Reservoir flood forecasting method based on physical constraints and space-time double-flow coupling

Abstract

The application discloses a reservoir flood forecasting method based on physical constraints and space-time double-flow coupling, comprising the following steps: S1, obtaining multi-source hydro-meteorological data; S2, decomposing the multi-source hydro-meteorological data into a historical state sequence and a future driving sequence; S3, extracting features of the historical state sequence data through a physical enhancement long short-term memory network of a historical inertia feature extraction branch, obtaining historical inertia features, and extracting features of the future driving sequence through a time domain convolution network of a future forced feature extraction branch, obtaining future forced features; S4, generating fusion features by weighting and fusing the historical inertia features and the future forced features; and S5, inputting the fusion features into a decoder to obtain a predicted water level increment, and superimposing the predicted water level increment to a current water level to obtain a prediction value. The application improves the timeliness of prediction, improves the prediction accuracy in the recession stage, and enhances the physical consistency of the prediction result.

Description

Technical Field

[0001] This invention relates to the field of water conservancy project safety monitoring and hydrological and water resources management technology, specifically to a method for high-precision real-time rolling forecasting of reservoir water level evolution by combining physical mechanism constraints and deep learning spatiotemporal modeling technology, and utilizing historical hydrological monitoring data and forecast meteorological scheduling data. Background Technology

[0002] As a crucial hub for flood control, water supply, and power generation in a river basin, accurate forecasting of reservoir water level changes is a prerequisite for making scientific scheduling decisions. With the intensification of global climate change, the suddenness and intensity of extreme rainfall events have increased significantly, placing extremely high demands on the response speed and forecast accuracy of reservoir flood control scheduling systems. Especially during the flood season, the evolution of reservoir water levels is influenced by complex nonlinear factors such as upstream rainfall runoff, inter-regional inflow, and artificial flood discharge scheduling. Achieving real-time water level forecasts with long lead times, high accuracy, and conformity to hydraulic laws remains a pressing challenge in the field of hydrological forecasting.

[0003] Existing reservoir flood forecasting methods are mainly divided into two categories: physical causal models and data-driven models. Traditional physical causal models, such as the Xin'anjiang model and the topographic index-based semi-distributed hydrological model (TOPMODEL), generalize the runoff generation and confluence processes based on hydrological mechanisms, with clear physical meaning. However, these models typically contain a large number of empirical parameters that need to be calibrated, are sensitive to changes in the underlying surface conditions of the watershed, and have high computational complexity, often failing to meet the demands of modern flood control scheduling for second-level real-time response. In recent years, data-driven models, represented by Long Short-Term Memory (LSTM) networks, have been widely used in hydrological time series forecasting due to their powerful nonlinear fitting capabilities. These models can automatically learn the patterns of water level changes from massive amounts of historical monitoring data, offering advantages such as fast computation speed and strong generalization ability.

[0004] However, existing data-driven forecasting models still have significant technical limitations in practical engineering applications. First, traditional models often adopt a "purely history-driven" approach, using only past rainfall and water level data to extrapolate the future, ignoring future rainfall information provided by numerical weather prediction (NWP) and the established reservoir scheduling plans. This often results in significant phase lag during the flood rise phase, failing to allow sufficient time for flood control and disaster relief operations. Second, general-purpose deep learning models are essentially "black box" models, lacking constraints from hydrophysical mechanisms. Model training focuses solely on minimizing fitting errors, easily leading to predictions that violate basic physical conservation laws. For example, they may predict spurious increases in water levels even without effective rainfall or inflow, severely impacting the confidence of dispatchers in the model's results. Furthermore, the gating mechanism of standard LSTM units, when handling long-duration flood events, suffers from rapid decay of the memory of high water storage potential over time. This results in simulations of the receding process often falling short of measured values, failing to accurately reflect the massive inertial characteristics of large reservoirs. Therefore, there is an urgent need to develop a new reservoir flood forecasting method that can effectively integrate future multi-source forcing information and has physical interpretability and constraint mechanisms.

[0005] The above content is only used to help understand the technical solution of the present invention and does not represent an admission that the above content is prior art. Summary of the Invention

[0006] The purpose of this invention is to provide a reservoir flood forecasting method based on physical constraints and spatiotemporal dual-flow coupling, aiming to overcome the technical defects of existing data-driven models in sudden flood forecasting, such as response lag, lack of physical mechanisms, and long-duration memory decay, and to achieve high-precision, long-term, real-time rolling forecasts of reservoir water level evolution that conform to hydraulic laws.

