A method and system for simulating river hydrological response by integrating vector autoregression and deep learning
By integrating vector autoregression and deep learning, the problem of difficulty in determining the propagation time lag and historical range of hydrological driving factors in river level simulation was solved, enabling rapid and accurate simulation of river level response and rapid evaluation of scheduling schemes.
Patent Information
- Application Number
- CN202611069349.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-17
- Publication Date
- 2026-11-17
AI Technical Summary
Existing methods for simulating river levels are ill-suited to changing river conditions. Model building and repeated calculations are time-consuming, making it difficult to quickly determine the impact of hydrological driving changes on river levels. Furthermore, the propagation time lag and effective historical range of different hydrological driving factors are difficult to determine, resulting in a lack of continuity and consistency in the water level response process.
A method combining vector autoregression and deep learning was adopted. The historical range of the input variables was determined by vector autoregression analysis, and the time-domain features of the hydrological process were extracted by the deep learning model to form a unified model input. The water level response process under different hydrological driving scenarios was simulated under unified boundary conditions.
It improves the accuracy and continuity of river water level response simulation, can quickly generate comparable water level response processes, provides a basis for reservoir scheduling, and reduces the impact of changes in data processing standards.
Smart Images

Figure CN123116901A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of hydrological simulation and water resource allocation, and in particular to a method and system for simulating river hydrological response by integrating vector autoregression and deep learning. Background Technology
[0002] With the expansion of reservoir operation and the increasing demands for watershed water resource management, rapidly determining the impact of hydrological changes on river levels has become a crucial foundation for flood control scheduling, water resource allocation, and river operation management. River levels are influenced by a combination of factors, including upstream inflow, reservoir discharge, tributary inflow, and existing hydrological conditions. The propagation time lag, response intensity, and duration of these factors vary, resulting in a distinct temporal and nonlinear response to river levels.
[0003] Existing river level simulations typically employ water level-discharge relationships, hydrological models, or hydrodynamic models. Water level-discharge relationships are difficult to adapt to changes in river conditions caused by riverbed scouring and deposition, cross-sectional changes, and the operation of hydraulic structures. Hydrological and hydrodynamic models usually require relatively complete river data and boundary conditions, and parameter calibration is necessary. In complex river networks or multi-scheme scheduling scenarios, model construction and repeated calculations take a long time, making it difficult to meet the requirements for rapid assessment.
[0004] Invention patent CN112001556A discloses a method for predicting downstream water levels in reservoirs based on a deep learning model. This method uses the maximum information coefficient to screen downstream water level-related factors, determines feature combinations through a genetic algorithm, and employs convolutional neural networks and long short-term memory networks to predict downstream water levels. However, this method primarily selects input information based on statistical correlation and model prediction error, and it does not clearly define the direction of influence, propagation lag, and effective historical range of different hydrological driving factors on the target section water level. Furthermore, it lacks targeted processing of the length and temporal location of historical information used for different hydrological variables, potentially introducing historical information with weak responses or omitting hydrological information with long-lasting effects. This method is mainly used to generate downstream water level prediction results, and there is still room for improvement in the continuous comparison of the river water level response process before and after hydrological driving adjustments.
[0005] Invention patent CN120068664A also discloses a deep learning method and device for assessing water level and flow changes in the middle and lower reaches of a watershed. This method combines hydrological data, flow reconstruction results, and flood evolution results to construct a surrogate model between reservoir regulation and water level and flow changes in the middle and lower reaches. The above method can improve the assessment efficiency of reservoir regulation scenarios, but the model input and training data are affected by the preceding flow reconstruction and flood evolution results. For the response time lag and effective historical range of various measured hydrological driving factors, a unified identification and input construction method has not yet been formed. This method focuses on assessing water level and flow changes under the influence of reservoir regulation. Further improvement is still needed in the hourly transition of the target section water level status in continuous scenario simulations and the boundary consistency control between different hydrological driving scenarios.
[0006] Existing technologies typically employ machine learning or deep learning models to directly establish the mapping relationship between historical hydrological data and target cross-sectional water levels. These methods usually use a uniform historical window for model input, making it difficult to account for the differences in propagation time lags of various hydrological driving factors. Furthermore, inconsistent data processing conditions during training, testing, and scenario simulation can lead to variations in the model input. Simultaneously, a single accuracy metric cannot comprehensively reflect the degree of fit, peak time, and sustained response characteristics of the water level process, and the generated water level results are difficult to directly support the comparison and selection of different hydrologically driven control schemes.
[0007] Therefore, determining the corresponding historical range of influence based on the temporal response characteristics of multiple hydrological driving factors to the target cross-section water level, forming a model input with consistent temporal meaning, taking into account both the local variation characteristics and continuous state evolution of river water level, and continuously simulating the water level response process corresponding to different hydrological driving scenarios under unified boundary conditions are technical problems that need to be solved in river hydrological response simulation. Summary of the Invention The main objective of this invention is to provide a method and system for simulating river hydrological response by integrating vector autoregression and deep learning. This addresses the problems in existing river hydrological response simulations, such as the difficulty in determining the direction of action, propagation time delay, and effective historical range of different hydrological driving factors; the difficulty of using a unified historical window to adapt to the temporal differences of different hydrological variables; and the lack of continuous connection and consistent boundary control of the target section water level under different hydrological driving scenarios.
[0008] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a method for simulating river hydrological response by integrating vector autoregression and deep learning, the method comprising the following steps: S1. Acquire external hydrological driving data and historical water level data of the target section, form an isochronous hydrological sequence and divide it into training and testing periods; stabilize the training sequence, perform vector autoregression analysis based on the stationary representation, and obtain the evaluation results of the candidate order information criterion and the directional impulse response of the input variables to the target section water level; S2. Based on the preset response stability conditions and information criterion tolerance, determine the order of the input variables by combining the evaluation results of the information criterion with the directional impulse response; extract historical windows of different lengths from the hydrological sequence of equal time axis, right-align them according to the target time and mark the effective positions, set the response weights according to the amplitude of the directional impulse response, determine the normalization parameters according to the hydrological data of the training period, form training and test samples containing model input and measured water level, and record the input processing configuration; S3. A deep learning model that integrates temporal feature extraction and hydrological state recursion is trained using training samples to generate simulated water levels at the target cross-section. S4. Establish a benchmark model using the same training and test samples and evaluate the benchmark model and the fusion deep learning model. Calculate the water level evaluation index based on the simulated water level and the measured water level. Determine the fusion deep learning model as the scene response model according to the preset model selection rules. Encapsulate the scene response model and input processing configuration into a river hydrological response model package. S5. Set the baseline hydrological driving data and adjust the hydrological driving data to maintain the consistency of the hydrological boundaries except for the input variables to be adjusted. Initialize the historical water level window with the measured water level before the scene. Form the scene input according to the input processing configuration in the model package. Call the scene response model to generate simulated water levels at each time step and backfill the scene historical water level window to form the river section water level response process. Determine the candidate hydrological driving control scheme according to the water level response process and the preset scheduling constraints.
[0009] In the preferred embodiment, step S1 includes: reading hydrological observation records according to the sampling time, verifying the monitoring location, physical quantity name and unit of measurement corresponding to the hydrological observation records; unifying the verified hydrological observation records to the same time reference, and forming an isochronous hydrological sequence according to the preset sampling interval; The training period and the testing period are divided according to the time sequence. During the training period, the augmented Dickey-Fuller test is performed on each hydrological variable. For hydrological variables that have not reached the stationarity condition, difference processing is performed until the corresponding stationary representation is obtained. The difference mapping corresponding to each hydrological variable is recorded, and the hydrological sequence with the original physical meaning is retained for subsequent sample construction. Based on the stationary representation of each hydrological variable, vector autoregressive models of different candidate orders are established. The Akaike information content criterion evaluation results of each candidate order are obtained by using a consistent effective evaluation interval and information criterion calculation method. Effective candidate models that meet the dynamic stability conditions are selected from the candidate vector autoregressive models. The impulse response reference model is determined from the effective candidate models, and the residual covariance of the impulse response reference model is orthogonally decomposed. Independent impulses of a preset scale are applied to each input variable according to a pre-fixed variable order, and the target section water level is set as the response variable. The directional impulse response, response peak value and response decay process of each input variable pointing to the target section water level are obtained.
[0010] In the preferred embodiment, determining the order of the input variables in step S2 includes: determining the relative response amplitude and adjacent response changes corresponding to each response order based on the peak value of the directional impulse response corresponding to each input variable; When the relative response amplitudes and adjacent response changes of multiple consecutive response steps satisfy the preset response amplitude condition and preset response change condition, the corresponding response step is determined as a response stability candidate step; when no response stability candidate step is formed within the preset investigation range, the response stability candidate step is determined based on the tail response state of multiple consecutive response steps. Based on the evaluation results of the Akaike information content criterion for each candidate order, candidate orders within the preset information loss tolerance range of the optimal information criterion value are determined; for each input variable, the corresponding input order is determined from the candidate orders that simultaneously satisfy the response stability condition and the information loss tolerance condition, and the input order is associated with the physical properties and sampling time range of the corresponding input variable.
[0011] In the preferred embodiment, further, step S2, forming training and testing samples includes: extracting historical windows before the target time from the isochronous hydrological sequence with original physical meaning according to the input order corresponding to each input variable; aligning historical windows of different lengths with the same target time as the right boundary, and setting a uniform input length according to the longest historical window; Valid location identifiers are used to distinguish the actual hydrological data locations and supplementary locations in each historical window, so that different input variables have consistent relative time meanings at the same input location; response weights are determined according to the distribution of directional impulse response amplitudes corresponding to each input variable at the valid historical location, so that each response weight is consistent with the corresponding input variable and historical location, and the supplementary location does not form a response weighting feature. The training period is divided into a model fitting period and a validation period according to the chronological order. The normalization parameters are determined based on the hydrological data corresponding to the effective locations within the model fitting period. The same normalization parameters are used to process the hydrological data during the model fitting period, validation period, testing period, and scenario simulation phase. The effective location identifier, response weight, and normalized hydrological data are used together to form the model input. The measured water level corresponding to the target time is set as the supervision target. Model fitting samples, validation samples, and test samples are formed respectively, and the corresponding input processing configurations are recorded.
[0012] In the preferred embodiment, step S3 includes: constructing a TCN-GRU model, performing dilated causal convolution processing on the model input formed in step S2 using a temporal convolutional neural network, passing hydrological time series features at different time scales through residual connections, and making the receptive field of the temporal convolutional neural network cover the longest historical window determined in step S2. The hydrological time-series features output by the temporal convolutional neural network are input into the gated recurrent unit in chronological order, and the hydrological state of the target river section is updated using the reset gate and update gate; the hydrological state corresponding to the last valid time position is input into the water level output layer to obtain the simulated water level of the target section; The training period is divided into model fitting data and validation data in chronological order. The TCN-GRU model parameters are updated according to the error between the model output water level and the corresponding measured water level. The model structure is determined, the model parameters are saved, and the model training is terminated according to the validation loss corresponding to the validation data. After the model training is completed, the temporal convolutional neural network structure parameters, gated recurrent unit structure parameters, water level output parameters, and trained model weights are locked for use with test data and scenario data.
