Intelligent flood forecasting method, device and readable storage medium based on FQI-DMTSNet adaptive generative adversarial mechanism

Abstract

This paper proposes an intelligent flood forecasting method, device, and readable storage medium based on the FQI-DMTSNet adaptive generative adversarial mechanism. This method fuses and assimilates preprocessed data from multiple sources, extracts temporal features using a dynamic layered multiscale time series perception network (DMTSNet), optimizes scheduling strategies using fit value iteration (FQI) reinforcement learning, and implements data enhancement and parameter migration in data-free areas using an adaptive feature-gated generative adversarial network (AFG-GAN), forming a closed-loop collaborative forecasting system. This method addresses the challenges of traditional models, such as their strong data dependence and poor adaptability to extreme scenarios, by enabling parameter generation in data-free areas, sample enhancement for extreme scenarios, and optimization of control strategies. This method improves the accuracy, adaptability, and practicality of flood forecasting, making it suitable for scenarios such as flood prediction, hydrological scheduling, and urban waterlogging warning.

Description

Technical Field

[0001] The present invention relates to the fields of hydrology and water resources management, intelligent early warning systems, artificial intelligence, and multi-source data fusion technology, and in particular to an intelligent flood forecasting method, device, and readable storage medium thereof based on an FQI-DMTSNet adaptive generative adversarial mechanism. Background Art

[0002] Current flood forecasting technologies primarily include emerging technologies such as physical mechanism models, statistical regression and time series models, machine learning and deep learning models, reinforcement learning, and adversarial neural networks. Physical models rely on large amounts of measured data and manual parameter adjustments, resulting in complex calculations and difficulty responding in real time. Statistical models struggle to handle nonlinear and non-stationary flood processes, and their predictive capabilities are limited under extreme weather conditions. Traditional machine learning and deep learning models lack constraints, are sensitive to data anomalies, and lack generalization capabilities. Emerging technologies such as reinforcement learning and GANs lack multi-model fusion and closed-loop optimization feedback mechanisms. As flood forecasting demands increase timeliness, accuracy, and the ability to handle complex scenarios, traditional "single model" and "static prediction" approaches are no longer sufficient, particularly in areas without data or under extreme weather conditions.

[0003] Therefore, there is an urgent need for an intelligent flood forecasting method that integrates multi-source data processing, dynamic decision optimization and generative adversarial enhancement capabilities to address the shortcomings of existing technologies. Summary of the Invention

[0004] The embodiments of the present invention provide an intelligent flood forecasting method, device and readable storage medium based on the FQI-DMTSNet adaptive generative adversarial mechanism, which addresses the common problems of current technologies such as strong dependence on high-quality historical data, complex parameter calibration, lack of dynamic feedback mechanism, poor adaptability to extreme scenarios and weak generalization ability in areas without data.

[0005] The core technology of this invention is to build a closed-loop collaborative intelligent flood forecasting system by integrating fitting value iteration (FQI) reinforcement learning, dynamic hierarchical multi-scale temporal perception network (DMTSNet) and adaptive feature gated generative adversarial network (AFG-GAN), combining multi-source data fusion, dynamic decision optimization and data generation mechanism.

[0006] In a first aspect, the present invention provides an intelligent flood forecasting method based on the FQI-DMTSNet adaptive generative adversarial mechanism, the method comprising the following steps:

[0007] Multi-source data fusion and assimilation: normalize the temporal and spatial resolutions of remote sensing, measured, and model data and perform Kalman filter correction to obtain pre-processed hydrological data;

[0008] Combined modeling based on FQI and DMTSNet: DMTSNet is used to extract multi-scale time series features of hydrological data, and the FQI reinforcement learning algorithm is used to construct a reward function that includes reservoir water level safety and downstream flood control safety to optimize flood scheduling strategies.

[0009] Data enhancement and parameter migration based on AFG-GAN: AFG-GAN performs dimension-level gated weighting on input features to achieve flood data enhancement and adaptive generation and migration of hydrological parameters in data-deficient areas;

[0010] Closed-loop collaborative forecasting: Collaboratively integrate multi-source data fusion results, DMTSNet feature extraction results, FQI scheduling strategies, and AFG-GAN data enhancement results to generate flood forecast results.

[0011] Furthermore, multi-source data fusion and assimilation specifically include:

[0012] The Kriging interpolation method is used to perform spatiotemporal interpolation of multi-source remote sensing precipitation data. By determining the weight coefficient of the data at each observation point to minimize the mean square error, eliminating null values ​​and aligning the spatiotemporal resolution, the precipitation data estimate at the location to be interpolated is obtained.

[0013] The Kalman filter is used to correct the deviation of the fused data, and the final estimated value is obtained by updating the state equation and observation equation. The state equation reflects the dynamic changes of the actual precipitation at different times, and the observation equation combines the observation value with the noise, and the weight of the predicted value and the observed value is weighed by the Kalman gain.

[0014] For example, the Kriging interpolation method is used to perform spatiotemporal interpolation on multi-source remote sensing precipitation data, remove null values ​​and align spatiotemporal resolution. The formula is:

[0015]

[0016] in is the estimated value of precipitation data at the location to be interpolated s; is the weight coefficient of the data of the i-th observation point, which is determined by minimizing the mean square error; is the precipitation value of the i-th valid observation point;

[0017] Kalman filtering is used to correct the deviation of the fusion data, and the state equation and the observation equation Update the calculation to get the final estimate ;

[0018] in, is the actual precipitation state variable at time t; is the actual precipitation state variable at time t-1; is the hydrological change rate at time t; is the Kalman gain; is the observed value at time t; is the observation noise; is the predicted value at time t, generated by the state equation.

[0019] Furthermore, the construction method of DMTSNet is:

[0020] The single dilated causal convolutional layer of the traditional temporal convolutional neural network is replaced by three parallel dilated causal convolutional layers with different dilation coefficients. The outputs of each layer are fused through a fully connected network, and the residual connection and parameterized ReLU activation function are combined to capture long-term temporal dependencies.