[0007] To achieve the above objectives, the technical solution of the present invention is as follows: The reservoir flood forecasting method based on physical constraints and spatiotemporal dual-flow coupling includes the following steps: S1. Obtain multi-source hydrological and meteorological data for the reservoir basin; S2. Decompose multi-source hydrological and meteorological data into historical state sequences and future driving sequences; S3. Through the historical inertia feature extraction branch, feature extraction is performed on the historical state sequence data to obtain historical inertia features. By using the future-forced feature extraction branch, the future-driven sequence is... Perform feature extraction to obtain future forced features. ; S4. Historical inertia characteristics and future compulsion characteristics fusion features are generated through weighted fusion. ; S5. Merge features Input to the decoder to obtain the predicted water level increment. The predicted value is obtained by adding it to the current water level. The formula is as follows: ,in for The water level in front of the dam is constantly monitored. This is the starting time of the current prediction. In step S3, the historical inertial feature extraction branch uses a Physically Augmented Long Short-Term Memory (PE-LSTM) network, while the future forced feature extraction branch uses a Temporal Convolutional Network (TCN).

[0008] Preferably, the input gate of the physically enhanced long short-term memory network is as follows:

[0009]

[0010]

[0011] In the formula, This is the active state of the input gate after physical enhancement. For the Sigmoid activation function, and Here is the weight matrix of the input gate. for The historical state of the input vector at each time step is a multi-source input vector. For the previous moment The hidden layer state vector, For bias terms, Represents the Hadama product; For the index of moments within a historical time sliding window, The current prediction start time, Indicates the first The average rainfall over the watershed at any given time. Indicates time The flow gain factor and These are the learnable weight matrix and bias term in a neural network, respectively. The length of the historical observation window. It serves as a proxy variable for soil saturation in the watershed.

[0012] Furthermore, during the model initialization phase, the forget gate bias parameters of the physical enhancement long short-term memory network are... Set as an interval Any constant within.

[0013] Furthermore, the model is trained under constraints using a hybrid loss function, which includes a water balance physical regularization loss value. The formula is as follows:

[0014]

[0015]

[0016] in, The batch size defined for each network training session; This is the sample index in the current batch. The current prediction start time, To predict the length of the window in the future, Index of moments for future forecast windows express Numerical weather forecast rainfall at any given time. express The reservoir's outflow is scheduled at any given time. Net flux, for Predicted water level at any given time for The water level in front of the dam is observed at all times. This represents the predicted total change in water level.

[0017] The hybrid loss function The following formula:

[0018] In the formula, For batch dimensions; For sample index; and The model is for the first The predicted water level and the actual observed water level output for each sample; This is the weighting balance coefficient.

[0019] This invention also provides a reservoir flood forecasting system based on physical constraints and spatiotemporal dual-flow coupling, including a data acquisition module, a data decomposition module, a dual-flow heterogeneous feature extraction module, a feature fusion module, a decoding module, and a prediction calculation module. The data acquisition module is used to acquire multi-source hydrological and meteorological data of the reservoir basin. The data decomposition module is used to decompose multi-source hydrological and meteorological data into historical state sequences and future driving sequences. The dual-stream heterogeneous feature extraction module includes a parallel historical inertia feature extraction branch and a future forced feature extraction branch. The historical inertia feature extraction branch is used to extract features from historical state sequence data to obtain historical inertia features. The future-driven feature extraction branch is used for future-driven sequences. Perform feature extraction to obtain future forced features. The historical inertial feature extraction branch employs a Physically Augmented Long Short-Term Memory (PE-LSTM) network, while the future forced feature extraction branch employs a Temporal Convolutional Network (TCN). The feature fusion module is used to dynamically calculate the contribution weights of the two branches and generate fused features through weighted fusion. , The decoding module is used to process the fused features. Decode the data and output the water level increment for the current prediction step. , The prediction calculation module calculates the water level increment based on the current prediction step. Calculate the predicted water level .

[0020] Preferably, the physically enhanced long short-term memory network modifies the calculation formula for the standard input gate as follows:

[0021]

[0022]

[0023] Preferably, during the model initialization phase, the forget gate bias parameters of the physically enhanced Long Short-Term Memory network are adjusted. Set as positive range Any constant value in it.