[0013] In the preferred scheme, step S4 includes: dividing the training samples into model fitting samples and validation samples according to the time sequence, and using the same model fitting samples, validation samples, test samples, input variables, target time and normalization parameters to establish a single GRU model, a single TCN model and a TCN-GRU model respectively. Each model is trained using model fitting samples, and the model structure and training parameters of each model are selected using validation samples. After the model parameters are fixed, the same test samples are used to generate the target cross-section simulated water level for each model. Based on the simulated water level at the target cross section of each model and the measured water level at the same time, the Nash efficiency coefficient, root mean square error, mean absolute error, peak time error, and Kling-Gupta efficiency coefficient are determined respectively. Based on the good fit of the Nash efficiency coefficient and the Kling-Gupta efficiency coefficient, and the water level simulation error represented by the root mean square error, mean absolute error and peak time error, the TCN-GRU model is determined to be the scenario response model. The input order, effective location identifier, response weight, normalization parameter, network structure parameter and trained model weight corresponding to each input variable are associated to form a river hydrological response model package, so that the same input processing configuration and model parameters can be called in the testing phase and the scenario simulation phase.
[0014] In the preferred scheme, step S5 includes: establishing a baseline hydrological driving scenario, an upward-adjusted hydrological driving scenario, and a downward-adjusted hydrological driving scenario, so that each scenario adopts the same scenario start time, simulation end conditions, non-adjusted hydrological boundary, and initial water level of the target section; Each scenario is set up with an independent historical water level cache, and the historical water level window is initialized with the measured water level before the scenario starts; the scenario input is formed based on the hydrological driving data, non-adjusted hydrological boundary data and historical water level cache of the corresponding scenario. The scene response model is called in the order of sampling time to generate the simulated water level at the current target time. The simulated water level at the current target time is written into the historical water level cache of the corresponding scene, and the updated historical water level cache is used as the scene input for the next target time, until the scene simulation ends. Compare the water level response processes of the adjusted hydrological-driven scenario and the baseline hydrological-driven scenario to determine the maximum water level change, response start time, peak response time, and response duration. The recovery condition is defined as the water level response returning to within the preset water level response threshold and continuously maintaining it for the preset sampling period. If the recovery condition is not yet met, the scenario simulation time is extended. Each hydrological driving scenario is evaluated based on the target control water level, allowable water level change, and allowable response duration. Hydrological driving scenarios that meet the scheduling constraints are identified as candidate hydrological driving control schemes, and scheduling auxiliary signals are generated. When the water level response exceeds any scheduling constraint, an over-limit warning signal is generated. The system receives confirmation information for candidate hydrological-driven control schemes, generates hydrological-driven control instructions based on the confirmation information, and sends the hydrological-driven control instructions to the reservoir gate control system to control the reservoir gate actuators to adjust the gate opening.
[0015] In a preferred embodiment, the present invention provides a river hydrological response simulation system integrating vector autoregression and deep learning, comprising a data analysis module, an input construction module, a model computation module, a model verification module, and a scenario simulation module, wherein: The data analysis module is used to acquire external hydrological driving data and historical water level data of the target section, form a hydrological sequence with an equal time axis and divide it into training and testing periods; the training sequence is stationary, and vector autoregression analysis is performed based on the stationary representation to obtain the evaluation results of the candidate order information criterion and the directional impulse response of the input variables to the water level of the target section; The input construction module is used to determine the order of input variables based on preset response stability conditions, information criterion tolerance, information criterion evaluation results, and directional impulse response; it extracts historical windows of different lengths from the hydrological sequence of equal time axis, aligns them according to the target time, and marks the effective positions; it sets response weights based on the amplitude of directional impulse response, determines normalization parameters based on hydrological data during the training period, and forms training samples, test samples, and input processing configurations. The model computation module is used to train a deep learning model that integrates temporal feature extraction and hydrological state recursion using training samples, and generates simulated water levels at target cross sections from the integrated deep learning model. The model verification module is used to establish a benchmark model using the same training and test samples and to evaluate the benchmark model and the fused deep learning model. It determines the water level evaluation index based on the simulated water level and the measured water level, determines the scene response model according to the preset model selection rules, and encapsulates the scene response model and input processing configuration into a river hydrological response model package. The scenario simulation module is used to set the baseline hydrological driving data and adjust the hydrological driving data, keeping the hydrological boundaries consistent except for the input variables to be adjusted. It initializes the historical water level window with the measured water level before the scenario starts. It forms the scenario input according to the input processing configuration in the river hydrological response model package, calls the scenario response model to generate simulated water levels time by time and backfills the historical water level window of the corresponding scenario, forming the river section water level response process. It determines the candidate hydrological driving control schemes based on the river section water level response process and scheduling constraints, generates scheduling auxiliary signals or over-limit warning signals, and generates hydrological driving control instructions after receiving the scheme confirmation information.
[0016] In a preferred embodiment, the present invention provides an electronic device, including a processor, a memory, a communication interface, and a communication bus, wherein the processor, memory, and communication interface are connected via the communication bus; the memory is used to store computer programs and model parameters and hydrological data required for river hydrological response simulation; the processor is used to read and execute the computer programs to implement the river hydrological response simulation method provided in the first aspect of the present invention; the communication interface is used to receive hydrological monitoring data and scheduling constraint information, and output river hydrological response results, scheduling auxiliary signals, over-limit warning signals, or hydrological drive control instructions.
[0017] In a preferred embodiment, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the river hydrological response simulation method provided in the first aspect of the present invention.
[0018] This invention provides a method and system for simulating river hydrological response by integrating vector autoregression and deep learning. It conducts vector autoregression analysis based on the stationary representation of hydrological sequences on equal time axes, and determines the historical range of each input variable by combining the evaluation results of information criteria and directional impulse response. This allows the direction of action, propagation lag, and response decay process, which are difficult to reflect by statistical correlation alone, to participate in the model input construction, reducing the introduction of invalid information and the omission of continuous response information caused by uniformly selecting historical information. Secondly, historical windows of different lengths are extracted according to the historical range of each input variable. By aligning the target time, marking the effective position, and setting the response weight, the hydrological time meaning of each input position is kept clear, and the real hydrological data is distinguished from the supplemented position. The time-domain variation characteristics of the hydrological process are extracted by using a fusion deep learning model and the hydrological state of the target river section is recursively deduced, thereby improving the ability to represent nonlinear water level changes and continuous state evolution under the combined action of multiple hydrological drivers. Furthermore, normalization parameters are determined based on the training period, and the input order, effective position, response weight, normalization parameters, and trained model parameters are associated and encapsulated. This allows the same input processing configuration and model version to be used in the training, testing, and scenario simulation stages, reducing the impact of changes in data processing methods on simulation results. A benchmark model is established using the same sample conditions, and multiple water level evaluation indicators are combined to verify the model, ensuring that the selection of the scenario response model can take into account the overall fit, average simulation error, and peak time deviation. Furthermore, to maintain consistency of non-adjusted hydrological boundaries in different hydrological driving scenarios, the historical water level window is initialized with the measured water level before the scenario begins, and the simulated water level generated at the current moment is backfilled into the input at subsequent moments, so that the water level status of the target section is inherited in the continuous simulation process. This forms a comparable river section water level response process, and candidate hydrological driving control schemes are determined by combining the water level change amplitude, response time, duration, recovery status and scheduling constraints, providing a basis for rapid judgment of scheduling schemes, over-limit early warning and control execution. Attached Figure Description
[0019] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a flowchart of a method for simulating river hydrological response that integrates vector autoregression and deep learning according to the present invention. Figure 2 This is a schematic diagram of a river hydrological response simulation system that integrates vector autoregression and deep learning according to the present invention. Figure 3 This is a schematic diagram showing the positional relationship between the water conservancy hub (reservoir) a and the hydrological station b in this invention; Figure 4 This is a graph showing the unit impulse response results of water conservancy hub a and hydrological station b analyzed by the vector autoregression method of this invention; Figure 5 This is a diagram showing the water level response results of hydrological station b simulated by the TCN-GRU model used in the method of this invention and the comparison model. Figure 6 This is a graph showing the water level response of hydrological station b in the scenario where the flow rate of water conservancy hub a is increased or decreased by 10,000 m³ / s at 20:00 on August 18, 2020, and continues for 3 days. Figure 7 This is a structural diagram of the temporal convolutional neural network-gated recurrent unit (TCN-GRU) model constructed for this invention. Detailed Implementation
[0020] Example 1 The following is in conjunction with the appendix Figure 1 To be continued Figure 6 This paper further explains a method for simulating river hydrological response by integrating vector autoregression and deep learning. This implementation takes reservoir a, downstream target hydrological station b, and tributary hydrological stations c and d that affect the water level of target station b as objects. It analyzes the directional lag response of each input variable to the target water level through historical hydrological sequence analysis. Based on the directional impulse response and the Akaike information content criterion, the historical input range of each variable is determined. Historical sequences of different lengths are constructed into a unified deep learning input tensor. The TCN-GRU model is used to establish the response mapping between the upstream discharge process, the tributary inflow process, the historical state of the target station, and the subsequent water level of the target station. After completing model training and testing, the discharge flow process of reservoir a is changed, recursively generating the corresponding water level response process of hydrological station b.
[0021] In this embodiment, the river hydrological response is characterized by the water level change process at hydrological station b; the outflow from reservoir a is denoted as... The tributary discharges at hydrological stations c and d are respectively denoted as... and The water level at hydrological station b is recorded as ;in, , and The unit is , The unit is The outflow from reservoir a and the tributary flows from hydrological stations c and d constitute external hydrological driving variables; the historical water level of hydrological station b constitutes the autoregressive input variable of the target station; and the water level of hydrological station b at the time to be simulated constitutes the model output variable. The training and scenario simulation phases use the same data structure. Each sample in the training phase takes the reservoir discharge flow, tributary flow, and historical water level of the target station before the target time as input and the water level of hydrological station b at the target time as output. The scenario simulation phase uses the same variable order, historical window length, mask position, response weight, and normalization parameters. The historical window of the target station before the scenario starts is initialized by the measured water level, and the historical window of the target station after the scenario starts is updated hourly by the water level already generated in the corresponding scenario. This processing enables the time-series mapping established during model training to be continuously used for multi-step water level response generation after changes in the reservoir discharge process.
[0022] The method includes the following steps: S1. Acquire external hydrological driving data and historical water level data of the target section, form an isochronous hydrological sequence and divide it into training and testing periods; stabilize the training sequence, perform vector autoregression analysis based on the stationary representation, and obtain the evaluation results of the candidate order information criterion and the directional impulse response of the input variables to the target section water level; S2. Based on the preset response stability conditions and information criterion tolerance, the order of the input variables is determined by combining the evaluation results of the information criterion with the directional impulse response; different lengths of historical windows are extracted from the hydrological sequence of equal time axis, right-aligned according to the target time and marked with effective positions, the response weight is set according to the amplitude of the directional impulse response, the normalization parameters are determined according to the hydrological data of the training period, and training and test samples containing model input and measured water level are formed, and the input processing configuration is recorded. S3. A deep learning model that integrates temporal feature extraction and hydrological state recursion is trained using training samples to generate simulated water levels at the target cross-section. S4. Establish a benchmark model using the same training and test samples and evaluate the benchmark model and the fusion deep learning model. Calculate the water level evaluation index based on the simulated water level and the measured water level. Determine the fusion deep learning model as the scene response model according to the preset model selection rules. Encapsulate the scene response model and input processing configuration into a river hydrological response model package. S5. Set the baseline hydrological driving data and adjust the hydrological driving data to keep the hydrological boundaries consistent except for the input variables to be adjusted. Initialize the historical water level window with the measured water level before the scene. Form the scene input according to the input processing configuration in the model package. Call the scene response model to generate simulated water levels at each time step and backfill the scene historical water level window to form the river section water level response process. Determine the candidate hydrological control schemes based on the water level response process and preset scheduling constraints.