[0021] Furthermore, the reward function of the FQI reinforcement learning algorithm is designed as:

[0022] The reward function comprehensively considers reservoir water level safety and downstream flood control safety, balancing the relative importance of the two through weight coefficients. Reservoir water level safety is associated with the flood control limit water level and the reservoir water level status at the next moment, while downstream flood control safety is associated with the warning flow threshold and the downstream flow determined by the dispatching action and environmental status. For example:

[0023]

[0024] in Limiting water levels in reservoirs for flood control; is the reservoir water level status at the next moment t+1; is the system state at time t+1; is the scheduling action taken at time t; is the warning flow threshold of the downstream river; The current dispatch action and environmental status The determined measured or predicted downstream flow; is a non-negative weight constant used to balance the relative importance of upstream and downstream objectives.

[0025] Furthermore, the feature gating mechanism of AFG-GAN specifically includes:

[0026] Insert a feature gating module before the discriminator to dynamically weight each dimension of the input feature to obtain the gated input;

[0027] Quantify the feature importance based on the gradient contribution of each feature dimension to the discriminator output, and construct a gating learning objective to optimize the gating weight;

[0028] The regional characteristic regularization term is introduced to optimize the parameter migration of the data-free area by measuring the similarity between the characteristics of the target area and the source area.

[0029] For example: insert a feature gating module before the discriminator to input samples Each dimension of the feature is dynamically weighted, and the gating function is , get the gate input ; Where d is the dimension of the original input feature vector x; is the gated input vector, whose dimension is consistent with the original input feature vector x, and is the optimized representation after feature filtering;

[0030] Based on gradient contribution Constructing Gated Learning Objectives , quantifies the importance of features; where, is the gradient contribution value of the i-th dimension feature, which represents the sensitivity of the discriminator output to the feature. ; The input vector after gated discriminator The output value of is the i-th dimension eigenvalue of the original input eigenvector x; is the weight output of the gating function for the i-th dimension feature;

[0031] Introducing regional characteristic regularization terms , optimize the parameter migration in the area without data; among them, is the sample feature vector of the target area, ,in is the eigenvalue of the i-th dimension; is the mean value of the i-th dimension feature of the source region, which represents the statistical characteristics of the source region.

[0032] Furthermore, the feature-gated adversarial loss function combines the discrimination results of the gated input under the real data distribution with the discrimination results of the gated input under the generated data distribution;

[0033] The feature gated adversarial loss function, gated learning objective, and regional feature regularization term are integrated into the original adversarial objective function. By balancing the relative importance of each loss term through hyperparameters, the training objective function of the improved adaptive gated AFG-GAN is obtained. For example:

[0034]

[0035] Among them, x is the distribution of the real data The hydrological feature vector of G(z) is the simulated hydrological data output by the generator;

[0036] Gating this feature against the loss function With gated learning objectives And the regional characteristic regularization term Integrating into the original adversarial objective function, we obtain the final improved adaptive gated AFG-GAN training objective function:

[0037]

[0038] in, , is a hyperparameter that balances the relative importance of each loss term.

[0039] Furthermore, it also includes adaptive search of DMTSNet's hyperparameters through a Bayesian optimization algorithm, with the objective function being the root mean square error or Nash-Sutcliffe efficiency.

[0040] In a second aspect, the present invention provides an intelligent flood forecasting device based on the FQI-DMTSNet adaptive generative adversarial mechanism, comprising:

[0041] The data acquisition and preprocessing module is used to perform temporal and spatial resolution normalization and Kalman filter correction on remote sensing, measured and model data to obtain preprocessed hydrological data;

[0042] A multi-source data fusion module is used to extract multi-scale time series features of hydrological data using DMTSNet, and then combine it with the FQI reinforcement learning algorithm to construct a reward function that includes reservoir water level safety and downstream flood control safety, thereby optimizing flood dispatch strategies.

[0043] The AFG-GAN generation module is used to perform dimension-level gated weighting on input features through AFG-GAN, enabling flood data enhancement and adaptive generation and migration of hydrological parameters in data-free areas;

[0044] The FQI decision feedback module is used to collaboratively integrate the multi-source data fusion results, DMTSNet feature extraction results, FQI scheduling strategies, and AFG-GAN data enhancement results to generate flood forecast results.

[0045] In a third aspect, the present invention provides an electronic device comprising a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to execute the above-mentioned intelligent flood forecasting method based on the FQI-DMTSNet adaptive generative adversarial mechanism.

[0046] In a fourth aspect, the present invention provides a readable storage medium storing a computer program, wherein the computer program includes a program code for controlling a process to execute a process, wherein the process includes an intelligent flood forecasting method based on the above-mentioned FQI-DMTSNet adaptive generation adversarial mechanism.

[0047] The main contributions and innovations of the present invention are as follows:

[0048] 1. Improved accuracy and adaptability: The combination of DMTSNet and FQI enables multi-scale time series feature extraction and dynamic scheduling optimization. Combined with Bayesian hyperparameter optimization, it improves the model's prediction accuracy for nonlinear and non-stationary flood processes and its generalization ability across different scenarios.

[0049] 2. Data-missing scenario processing capability: AFG-GAN uses a feature gating mechanism to screen key hydrological factors, enabling the adaptive generation and migration of hydrological parameters in data-free areas, thus alleviating the problem of data scarcity.

[0050] 3. Enhanced robustness to extreme weather: AFG-GAN dynamically weights the abnormal feature channels of extreme samples, improving the discriminator's ability to identify low-probability events and reducing the mode collapse problem of traditional GAN.

[0051] 4. Closed-loop decision-making mechanism: Through the integrated framework of "forecasting-decision-feedback", the coordinated optimization of flood forecasting and scheduling strategies is achieved to meet the needs of real-time emergency response.

[0052] The details of one or more embodiments of the invention are set forth in the accompanying drawings and the description below so that other features, objects, and advantages of the invention are more readily apparent. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0054] Figure 1 is a flow chart of an intelligent flood forecasting method based on the FQI-DMTSNet adaptive generative adversarial mechanism according to an embodiment of the present invention;

[0055] Figure 2 2. It is a framework diagram of a multi-source data fusion module according to an embodiment of the present invention;

[0056] Figure 3 1 is a diagram of the AFG-GAN data enhancement and parameter migration generation framework according to an embodiment of the present invention;

[0057] Figure 4 is a schematic diagram of DMSTNet according to an embodiment of the present invention;

[0058] Figure 5 2. This is a diagram of the FQI reinforcement learning and DMTSNet dynamic scheduling framework according to an embodiment of the present invention;

[0059] Figure 6 is a general framework diagram according to an embodiment of the present invention;

[0060] Figure 7 FIG. 4 is a schematic diagram of the hardware structure of an electronic device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0061] Exemplary embodiments will be described in detail herein, with examples illustrated in the accompanying drawings. In the following description, when referring to the drawings, identical numerals in different figures represent identical or similar elements, unless otherwise indicated. The implementations described in the following exemplary embodiments are not intended to represent all implementations consistent with one or more embodiments of this specification. Rather, they are merely examples of apparatuses and methods consistent with certain aspects of one or more embodiments of this specification, as detailed in the appended claims.