[0024] Preferably, the system further includes a hybrid loss function for constrained training of the model, the hybrid loss function including a water balance physical regularization loss value. The formula is as follows:

[0025]

[0026]

[0027] In addition, the present invention also provides a reservoir flood forecasting device based on physical constraints and spatiotemporal dual-flow coupling, comprising: a memory, a processor, and a reservoir flood forecasting program based on physical constraints and spatiotemporal dual-flow coupling stored in the memory and executable on the processor, wherein the reservoir flood forecasting program based on physical constraints and spatiotemporal dual-flow coupling is configured with a method for implementing reservoir flood forecasting based on physical constraints and spatiotemporal dual-flow coupling.

[0028] The present invention also provides a storage medium storing a reservoir flood forecasting program based on physical constraints and spatiotemporal dual-flow coupling. When executed, the reservoir flood forecasting program based on physical constraints and spatiotemporal dual-flow coupling implements a reservoir flood forecasting method based on physical constraints and spatiotemporal dual-flow coupling.

[0029] By adopting the above technical solution, the present invention has the following beneficial technical effects: 1. This invention alleviates the phase lag problem in flood forecasting by introducing future forcing information. It utilizes a temporal convolutional network (TCN) branch to process numerical weather prediction (NWP) and reservoir scheduling plans, enabling the model to perceive rainfall and outflow trends within the forecast period. Compared to traditional machine learning models that rely solely on historical data, this method can respond to the flood rise process earlier, improving the timeliness of water level trend prediction within the forecast period.

[0030] 2. This invention improves the model's ability to simulate long-duration hydrological processes and enhances the prediction accuracy during the recession phase. By introducing soil saturation feedback and hydraulic damping bias through PE-LSTM units, the inertia of reservoir storage and the nonlinear runoff characteristics are simulated. This mechanism effectively slows down the model's forgetting of previous high-water levels during long-sequence calculations, thus maintaining a good fit during the recession phase of long-duration floods and correcting the problem of traditional models predicting low values.

[0031] 3. This invention enhances the physical consistency of forecast results and improves the engineering practical value of the model. By introducing water balance regularization constraints into the loss function, the optimization direction of the model's parameters is guided to conform to the physical conservation laws. This mechanism effectively suppresses non-physical water level fluctuations (such as sudden rises without cause) caused by data noise or overfitting, improving the interpretability and stability of the forecast results. Attached Figure Description

[0032] Figure 1 This is a flowchart of the present invention.

[0033] Figure 2 This is a system structure diagram of the present invention.

[0034] Figures 3(a)-3(d) show the water level process prediction results of the present invention in four typical flood events in the Meishan Reservoir (a large reservoir in Anhui Province). Detailed Implementation

[0035] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0036] Example 1: As Figure 1 As shown, the present invention provides a reservoir flood forecasting method based on physical constraints and spatiotemporal dual-flow coupling, comprising the following steps: S1. Obtain multi-source hydrological and meteorological data for the reservoir basin.

[0037] Specifically, time-series data of the target watershed can be obtained through an automatic water and rainfall monitoring system and a meteorological forecasting interface. This includes, for example, four types of historical state variables: average rainfall across the watershed. Reservoir inflow Outbound flow and the water level in front of the dam ; and two types of future driving variables: numerical weather prediction (NWP) rainfall. and reservoir scheduling plan outflow .

[0038] Preferably, if the sampling frequency of the original data is inconsistent (e.g., rainfall is 1 hour and water level is 5 minutes), linear interpolation can be used to resample all data to a uniform time resolution of 1 hour.

[0039] Preferably, to eliminate the impact of differences in physical dimensions on the convergence speed of the model's gradient descent (e.g., water level values ​​are on the order of hundreds of meters, while rainfall values ​​are on the order of millimeters), a min-max normalization method is used to map all feature variables to... The dimensionless interval. The normalization transformation formula is as follows:

[0040] In the formula, Indicates the first The feature in the first The original observations at each time point, The normalized value. Indicates the first The minimum value among all original observations of each feature. Indicates the first The maximum value among all original observations for each feature. It is important to note that after the model outputs the prediction results, the inverse transformation of this formula must be used to restore the predicted values ​​to the actual physical quantities.