[0023] In the preferred embodiment, step S1 includes the following steps: S11. Establishing isochronous hydrological sequences The hydrological data access terminal reads the historical outflow of reservoir a from the reservoir scheduling and operation database. Historical flow data for hydrological stations C and D are read from the hydrological station data management equipment. , and the historical water level of hydrological station b The historical outflow of reservoir a is collected by the flow monitoring device installed at the outflow section of the reservoir and written into the reservoir scheduling and operation database; the historical flow of hydrological stations c and d is collected by the flow monitoring equipment of the corresponding hydrological stations; and the historical water level of hydrological station b is collected by the water level monitoring equipment installed at the target section. Each hydrological observation record is associated with the sampling time, monitoring location, physical quantity name and unit of measurement. The hydrological data access terminal reads the above hydrological observation records according to the sampling time, and after verifying the monitoring location, physical quantity name, and unit of measurement, establishes a common time index according to the unified sampling time: (1); (2); in, This represents the common time index corresponding to all hydrological observation records. Indicates the first Each sampling time, This indicates the total number of sampling times included in the common time index. This indicates a fixed sampling interval; the discharge flow of reservoir a at the same sampling time. Hydrological station C flow Hydrological station flow rate and the water level at hydrological station B Together they represent the river hydrological state at the corresponding sampling time; This embodiment uses a 6-hour sampling interval, that is: (3); Other equal-time sampling intervals can also be set based on the hydrological data recording frequency and the response duration of the target river section; the sampling interval used should ensure that the upstream discharge changes and the main response processes formed by the upstream discharge changes at the downstream target station are continuously recorded; when using other sampling intervals, the subsequent stationarity test, vector autoregression analysis, input order determination and model training should be re-executed according to the corresponding sampling interval; Construct time intervals according to the order of reservoir discharge variables, tributary boundary variables, and target water level variables. Hydrological state vector: (4); On the common timeline, each set of river hydrological states is bound to a unique sampling time. When there are differences in time zone, starting time, or recording frequency among the recording times of reservoir a, hydrological station c, hydrological station d, or hydrological station b, the hydrological data access terminal first unifies all hydrological observation records to the same time reference, and then forms an equally spaced hydrological sequence according to the preset sampling interval. The hydrological observation records entering this method are valid records that have undergone hydrological business quality control, so that each variable in the same hydrological state vector represents the hydrological boundary and target station state of the same time period. The discharge flow from reservoir a reflects the upstream control boundary of the main river channel; the flow rates at hydrological stations c and d reflect the tributary boundaries upstream of the target station or in the relevant confluence area; and the historical water level at hydrological station b reflects the comprehensive storage and discharge status of the target river section before the time to be simulated. By incorporating the above hydrological variables into the hydrological state vector, the water level generation process at the target station can simultaneously express the effects of reservoir discharge, tributary inflow, and the continuity of the target river section's state. Appendix Figure 4 Based on the analysis results of reservoir a and hydrological station b in Table 1, we further construct the a-b hydrological state sub-vector: (5); Hydrological station C and hydrological station D are in accordance with the appendix Figure 3 The positional relationship shown is used as the boundary of an independent tributary; when only the discharge flow of reservoir a is adjusted later, the flow of hydrological stations c and d follows the corresponding historical baseline process; when a tributary flow scenario is given separately, the given tributary flow process is written into the corresponding hydrological variable position. To maintain a one-to-one correspondence between hydrological variables, input order, impulse response, and model input channels, a set of variable metadata is established: (6); in, Record Variable names, units of measurement, monitoring locations, and reservoir discharge driving attributes; and Record separately and Variable names, units of measurement, tributary locations, and tributary boundary attributes; Record The variable names, units of measurement, target station locations, target water level response attributes, and historical autoregressive attributes are specified; subsequent VAR coefficients, impulse responses, input order, effective historical locations, and response weights are all based on the variable metadata set. The determined hydrological variables are arranged in order; The original training period and the original testing period are divided according to their chronological order. In this embodiment, historical data from January 1, 2010 to December 30, 2020 is used as the original training data, and historical data from January 1, 2021 to August 15, 2022 is used as the original testing data. The time index corresponding to the original training data is denoted as... The time index corresponding to the original test data is denoted as Stationarity testing, difference order determination, VAR parameter estimation, AIC calculation, and impulse response analysis are all performed in [the relevant area]. Complete within the scope, For use in subsequent independent model testing; The boundary between the training and testing periods is fixed before statistical analysis is performed; the difference mapping, candidate order, impulse response, stability threshold, and AIC tolerance determined by the training period remain fixed during the testing phase; the time boundaries of the training and testing periods are used to determine subsequent normalization parameters and model structure parameters, so that the testing period only serves the function of model performance verification; for other hydrological data, corresponding training and testing boundaries can be set according to the data age and hydrological process coverage, and the sampling time of the training samples should be earlier than that of the testing samples. S12, Hydrological sequence stabilization processing For the state vector The first in Hydrological variables The augmented Dickey-Fuller test was performed during the original training period; the ADF regression form is as follows: (7); in, , Represents the intercept term. Represents the coefficient of the time trend term. This represents the unit root test coefficient. This represents the number of difference lag terms used to absorb residual autocorrelation. Indicates the first The coefficients of the lag terms. This represents the ADF regression residuals; The hypothesis to be tested is: (8); (9); The corresponding ADF statistic Value and preset significance level Comparison; This embodiment can make When the tested sequence reaches the stationarity condition, the current sequence representation is retained; when the tested sequence has not yet reached the stationarity condition, the sequence is differencing and the ADF test is continued until a stationary representation is formed. Number of differenced lag terms in ADF regression The significance level can be determined based on the residual autocorrelation test or the information criterion during the training period; the same significance level is applied to each variable, and the difference times are recorded separately, so that the reservoir discharge, tributary discharge and target station water level can form a stationary representation according to their respective statistical properties; the stationarity test results are used for VAR coefficient estimation and impulse response analysis, and the original physical quantity sequence is retained for subsequent deep learning sample construction and actual water level output; The first The number of differences used to achieve stationarity for each variable is denoted as . Its stationary representation is: (10); When the original sequence satisfies the stationarity condition and When the first-order difference sequence satisfies the stationarity condition, and Each variable is assigned a different difference order, thus establishing a set of difference mappings: (11); Difference Map Set Used to record the correspondence between the VAR stationary representation and the original physical quantity sequence; VAR analysis is performed on the stationary sequence, and subsequent TCN-GRU inputs are still obtained from the original flow and water level sequences at the same lag position, so that the model input and output maintain the original physical units; Calculate the maximum value of the degree of difference for each variable: (12); and remove the previous one from the original training index. At a point where not all difference variables can be formed simultaneously, a common effective training index is obtained. According to the variable metadata set Construct the complete stationary state vector: (13); (14); And the a-b stationary state subvectors: (15); By aligning with a common effective training index, each stationary variable corresponds to the same sampling time in the same state vector; The starting time of the publicly valid training index is constrained by both the maximum difference order and the subsequent candidate VAR order; estimation When using a VAR model of order 1, only select models with... The target time is a complete historical stationary state vector; therefore, each candidate order uses effective observations that are time-continuous and have complete variables; when comparing different candidate orders using AIC, the same effective evaluation interval is preferred, or a consistent sample size processing rule is used in AIC calculation to maintain a consistent comparison caliber between candidate models. S13, Candidate VAR Model and AIC Calculation Establish candidate VAR models within the public effective training period; denote the set of candidate orders as: (16); in, This indicates the maximum candidate order; Table 1 lists candidate orders a-b from 0 to 14, therefore: (17); For each candidate order Estimate the corresponding VAR model: (18) in, express The intercept vector of the VAR model of order 1. Indicates the first The coefficient matrix of each lag position. Indicates the first The stationary state vector of each historical sampling step. This represents the residual vector; the same method is used to estimate the a-b state subvectors. , and residual vector ; coefficient matrix The row index represents the response variable, and the column index represents the first column. Historical driving variables; variable metadata collection Used to fix the meaning of rows and columns in a matrix, so that , , and The arrangement remains consistent across different candidate models; the a-b state sub-vectors are used to form the append... Figure 4 The a-b analysis results corresponding to Table 1 are used to establish a response analysis with the same direction for tributary flow and historical water level at the target station. Candidate order The corresponding baseline model only includes the intercept term and the residual structure; with As the length of the historical data increases, the model can use longer historical sequences, and the number of free parameters increases accordingly. The likelihood term in AIC reflects the degree to which the candidate model fits the training data, and the parameter penalty term constrains the model complexity. By recording the AIC values step by step, the quantitative relationship between the increase in historical length and model complexity can be obtained. Record the residual covariance matrix of each candidate model Maximum likelihood value and the number of free parameters And calculate the Akaike information content criterion: (19); in, express The AIC value of the candidate models of each order; the same sample range, parameter counting method and AIC calculation method are used for each candidate order; Construct the adjoint matrix based on the coefficient matrix of the candidate model and examine its eigenvalues; construct a valid candidate set from the candidate orders whose eigenvalues satisfy the dynamic stability condition. The corresponding model is denoted as: (20); In this embodiment, the AIC calculation results of the candidate VAR model are shown in Table 1: Table 1 AIC Calculation Results
[0024] Among the candidates listed in Table 1, the minimum AIC value is: (twenty one); The corresponding order is 14; the difference between the 8th order AIC value and the minimum AIC value is: (twenty two); An ordered mapping is constructed between each effective candidate order and the AIC value: (twenty three); The 14th order in Table 1 is used to provide a reference for the minimum AIC value. The difference between the 8th order and this minimum value is used for subsequent AIC tolerance judgment. The final input order is determined by combining the impulse response stability and the degree to which the AIC is close to the optimal value. In Table 1, the AIC value for the 8th order is 7.649, and the AIC value for the 14th order is 7.628; the difference of 0.021 is used to express the degree of information loss of the 8th order relative to the minimum AIC value in the table; in the subsequent order selection process, smaller orders are preferentially selected within the candidate range where the impulse response has already met the stability condition, and the difference between the smaller order and the minimum AIC value is limited by the AIC tolerance; accordingly, in this embodiment, the 8th order is determined as the input order that meets the continuous response stability condition and the information criterion tolerance condition; S14. Calculation of Target Water Level Oriented Impulse Response A reference model for impulse response calculation is determined from the valid candidate VAR models. The order of the reference model is denoted as . Convert the reference model to a moving average representation, let And recursively: (twenty four); (25); in, Indicates the impact process The propagation matrix after each sampling step This indicates the upper limit of the impulse response assessment; in this embodiment, it can be set to... To cover the attached Figure 4 The response time domain is shown. Decompose the residual covariance matrix of the reference model: (26); in, Let represent the lower triangular decomposition matrix that follows the order of the variables; Indicates the first The choice vector of the shock variables, let Indicates the target water level If the chosen vector is , then the first... The variables affect the target water level on the [number]th [day]. The directional impulse response for each sampling step is: (27); in, This indicates a pre-fixed impact scale; this embodiment uses a unit impact, i.e. And set the impact of the remaining contemporaneous structures to zero; Unit impacts were applied to the discharge flow of reservoir a, the flow rate of hydrological station c, the flow rate of hydrological station d, and the historical water level of hydrological station b, respectively, and the following results were obtained: (28); (29); (30); (31); The above response sequences are used to determine , , and The historical input order; for each response source Calculate the absolute peak value of the response: (32); When calculating the directional response of a certain driving variable, the target extraction vector is... Always corresponding to the water level position of hydrological station b, the impact selection vector According to the variable metadata set Switch to the corresponding variable position; the resulting four sets of response curves have the same number of response steps and a unified target