[0062] It should be noted that in other embodiments, the steps of the corresponding method are not necessarily performed in the order shown and described in this specification. In some other embodiments, the method may include more or fewer steps than those described in this specification. In addition, a single step described in this specification may be broken down into multiple steps for description in other embodiments, and multiple steps described in this specification may be combined into a single step for description in other embodiments.

[0063] Existing technologies have problems such as reliance on high-quality historical data, lack of dynamic feedback mechanisms, and weak generalization capabilities in areas with incomplete data, no data, and extreme weather conditions, making it difficult to achieve high-precision, real-time flood forecasting and scheduling optimization.

[0064] Based on this, the present invention adaptively generates an adversarial mechanism based on FQI-DMTSNet to solve the problems existing in the prior art.

[0065] Example 1

[0066] The present invention aims to propose an intelligent flood forecasting method based on the FQI-DMTSNet adaptive generative adversarial mechanism. Figure 1-Figure 5 As shown, the method includes the following steps:

[0067] Step 1: Multi-source data fusion and assimilation: normalize the temporal and spatial resolutions of remote sensing, measured, and model data and perform Kalman filter correction to obtain pre-processed hydrological data;

[0068] In this embodiment, the premise of flood forecasting is to obtain high-quality, continuous and spatially consistent input data. Figure 2 As shown, it can be integrated into a multi-source data fusion module. The goal of this invention is to first address the problems of uneven temporal and spatial resolution, many missing values, and high data noise in multi-source precipitation data (such as ground-based weather stations, weather radars, satellite remote sensing, and numerical model outputs). By building a stable and efficient data fusion and assimilation mechanism. The details are as follows:

[0069] (1) Preliminary screening

[0070] Unify precipitation information of different spatial scales and different time granularities to the target spatiotemporal scale. Assume that multi-source remote sensing precipitation data is X={ x 1, x 2,⋯, x n}, after removing the null values, perform spatiotemporal interpolation using the Kriging interpolation method:

[0071]

[0072] in, is the estimated value of precipitation data at the location to be interpolated, s (the result after spatiotemporal normalization); s usually represents a spatial location (such as longitude and latitude coordinates) and may also imply a temporal dimension (such as interpolating a spatial point at a certain moment after normalization to the target time granularity). This value is derived from multi-source observation data using the Kriging interpolation method and is used to fill in gaps or unify spatiotemporal resolution. is the weight coefficient of the i-th observation point data, which is determined by minimizing the mean square error (MSE); is the precipitation value at the i-th valid observation point; n is the number of valid observation points. Data sources include ground meteorological stations, radars, satellite remote sensing, etc. The original data may have different temporal and spatial resolutions and need to be unified to the target scale through this interpolation process.

[0073] (2) Data null value processing and Kalman filter correction

[0074] The Kalman filter is used to correct the deviation of the multi-source data fusion results to improve the accuracy of flood estimation. The state variable is defined as the actual precipitation , the observation equation is , the state transition is:

[0075] a. Equation of state:

[0076]

[0077] in, is the true precipitation state variable at time t (i.e., the theoretical value not affected by observation noise). This variable is the core state variable of the Kalman filter and is used to recursively update the precipitation estimation value at different times; is the actual precipitation state variable at time t-1, and the dynamic modeling of the precipitation process is realized through the recursive relationship of the time series; It is the random noise in the state transition process at time t (such as random fluctuations in rainfall and hydrological uncertainties not captured by the model). Its statistical characteristics (such as variance) can be estimated based on historical data and used to quantify the random changes in precipitation at adjacent moments rather than the deterministic change trend.

[0078] b. Observation equation:

[0079]

[0080] in is the observed value at time t (such as measured water level, rainfall, flow, etc.); is the observation noise;

[0081] c. Update calculation;

[0082]

[0083] in, is the prediction error covariance, quantifying the predicted value The uncertainty of the prediction result is lower. To measure the noise variance, characterize the observation The noise level (such as sensor measurement error, remote sensing data error, etc.). The larger the value, the lower the reliability of the observation data; is the Kalman gain, which determines the weight of fusion, that is, the weight used to weigh the predicted value and the observed value in the final estimation, and its value range is [0, 1]. ,but , indicating that the prediction error is large and the observed value is trusted first; if ,but , indicating that the observation noise is large and the predicted value is trusted first.

[0084]

[0085] in; is the predicted value at time t, generated by the state equation; is the final estimate (i.e., the posterior estimate).

[0086] In this way, by recursively inferring the dynamic changes of true precipitation through the state equation and combining it with the noise correction of the observation equation, dynamic compensation for the spatiotemporal inconsistency and observation errors of multi-source remote sensing precipitation data can be achieved; the nonlinear trend of the precipitation process can be captured (through Adaptive estimation of ), providing more accurate input for subsequent DMTSNet temporal feature extraction.

[0087] Step 2: Combined modeling based on FQI (Fitted Q-Iteration) and DMTSNet (Dynamically-stratified Multi-scale Temporal Sensing Network): DMTSNet is used to extract multi-scale temporal features of hydrological data. Combined with the FQI reinforcement learning algorithm, a reward function is constructed that incorporates reservoir water level safety and downstream flood control safety to optimize flood control strategies.