[0041] S2. Decompose multi-source hydrological and meteorological data into historical state sequences reflecting the water storage status of the basin and future driving sequences reflecting external meteorological scheduling.

[0042] Specifically, the sliding window technique can be used to construct the time series samples required for the model, and then the time series samples can be decoupled into two sequences—the historical state sequence. With future driving sequence .

[0043] This embodiment provides a detailed process as follows (other methods and processes can also be used by those skilled in the art): Setting the length of the historical observation window. This embodiment sets Hours, set the length of the future forecast window This embodiment sets Hours. For any given time. The corresponding time segments are extracted from the normalized full dataset as time-series samples. If historical data is missing, the samples at that time point are removed to ensure the completeness of the training data.

[0044] The extracted time segments are then decoupled into two independent sequences, which can be represented using a feature matrix. For example, a historical state sequence matrix can be defined. , dimension This embodiment is Characterizing the current water storage potential energy state of the reservoir and the inertial effect generated by previous runoff:

[0045] Define the future driving sequence matrix , dimension This embodiment is Characterizes the external forcing forces applied to the reservoir system during the foreseeable future period:

[0046] S3. The historical state sequence and the future driving sequence are feature-encoded by a dual-stream heterogeneous feature extraction network, wherein the dual-stream heterogeneous feature extraction network includes a parallel historical inertial feature extraction branch and a future forced feature extraction branch.

[0047] The historical inertia feature extraction branch employs a Physically Enhanced Long Short-Term Memory (PE-LSTM) network to extract features from historical state sequence data, thereby obtaining historical inertia features. The number of hidden units in the network is set to 64. This value is a standard hyperparameter setting in this embodiment, and those skilled in the art can adaptively adjust it according to the actual data scale.

[0048] To address the issues of standard LSTMs easily forgetting high-water level characteristics and lacking flow generation mechanisms, the internal structure of LSTMs was physically modified.

[0049] This includes: (a) constructing a soil saturation feedback mechanism. The cumulative rainfall within a historical observation window (24 hours in this example) is calculated and used as a proxy variable for watershed soil saturation. The formula is as follows:

[0050] in For the index of moments within a historical time sliding window, The current prediction start time, Indicates the first The average rainfall over the watershed at any given time.

[0051] The current generation gain factor was constructed using the hyperbolic tangent function. The formula is as follows:

[0052] In the formula, Indicates time The flow gain factor and These are the learnable weight matrix and bias term in a neural network, respectively. This is the starting time for the current prediction.

[0053] Subsequently, this factor is applied to the input gate of a standard LSTM. The calculation formula for the standard input gate is modified to be the input gate for the historical inertia feature extraction branch of this invention. :

[0054] In the formula, The input gate is in an active state after physical enhancement; Use the Sigmoid activation function; and This is the weight matrix of the input gate; for The historical state of the input vector at any given time is a multi-source input vector. For the previous moment The hidden layer state vector; For bias terms; This represents the Hadamard product (i.e., matrix element-wise multiplication).

[0055] This mechanism enables the input gate to dynamically increase its openness to current rainfall information as cumulative rainfall increases, thus simulating nonlinear runoff generation.

[0056] (b) Initialize the hydraulic damping bias. In order to simulate the long-term memory characteristics of a large reservoir during the water receding process, a special initialization strategy is implemented for the forget gate of PE-LSTM.

[0057] During the model initialization phase, the forget gate bias parameters are explicitly set. Set as positive range A certain constant within, preferably in this embodiment. .

[0058] The formula for the forget gate activation value in standard LSTM is:

[0059] In the formula: and Here is the weight matrix for the forget gate. To offset the forget gate, This is the Sigmoid activation function.

[0060] when At that time, the initial output of the Sigmoid function is close to This means that, without any input, the model retains 82% of its previous state by default. This highly damped initialization effectively prevents gradient vanishing and ensures long-term memory of the flood peak level.

[0061] After iterative calculations using the aforementioned internal mechanism, the PE-LSTM branch reaches its final time step within the historical window (i.e., ...). At time (time), output the final hidden state vector. This invention defines it as a historical inertia feature. ,Right now:

[0062] This feature vector encapsulates the water storage inertia that the basin has evolved over the past 24 hours.

[0063] The future-driven feature extraction branch employs a temporal convolutional network (TCN) for future-driven sequences. Perform feature extraction to obtain future forced features. .