end, and can respectively reflect the dynamic impact of reservoir discharge, the inflow of the two tributaries, and the historical status of the target station on the subsequent target water level; absolute peak response This is used to eliminate the influence of different variable dimensions and shock response amplitudes; subsequent stability assessments are based on... The normalized amplitude and adjacent changes are judged separately for each variable; the sign of the impulse response is used to indicate the direction of the target water level response, and the absolute value is used to determine the degree of response attenuation and the input position weight; after the weights are constructed, the original flow rate and water level values are still entered into the deep learning model according to their respective physical quantities. Appendix Figure 4 Give the unit impulse response from reservoir a to hydrological station b; (Appendix) Figure 4 The data shows that the impulse response value of hydrological station b reaches its maximum in the first order. In this embodiment, the sampling interval is 6 hours, so the first order corresponds to the first 6-hour sampling step after the change in the discharge of reservoir a. The response then gradually decays and enters a stable range near the 8th order. This response relationship is used to characterize the directional lag relationship between the discharge flow of reservoir a and the water level of hydrological station b under historical data conditions. The change in the outflow from reservoir a forms the maximum response at the first order, indicating that the upstream outflow information from the most recent sampling step has a high response weight to the target water level; it enters a stable range near the eighth order, indicating that when adding earlier historical values, the magnitude of the new response has entered a low and gradual range; this response curve then enters step S2 together with the AIC sequence in Table 1 to determine the historical order that can cover the main dynamic response and control the input length; The statistical analysis results generated in step S1 are kept correlated according to the variable metadata set; for the discharge flow of reservoir a, its response sequence, response peak value, candidate AIC value and sampling interval are saved; for the flow of hydrological stations c and d and the historical water level of hydrological station b, the same type of data are saved respectively; when reading a variable in step S2, the physical name, response curve, peak amplitude and candidate order range of the variable are obtained at the same time, so that the stability judgment of each variable is independent of each other and the final order can be accurately returned to the corresponding input channel; The differential mapping set is retained after the impulse response calculation is completed; the statistical order represents the number of sampling steps between each other on the stationary sequence, and the corresponding physical history duration is obtained by multiplying the statistical order by the sampling interval; when constructing TCN-GRU samples, the same number of sampling steps is used to extract the original flow or water level sequence; through this mapping, the VAR stage can estimate dynamic relationships under stationary conditions, the deep learning stage can directly process numerical values with practical hydrological meaning, and the final output of the model retains the water level unit; The impact scale, variable order, and upper limit of response are fixed during the training period; for the same variable, different response orders use the same impact scale; for different variables, the pre-set unit impact or standardized impact rules are followed; other contemporaneous structural impacts are kept to zero, so that each directional response curve has a clear impact source; subsequent response weights are obtained by normalizing the corresponding curve itself, so the weight size expresses the relative response intensity of each historical position within the variable; In the preferred embodiment, step S2 includes the following steps: S21. Determining the candidate order for a stable response For variables Based on its directional impulse response and absolute peak response Calculate the normalized response amplitude; to keep the denominator stable, set a positive number. And calculate: (33); No. The relative amplitude of the first-order response is: (34); The changes in response between adjacent orders are: (35); (36); Set the continuous stability order Response amplitude threshold and adjacent response change threshold This embodiment can be adopted. , and For candidate order When from the first Continuous from the beginning of the order The order simultaneously satisfies: (37); (38); as well as: (39); (40); Then Join the A stable candidate set of responses for each variable: (41); When no candidate order satisfying the two thresholds has been formed within the preset observation interval, the continuous order is calculated. Maximum tail response metric of order: (42); And will make Add the order of obtaining the minimum value This supplementary rule ensures that each input variable forms a definite and stable candidate range. continuous The order condition is used to prevent a single order from briefly falling into the threshold range due to local fluctuations; the relative amplitude of the response. Constrain the magnitude of the residual response and the change in response between adjacent orders. The smoothness of the constraint response curve; when both satisfy the threshold simultaneously within multiple consecutive orders, the corresponding order can be used as the stable starting point of the variable's historical range of action; , and All of these are determined during the training period and remain fixed during the verification, testing, and scenario simulation phases. For tributary variables with small response amplitudes This ensures that the normalized calculation has a stable denominator; for variables with long response durations, the stable candidate order can be increased accordingly; thus forming , , and Each variable carries its own hysteresis propagation characteristics; S22. Variable-specific order selection under AIC tolerance constraints According to AIC ordered mapping Calculate candidate order The difference between the minimum AIC value and the minimum AIC value: (43); Set AIC tolerance and in the A joint candidate set is formed within the stable candidate set of the responses of each variable: (44); From the joint candidate set Choose the smallest order: (45); Therefore, the following can be determined: (46); in, This indicates the input order of the discharge flow from reservoir a. and These represent the flow input orders of hydrological stations c and d, respectively. This indicates the historical water level input order for hydrological station b; For the original embodiments a-b, see Appendix Figure 4 The response shows that it enters a stable region near the 8th order, and the difference between the 8th order AIC value and the minimum AIC value in Table 1 is 0.021; when And when the 8th order is the smallest order that satisfies the continuous stability condition, determine: (47); Based on the 6-hour sampling interval in this embodiment, the actual historical window corresponding to the discharge flow of reservoir a is: (48); The actual historical windows for the other variables are as follows: (49); Each variable uses its own directional response to determine its historical input range, thereby preserving the propagation duration corresponding to different hydrological driving quantities and the historical state of the target station; AIC tolerance A value of 2 can be used, or a smaller value predetermined during the training period can be used; the tolerance value should be chosen so that the candidate order for stable response and the candidate order for AIC have an intersection; after determination , , and Simultaneously, its statistical order and actual time length are recorded, and compared with the variable metadata set. Binding; In subsequent sample construction, each order is used to extract the historical sequence of the corresponding variable, so that the variable order is consistent with the tensor channel; In the original a-b analysis, the 8th order enters the stable response interval, and the difference between its AIC value and the minimum value in the table is 0.021. This difference is within the example tolerance range, so the 8th order can enter the joint candidate set. If the 8th order is the minimum order in the joint candidate set, then the discharge flow of reservoir a adopts a 48-hour historical window. The discharges of hydrological stations c and d and the historical water level of hydrological station b are determined according to the same rules to determine the corresponding windows, and their specific values are generated by their respective response curves and the same AIC tolerance condition. The order and decision parameters used in the order selection process are recorded as follows: (50); The process remains consistent across the sample construction, model training, testing, and scenario simulation phases. S23, Differentiated History Window Alignment Calculate the maximum value among the input orders of each variable: (51); For the target time Extract each variable separately Recently Each sample value: (52); in, , , , Arrange the historical windows of each variable according to the same target time. Right-align to length Construct the input tensor : (53); in, The rightmost position of the tensor corresponds to the most recent sampled value before the target time, and the values to the left correspond to earlier historical values. Construct effective bit mask : (54); The location corresponds to actual historical hydrological data. The position corresponds to the filling position set with a uniform tensor length; target time The supervised output is: (55); Before selecting the target time from the public timeline, at least the following must be included: At each sampling step, a usable sample set is formed; then according to... and Maintain the time correspondence between training samples, test samples, and target water levels; After right alignment, the rightmost valid position of each variable corresponds to... At that time, the second valid position corresponds to At time; variables with smaller input order retain their valid windows only on the right side of the tensor, and the left-side padding positions are marked by a mask; this arrangement ensures that the same tensor row position has a consistent relative time meaning and enables the convolution kernel to process hydrological information of different variables at the same lag position along the time direction. The effective bitmask is entered into the TCN-GRU model as an independent feature channel; thus, the real hydrological observation with a value of zero and the zero-fill position with a uniform length have different mask states; the subsequent element-wise multiplication prevents the filled position from forming a response weighted feature, and the mask channel further provides the network with the effective window boundaries of each variable. S24, Response Weight Mapping and Normalization The first Convert the absolute values of the impulse responses of each variable into hysteresis position weights: (56); (57); in, It is a positive number; according to the right alignment of the history window, it will be... Write into the unified weight matrix : (58); This ensures that each response weight is consistent with the variables and hysteresis positions in the input tensor; The model fitting period and validation period are divided chronologically within the original training period; the model fitting period is located in the first part, and the validation period is located in the second part; the validation sample can account for 10% to 20% of the original training sample; for other implementation data, the model development data and test data can also be divided chronologically in a 7:3 ratio; this embodiment continues to use January 1, 2010 to December 30, 2020 as the model development period, and January 1, 2021 to August 15, 2022 as the independent testing period; Calculate the first based only on the valid input positions during the model fitting period. Minimum value of each variable and maximum value And normalize it: (59); in, The normalized stability constant; the effective bit mask keeps the fill positions zero; The minimum water level at hydrological station b within the specified time period was obtained by fitting the model. and maximum value Processing and monitoring objectives: (60); Combine the normalized input, effective bitmask, and response weights element by element: (61); And and Connect along the feature channel direction: (62); As a direct input to the TCN-GRU model, As a normalization supervision target, the input order, mask, response weight, and normalization parameters remain consistent throughout the verification, testing, and scheduling phases. Response weight Normalization is performed based on the absolute value of the impulse response within the effective window of each variable; lag positions with larger response amplitudes are given higher weights, and lag positions with smaller response amplitudes are given lower weights; these weights participate in the formation of model input features and, together with the input order, maintain the correspondence between VAR analysis results and TCN convolution positions; The normalized minimum and maximum values are calculated from the effective locations during the model fitting period; the validation, testing, and scenario data are all transformed using the same set of parameters, and the model output also uses the same target water level parameter to restore the actual units; therefore, Each value in the data can be traced back to the original physical quantity, normalization parameter, mask state, and response weight; The normalization parameters are recorded as follows: (63); in, ; Verification, testing, and scenario input are all called Complete the numerical conversion; Sample construction slides progressively according to the target time; for adjacent target times... and Each variable's historical window moves forward one sampling step synchronously, removing the oldest historical value and adding the latest available historical value; the target station's historical water level window only contains the water level before the target time, and the supervised target uses the water level at the current target time; this time arrangement ensures that the training samples are consistent with the subsequent scene recursion process; The fitted, validation, and test samples all use the same tensor shape; differences in input order are expressed through a mask matrix, and the number of variables and the order of tensor channels remain fixed; therefore, the TCN-GRU model can use the same network structure to process data from all time periods; for other implementation data where the input order is redefined, the model is recalculated. The mask and weights are set, and the TCN receptive field is set according to the new maximum window length; The target station water level is used as both historical input and monitoring output with clear time boundaries; prediction At that time, the input tensor contains to The supervisory target is During the training and testing phases, historical water levels are obtained from the measured sequence. During the scenario phase, historical water levels are obtained from the measured sequence before the scenario begins and from the model output of the corresponding scenario after the scenario begins. This time indexing rule is saved when the sample is constructed and is directly called by step S5. The response weight matrix can be used as a fixed multiplicative weight to participate in the formation of input features, or it can be set as a trainable parameter based on a fixed initial value. When the trainable method is used, the initial value of the weight is still determined by the directional impulse response, and the position weight is always zero by mask constraint. After training, the updated effective position weight is saved. Regardless of whether the fixed method or the trainable method is used, the row position of the weight matrix is consistent with the corresponding lag order. In the preferred embodiment, step S3 includes the following steps: S31. Constructing the TCN feature extraction network Input tensor Input a temporal convolutional neural network; the TCN