[0088] In this embodiment, this step aims to intelligently optimize reservoir operation strategies through reinforcement learning and, in combination with DMTSNet, model the time-dependent information in forecast scenarios to avoid overfitting of the model, balance long-term goals with short-term accuracy, and enhance the intelligent integration of operation and forecasting. The details are as follows:

[0089] (1) Reward function design

[0090] Combining the dual goals of flood control: upstream reservoir safety (not exceeding flood limit water level) and downstream flood control safety (not exceeding warning flow), the reward function is as follows:

[0091]

[0092] in The flood control limit water level of the reservoir (also called the flood limit water level); The reservoir water level status at the next moment t+1 (prediction or simulation result); is the system state at time t+1 (global state vectors such as water level, reservoir capacity, and downstream flow); is the dispatching action taken at time t (such as flood discharge, gate opening, etc.); is the warning flow threshold of the downstream river (i.e. the upper limit of downstream flood control safety target); The current dispatch action and environmental status The determined measured or predicted downstream flow; is a non-negative weight constant used to balance the relative importance of upstream and downstream objectives.

[0093] (2) Bellman Equation and Strategy Optimization

[0094] The Bellman equation is a necessary condition for dynamic programming. It ensures the realization of the global optimal solution by converting the multi-stage decision problem into a recursive relationship of subproblems.

[0095] Basic Equation: State-Value Function V ( s ) is expressed as the expected sum of the current reward and the future discounted reward:

[0096]

[0097] in, is the value function of state s, indicating that starting from state s, following the strategy The expected cumulative reward that can be obtained; s is the current system state (such as reservoir water level, rainfall, river flow in flood forecasting, etc.); a is the action space (such as reservoir discharge flow decision, scheduling plan, etc.); is the policy function, which represents the probability of selecting action a in state s; r is the immediate reward (such as upstream reservoir safety reward, downstream flood control safety reward); is the probability of obtaining reward r after executing action a in state s. s' is the state at the next moment (such as the new water level or flow rate after executing action a). is the state transition probability, which indicates the probability of transitioning to state s' after executing action a in state s; Discount factor , used to measure the importance of future rewards ( The closer it is to 0, the more attention is paid to current rewards; the closer it is to 1, the more attention is paid to long-term rewards).

[0098] The optimal Bellman equation is:

[0099]

[0100] in, is the optimal state value function, which represents the maximum expected cumulative reward that can be obtained when taking the optimal strategy starting from state s; The immediate reward (such as upstream and downstream safety rewards) obtained by performing action a in state s; For the next state The optimal value reflects the long-term impact of the decision.

[0101] The value function corresponding to the strategy satisfies (decomposing the value of the current action into the recursion of "immediate reward" and "future optimal decision value"):

[0102]

[0103] in, is the state at time t Next action The optimal action value function represents the expected long-term cumulative reward brought by the action; In state Execute an action immediate rewards received (e.g., weighted rewards for upstream reservoir safety and downstream flood control safety); To perform an action The next state at the moment of transfer (such as the new water level after the reservoir discharge is adjusted, the river flow, etc.); is the action at time t+1 (the scheduling strategy at the next moment), It means choosing the best action that maximizes future value; The next moment state Next action The optimal action-value function.

[0104] Optimal strategy:

[0105]

[0106] in, is the optimal strategy, that is, the scheduling strategy sequence that maximizes the cumulative reward from the initial state to the terminal state T; To maximize the strategy set p, we need to find the strategy that maximizes the objective function. is the cumulative discount reward within a limited time domain, where:

[0107] t is the time step (t=1 is the initial time, t=T is the end time); Discount factor , used to decay the weight of future rewards ( The closer it is to 1, the more focused it is on long-term rewards); is the reward function at time t, measuring the state Next action immediate benefits (such as upstream and downstream safety rewards).

[0108] (3) DMTSNet modeling

[0109] like Figure 4 and Figure 5 As shown in Figure 1, this paper innovatively improves the original TCN (Time Convolutional Network), designing the original single dilated causal convolution structure into multiple parallel causal convolution paths with different dilation rates to enhance the perception of multi-scale time-dependent information. Each parallel path extracts feature information at a different time scale, and ultimately their outputs are fused and integrated into a model through a fully connected layer. The details are as follows:

[0110] 1. Causal Convolution:

[0111] Ensure that the output at the current moment depends only on the input at the current moment and before (satisfying temporal causality):

[0112]

[0113] Where y(t) is the output at the current time t (such as the water level forecast in flood forecasting); x(ti) is the input data at the past time ti (such as historical rainfall, water level observations), where:

[0114] i=0 corresponds to the current input x(t);

[0115] i=1 corresponds to the input x(t-1) at time t-1, and so on;

[0116] f(i) is the parameter of the convolution kernel at position i, which is used to weight the input features at different times; k is the size of the convolution kernel, which determines the length of the model's dependence on historical data (for example, k=3 means dependence on the input of the current and previous two moments).

[0117] 2. Dilated Convolution:

[0118] It is used to increase the receptive field without increasing the number of convolutional layers. The formula is:

[0119]

[0120] d is the dilation rate, which controls the skipped input interval; when d =1 is ordinary convolution, d When >1, it is dilated convolution. Multiple parallel dilated causal convolution layers use different dilation rates (such as d =1,2,3) can extract multi-scale temporal features. The actual time range covered by the convolution kernel is [td(k-1),t], and the equivalent receptive field size is 1+(k-1)d (compared to the k of ordinary convolution).

[0121] 3. Residual Connection:

[0122] The output of each layer is added to the input to alleviate the gradient vanishing problem of deep networks:

[0123]

[0124] The input signal x(t) first passes through the first convolution kernel With bias , and then through the ReLU activation function, we get the intermediate features: intermediate feature 1 = ;

[0125] Then pass through the second convolution kernel With bias , activated by ReLU again, the output of the main path is obtained: Main path output = (where * represents the convolution operation and the ReLU activation function is used to introduce nonlinearity).

[0126] Directly add the output of the main path to the original input x(t) to get the final output ;

[0127] This step is called residual learning, meaning the network learns the difference (residual) between the input x(t) and the target output. This formula describes the core structure of the residual connection in deep learning, commonly found in the residual block of the ResNet network. Its design aims to alleviate the vanishing gradient problem during deep network training by directly adding input to output through skip connections.

[0128] 4. Receptive field:

[0129] The total receptive field of a DMTSNet layer (with expansion) for the input is:

[0130]

[0131] Where R is the total receptive field size of the DMTSNet layer; k is the size of the basic convolution kernel (such as 3×3, 5×5, usually an odd number); is the dilation rate of the i-th layer dilated convolution , represents the sampling interval (usually 2 i ); L is the number of layers in DMTSNet (i.e., the number of dilated convolution layers).