[0064] The specific feature extraction steps of temporal convolutional networks are common knowledge, and this invention only provides a brief introduction. Those skilled in the art can set the specific parameters according to actual conditions, and the invention is not limited to the description in this embodiment. Construct a one-dimensional causal convolutional layer (Causal Conv1d), set the kernel size to 3, and the number of output channels to 32.

[0065] To ensure causality (i.e., the time within the foreseeability period) The convolution output does not depend on (The following information) performs left zero padding on the input sequence, with a padding length of... For any time within the forecast period (in Convolution operation is expressed as:

[0066] In the formula, The time after extraction through the convolutional layer Feature map vectors; This is the local temporal index within the convolution kernel (values ​​0, 1, 2). The kernel weight matrix; For a moment The future is driven by multi-source input; is the convolution bias; ReLU is the linear rectified activation function.

[0067] Subsequently, an adaptive max pooling layer is used to extract the extreme values ​​from the variable-length convolutional feature maps along the time dimension, and these values ​​are aggregated into a vector of fixed dimensions. Its operation process can be represented as:

[0068] Mathematically, this vector extracts the maximum response value of each feature component within the future window. Physically, it successfully captures the maximum rainfall intensity and maximum flood discharge within the forecast period, which is the future forced feature.

[0069] S4. Feature Fusion: Based on the current watershed hydraulic response state, the contribution weights of the two branches are dynamically calculated, and a fused feature is generated through weighted fusion. .include: S41. Feature splicing and gating coefficient calculation.

[0070] The historical inertia feature extraction branch outputs the historical inertia features. Future forced features of TCN output The joint feature vector is obtained by concatenating the features along the channel dimension. This joint feature vector is then mapped to a scalar using a fully connected (dense) layer and passed through a sigmoid activation function to obtain the gating coefficients. :

[0071] In the formula, To integrate the gating coefficients, their range is: ; This represents the concatenation operation of feature vectors. and These are the learnable weight matrix and bias term of the adaptive weighted network, respectively.

[0072] S42. Weighted fusion.

[0073] To unify the dimensions, the forced features are first passed through a linear layer. Mapping to 64 dimensions yields the enhanced forced features. Then, a weighted summation is performed to generate the fused features. :

[0074] S5. Decode and calculate the predicted values. Fuse features. Input to a fully connected decoder, output the water level increment at the current prediction step. The predicted value is obtained by adding it to the current water level. The formula is expressed as:

[0075] In the formula, for The water level in front of the dam is constantly monitored. This represents the increment of water level change output by the decoder.

[0076] The above only provides the predicted value at time t+1. An auto-regressive rolling strategy can be used to generate a prediction covering the entire forecast period. The continuous predicted water level process line is obtained. The specific iterative process is as follows: (1) Initial time inference: At the current time , will be built in real time and Input the network and output the predicted water level at the next time step. .

[0077] (2) State update: In order to perform multi-step prediction, As "pseudo-observations", they are filled into the historical state sequence. At the end, remove the earliest one. Time-based data, maintaining a constant time window length; simultaneously, future driving sequences. Swipe back one step to read to Real-time weather and dispatch data.

[0078] (3) Iterative Loop: Utilize the updated historical state sequence and future driving sequence for the next round of reasoning, executing steps S3-S5. This iterative loop continues, generating... Connecting these consecutive predicted values ​​yields a water level evolution process line that conforms to physical constraints and covers the entire forecast period, which can be used for flood control scheduling decision support.

[0079] Preferably, this embodiment also constructs a hybrid loss function that includes a physical consistency constraint term, and uses historical data to train the model under constraints. Specifically, this includes: (a) Construction of water balance physical regularization loss values: First, calculate the total water level change predicted by the model during the forecast period. Its formula is:

[0080] Simultaneously, based on the predicted NWP rainfall and planned outflows during the forecast period, the net flux proxy index is calculated. :

[0081] According to the generalized water balance principle, the trend (sign) of water level change should be consistent with the direction of net flow. A penalty function is constructed that incurs a loss only when the signs of the two changes are opposite. The penalty function is defined as follows:

[0082] In the formula, This represents the physical regularization loss value for the water balance of this batch. The batch size defined for each network training session; This refers to the sample index in the current batch; Ensure that when the water level changes With net flux When the signs are the same (the product is positive, and becomes negative after substituting a negative sign), the function output 0 does not incur a penalty; a positive gradient penalty is only incurred when the directions are opposite.