performs one-dimensional causal convolution along the temporal direction, so that the convolutional features at each time position are formed by the historical input at that position and before it; the... The dilated causal convolution of each residual block is represented as: (64); in, This indicates the output of the previous residual block. Indicates the convolution kernel weights. Indicates the kernel length. Indicates the first The expansion coefficient of each residual block, Indicates bias. Indicates the time position in the input window; Perform weight normalization on the convolution parameters, and then apply activation and Dropout: (65); in, This indicates weight normalization. This represents the activation function. Indicates the Dropout ratio; adds the convolution result to the residual branch: (66); in, This represents the channel mapping of the residual branch; when the number of input channels is the same as the number of output channels, an identity mapping is used; when the number of channels is different, an identity mapping is used. Convolution completes dimension matching; When each residual block contains two causal convolutions with the same dilation coefficient, the receptive field of the TCN is: (67); in, Indicates the number of residual blocks; set the TCN structure parameters to: (68); This embodiment can use the convolution kernel length. The dilation coefficients are 1, 2, and 4, and there are 32 or 64 convolution channels. The dropout ratio can be determined within the range of 0.1 to 0.3. For the original 8th-order input window of a-b, the following is adopted: And when the expansion coefficients are 1, 2, and 4, the receptive field is: (69); It can cover 8th-order historical input; as the uniform tensor length increases, residual blocks with expansion coefficients of 8, 16, etc. can be added until the receptive field coverage condition is met. The output of the last residual block of TCN forms a feature sequence in chronological order: (70); In TCN, causal convolutions preserve the prediction time boundaries, and convolution computations only access input positions prior to the target time. The dilation coefficient increases layer by layer, allowing fewer network layers to cover a longer historical window. Residual connections provide a common propagation path for shallow local features and deep long-term features. Weight normalization and Dropout are used to stabilize the training process and constrain model complexity. Receptive field conditions are also included. Ensure that the earliest valid historical position retained in step S2 can influence the TCN output features; When the input order of different variables changes, the uniform tensor length The data is updated accordingly, and the receptive field is re-verified based on this data. The number of TCN convolution channels, the number of residual blocks, and the dilation coefficient can be selected on the validation data. The selected parameters are saved together with the trained convolution weights for direct use in testing and scheduling scenarios. Record the TCN structure parameters as follows: (71); in, Indicates the number of convolution channels; S32, GRU timing status update and water level output TCN feature sequence Enter GRU sequentially according to the time position; for the time position Calculate the reset gate: (72); Compute update gate: (73); Calculate candidate hidden states: (74); And update the current hidden state: (75); in, This indicates that the door is being reset. Indicates an update to the door. Indicates the candidate hidden state. Indicates the current hidden state. , and This represents the input weight matrix. , and Represents the cyclic weight matrix. , and Represents the bias vector; Input the hidden state at the last valid time position into the fully connected output layer to obtain the normalized simulated water level: (76); Inverse normalization is performed using the water level normalization parameters determined during the model fitting period: (77); This leads to the TCN-GRU model: (78); in, This represents all trainable parameters for the TCN, GRU, and fully connected output layer; the number of hidden units in the GRU can be 32 or 64, and the specific setting should be determined based on the validation data. The reset gate is used to adjust the proportion of the previous hidden state participating in the candidate state calculation, and the update gate is used to adjust the proportion of the previous hidden state and the current candidate state in the new hidden state. The TCN has converted the multivariate inputs at each lag position into convolutional features, and the GRU further updates the hidden state along the time position, so that the local propagation features and the state evolution relationship within the entire historical window can be entered into the target water level output together. The fully connected output layer maps the hidden state at the last valid time position to a single normalized water level value; after denormalization, it yields a value in meters. This value is used to compare with the measured water level during the training and testing phases, and is used to write the historical water level cache of the corresponding scenario during the scenario phase. The GRU and output layer structure parameters are recorded as follows: (79); in, This represents the number of hidden units in the GRU; the weight matrix described above is updated during training and serves as... The components are saved; S33, Model Training, Validation, and Parameter Locking The model development data from January 1, 2010 to December 30, 2020 was divided into model fitting data and validation data according to time sequence. The model fitting data was located in the first part of the time series, and the validation data was located in the second part. The model parameters were updated using the model fitting data, and the early stopping time and model structure parameters were determined using the validation data. Using the normalized water level as the monitoring target, calculate the batch root mean square error loss: (80); in, Indicates the number of samples in the current batch. Indicates the model output, This represents the corresponding normalized measured water level; This embodiment can use the Adam optimizer, and the initial learning rate can be set to... The batch size can be set to 32 or 64, the maximum number of training epochs can be set to 200, and the early stopping patience value can be set to 20. After each training epoch is completed, the validation loss is calculated using the validation data. When the validation loss reaches the current minimum value, the model parameters are saved. (81); in, Indicates the first The model parameters at the end of each training epoch; if the loss does not decrease after a preset number of consecutive epochs, training is stopped and the parameters are restored. The learning rate can be reduced according to a preset ratio after the validation loss reaches a plateau. The completed training model is denoted as Input order , , and Valid bitmask Response weight matrix The normalized parameters of the fitting time period and the TCN-GRU network parameters are stored together for test data and subsequent scheduling scenario data processing. During training, the time window order within each sample remains unchanged; validation data always maintains independent time boundaries; early stopping data is saved. This includes TCN convolutional kernels, residual mappings, GRU gating matrices, and fully connected output layer parameters; the validation phase is used for model selection, and the testing phase performs independent evaluation after all parameters are locked. The model parameter package simultaneously records the optimizer used, initial learning rate, batch size, maximum number of training epochs, early stopping patience value, TCN receptive field, and number of GRU hidden units. Using this parameter package, the model calculation process can be recovered under the same input data structure, and the same set of model parameters can be used in the baseline and adjustment scenarios. S34. Water level simulation and index calculation during the test period Using the completed training model The test data from January 1, 2021 to August 15, 2022 were processed to obtain the simulated water level sequence for the test period; the simulated water level was matched point by point with the measured water level at the same time, and NSE, RMSE, MAE, KGE and TPE were calculated. The Nash efficiency coefficient is: (82); The root mean square error is: (83); The mean absolute error is: (84); The Kling-Gupta efficiency coefficient is: (85); in: (86); (87); Indicates the number of evaluation samples. This represents the average measured water level. This represents the correlation coefficient between the simulated water level and the measured water level. and These represent the standard deviations of the simulated water level and the measured water level, respectively. and These represent the average values of the simulated water level and the measured water level, respectively. Determine the measured peak time and simulated peak time And calculate the normalized peak time error: (88); in, This indicates the total duration of the evaluation period; when there are multiple independent flood events, the TPE of each event can be calculated separately and the average value can be taken. NSE is used to evaluate the overall fit of simulated water level to the measured water level change process; RMSE assigns higher weight to larger water level errors; MAE represents the average level of absolute error at each time step; KGE comprehensively reflects correlation, fluctuation amplitude, and mean ratio; TPE is used to reflect the deviation at the time of peak occurrence; all indicators are calculated using the same test sample range and the same simulated water level sequence. During the testing period, simulated water level, measured water level, evaluation index and time index are saved together; subsequent model comparisons use the same test boundary and index formula, so that the performance differences between GRU, TCN and TCN-GRU correspond to their respective model structures and feature processing processes. During the model training phase, the hierarchical relationship between data processing parameters and network parameters is further maintained; input order, mask position, and initial response weights belong to sample structure parameters; kernel length, dilation coefficient, number of convolution channels, and receptive field belong to TCN structure parameters; number of hidden units and gating matrix dimension belong to GRU structure parameters; learning rate, batch size, maximum number of epochs, and early stopping patience value belong to training control parameters; all types of parameters are recorded separately and loaded together when the model is used for testing or scenario simulation. Validation data is used to compare candidate network structures and training epochs; candidate structures may include different numbers of convolutional channels, different numbers of GRU hidden units, and Dropout ratios within a preset range; each candidate structure is trained using the same fitted data and validation data, its optimal parameters are determined according to the validation loss, and then the network structure used for testing is determined according to the validation metrics; after the structure selection is completed, the test data is input only once to form independent test results; The dimensions of the TCN output feature sequence and the GRU hidden state are determined by the network structure parameters; the time length of the TCN feature sequence is consistent with the length of the unified tensor, and the number of feature channels is determined by the last TCN residual block; the GRU reads these time positions one by one and inputs the final hidden state into the water level output layer; this connection method enables all time position features generated by the TCN to participate in the final water level generation through GRU state updates; Each simulated water level in the test results can be traced back to the corresponding original input window; based on the target time, the discharge flow of reservoir a, the flow of hydrological stations c and d, and the historical water level of hydrological station b can be obtained at that time, and the mask position, response weight, normalization parameters, and model output can be further obtained; this correspondence is used to verify the sample range of abnormal simulation points, peak times, and evaluation indicators; In the preferred embodiment, step S4 includes the following steps: S41. Comparison Model Settings Under the same model fitting data, validation data, test data, input variables, target time, and normalization parameters, a single GRU model, a single TCN model, and a TCN-GRU model were established respectively. The GRU model uses the same historical input tensor as the fusion model and generates the target water level through GRU. The TCN model uses the same historical input tensor and generates the target water level through dilated causal convolution and output layer. The TCN-GRU model uses TCN to extract time-domain features, and then uses GRU to update the hydrological status and output the target water level. All three models used validation data to select parameters and, after fixing the parameters, processed the same test period, resulting in: (89); Among them, superscript This represents the TCN-GRU model; all three sets of simulated water levels are aligned with the measured water levels step by step according to the test time index. A single GRU model was used to test the ability of gated cyclic state updates to simulate the target water level; a single TCN model was used to test the ability to extract temporal convolutional features; and a TCN-GRU model was used to test the ability to simulate water level after the two types of structures are combined in a predetermined order. All three models used the same target variable and test period, and the units and calculation directions of the evaluation indicators were kept consistent. S42, Evaluation Data Verification Substitute the simulated water levels of the three models into the unified evaluation formula in step S34; based on the connection order of the fusion model, which first performs TCN feature extraction and then GRU state update, the fusion model is denoted as "TCN-GRU"; the evaluation results of each model are shown in Table 2. The data in Table 2 is as follows: Table 2 Evaluation Indicators of Hydrological Station b-level Model under Three Models