[0132] The receptive field (RF) refers to the area of ​​the original input corresponding to the output of a node in a neural network layer, reflecting the range of input information that the node can "see." Dilated (atrous) convolution inserts "holes" (spaced by the dilation rate d) between the elements of a standard convolution kernel, expanding the receptive field without increasing the number of parameters. For example, a 3×3 convolution kernel with a dilation rate of 2 is equivalent to a 5×5 receptive field, but the number of parameters remains 3×3.

[0133] (4) Bayesian optimization module

[0134] In the process of hyperparameter selection for flood forecasting, Bayesian optimization is introduced to automatically adjust the number of convolutional layers, kernel width, learning rate and other parameters of DMTSNet to optimize the generalization performance.

[0135] The Bellman equation is an important equation in reinforcement learning (FQI), mainly for policy optimization; Bayesian is used to optimize the hyperparameters of the neural network (the neural network here is DMTSNet) to improve the model fit and does not involve policy optimization. In fact, it can be understood as follows: DMTSNet (used for hydrological feature extraction) provides state expression for FQI. FQI (based on the Bellman equation) is used for policy learning for optimized scheduling and uses DMTSNet output. Finally, Bayesian automatically selects DMTSNet hyperparameters, that is, tunes the hyperparameters in the DMTSNet network. Only by working together can the model's predictive ability be improved. The details are as follows:

[0136] 1. The goal of Bayesian optimization:

[0137]

[0138] in, are the hyperparameters of the model (such as learning rate, network depth, Dropout rate), and is the optimal parameter vector, that is, the model parameter combination obtained by optimization (such as the weight matrix and bias term of the neural network); is the hyperparameter space; is the flood forecast model trained with hyperparameters; To evaluate the loss function (such as RMSE, MAE, NSE).

[0139] 2. Gaussian Process (GP):

[0140] Bayesian optimization often uses Gaussian processes to model the loss function:

[0141]

[0142] in, is the mean function (usually set to 0); The kernel function (such as RBF kernel) represents the similarity between different parameters.

[0143] The core value of Gaussian process in Bayesian optimization lies in transforming the optimization of deterministic loss function into uncertainty reasoning in probability space - capturing loss trend through mean function, quantifying parameter correlation through covariance function, and finally balancing exploration and utilization through acquisition function to minimize the number of evaluations to find the optimal parameters. This mechanism can not only accelerate model training in computationally intensive optimization tasks such as flood forecasting, but also incorporate hydrological and physical constraints, providing a more efficient parameter optimization solution for intelligent flood control scheduling.

[0144] 3. Collection function:

[0145] The strategy for selecting the next evaluation point in each round is often based on Expected Improvement (EI), a key strategy for balancing exploration and exploitation. Its core idea is to quantify the mathematical expectation of the "improvement" that can be obtained when evaluating at a new point:

[0146]

[0147] in, For new points The objective function value at (its value is unknown, and is distributed through GP prediction. It is the loss function or evaluation index in flood prediction and can reflect the performance of the scheduling strategy or model parameters. max(f min -f(θ),0) represents the improvement that the new point θ can bring relative to the current optimal value. If f(θ)≥f min , the improvement is 0; if f(θ) <f min , the improvement is f min -f(θ)).

[0148] The closed form of EI is:

[0149]

[0150] in, ; , are the mean and standard deviation predictions of the Gaussian process, respectively; are the CDF (cumulative distribution function) and PDF (probability density function) of the normal distribution, respectively.

[0151] Among them, the utilization terms of the closed expression of EI are: Reflecting the “expected improvement potential” of the current forecast: If (i.e., predicting new points is better), and z is large (low uncertainty), then , using the dominant term, tends to choose this point (using the current optimal prediction); if ,but But due to exist hour , the utilization term may be negative, but is compensated by the second term.

[0152] EI closed expression exploration items: Reflecting the “potential improvement opportunities” brought about by uncertainty: Large (high uncertainty), even if , exist Reaching the maximum value ,The exploration term may dominate and encourage exploration of new areas with high uncertainty;

[0153] It decays exponentially as z increases, which means that only at z.

[0154] This expression, through analytical simplification of probability integrals, decomposes "expected improvement" into "weighted prediction bias" and "compensation for uncertainty," providing a computationally efficient and theoretically rigorous evaluation strategy for Bayesian optimization. In computationally intensive optimization tasks like flood forecasting, the closed form of EI can rapidly screen high-potential parameter points while avoiding local optima through uncertainty quantification. This complements Gaussian process modeling to support intelligent hyperparameter optimization.

[0155] 4. Common loss functions (objective functions) for flood forecasting:

[0156] In practice, Bayesian optimization uses one of the following metrics as the loss function:

[0157] Root Mean Square Error (RMSE):

[0158]

[0159] in, is the true value of the i-th sample (such as the actually observed water level or flow); is the predicted value of the model for the i-th sample; n is the number of samples (such as the length of the time series of flood forecasts).

[0160] Nash-Sutcliffe Efficiency (NSE):

[0161]

[0162] The objective function can be minimized RMSE or maximized NSE. Perfect prediction (the predicted value is exactly the same as the true value); The error in the model's prediction is equal to the error in "always predicting the mean" (the model has no real predictive power); The model prediction is worse than "using the mean directly" (meaningless and needs improvement).

[0163] Step 3: Data enhancement and parameter transfer based on AFG-GAN: AFG-GAN performs dimension-level gated weighting on input features to achieve flood data enhancement and adaptive generation and transfer of hydrological parameters in data-deficient areas;

[0164] Flood forecasting often faces challenges such as insufficient data samples, uneven data distribution, difficulty covering areas without data, and data outliers. Traditional single-discriminator GAN architectures suffer from feature redundancy when dealing with multi-source data, making it difficult to effectively focus on sensitive information and inferring parameters in unobserved areas. This can lead to a significant performance degradation during the generalization phase.

