the first predicted water level series, the prediction error series, and the rainfall data. To improve generalization, dropout is added after the hidden layer.

[0089] Furthermore, training data is prepared, including a sequence of predicted water levels, error sequences, predicted values, and rainfall data for the corresponding time period, and these data are then unified and normalized. The first and second computational model output sequences, as well as the rainfall sequence, are then input into a multi-layer perceptron network. Through the network's forward computation, a set of dynamic weight coefficients corresponding to each sample is obtained. The corresponding original first predicted water level sequence, predicted error sequence, and rainfall data are then obtained. Using the corresponding weights output by the MLP, these three values are weightedly combined to calculate the predicted water level value combination for each sample, yielding a second predicted water level sequence. The resulting second predicted water level sequence is then compared with the corresponding actual observed water level sequence, and a loss value, measuring the current prediction error, is calculated using the mean squared error. Based on this calculated loss value, a backpropagation algorithm is applied. Starting from the loss value, along the reverse path of the neural network, the gradient of the loss value relative to all trainable parameters in the multilayer perceptron network is calculated. The gradient indicates how each parameter should be adjusted to reduce the loss function value. Based on the Adam gradient descent optimization algorithm, all trainable parameters of the multilayer perceptron network are calculated and updated. A preset number of training iterations are completed on the MLP to obtain the trained MLP model.

[0090] Furthermore, in the actual prediction stage, for each time point that needs to be predicted, the first predicted water level sequence, prediction error sequence and rainfall data at the time point are obtained, and the data are input into the trained MLP network to obtain a set of dynamic weights for the time point. The dynamic weights generated by the MLP are used to dynamically weight the first predicted water level sequence, prediction error sequence and rainfall data to obtain the second predicted water level sequence.

[0091] By automatically learning and adjusting the contribution of each information source to the final prediction results, a more sophisticated and effective fusion strategy is achieved than fixed weights or simple addition. This data-driven dynamic multi-weight combination method can further improve the accuracy and robustness of the overall flood simulation.

[0092] S70. The second predicted water level sequence is passed as input to the hydrodynamic numerical calculation model, and the predicted data is obtained through flow field calculation and analysis.

[0093] In this example, the river channel in Basin A is relatively regular and the primary concern is flood evolution within the channel. Therefore, a one-dimensional hydrodynamic model (HEC-RAS) was selected for subsequent computational analysis. The HEC-RAS model must be pre-constructed and parameterized based on the actual geographic information of Basin A. This includes inputting channel geometry data (e.g., cross-sectional shape, cross-sectional spacing, and channel centerline coordinates, obtained through field measurements or remote sensing). Cross-sectional measurements can be performed using equipment such as total stations, GPS-RTK, LiDAR point cloud data, and aerial photogrammetry. Flow resistance parameters (e.g., estimated or measured Manning's roughness coefficient n for each river section or section) are then determined by referring to typical empirical roughness values for different riverbed materials, vegetation cover types, and channel morphologies, as provided in relevant hydraulic engineering manuals and literature. Empirical estimates are then made based on on-site surveys and observations of the actual river channel and floodplain conditions. Finally, initial roughness values are assigned to different land cover types within the study area using remote sensing imagery or land use maps.

[0094] Furthermore, possible hydraulic structures such as bridges, culverts, and weirs in the model are defined. In the constructed HEC-RAS model, one or more calculation sections are pre-specified for which calculation results need to be output and flood characteristics analyzed (for example, these could be sections where important downstream cities are located, sections at specific hydrological stations, or other locations where flood risk assessment is required).

[0095] Furthermore, the training process of the hydrodynamic data calculation model includes running the model using selected historical flood event data as the input boundary conditions of the model, comparing the water level process simulated by the model at the internal observation point with the measured process at that point, and using the Nash efficiency coefficient to calculate the error index between the two. According to the error size and the matching between the simulation process and the measured process, the model parameters are systematically adjusted, and the model is repeatedly run and the parameters are adjusted. This process is carried out by combining genetic algorithm and particle swarm optimization until the simulation results and the measured results of the internal observation point reach the preset satisfaction standard.