[0025] Among them, the larger the NSE and KGE values, the higher the consistency between the simulated sequence and the measured sequence; the smaller the RMSE and MAE values, the lower the water level error; and the smaller the TPE value, the closer the simulated peak time is to the measured peak time. S43. Calculation of the percentage change in evaluation indicators For NSE and KGE, the improvement of TCN-GRU relative to the baseline model is calculated using the following formula: (90); For RMSE, MAE, and TPE, the reduction rate is calculated using the following formula: (91); According to the values in Table 2, the NSE of TCN-GRU is approximately 11.25% higher than that of GRU, the RMSE is approximately 68.12% lower, the MAE is approximately 67.31% lower, the TPE is approximately 54.55% lower, and the KGE is approximately 8.31% higher. Compared to TCN, TCN-GRU improves NSE by approximately 1.85%, reduces RMSE by approximately 24.14%, reduces MAE by approximately 22.73%, reduces TPE by approximately 28.57%, and improves KGE by approximately 0.32%; all the above ratios are based on the values in Table 2 and are rounded to two decimal places. The above proportions are used to express the relative relationships between the values already in Table 2; NSE and KGE use the improvement rate formula, while RMSE, MAE, and TPE use the reduction rate formula; the benchmark value of each proportion is taken from the same index of the corresponding comparison model, so as to keep the calculation direction consistent with the index evaluation direction. S44, Response Model Parameter Encapsulation According to Table 2, the NSE and KGE of TCN-GRU are higher than the corresponding metrics of GRU and TCN, while its RMSE, MAE, and TPE are lower than the corresponding metrics of GRU and TCN. The trained TCN-GRU model is identified as the scene response model and encapsulated as follows: (92); in, Record the input order, stability threshold, and AIC tolerance for each variable. Record the normalized parameters during the model fitting period. Record the convolution kernel, dilation coefficient, number of channels, Dropout, and receptive field. Record the number of hidden units and network layer settings. Record all network weights after training; The model package is also associated with the test time boundary, the evaluation data in Table 2, and the calculation method of the indicators; step S5 directly reads the model package to build the scene input and generate the target water level, so that the scene stage uses the input structure, normalization rules and model parameters that have been verified; During the model package loading process, the variable metadata set is first read and the input channel order is restored. Then, the input order of each variable and the uniform tensor length are read. Subsequently, the mask, response weights and normalization parameters are loaded. Finally, the TCN-GRU network structure and training weights are restored. After the above loading is completed, the target water level can be generated for any continuous hydrological sequence that meets the same variable definition. Table 2 establishes a connection between the evaluation data and the model package to confirm the training data boundaries, test data boundaries, and evaluation results corresponding to the current model; the model identifier used in the scenario simulation is consistent with the TCN-GRU model in Table 2; if the model is retrained in other river sections, the variable metadata, order selection results, normalization parameters, network weights, and evaluation data are saved for each river section to maintain the independence of the boundaries between different model packages; Model verification also includes consistency between the input tensor length and the TCN receptive field, consistency between the model output dimension and the target water level dimension, and consistency between the inverse normalization parameter and the water level variable of hydrological station b. After the above verification, the model package enters step S5. Step S5 only changes the scene input sequence and keeps the model structure and parameters locked, so that different release schemes can be compared under the same model conditions. Based on the unified evaluation results, the trained TCN-GRU model will be... The model was determined to be a response model for scheduling scenarios, and the input order, mask, response weights, normalization parameters, TCN structure parameters, GRU structure parameters, and trained model weights were collectively encapsulated into a river hydrological response model package. ; In the preferred embodiment, step S5 includes the following steps: S51, Construction of Baseline Scenario and Downward Adjustment Scenario Obtain historical discharge flow from reservoir a Historical flow rates of hydrological stations c and d and And the historical water level of hydrological station b before the scene begins. ; Set the historical discharge flow of reservoir a as the baseline scenario: (93); Set the start time of the scene Adjustment of outflow rate and adjust duration The scenario for increasing the outflow is: (94); Scenarios for reducing outflow: (95); The three scenarios use the same tributary boundary: (96); (97); Three scenes in Previously, the same historical water level at hydrological station b was used; this embodiment uses the following scenario parameters: (98); (99); (100); The sampling interval is 6 hours, therefore the duration step count is adjusted as follows: (101); The scenario simulation scope includes the period of leakage adjustment and the response extension period after the adjustment ends; the response extension time can be set to no less than When the scene water level difference is still higher than the preset water level response threshold at the end of the extended time period; At that time, continue with Extend the simulation period by unit until the water level difference is continuous. The sampling time is lower than ; The baseline scenario, the increased flow scenario, and the decreased flow scenario all use the same start time, simulation end conditions, and initial water level at the target station. The increased and decreased flow scenarios only replace the outflow of reservoir a during the adjustment period, and restore the corresponding historical baseline outflow process after the adjustment period ends. The tributary flow uses the same time series in the three scenarios, thus forming a scenario comparison with a unified external boundary. The extended response period is used to cover water level changes that continue to propagate in the river channel after the discharge adjustment has ended; It provides a minimum extension length, a water level response threshold, and a number of consecutive recovery steps to provide dynamic termination conditions; the resulting scenario time range can simultaneously cover the discharge adjustment process and the subsequent recovery process of the target station. S52, Scene water level is generated recursively time-by-time. Scene type is denoted as Determine the target time according to the time range of the scene: (102); (103); For any scenario and target time Extract the most recent water from reservoir a. The recent discharge flow and hydrological station C The flow rate and hydrological station d recently The flow rate and hydrological station b recently Water level; The historical water level of the target station before the scene begins uses the measured historical water level; the historical water level of the target station after the scene begins uses the water level already generated by the model for the corresponding scene. (104); This allows for the construction of a historical water level window for the target station: (105); The baseline scenario, the increased flow scenario, and the decreased flow scenario each maintain their own recursive water level cache; following the right alignment rule in step S23, the history windows of each variable are written into the scenario input tensor. Using the same effective bit mask Identify the actual historical location and use the normalized parameters determined by the model fitting period to process each valid input; For the target time The discharge flow window of reservoir a is determined according to the scenario. Read The flow windows of hydrological stations c and d always read the common tributary boundary, while the water level window of the target station reads the water level buffer of the scenario itself. The input tensors of the three scenarios have the same shape and the same variable order. The difference lies in the outflow of the scenario and the recursive target water level formed by this difference. In the first simulation step after the scenario begins, the target station's historical water level window is entirely derived from the measured water level before the scenario begins; after the model generates the first scenario water level, it is written to the cache; the target station's historical water level window in the second simulation step is composed of the earlier measured water level and the first model water level; as the recursion progresses, the model water level after the scenario begins gradually replaces the measured historical position in the window until the entire target station's historical water level window is composed of the model output of the corresponding scenario. Using response weight matrix Forming a scene feature tensor: (106); And form the model input: (107); The scene input is fed into the fully trained TCN-GRU model: (108); And perform inverse normalization: (109); Will Write the water level to the corresponding scenario's water level cache, and calculate... The time is used as an input value in the historical water level window of the target station; the above process is repeated in chronological order to obtain the benchmark water level sequence. Water level sequence in scenarios with increased discharge flow and water level sequence in scenarios of reduced outflow ; The baseline water level sequence is also generated recursively, so that the baseline scenario and the adjustment scenario have the same historical water level source of the target station; the water level difference at each time step is calculated by two simulation sequences using the same recursive rule, so as to keep the scenario comparison caliber consistent; each scenario water level retains its target time, input tensor, normalized input and previous water level buffer position, which is used to trace the complete recursive process. S53, Calculation of Scene Water Level Response Indicators Calculate the water level difference between the increased discharge scenario and the baseline scenario: (110); And reducing the water level difference between the outflow scenario and the baseline scenario: (111); The maximum water level rise is: (112); The maximum water level drop is: (113); Determine the corresponding peak time: (114); (115); And calculate the peak response time relative to the start of the scene: (116); (117); The scene water level difference first reaches the response threshold. The moment is determined as the response start time: (118); Restore the water level difference to Within and continuously maintain The first sampling moment is determined as the response end moment. The response duration is: (119); Response threshold The value can be set based on the water level observation resolution or the error level of the model fitting period, and the same value should be used in all three scenarios; Continuous recovery steps Used to confirm that the water level difference has continuously returned to the threshold range, reducing the impact of a short-term drop at a single sampling point on the end of the response; the maximum water level change reflects the magnitude of the scenario's impact, the peak response time reflects the time it takes for the adjustment effect to be transmitted to the extreme value of the target station, and the response duration reflects the range of the scenario's impact. In this embodiment, the discharge flow rate at 20:00 on August 18, 2020 is increased. And after three consecutive days, the water level at hydrological station B rose by 1.56m; this was reduced compared to the same historical discharge flow. And after 3 days, the water level at hydrological station b dropped by 1.83m, that is: (120); The reference water level, the water level under the increased flow scenario, the water level under the decreased flow scenario, and the water level difference between the two sets are indexed according to the same time period to form an appendix. Figure 6 The corresponding scenario response curve; scenario parameters, recursive input, recursive water level output, maximum water level change, peak time and response duration together form the river hydrological response result; through the above steps, historical hydrological data, VAR directional response, variable specific input order, differentiated historical window, TCN-GRU model and reservoir discharge scenario water level response form a continuous data processing relationship. S54, Forming Scene Hydrological Response Results Based on the scenario time index, the hydrological station b water levels corresponding to the baseline scenario, the scenario with increased discharge flow, and the scenario with decreased discharge flow are combined as follows: (121); in, Indicates the baseline scenario at the target time. The corresponding simulated water level at hydrological station b, This indicates a scenario where the outflow is increased at the target time. The corresponding simulated water level at hydrological station b, This indicates a scenario where the outflow is reduced at the target time. The corresponding simulated water level at hydrological station b, This represents the target water level records for three scenarios at the same target time. Will Water level difference compared to scenarios with increased outflow and reducing the water level difference in scenarios involving downstream flow. By associating the same target time, the time process of forming the benchmark water level, the water level in the scenario of increasing the discharge flow, the water level in the scenario of decreasing the discharge flow, and the water level difference between the two adjustment scenarios relative to the benchmark scenario is formed; further, the discharge flow adjustment amount, adjustment duration, maximum water level change, response start time, peak response time, and response duration are written into the result record of the corresponding scenario. The river hydrological response results include the river hydrological response model package. The system includes scene boundaries, recursive inputs and water level buffers for three scenes, three sets of target water level sequences, two sets of water level difference sequences, and corresponding water level response indicators. The river hydrological response results are used to compare the target station water level changes corresponding to different reservoir discharge processes. The system can also trace the generation process of scene water level response based on the saved variable input order, effective historical location, response weight, normalization parameters, and time-by-time model output. The target water level sequences for the three scenarios use the same sampling time and sequence length. When plotting the scenario water level response curves, the scenario time is used as the horizontal axis and the water level at hydrological station b is used as the vertical axis. The baseline scenario water level, the water level of the scenario with increased discharge flow, and the water level of the scenario with decreased discharge flow are displayed simultaneously. The same scenario time is used as the horizontal axis to display the water level difference of the scenario with increased discharge flow and the water level difference of the scenario with decreased discharge flow. Based on the scenario water level response curves, the time when the water level difference first reaches the response threshold after the start of discharge flow adjustment, the time when the maximum water level change occurs, and the water level recovery process after the end of discharge flow adjustment are identified. In the scenario results record, the discharge flow adjustment is kept in flow rate units, the maximum water level change is kept in water level units, and the peak response time and response duration are converted to hours or days based on the sampling interval. In this embodiment, the adjustment amounts for both the increased and decreased discharge flow scenarios are... The adjustment period is 3 days, and the maximum water level changes at hydrological station b are 1.56m rise and 1.83m fall, respectively; the above data and the corresponding scene boundary are written into the scene result record. When comparing multiple reservoir discharge scenarios, keep the tributary boundaries of hydrological stations c and d consistent with the initial water level of hydrological station b, and replace the discharge flow rate of reservoir a scenario accordingly. The scenario water level is recursively generated repeatedly. Each discharge scheme forms an independent target water level buffer and water level difference sequence, which are then compared according to the maximum water level change, peak response time, and response duration. For cases where a tributary inflow process is given separately, the given tributary flow sequence is written into the corresponding scenario boundary, and the same variable input order, effective historical position, response weight, and model parameters are used to complete the scenario water level recursively generated. In step S1, the differential mapping, AIC sequence, and directional impulse response generated are converted into variable-specific input order, effective historical position, and response weight in step S2. In step S3, the model input generated in step S2 is converted into TCN time series features, GRU hydrological state, and target water level. In step S4, the model trained in step S3 undergoes unified testing and comparative verification. In step S5, the river hydrological response model package encapsulated in step S4 is used for recursive water level generation in the baseline and adjustment scenarios. The intermediate results formed in each stage have subsequent usage positions, and finally, the complete water level response process of hydrological station b under the changing conditions of reservoir discharge process is obtained. In the