[0165] For this reason, Figure 3 As shown in the figure, this paper proposes an adaptive feature-gated generative adversarial network (AFG-GAN)-based approach. This approach embeds a feature dimension gating mechanism within the discriminator architecture and dynamically adjusts the weights of each dimension of the input vector by constructing a dynamic adaptive feature selection function. This approach aims to address core issues commonly encountered in flood forecasting, such as a scarcity of extreme samples, uneven data distribution, and difficulty modeling data-free areas. Specifically, the improvement involves inserting a feature gating module before the fully connected layer of the discriminator to dynamically weight each dimension of the input feature vector. Furthermore, a feature gradient backpropagation path is introduced into the gating module to quantify the contribution of each dimension to the discrimination result. Finally, a regional characteristic regularization term is introduced during the gating parameter training phase. When generating hydrological parameters in data-free areas, the gating mechanism automatically selects key transferable feature channels, enabling the generator to deduce reasonable parameters based on limited data (such as observations from adjacent watersheds), thereby improving the model's generalization capabilities in data-scarce scenarios. AFG-GAN realizes the collaborative generation capabilities of the model in flood forecasting, combining "global + local" and "modeling + compensation", further improving the stability and robustness of the model in complex environments. The details are as follows:

[0166] (1) Feature gating mechanism:

[0167] Add a feature gating module before the discriminator to input samples Each dimension of the feature in is dynamically weighted to obtain the gated input x′.

[0168] Define the feature gating function:

[0169]

[0170] in, It is a feature gating function, the input is a high-dimensional feature vector x, and the output is a weight vector of equal dimension , used to dynamically weight the features of each dimension of x; is the i-th dimension output of the gating function , represents the weight coefficient of the i-th dimension in the input feature vector x; d is the dimension of the original input feature vector x (for example, in flood forecasting, d can include dimensions of multiple source data such as rainfall, evaporation, water level, and NDVI).

[0171] Get the gated input vector:

[0172]

[0173] in, is the input vector after gating, and its dimension is consistent with the original input feature vector x. It is the optimized representation after feature screening.

[0174] Replace the original discriminator input and change the original Replace with:

[0175]

[0176] in, The input vector after gated discriminator The output value of .

[0177] (2) Gradient-guided feature contribution mechanism:

[0178] Gradient information is introduced to measure the sensitivity of each feature to the discriminator output and is used to supervise the direction of training the gating function, making it more focused on effective features.

[0179] Calculate the gradient contribution for each dimension:

[0180]

[0181] Construct a gated learning objective so that the gate function Close to softmax normalization :

[0182]

[0183] in, is the gradient contribution value of the i-th dimension feature, which represents the sensitivity of the discriminator output to the feature. , the larger the value, the greater the influence of this feature on the discriminator's judgment of "data authenticity"; is the i-th dimension eigenvalue of the original input eigenvector x (such as rainfall, water level, etc.); is the weight output of the gating function for the i-th dimension feature.

[0184] (3) Regional characteristic regularization term:

[0185] In areas without data, the gating function should tend to select transferable features that are similar in both the source and target regions. To this end, a regional regularization term is introduced to prevent inconsistent features from perturbing the discriminator.

[0186] Assume that the target area sample is , the mean characteristic value of the source region is , then the regional regularization term is:

[0187]

[0188] Where sim(⋅) can use cosine similarity or inverse Euclidean distance function to express similarity; is the sample feature vector of the target area, ,in is the eigenvalue of the i-th dimension; is the mean value of the i-th dimension feature of the source region, which represents the statistical characteristics of the source region.

[0189] The feature gated adversarial loss function is:

[0190]

[0191] Among them, x is the distribution of the real data The hydrological feature vector of ; G(z) is the simulated hydrological data output by the generator.

[0192] Step 4: Closed-loop collaborative forecasting: Collaboratively integrate the multi-source data fusion results, DMTSNet feature extraction results, FQI scheduling strategy and AFG-GAN data enhancement results to generate flood forecast results.

[0193] In this embodiment, the multi-source data fusion results, DMTSNet feature extraction results, FQI scheduling strategy and AFG-GAN data enhancement results are collaboratively integrated to construct the AFG-GAN objective function:

[0194] Integrating the three parts of step 3 into the original adversarial objective function, we obtain the training objective function of the improved adaptive gated AFG-GAN:

[0195]

[0196] in, is the adversarial loss of the discriminator acting on the gated input; Contribute supervision loss to features; Regularize the migration characteristics for data-free areas; are two regular weight hyperparameters that balance the relative importance of each loss term.

[0197] The present invention optimizes the closed-loop feedback in the process, and the generator is based on the target area characteristics. Generate hydrological parameters; the discriminator combines the gated input to evaluate the authenticity of the generated data and calculate and ; Regional regularization term Force the features with high gate weights to be similar to the source area, and back propagate to update the parameters of the generator and discriminator; iterate and optimize until the generated data satisfies both authenticity ( ), key feature focus ( ), and has cross-regional similarity ( ).

[0198] In summary, the present invention creatively designs an Adaptive Feature Gating Generative Adversarial Network (AFG-GAN), specifically for flood data enhancement and data-free area hydrological parameter migration generation tasks. The core innovation lies in the implantation of a feature dimension-level gating control mechanism in the discriminator module. By constructing a dynamic adaptive feature selection function, the weights of each dimension of the input vector can be adjusted in real time. This mechanism embeds a dimension-level gating unit in the discriminator architecture and realizes dynamic adjustment of feature weights through a data-driven adaptive screening strategy. The gating function automatically identifies and enhances feature dimensions with high discriminant value based on the statistical characteristics of the input samples (such as mean and variance distribution), temporal dependency structure (autocorrelation coefficient), and degree of anomaly (outlier quantification). In addition, based on the underlying surface properties of the receiving basin (such as soil permeability and vegetation coverage) and regional climate parameters (such as evaporation capacity and pressure field distribution), AFG-GAN screens cross-regional universal hydrological characteristics through a feature gating mechanism, realizing the adaptive deduction and migration generation of hydrological parameters in areas without measured data, thereby improving the generalization ability of the model.