[0096] To further verify the model's reliability, measured data from one or more independent historical flood events at different stages are used for model validation. While maintaining the established model parameters, the model is run using the data from the validation event as boundary conditions. The model's simulated output at internal observation points is compared with the measured process of the validation event, and an error index is calculated using the Nash efficiency coefficient. If the error index during the validation phase meets the preset accuracy requirements, it indicates that the hydrodynamic model has good predictive capabilities and can be used for subsequent automated flood simulations.

[0097] Furthermore, the second predicted water level sequence generated in the previous step is used as the primary upstream input boundary condition to drive the HEC-RAS model. The water level sequence is converted into a flow process according to the water level-flow comparison table. The time range and calculation time step of the HEC-RAS model's unsteady flow calculation are set, and the model is then started to perform unsteady flow hydrodynamic calculations. After the HEC-RAS model calculations are completed, the model outputs detailed results of the water level, flow, and flow velocity at all predetermined locations over time. The complete water level time series corresponding to the predetermined section is then extracted from the model output result database. The maximum value in the sequence is found and recorded as the predicted flood peak water level for the section. The time point when the maximum value first appears is found and recorded as the predicted flood peak arrival time for the section.

[0098] Compared with directly using the second predicted water level sequence to determine the peak water level and peak arrival time, the use of the second predicted water level sequence and hydrodynamic model can focus on the water level changes at multiple predetermined locations, and go deep from statistical law deduction to physical process simulation, thereby obtaining a more comprehensive, more accurate, and more physically meaningful understanding and prediction results of the flood evolution process in the entire basin.

[0099] Furthermore, the peak water levels reached at each calculated section during the entire flood process, as well as the longitudinal water surface lines formed by the highest water levels of these sections, are obtained. The highest water levels on these sections are interpolated to the entire potentially flooded area to obtain a continuous spatial water surface elevation distribution. Next, digital elevation model (DEM) data covering the study area is obtained. The DEM data must have sufficient resolution and be consistent with the coordinate system of the hydrodynamic model. DEM data can be derived from aerial photogrammetry and publicly available regional DEM datasets. The maximum water surface elevation data of the spatial water surface elevation distribution is overlaid with the DEM data in a grid or spatial manner. For each grid cell of the DEM, the maximum water surface elevation value during the flood period and the ground elevation value at that location are obtained. If the maximum water surface elevation is greater than the ground elevation, the cell is considered flooded. For flooded cells, the flooding depth is obtained. If not flooded, the water depth is recorded as 0.

[0100] Furthermore, all grid cells determined to be flooded are visualized to generate a flood range map reflecting the area with the greatest potential impact of this flood event. The maximum flood depth value calculated for each flooded cell is rendered using different colors or grayscale levels to generate a flood depth distribution map, which intuitively shows the severity of flooding in different areas.

[0101] By outputting a flood depth distribution map that can be used for decision support to quantify the severity of flooding in different areas, compared with the simplified method of static mapping based only on a single point water level, it can more realistically reflect the spatial changes of the actual water surface line, thereby providing more accurate and reliable flooding results.

[0102] The present invention proposes an automated deduction method for basin flood processes based on time series analysis, which uses adaptive wavelet transform to automatically process data and extract key multi-scale dynamic features and trend information, thereby improving adaptability to different data. Then, by combining the ARIMA linear model with the nonlinear error correction model, the limitations of a single model are effectively overcome, and the simulation accuracy of the water level time series is improved. Dynamic weighting further intelligently fuses the model output based on real-time information, thereby enhancing the robustness of the results. Finally, by using precise time series results to drive the hydrodynamic model of the physical process, the prediction is extended from a single point to space, and the flood peak information of multiple key locations and the intuitive inundation range and water depth distribution map are automatically obtained. This method has significantly improved the degree of automation, prediction accuracy, comprehensiveness of results and physical significance, providing more powerful and intelligent technical support for flood risk management and emergency decision-making.

[0103] Example 2

[0104] For Basin B, the demand is to more accurately predict the future flood peak arrival time, flood peak water level and inundation method. Basin B has wide and shallow river channels, wide shoals, complex terrain and changeable water flow directions. Therefore, an automated deduction method for basin flood processes based on time series analysis proposed in this invention is introduced.