specific execution process, a unique version identifier is established for each model version, and the model version identifier is associated with the training time boundary, variable metadata set, difference mapping, order selection parameters, normalization parameters, and network weights; when the scenario simulation starts, all parameters under the same model version are read to ensure that the input sequence, model structure, and output recovery rules are consistent; when the model version is updated, the corresponding test evaluation record is regenerated, and the scenario simulation is executed using the updated complete river hydrological response model package; Before hydrological data is input into the model, input channels are validated according to the variable metadata set; input channel validation includes verifying variable names, units of measurement, sampling intervals, time index ranges, and target station identifiers; the discharge flow from reservoir a is maintained. The flow rates at hydrological stations c and d are measured in the same unit, while the water level at hydrological station b is maintained at a constant level. Units; When the sampling interval of the hydrological data to be processed is consistent with the sampling interval recorded in the river hydrological response model package, a historical input window is constructed according to the determined variable input order; When the hydrological data to be processed uses other sampling intervals, VAR analysis, joint order selection and model training are re-executed to form a river hydrological response model package corresponding to the corresponding sampling interval; The normalization parameters determined during the model fitting period are continuously used during the scenario simulation. When the scenario discharge exceeds the maximum value or falls below the minimum value during the model fitting period, the normalized discharge continues to be calculated according to the linear transformation relationship determined during the model fitting period, and the actual scenario discharge, normalized discharge, and model output are saved together. The above records are used to characterize the position of the scenario input relative to the model training data range and to provide data basis for interpreting the scenario results. The target water level denormalization always uses the normalization parameters of the model fitting period corresponding to hydrological station b. The recursive scenario calculation is executed step by step according to the sampling time sequence. After the simulated water level at any target time is generated, the simulated water level is written to the water level cache of the corresponding scenario, and then the process is advanced to the next target time. Independent water level caches are set for the baseline scenario, the increased discharge flow scenario, and the decreased discharge flow scenario, so that the simulated water level of the increased discharge flow scenario only enters the subsequent historical water level input of the increased discharge flow scenario, the simulated water level of the decreased discharge flow scenario only enters the subsequent historical water level input of the decreased discharge flow scenario, and the simulated water level of the baseline scenario only enters the subsequent historical water level input of the baseline scenario. This maintains the continuity of hydrological status within each scenario and maintains data isolation between different scenarios. When calculating the scene water level difference, the target time corresponding to the two target water level sequences to be compared is verified to be consistent, and then the subtraction is performed time by time. The maximum water level rise is obtained from the water level difference of the scene with increased discharge flow, and the maximum water level drop is obtained from the water level difference of the scene with decreased discharge flow. The peak time corresponds to the sampling time when the extreme value of water level change is obtained. When multiple sampling times obtain the same extreme value, the sampling time that reaches the corresponding extreme value earliest is selected as the peak response time, and all parallel extreme value times are retained in the scene result record. The same water level response threshold is used for the start and end times of the response; the start time of the response corresponds to the sampling time when the scene water level difference first reaches the water level response threshold, and the end time of the response corresponds to the first sampling time when the scene water level difference returns to within the water level response threshold and continuously maintains the preset sampling steps; the response duration covers the complete interval from the start of the identifiable response to the continuous recovery of the scene water level difference; for scenarios that are still in the response state at the end of the response extension period, the tributary boundaries of hydrological stations c and d and the baseline discharge process of reservoir a are added, and the subsequent target water level of hydrological station b is recursively generated until the response end condition is met; The river hydrological response results include both scenario-specific extreme water level values and complete target water level time series. The complete target water level time series is used to characterize the water level response propagation, peak formation, and recovery process, while extreme water level changes and response time indicators are used to form quantitative comparisons between different discharge scenarios. The reservoir dispatch control terminal acquires the river hydrological response results and dispatch constraint information applicable to the target river section; the dispatch constraint information includes the target control water level of hydrological station b. Allowable water level variation and allowed response duration The reservoir dispatch control terminal compares the target water level sequence, maximum water level change, and response duration for each discharge scenario with the dispatch constraint information to generate an evaluation result for the discharge scenario. ; Evaluation results of leakage scenarios Record the scenario identifier, discharge flow adjustment amount, adjustment duration, maximum water level change at the target station, peak response time, response duration, and compliance status of scheduling constraints; the reservoir scheduling and control terminal evaluates the discharge scenario based on the results. Different reservoir release scenarios are ranked, and release scenarios that meet the scheduling constraints are marked as candidate release schemes; The reservoir dispatch and control terminal evaluates the results of the discharge scenario. Generate discharge scheduling auxiliary signals And will release the dispatch auxiliary signal The information is sent to the dispatch information display device; the dispatch information display device then uses the released dispatch auxiliary signal. Displays candidate discharge schemes, the corresponding discharge flow process of reservoir a, the water level response process of hydrological station b, and various water level response indicators; When the target water level for any discharge scenario exceeds the target control water level The maximum water level change exceeds the allowable water level change. Or the response duration exceeds the allowed response duration. At that time, the reservoir dispatch control terminal generates an early warning signal for water level exceeding the limit. And will issue an early warning signal for water level exceeding the limit. Send to the dispatch information display device to trigger visual and audible warnings for the corresponding leakage scenario; Authorized dispatchers determine the actual release plan based on candidate release plans, river hydrological response results, and applicable reservoir operation procedures. The reservoir operation control terminal receives confirmation of the actual release plan and generates release control commands based on the confirmed plan. and will release control commands The command is sent to the reservoir gate control system; the reservoir gate control system then responds to the discharge control command. The gate actuator of the reservoir is controlled to adjust the gate opening, so that reservoir a can carry out the corresponding discharge process according to the confirmed actual discharge plan.
[0026] In a preferred embodiment, this example provides a river hydrological response simulation system that integrates vector autoregression and deep learning. The river hydrological response simulation system executes the river hydrological response simulation method described in Example 1, and includes a data analysis module, an input construction module, a model computation module, a model verification module, and a scenario simulation module, wherein: The data analysis module uses an Advantech IPC-610L industrial control computer. Through the communication interface, it reads the historical outflow from reservoir a (reservoir scheduling and operation database) and the flow rates from hydrological stations c and d, and the water level from hydrological station b (hydrological station data management equipment). It verifies the hydrological observation records according to sampling time, monitoring location, physical quantity name, and unit of measurement, establishes a unified timeline, and divides the time into training and testing periods. It performs augmented Dickey-Fuller tests on each hydrological sequence, and performs differencing on hydrological sequences that do not meet the stationarity condition. It establishes vector autoregressive models of different candidate orders, calculates the Akaike information criterion value and directional impulse response, and forms the response amplitude, response decay process, and adjacent order response changes for each hydrological variable pointing to the subsequent water level of the hydrological station. The input construction module determines the candidate order of response stability based on the response amplitudes of multiple consecutive sampling orders and the changes in adjacent responses. Then, based on the minimum value of the Akaike information content criterion and the preset information loss tolerance, it determines the input order for the discharge flow of reservoir a, the flow rate of hydrological station c, the flow rate of hydrological station d, and the historical water level of hydrological station b, respectively. This embodiment is based on the appendix... Figure 4 Table 1 defines the input order of the discharge flow from reservoir a as order 8, corresponding to a 48-hour historical range. The input construction module uses the same time to be simulated as the right boundary of the historical window of each hydrological variable, writes historical windows of different lengths into the input position of uniform length, uses effective position identifiers to distinguish between the real hydrological observation position and the supplementary position, sets the response weight of each historical position according to the amplitude of the directional impulse response, and uses the maximum and minimum values of the training period to process the training data, test data and scenario data. The model computation module uses a Dell PowerEdge R750 server and an NVIDIA RTX A6000 graphics processor to perform model training and water level simulation. The module uses a TCN model to perform dilated causal convolution processing on the hydrological input features, passing hydrological features at different time scales through residual connections. Then, a GRU model updates the hydrological status of the target river segment according to the sampling time sequence and outputs the simulated water level at hydrological station b. The module adjusts the TCN model parameters, GRU model parameters, and water level output parameters using mean squared error or root mean square error, and saves the training parameters and terminates model training based on the validation loss. The model verification module uses the same test period to test the GRU model, TCN model, and TCN-GRU model respectively, and calculates the Nash efficiency coefficient, root mean square error, mean absolute error, peak time error, and Kling-Gupta efficiency coefficient. The model verification module determines the TCN-GRU scenario response model based on the unified evaluation results, and saves the input order of each hydrological variable, effective location identifier, response weight, normalization parameter, network structure parameter, and training weight to form a river hydrological response model package. The scenario simulation module establishes a baseline scenario, an increased discharge flow scenario, and a decreased discharge flow scenario. Under the same scenario start time, tributary boundary, and initial water level conditions at the target station, the discharge process of reservoir a is adjusted. In this embodiment, 20:00 on August 18, 2020, is used as the scenario start time. The discharge flow of reservoir a is increased or decreased by 10,000 cubic meters per second and continues for 3 days. The scenario simulation module sets up independent water level caches for each scenario, initializes the historical water level window using the measured water level before the scenario starts, calls the river hydrological response model package hourly to generate the simulated water level of hydrological station b, and writes the current simulated water level into the water level cache of the corresponding scenario for continued use in the next moment. The scenario simulation module calculates the difference between the adjusted scenario water level and the baseline scenario water level, determining the maximum water level change, response start time, peak response time, and response duration. In this embodiment, the maximum water level rise corresponding to increased discharge flow is 1.56 meters, and the maximum water level drop corresponding to decreased discharge flow is 1.83 meters. The scenario simulation module connects to the reservoir dispatch control terminal and the dispatch information display device via an Ethernet interface. It sends candidate release plans, the release process of reservoir a, and the water level response process of hydrological station b to the dispatch information display device. When the water level response exceeds the dispatch constraints, it sends an over-limit warning signal. After the reservoir dispatch control terminal confirms the actual release plan, it generates a release control command and sends the release control command to the reservoir gate control system, causing the reservoir gate actuator to adjust the gate opening.
[0027] In a preferred embodiment, this embodiment provides an electronic device, including a processor, a memory, a communication interface, and a communication bus. The communication bus connects the processor, the memory, and the communication interface. The memory stores computer programs, model parameters, and hydrological data required for river hydrological response simulation. The processor reads and executes the computer programs to implement the aforementioned river hydrological response simulation method. The communication interface receives flow monitoring data, water level monitoring data, and scheduling constraint information, and outputs hydrological response results, scheduling auxiliary signals, over-limit warning signals, or discharge control commands.
[0028] In a preferred embodiment, this embodiment provides a non-transitory computer-readable storage medium that stores a computer program. When the processor executes the computer program, it implements the above-mentioned river hydrological response simulation method. The computer-readable storage medium includes at least one of a read-only memory, a random access memory, a solid-state memory, a disk, or an optical disk.
[0029] The above embodiments are merely preferred technical solutions of the present invention and should not be considered as limitations on the present invention. The scope of protection of the present invention should be limited to the technical solutions described in the claims, including equivalent substitutions of the technical features described in the claims. That is, equivalent substitutions and improvements within this scope are also within the scope of protection of the present invention.