[0199] Example 2

[0200] like Figure 6 As shown, based on the same concept, the present invention also proposes an intelligent flood forecasting device based on the FQI-DMTSNet adaptive generative adversarial mechanism, including:

[0201] The data acquisition and preprocessing module is used to perform temporal and spatial resolution normalization and Kalman filter correction on remote sensing, measured and model data to obtain preprocessed hydrological data;

[0202] A multi-source data fusion module (also including a DMTSNet predictive modeling module) is used to extract multi-scale time series features of hydrological data using DMTSNet, and then use the FQI reinforcement learning algorithm to construct a reward function that includes reservoir water level safety and downstream flood control safety to optimize flood scheduling strategies.

[0203] The AFG-GAN generation module is used to perform dimension-level gated weighting on input features through AFG-GAN, enabling flood data enhancement and adaptive generation and migration of hydrological parameters in data-free areas;

[0204] The FQI decision feedback module (which also includes the automatic scheduling and warning issuance module) is used to collaboratively integrate the multi-source data fusion results, DMTSNet feature extraction results, FQI scheduling strategies, and AFG-GAN data enhancement results to generate flood forecast results (and issue warnings).

[0205] Example 3

[0206] This embodiment also provides an electronic device, referring to Figure 7 , includes a memory 404 and a processor 402, wherein the memory 404 stores a computer program, and the processor 402 is configured to run the computer program to perform the steps in any of the above method embodiments.

[0207] Specifically, the processor 402 may include a central processing unit (CPU), or an application specific integrated circuit (ASIC), or may be configured to implement one or more integrated circuits for implementing the embodiments of the present invention.

[0208] Memory 404 may include a large-capacity memory 404 for data or instructions. By way of example, and not limitation, memory 404 may include a hard disk drive (HDD), a floppy disk drive, a solid-state drive (SSD), flash memory, an optical disk, a magneto-optical disk, a magnetic tape, or a Universal Serial Bus (USB) drive, or a combination of two or more of these. Where appropriate, memory 404 may include removable or non-removable (or fixed) media. Where appropriate, memory 404 may be internal or external to the data processing device. In certain embodiments, memory 404 is non-volatile memory. In certain embodiments, memory 404 includes read-only memory (ROM) and random access memory (RAM). Where appropriate, the ROM may be a mask-programmed ROM, a programmable ROM (PROM), an erasable PROM (EPROM), an electrically erasable PROM (EEPROM), an electrically alterable ROM (EAROM) or a flash memory (FLASH), or a combination of two or more of these. In appropriate circumstances, the RAM may be a static random access memory (SRAM) or a dynamic random access memory (DRAM), wherein the DRAM may be a fast page mode dynamic random access memory 404 (FPMDRAM), an extended data output dynamic random access memory (EDODRAM), a synchronous dynamic random access memory (SDRAM), etc.

[0209] The memory 404 may be used to store or cache various data files required for processing and / or communication, as well as possible computer program instructions executed by the processor 402 .

[0210] The processor 402 reads and executes computer program instructions stored in the memory 404 to implement any one of the intelligent flood forecasting methods based on the FQI-DMTSNet adaptive generative adversarial mechanism in the above embodiments.

[0211] Optionally, the electronic device may further include a transmission device 406 and an input / output device 408 , wherein the transmission device 406 is connected to the processor 402 , and the input / output device 408 is connected to the processor 402 .

[0212] Transmission device 406 can be used to receive or transmit data via a network. Specific examples of such networks may include wired or wireless networks provided by the electronic device's communications provider. In one embodiment, the transmission device includes a network interface controller (NIC), which can be connected to other network devices via a base station to enable communication with the Internet. In another embodiment, transmission device 406 can be a radio frequency (RF) module, which is used to communicate with the Internet wirelessly.

[0213] The input / output device 408 is used to input or output information.

[0214] Example 4

[0215] This embodiment also provides a readable storage medium, which stores a computer program. The computer program includes program code for controlling a process to execute a process. The process includes an intelligent flood forecasting method based on the FQI-DMTSNet adaptive generation adversarial mechanism according to embodiment one.

[0216] It should be noted that the specific examples in this embodiment can refer to the examples described in the above embodiments and optional implementation modes, and this embodiment will not be repeated here.

[0217] In general, various embodiments may be implemented in hardware or dedicated circuitry, software, logic, or any combination thereof. Some aspects of the invention may be implemented in hardware, while other aspects may be implemented in firmware or software executed by a controller, microprocessor, or other computing device, but the invention is not limited thereto. Although various aspects of the invention may be shown and described as block diagrams, flow charts, or using some other graphical representation, it should be understood that, as non-limiting examples, the blocks, devices, systems, techniques, or methods described herein may be implemented in hardware, software, firmware, dedicated circuitry or logic, general-purpose hardware or a controller or other computing device, or some combination thereof.

[0218] The embodiments of the present invention may be implemented by computer software that is executable by a data processor of a mobile device, such as in a processor entity, or by hardware, or by a combination of software and hardware. Computer software or programs (also referred to as program products) including software routines, applets and / or macros may be stored in any device-readable data storage medium, and they include program instructions for performing specific tasks. A computer program product may include one or more computer executable components that are configured to perform an embodiment when the program is run. One or more computer executable components may be at least one software code or a portion thereof. In addition, it should be noted at this point that, for example, Figure 1 Any block of the logic flow in the program may represent program steps, or interconnected logic circuits, blocks and functions, or a combination of program steps and logic circuits, blocks and functions. The software may be stored on physical media such as memory chips or memory blocks implemented within the processor, magnetic media such as hard disks or floppy disks, and optical media such as, for example, DVDs and their data variants, CDs, etc. Physical media are non-transitory media.

[0219] Those skilled in the art should understand that the technical features of the above embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0220] The above embodiments merely illustrate several embodiments of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art would be able to make various modifications and improvements without departing from the spirit of the present invention, all of which fall within the scope of the present invention. Therefore, the scope of the present invention shall be determined by the appended claims.