[0105] First, hourly water level data and corresponding rainfall data for Basin B over the past ten years were collected and preprocessed. The data was then divided into a training set (the first 95% of the time period), a validation set (the 95% to 98% period), and a test set (the 98% to 100% period). An adaptive wavelet transform was then used to automatically determine the threshold and wavelet basis functions for the training set. The cumulative energy percentage rule (with a threshold of 99.5%) was then applied to automatically determine the final number of decomposition levels. In this example, energy calculation and cumulative percentage analysis were performed on the 10-year training data set. It was determined that the cumulative energy percentage exceeded 99.5% for the first time at the ninth decomposition level. Therefore, in this example, the final number of decomposition levels used in this wavelet transform was automatically determined as nine.

[0106] Furthermore, the intermediate frequency wavelet feature sequence and the transformation residual sequence are extracted as data features. The training set, denoised by wavelet transform, is then input into the ARIMA model for parameter training and fitting. The trained and fitted ARIMA model is then used to predict the training set, yielding an error sequence. Finally, the intermediate frequency wavelet feature sequence, transformation residual sequence, and rainfall sequence are input into a second computational model for training, and the model is regressed using the error sequence as the target output.

[0107] Furthermore, a multi-layer perceptron (MLP) is trained to perform a weighted fusion of the first computational model and the second trained model to generate a second predicted water level series. This simulated time series is then passed to a hydrodynamic numerical computation model. Because Basin B features wide and shallow river channels, wide shoals, complex terrain, and variable flow directions, this embodiment utilizes a two-dimensional hydrodynamic model. Specifically, the two-dimensional hydrodynamic calculation module HEC-RAS2D provided by the HEC-RAS software is used. The two-dimensional model must be pre-constructed based on a digital elevation model (DEM) covering the study area and set parameters such as roughness. During the computation, the final simulated water level time series obtained in the previous step serves as the upstream input boundary condition for the HEC-RAS2D model. After setting other necessary boundary conditions, the model is run with a two-dimensional (or coupled) unsteady flow simulation.

[0108] After the simulation is complete, HEC-RAS2D outputs information on the spatiotemporal distribution of water levels, depths, and flow velocities within the calculation area. By extracting and analyzing the time series of predetermined points of interest, peak flood levels and arrival times at these locations are determined. By overlaying the spatial distribution of the model's maximum water surface elevation with the DEM data, maps of the maximum inundation range and depth are determined.

[0109] To verify the effectiveness of the method of the present invention, an experimental comparison was conducted between the method of the present invention and the original prediction method of Basin B on the test set. The absolute error of the forecast was calculated by subtracting the measured value of each hourly hydrological element from the calculated value. The relative error was expressed as a percentage by dividing the absolute error by the measured value. The degree of agreement between the calculation process and the measured process of the hydrological element can be measured using the coefficient of certainty as an indicator. The specific calculation formula is as follows:

[0110]

[0111] represents the coefficient of certainty, Represents each time point, represents the predicted value at each time point, Indicates the actual test value at each time point, It represents the mean of the measured value in this time period, and n represents the length of the time period.

[0112] Table 2 shows the comparison data of flood peak prediction time and water level. It can be seen from the table that the method of the present invention significantly reduces the error in flood peak prediction time and flood peak water level prediction compared with the original prediction method of basin B.

[0113] Table 2 Comparison of flood peak prediction time and water level

[0114]

[0115] In addition, to further demonstrate the effectiveness of the method of the present invention, the data on the test set were compared using the certainty coefficient. The verification results showed that the certainty coefficient of the method of the present invention was 0.86, and the certainty coefficient of the original method B was 0.75. The prediction results of the method of the present invention were significantly better than those of the original method B, verifying the effectiveness and advancement of the method of the present invention in realizing high-precision automated deduction of flood processes.