Claims
1. A method for simulating river hydrological response by integrating vector autoregression and deep learning, characterized in that, Includes the following steps: S1. Acquire external hydrological driving data and historical water level data of the target section, form an isochronous hydrological sequence and divide it into training and testing periods; The training sequence is stationary, and vector autoregression analysis is performed based on the stationary representation to obtain the evaluation results of the information criterion for the candidate order and the directional impulse response of the input variables to the target section water level. S2. Determine the order of the input variables based on the preset response stability conditions and information criterion tolerance, combined with the evaluation results of the information criterion and the directional impulse response; Historical windows of different lengths are extracted from the hydrological sequence of equal time axis, right-aligned according to the target time and marked with effective positions, response weights are set according to the amplitude of directional impulse response, normalization parameters are determined according to the hydrological data of the training period, training and test samples containing model input and measured water level are formed, and input processing configuration is recorded. S3. A deep learning model that integrates temporal feature extraction and hydrological state recursion is trained using training samples to generate simulated water levels at the target cross-section. S4. Establish a benchmark model using the same training and test samples and evaluate the benchmark model and the fusion deep learning model. Calculate the water level evaluation index based on the simulated water level and the measured water level. Determine the fusion deep learning model as the scene response model according to the preset model selection rules. Encapsulate the scene response model and input processing configuration into a river hydrological response model package. S5. Set the baseline hydrological driving data and adjust the hydrological driving data to maintain the consistency of the hydrological boundaries except for the input variables to be adjusted. Initialize the historical water level window with the measured water level before the scene. Form the scene input according to the input processing configuration in the model package. Call the scene response model to generate simulated water levels at each time step and backfill the scene historical water level window to form the river section water level response process. Determine the candidate hydrological driving control scheme according to the water level response process and the preset scheduling constraints.
2. The river hydrological response simulation method according to claim 1, characterized in that, Step S1 specifically includes: Hydrological observation records are read according to the sampling time, and the monitoring location, physical quantity name and unit of measurement corresponding to the hydrological observation records are verified. The verified hydrological observation records are unified to the same time reference and formed into an equal time axis hydrological sequence according to the preset sampling interval. The training period and the testing period are divided according to the time sequence. During the training period, the augmented Dickey-Fuller test is performed on each hydrological variable. For hydrological variables that do not reach the stationarity condition, the difference processing is performed until the corresponding stationary representation is obtained. The difference mapping corresponding to each hydrological variable is recorded, and the hydrological sequence with the original physical meaning is retained for subsequent sample construction. Based on the stationary representation of each hydrological variable, vector autoregressive models of different candidate orders are established. The Akaike information content criterion evaluation results of each candidate order are obtained by using a consistent effective evaluation interval and information criterion calculation method. Effective candidate models that meet the dynamic stability conditions are selected from the candidate vector autoregressive models. The impulse response reference model is determined from the effective candidate models. The residual covariance of the reference model is orthogonally decomposed. Independent impulses of a preset scale are applied to each input variable according to a pre-fixed variable order. The target section water level is set as the response variable. The directional impulse response, response peak value and response decay process of each input variable pointing to the target section water level are obtained.
3. The river hydrological response simulation method according to claim 1, characterized in that, Determining the order of the input variables in step S2 specifically includes: Based on the peak values of the directional impulse responses corresponding to each input variable, determine the relative response amplitude and adjacent response changes for each response order; When the relative response amplitudes and adjacent response changes of multiple consecutive response steps satisfy the preset response amplitude condition and preset response change condition, the corresponding response step is determined as a response stability candidate step; when no response stability candidate step is formed within the preset investigation range, the response stability candidate step is determined based on the tail response state of multiple consecutive response steps. Based on the evaluation results of the Akaike information content criterion for each candidate order, the candidate orders that fall within the preset information loss tolerance range of the optimal information criterion value are determined. For each input variable, the corresponding input order is determined from the candidate orders that simultaneously satisfy the response stability condition and the information loss tolerance condition, and the input order is associated with the physical properties and sampling time range of the corresponding input variable.
4. The river hydrological response simulation method according to claim 3, characterized in that, Step S2, which involves forming training and testing samples, specifically includes: Based on the input order corresponding to each input variable, historical windows before the target time are extracted from the hydrological sequence with the original physical meaning on the same time axis. Historical windows of different lengths are aligned with the same target time as the right boundary, and a uniform input length is set according to the longest historical window. Valid location identifiers are used to distinguish the actual hydrological data location and the supplementary location in each historical window, so that different input variables have a consistent relative time meaning at the same input location; The response weights are determined according to the amplitude distribution of the directional impulse response corresponding to each input variable at its effective historical position, so that each response weight is consistent with the corresponding input variable and historical position, and the fill-in position does not form a response weighting feature. The training period is divided into a model fitting period and a validation period according to the time sequence. The normalization parameter is determined based on the hydrological data corresponding to the effective locations within the model fitting period. The same normalization parameter is used to process the hydrological data of the model fitting period, validation period, testing period and scenario simulation stage. The effective location identifier, response weight, and normalized hydrological data are used together to form the model input. The measured water level corresponding to the target time is set as the supervision target. Model fitting samples, validation samples, and test samples are formed respectively, and the corresponding input processing configurations are recorded.
5. The river hydrological response simulation method according to claim 1, characterized in that, Step S3 specifically includes: A TCN-GRU model is constructed. The TCN is used to perform dilated causal convolution processing on the model input formed in step S2. The hydrological time series features at different time scales are passed through residual connections, and the receptive field of the TCN covers the longest historical window determined in step S2. The hydrological time-series features output by TCN are input into GRU in chronological order. The hydrological state of the target river section is updated using the reset gate and update gate. The hydrological state corresponding to the last valid time position is input into the water level output layer to obtain the simulated water level of the target section. The training period is divided into model fitting data and validation data in chronological order. The TCN-GRU model parameters are updated according to the error between the model output water level and the corresponding measured water level. The model structure is determined, the model parameters are saved, and the model training is terminated according to the validation loss corresponding to the validation data. After the model training is completed, the TCN structure parameters, GRU structure parameters, water level output parameters, and trained model weights are locked for use with test data and scenario data.
6. The river hydrological response simulation method according to claim 1, characterized in that, Step S4 specifically includes: The training samples are divided into model fitting samples and validation samples according to the time sequence. Using the same model fitting samples, validation samples, test samples, input variables, target time and normalization parameters, a single GRU model, a single TCN model and a TCN-GRU model are established respectively. Each model is trained using model fitting samples, and the model structure and training parameters of each model are selected using validation samples. After the model parameters are fixed, the same test samples are used to generate the target cross-section simulated water level for each model. Based on the simulated water level at the target cross section of each model and the measured water level at the same time, the Nash efficiency coefficient, root mean square error, mean absolute error, peak time error, and Kling-Gupta efficiency coefficient are determined respectively. Based on the good fit of the Nash efficiency coefficient and the Kling-Gupta efficiency coefficient, and the water level simulation error represented by the root mean square error, mean absolute error and peak time error, the TCN-GRU model is determined to be the scenario response model. The input order, effective location identifier, response weight, normalization parameter, network structure parameter and trained model weight corresponding to each input variable are associated to form a river hydrological response model package, and the same model version input processing configuration and model parameters are called in the testing phase and the scenario simulation phase.
7. The river hydrological response simulation method according to claim 1, characterized in that, Step S5 specifically includes: Establish a baseline hydrological driving scenario, an upward-adjusted hydrological driving scenario, and a downward-adjusted hydrological driving scenario, so that each scenario adopts the same scenario start time, simulation end conditions, non-adjusted hydrological boundary, and initial water level of the target section; Each scenario is set up with an independent historical water level cache. The historical water level window is initialized with the measured water level before the scenario starts. The scenario input is formed based on the hydrological driving data, non-adjusted hydrological boundary data and historical water level cache of the corresponding scenario. The scene response model is called in the order of sampling time to generate the simulated water level at the current target time. The simulated water level at the current target time is written into the historical water level cache of the corresponding scene, and the updated historical water level cache is used as the scene input for the next target time, until the scene simulation ends. Compare the water level response processes of the adjusted hydrological-driven scenario and the baseline hydrological-driven scenario to determine the maximum water level change, response start time, peak response time, and response duration. The recovery condition is defined as the water level response returning to within the preset water level response threshold and continuously maintaining it for the preset sampling period. If the recovery condition is not yet met, the scenario simulation time is extended. Each hydrological driving scenario is evaluated based on the target control water level, allowable water level change, and allowable response duration. Hydrological driving scenarios that meet the scheduling constraints are identified as candidate hydrological driving control schemes, and scheduling auxiliary signals are generated. When the water level response exceeds any scheduling constraint, an over-limit warning signal is generated. The system receives confirmation information for candidate hydrological-driven control schemes, generates hydrological-driven control instructions based on the confirmation information, and sends the hydrological-driven control instructions to the reservoir gate control system to control the reservoir gate actuators to adjust the gate opening.
8. A river hydrological response simulation system integrating vector autoregression and deep learning, characterized in that, It includes a data analysis module, an input construction module, a model calculation module, a model verification module, and a scene simulation module, among which: The data analysis module is used to acquire external hydrological driving data and historical water level data of the target section, form a hydrological sequence with an equal time axis and divide it into training and testing periods; the training sequence is stationary, and vector autoregression analysis is performed based on the stationary representation to obtain the evaluation results of the candidate order information criterion and the directional impulse response of the input variables to the water level of the target section; The input construction module is used to determine the order of input variables based on preset response stability conditions, information criterion tolerance, information criterion evaluation results and directional impulse response, extract historical windows of different lengths from the hydrological sequence of equal time axis, align them according to the target time and mark the effective positions, set response weights based on the amplitude of directional impulse response, determine normalization parameters based on hydrological data of the training period, and form training samples, test samples and input processing configuration. The model computation module is used to train a deep learning model that integrates temporal feature extraction and hydrological state recursion using training samples, and generates simulated water levels at target cross sections from the integrated deep learning model. The model verification module is used to establish a benchmark model using the same training and test samples and to evaluate the benchmark model and the fused deep learning model. It determines the water level evaluation index based on the simulated water level and the measured water level, determines the scene response model according to the preset model selection rules, and encapsulates the scene response model and input processing configuration into a river hydrological response model package. The scenario simulation module is used to set the baseline hydrological driving data and adjust the hydrological driving data, keeping the hydrological boundaries consistent except for the input variables to be adjusted. It initializes the historical water level window with the measured water level before the scenario starts, forms the scenario input according to the input processing configuration in the river hydrological response model package, calls the scenario response model to generate simulated water levels at each time step and backfills the historical water level window of the corresponding scenario, forming the river section water level response process, determines candidate hydrological driving control schemes based on the river section water level response process and scheduling constraints, generates scheduling auxiliary signals or over-limit warning signals, and generates hydrological driving control instructions after receiving scheme confirmation information.
9. An electronic device, characterized in that, It includes a processor, a memory, a communication interface, and a communication bus, with the processor, memory, and communication interface connected via the communication bus; the memory is used to store computer programs as well as model parameters and hydrological data required for river hydrological response simulation, and the processor is used to read and execute the computer programs to implement the river hydrological response simulation method according to any one of claims 1 to 7; The communication interface is used to receive hydrological monitoring data and scheduling constraint information, and output river hydrological response results, scheduling auxiliary signals, over-limit early warning signals or hydrological driving control commands.
10. A computer-readable storage medium, characterized in that, A computer-readable storage medium stores a computer program that, when executed by a processor, implements the river hydrological response simulation method according to any one of claims 1 to 7.
Citation Information
Patent Citations
Reservoir downstream water level prediction method based on deep learning model
CN112001556A
Deep learning method and device for evaluating watershed middle and downstream water level flow change
CN120068664A