[0083] (b) Construction of the hybrid loss function: Define the total loss function The weighted sum of the data-driven mean squared error (MSE) and the water balance physical regularization loss:

[0084] In the formula, Let this be the total loss function for network training iterations; Batch Size; For sample index; and The model is for the first The predicted water level and the actual observed water level output for each sample; The weighting balance coefficient for the physical regularization term (preferred in this embodiment) This is used to adjust the ratio of data-driven error to physical mechanism constraints.

[0085] In this embodiment, the physical weighting coefficient The initial learning rate is set to 0.2. The Adam optimizer is used for end-to-end parameter iteration on the two-stream heterogeneous network, with the initial learning rate set to... Train until the loss function converges to obtain the optimized model parameters. In parameters Once convergence is achieved, the real-time forecasting application phase can begin.

[0086] To verify the practical application effect of the reservoir flood forecasting method based on physical constraints and spatiotemporal dual-flow coupling proposed in this invention, this embodiment selects four typical floods from the Meishan Reservoir (a large (1) type) in Anhui Province during the flood season of 2024 to 2025 for retrospective testing, and compares the simulation accuracy of the method of this invention with that of the Xinanjiang Model, a physical cause model commonly used in the water conservancy industry.

[0087] The specific results are shown in Table 1, with Nash efficiency coefficient (NSE) and mean absolute error (MAE) used as the main evaluation indicators. The accuracy data for the Xin'anjiang model is taken from the average level of the reservoir's historical forecast system (NSE is typically between 0.80 and 0.85). The test results of the method of this invention are based on four typical flood events shown in Figures 3(a)-3(d). Figures 3(a)-3(d) show the prediction results of a single prediction using the method of this invention in the computer system. The data in Table 1 are the mean (Nash efficiency coefficient) and range (mean absolute error) of 10 prediction results.

[0088] In the figures, the solid black line represents the actual observed water level in front of the dam, the dashed red line represents the rolling predicted water level of the model of this invention, and the blue bar chart at the top represents the rainfall process. The text box displays the Nash efficiency coefficient (NSE) and mean absolute error (MAE) of the prediction for this event, and the red box indicates the actual observed peak water level. As shown in Figures 3(a)-3(d), under different rainfall conditions and flood initiation conditions, the predicted water level curve output by this invention closely matches the measured water level curve, with no obvious phase lag, proving that the method of this invention has extremely high forecast accuracy and timeliness.

[0089] Table 1

[0090] Example 2: Additionally, this invention also provides a reservoir flood forecasting system based on physical constraints and spatiotemporal dual-flow coupling, such as... Figure 2 As shown, it includes a data acquisition module, a data decomposition module, a dual-stream heterogeneous feature extraction module, a feature fusion module, a decoding module, and a prediction calculation module.

[0091] The data acquisition module is used to acquire multi-source hydrological and meteorological data of the reservoir basin.

[0092] The data decomposition module is used to decompose multi-source hydrological and meteorological data into a historical state sequence reflecting the water storage status of the basin and a future driving sequence reflecting external meteorological scheduling.

[0093] The dual-stream heterogeneous feature extraction module includes a parallel historical inertia feature extraction branch and a future forced feature extraction branch. The historical inertia feature extraction branch is used to extract features from historical state sequence data to obtain historical inertia features. The future-driven feature extraction branch is used for future-driven sequences. Perform feature extraction to obtain future forced features. .

[0094] The historical inertial feature extraction branch employs a Physically Augmented Long Short-Term Memory (PE-LSTM) network, while the future forced feature extraction branch employs a Temporal Convolutional Network (TCN).

[0095] The physically enhanced long short-term memory network modifies the calculation formula for the standard input gate as follows:

[0096] Forget gate bias parameters Set as positive range Any value in, preferably .

[0097] The feature fusion module is used to dynamically calculate the contribution weights of the two branches and generate fused features through weighted fusion. .

[0098] The decoding module is used to process the fused features. Decode the data and output the water level increment for the current prediction step. .

[0099] The prediction calculation module calculates the water level increment based on the current prediction step. Calculate the predicted water level .