Claims

1. An intelligent flood forecasting method based on the FQI-DMTSNet adaptive generative adversarial mechanism, characterized by: The following steps are involved: Multi-source data fusion and assimilation: normalize the temporal and spatial resolutions of remote sensing, measured, and model data and perform Kalman filter correction to obtain pre-processed hydrological data; Combined modeling based on FQI and dynamic hierarchical multi-scale time-series perception network: The dynamic hierarchical multi-scale time-series perception network is used to extract the multi-scale time-series features of hydrological data. Combined with the FQI reinforcement learning algorithm, a reward function that includes reservoir water level safety and downstream flood control safety is constructed to optimize flood scheduling strategies. Data enhancement and parameter transfer based on adaptive feature-gated generative adversarial networks: Dimension-level gating and weighting of input features is performed through adaptive feature-gated generative adversarial networks to achieve flood data enhancement and adaptive generation and transfer of hydrological parameters in data-deficient areas; Closed-loop collaborative forecasting: Collaboratively integrates multi-source data fusion results, dynamic layered multi-scale time-series perception network feature extraction results, FQI scheduling strategies, and adaptive feature-gated generative adversarial network data enhancement results to generate flood forecast results; Among them, the construction method of the dynamic hierarchical multi-scale time series perception network is: The single dilated causal convolutional layer of the traditional temporal convolutional neural network is replaced by three parallel dilated causal convolutional layers with different dilation coefficients. The outputs of each layer are fused through a fully connected network, and residual connections and parameterized ReLU activation functions are combined to capture long-term temporal dependencies. The feature gating mechanism of the adaptive feature gating generative adversarial network specifically includes: Insert a feature gating module before the discriminator to dynamically weight each dimension of the input feature to obtain the gated input; Quantify the feature importance based on the gradient contribution of each feature dimension to the discriminator output, and construct a gating learning objective to optimize the gating weight; A regional characteristic regularization term is introduced to optimize the parameter migration of the data-free area by measuring the similarity between the characteristics of the target area and the source area; Among them, the feature gated adversarial loss function combines the discrimination results of the gated input under the real data distribution and the discrimination results of the gated input under the generated data distribution; The feature-gated adversarial loss function, gated learning objective and regional feature regularization term are integrated into the original adversarial objective function. The relative importance of each loss term is balanced through hyperparameters to obtain the training objective function of the final improved adaptive gated adaptive feature-gated generative adversarial network.

2. The intelligent flood forecasting method based on the FQI-DMTSNet adaptive generative adversarial mechanism according to claim 1, characterized in that: Multi-source data fusion and assimilation specifically include: The Kriging interpolation method is used to perform spatiotemporal interpolation of multi-source remote sensing precipitation data. By determining the weight coefficient of the data at each observation point to minimize the mean square error, eliminating null values ​​and aligning the spatiotemporal resolution, the precipitation data estimate at the location to be interpolated is obtained. The Kalman filter is used to correct the deviation of the fused data, and the final estimated value is obtained by updating the state equation and observation equation. The state equation reflects the dynamic changes of the actual precipitation at different times, and the observation equation combines the observation value with the noise, and the weight of the predicted value and the observed value is weighed by the Kalman gain.

3. The intelligent flood forecasting method based on the FQI-DMTSNet adaptive generative adversarial mechanism according to claim 1, characterized in that: The reward function of the FQI reinforcement learning algorithm is designed as: The reward function comprehensively considers reservoir water level safety and downstream flood control safety, balancing the relative importance of the two through weight coefficients. Reservoir water level safety is associated with the flood control limit water level and the reservoir water level status at the next moment, while downstream flood control safety is associated with the warning flow threshold and the downstream flow determined by the scheduling action and environmental conditions.

4. The intelligent flood forecasting method based on the FQI-DMTSNet adaptive generative adversarial mechanism according to any one of claims 1 to 3, characterized in that: It also includes adaptive search for hyperparameters of dynamic hierarchical multi-scale time-aware networks through Bayesian optimization algorithm, with the objective function being root mean square error or Nash-Sutcliffe efficiency.

5. An intelligent flood forecasting device based on the FQI-DMTSNet adaptive generative adversarial mechanism, characterized by: include: The data acquisition and preprocessing module is used to perform temporal and spatial resolution normalization and Kalman filter correction on remote sensing, measured and model data to obtain preprocessed hydrological data; A multi-source data fusion module is used to extract the multi-scale time series features of hydrological data using a dynamic hierarchical multi-scale time series perception network. This module is then combined with the FQI reinforcement learning algorithm to construct a reward function that includes reservoir water level safety and downstream flood control safety, thereby optimizing flood control strategies. An adaptive feature-gated generative adversarial network generation module is used to perform dimension-level gated weighting on input features through an adaptive feature-gated generative adversarial network, enabling flood data enhancement and the adaptive generation and migration of hydrological parameters in data-free areas; The FQI decision feedback module is used to collaboratively integrate the multi-source data fusion results, the dynamic layered multi-scale time series perception network feature extraction results, the FQI scheduling strategy, and the adaptive feature gated generative adversarial network data enhancement results to generate flood forecast results; Among them, the construction method of the dynamic hierarchical multi-scale time series perception network is: The single dilated causal convolutional layer of the traditional temporal convolutional neural network is replaced by three parallel dilated causal convolutional layers with different dilation coefficients. The outputs of each layer are fused through a fully connected network, and residual connections and parameterized ReLU activation functions are combined to capture long-term temporal dependencies. The feature gating mechanism of the adaptive feature gating generative adversarial network specifically includes: Insert a feature gating module before the discriminator to dynamically weight each dimension of the input feature to obtain the gated input; Quantify the feature importance based on the gradient contribution of each feature dimension to the discriminator output, and construct a gating learning objective to optimize the gating weight; A regional characteristic regularization term is introduced to optimize the parameter migration of the data-free area by measuring the similarity between the characteristics of the target area and the source area; Among them, the feature gated adversarial loss function combines the discrimination results of the gated input under the real data distribution and the discrimination results of the gated input under the generated data distribution; The feature-gated adversarial loss function, gated learning objective and regional feature regularization term are integrated into the original adversarial objective function. The relative importance of each loss term is balanced through hyperparameters to obtain the training objective function of the final improved adaptive gated adaptive feature-gated generative adversarial network.

6. An electronic device comprising a memory and a processor, characterized in that: The memory stores a computer program, and the processor is configured to run the computer program to execute the intelligent flood forecasting method based on the FQI-DMTSNet adaptive generative adversarial mechanism according to any one of claims 1 to 4.

7. A readable storage medium, characterized in that: The readable storage medium stores a computer program, which includes a program code for controlling a process to execute a process, and the process includes the intelligent flood forecasting method based on the FQI-DMTSNet adaptive generation confrontation mechanism according to any one of claims 1 to 4.

Images