[0116] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for automatic deduction of basin flood process based on time series analysis, characterized by: include: Obtain historical hydrological data of the basin, and select part of the historical hydrological data of the basin as a training set; Adaptive wavelet transform is applied to the training set to automatically determine the transform parameters and perform denoising operations to obtain a denoised training set, and the intermediate frequency wavelet feature sequence and transform residual sequence are extracted as data features; Based on the denoising training set, a first calculation model is established to generate a first predicted water level sequence; based on the first predicted water level sequence and the actual water level sequence, an error sequence is obtained; A second computing model is constructed using the data features and rainfall data, and the error sequence is fitted using the second computing model to obtain a predicted error sequence. The second computing model includes: a rainfall encoder, configured to receive a rainfall sequence as input, processed by a first recurrent neural network layer, and generate a first hidden state sequence; an intermediate frequency encoder, configured to receive an intermediate frequency wavelet feature sequence as input, processed by a second recurrent neural network layer, to generate a second hidden state sequence; a residual encoder, configured to receive a transformed residual sequence as input, processed by a third recurrent neural network layer, and generate a third hidden state sequence; a feature fusion unit, connected to the output ends of the three input coding branches, configured to combine the first, second and third hidden state sequences at corresponding time steps to generate a combined feature sequence; a temporal attention module, which analyzes the combined feature sequence through a temporal attention mechanism to generate the above The following feature sequence is described, and the context feature sequence is the result of weighted aggregation of historical information; the prediction output unit, a multi-layer perceptron, whose input end is connected to the output end of the temporal attention module, is configured to calculate and output a prediction error sequence based on the context vector sequence; the first predicted water level sequence, the prediction error sequence and the rainfall data are weightedly fused to generate a second predicted water level sequence; the second predicted water level sequence is used as input, combined with the hydrodynamic numerical calculation model, and flood prediction data is obtained through flow field calculation and analysis. The prediction data obtained through flow field calculation and analysis includes: using the second predicted water level sequence as the input condition of the hydrodynamic numerical calculation model to drive it, and processing the spatiotemporal change results output by the hydrodynamic numerical calculation model to determine the flood peak arrival time and flood peak water level at a predetermined location in the basin. The spatiotemporal change results include spatial distribution data of water level and water depth.

2. The method for automatically simulating a river basin flood process based on time series analysis according to claim 1 is characterized in that: The transformation parameters include wavelet transformation threshold, wavelet basis function type and wavelet decomposition layer number. The specific steps for automatically determining the transformation parameters are as follows: quantifying the data discreteness of the highest frequency wavelet detail coefficient of the training set, automatically generating a signal background noise intensity estimate based on the discreteness, and calculating a specific threshold value of the wavelet coefficient based on the noise intensity estimate and a general threshold calculation formula; The wavelet basis function type is selected by matching the skewness and sample entropy values of the training set with predefined criteria; the wavelet decomposition level is determined by analyzing the time length distribution of the data and applying the rule of cumulative energy percentage.

3. The method for automatically simulating a river basin flood process based on time series analysis according to claim 2 is characterized in that: The intermediate frequency wavelet feature sequence acquisition process includes: calculating the local energy value of the wavelet detail coefficient of the selected intermediate decomposition level within the sliding window to form an intermediate frequency wavelet feature sequence that changes with time.

4. The method for automatically simulating a river basin flood process based on time series analysis according to claim 1 is characterized in that: The transformation residual sequence includes: the remaining portion after all detail components of each layer extracted in the wavelet transformation process are subtracted from the original signal sequence.

5. The method for automatic deduction of basin flood process based on time series analysis according to claim 1 is characterized in that: The error sequence acquisition process includes: subtracting the first predicted water level sequence from the actual water level observation value to obtain the error sequence.

6. The method for automatic deduction of basin flood process based on time series analysis according to claim 1 is characterized in that: The weighted fusion of the first predicted water level sequence, the error sequence predicted value and the rainfall data comprises: According to the predicted water level sequence, prediction error sequence and rainfall data of the current time step, a set of dynamic weight coefficients is calculated; The dynamic weight coefficient is used to weight the first predicted water level sequence, the prediction error sequence and the rainfall data to generate a second predicted water level sequence.

7. The method for automatic deduction of basin flood process based on time series analysis according to claim 1 is characterized in that: The prediction data obtained by flow field calculation and analysis also includes: Based on the spatial distribution data of the maximum water surface elevation during floods output by the hydrodynamic numerical calculation model and combined with the digital elevation model, the submerged water depth distribution map in the study area is determined.

Images