[0100] The system also includes a hybrid loss function for constrained training of the model, the hybrid loss function including the water balance physical regularization loss value. The formula is as follows:

[0101]

[0102]

[0103] In the formula, This represents the physical regularization loss value for the water balance of this batch. The batch size defined for each network training session; This refers to the sample index in the current batch; Ensure that when the water level changes With net flux When the signs are the same (the product is positive, and becomes negative after substituting a negative sign), the function outputs 0 and no penalty is generated. A positive gradient penalty is only generated when the directions are opposite.

[0104] Furthermore, this embodiment is merely a basic description of the reservoir flood forecasting system based on physical constraints and spatiotemporal dual-flow coupling of the present invention. Technical details not described in detail in this embodiment can be found in the methods provided in any embodiment of the present invention, and will not be repeated here.

[0105] Example 3: Those skilled in the art will clearly understand that the systems and methods of the above embodiments can be implemented using software plus necessary general-purpose hardware platforms. Of course, they can also be implemented using hardware, but in many cases, the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as read-only memory (ROM) / RAM, magnetic disk, optical disk), and includes several instructions to cause a terminal device (which may be a mobile phone, computer, node packaging device, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0106] Therefore, the present invention also provides a reservoir flood forecasting device based on physical constraints and spatiotemporal dual-flow coupling, comprising: a memory, a processor, and a reservoir flood forecasting program based on physical constraints and spatiotemporal dual-flow coupling stored in the memory and executable on the processor, wherein the reservoir flood forecasting program based on physical constraints and spatiotemporal dual-flow coupling is configured with a method for implementing reservoir flood forecasting based on physical constraints and spatiotemporal dual-flow coupling.

[0107] In addition, the present invention also provides a storage medium storing a reservoir flood forecasting program based on physical constraints and spatiotemporal dual-flow coupling.

[0108] In reality, when deploying equipment or programs, a program may execute all steps or only one step, and all steps may be achieved through the cooperation of multiple programs. Therefore, the reservoir flood forecasting program based on physical constraints and spatiotemporal dual-flow coupling, when executed, realizes all or a certain process of the reservoir flood forecasting method based on physical constraints and spatiotemporal dual-flow coupling.

[0109] The above are merely preferred embodiments of the present invention and do not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention specification, or direct or indirect applications in other related technical fields, are included within the patent protection scope of the present invention.

Claims

1. A reservoir flood forecasting method based on physical constraints and spatiotemporal dual-flow coupling, characterized in that, Includes the following steps: S1. Obtain multi-source hydrological and meteorological data for the reservoir basin, including historical state variables: average rainfall over the basin area. Reservoir inflow Outbound flow and the water level in front of the dam ; as well as Future driving variables: Numerical weather forecast rainfall and reservoir scheduling plan outflow ; S2. Decompose multi-source hydrological and meteorological data into historical state sequences and future driving sequences; S3. Through the historical inertia feature extraction branch, feature extraction is performed on the historical state sequence data to obtain historical inertia features. By using the future-forced feature extraction branch, the future-driven sequence is... Perform feature extraction to obtain future forced features. ; S4. Historical inertia characteristics and future compulsion characteristics fusion features are generated through weighted fusion. ; S5. Merge features Input to the decoder to obtain the predicted water level increment. The predicted value is then added to the current water level. ; In step S3, the historical inertial feature extraction branch uses a physically enhanced long short-term memory network, and the future forced feature extraction branch uses a temporal convolutional network. The input gate of the physically enhanced long short-term memory network is as follows: In the formula, This is the active state of the input gate after physical enhancement. For the Sigmoid activation function, and Here is the weight matrix of the input gate. for The historical state of the input vector at each time step is a multi-source input vector. For the previous moment The hidden layer state vector, For bias terms, Represents the Hadama product; For the index of moments within a historical time sliding window, The current prediction start time, The length of the historical observation window. Indicates the first The average rainfall over the watershed at any given time. Indicates time The flow gain factor and These are the learnable weight matrix and bias term in a neural network, respectively. It serves as a proxy variable for soil saturation in the watershed.

2. The reservoir flood forecasting method based on physical constraints and spatiotemporal dual-flow coupling as described in claim 1, characterized in that, During the model initialization phase, the forget gate bias parameters of the Long Short-Term Memory network are physically enhanced. Set as an interval Any constant within.

3. The reservoir flood forecasting method based on physical constraints and spatiotemporal dual-flow coupling as described in claim 1, characterized in that, The model is trained under constraints using a hybrid loss function, which includes a water balance physical regularization loss value. The formula is as follows: in, The batch size defined for each network training session. This is the sample index in the current batch. The current prediction start time, To predict the length of the window in the future, Index of moments for future forecast windows express Numerical weather forecast rainfall at any given time. express The reservoir's outflow is scheduled at any given time. Net flux, for Predicted water level at any given time for The water level in front of the dam is constantly monitored. This represents the predicted total change in water level.

4. The reservoir flood forecasting method based on physical constraints and spatiotemporal dual-flow coupling as described in claim 3, characterized in that, The hybrid loss function The following formula: In the formula, For batch dimensions; For sample index; and The model is for the first The predicted water level and the actual observed water level output for each sample; This is the weighting balance coefficient.

5. A reservoir flood forecasting system based on physical constraints and spatiotemporal dual-flow coupling, characterized in that, It includes a data acquisition module, a data decomposition module, a two-stream heterogeneous feature extraction module, a feature fusion module, a decoding module, and a prediction calculation module. The data acquisition module is used to acquire multi-source hydrological and meteorological data of the reservoir basin, including historical state variables such as basin-wide average rainfall. Reservoir inflow Outbound flow and the water level in front of the dam ; and future driving variables: numerical weather forecast rainfall. and reservoir scheduling plan outflow , The data decomposition module is used to decompose multi-source hydrological and meteorological data into historical state sequences and future driving sequences. The dual-stream heterogeneous feature extraction module includes a parallel historical inertia feature extraction branch and a future forced feature extraction branch. The historical inertia feature extraction branch is used to extract features from historical state sequence data to obtain historical inertia features. The future-driven feature extraction branch is used to extract features from the future-driven sequence to obtain future-driven features. The historical inertia feature extraction branch employs a physically enhanced long short-term memory network, while the future forced feature extraction branch employs a temporal convolutional network. The feature fusion module is used to dynamically calculate the contribution weights of the two branches and generate fused features through weighted fusion. , The decoding module is used to process the fused features. Decode the data and output the water level increment for the current prediction step. , The prediction calculation module calculates the water level increment based on the current prediction step. Calculate the predicted water level ; The physically enhanced long short-term memory network modifies the calculation formula for the standard input gate as follows: In the formula, This is the active state of the input gate after physical enhancement. For the Sigmoid activation function, and Here is the weight matrix of the input gate. for The historical state of the input vector at each time step is a multi-source input vector. For the previous moment The hidden layer state vector, For bias terms, Represents the Hadama product; For the index of moments within a historical time sliding window, The current prediction start time, The length of the historical observation window. Indicates the first The average rainfall over the watershed at any given time. Indicates time The flow gain factor and These are the learnable weight matrix and bias term in a neural network, respectively. It serves as a proxy variable for soil saturation in the watershed.

6. The reservoir flood forecasting system based on physical constraints and spatiotemporal dual-flow coupling as described in claim 5, characterized in that, The system also includes a hybrid loss function for constrained training of the model, the hybrid loss function including the water balance physical regularization loss value. The formula is as follows: in, The batch size defined for each network training session. This is the sample index in the current batch. The current prediction start time, To predict the length of the window in the future, Indexing moments for future forecast windows. express Numerical weather forecast rainfall at any given time. express The reservoir's outflow is scheduled at all times. Net flux, for Predicted water level at any given time for The water level in front of the dam is constantly monitored. This represents the predicted total change in water level.

7. A reservoir flood forecasting device based on physical constraints and spatiotemporal dual-flow coupling, characterized in that, include: The system includes a memory, a processor, and a reservoir flood forecasting program based on physical constraints and spatiotemporal dual-flow coupling, which is stored in the memory and can run on the processor. The reservoir flood forecasting program based on physical constraints and spatiotemporal dual-flow coupling is configured to implement the reservoir flood forecasting method based on physical constraints and spatiotemporal dual-flow coupling as described in any one of claims 1-4.

8. A storage medium, characterized in that, The storage medium stores a reservoir flood forecasting program based on physical constraints and spatiotemporal dual-flow coupling. When executed, the reservoir flood forecasting program based on physical constraints and spatiotemporal dual-flow coupling implements the reservoir flood forecasting method based on physical constraints and spatiotemporal dual-flow coupling as described in any one of claims 1-4.

Images