Flood disaster forecasting method based on spatio-temporal feature fusion

Abstract

The invention discloses a flood disaster forecasting method based on spatio-temporal feature fusion, and relates to the technical field of flood forecasting, the flood disaster forecasting method based on spatio-temporal feature fusion comprises the following steps: S1, collecting data for data preprocessing; s2, rainfall space-time parameters and runoff space-time parameters are extracted to generate a flood influence parameter set; s3, presetting a river bearing rule, and establishing a river bearing model based on the river bearing rule and the river structure parameters; s4, performing training optimization on the river bearing model, and generating a flood forecasting scheme; and S5, verifying and optimizing the flood forecasting scheme, and outputting the flood forecasting scheme as a flood disaster forecasting method. According to the method, the accuracy of a flood forecasting scheme is improved by combining spatial-temporal feature fusion and a data driving algorithm, and the generated flood influence parameter set can comprehensively and accurately reflect various factors influencing flood generation by extracting spatial-temporal parameters of a river structure, rainfall and runoff in a multi-dimensional manner.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of flood prediction, in particular, to a flood disaster prediction method based on spatio-temporal feature fusion. BACKGROUND

[0002] The significance of flood disaster prediction mainly lies in disaster prevention and reduction, social and economic development, and ecological safety. Flood is one of the most common and most harmful natural disasters, which is strong in suddenness and wide in influence range. If it cannot be predicted and responded in time, it is easy to cause heavy casualties and property losses. Through scientific flood disaster prediction, the precipitation trend, water level change and possible influence range of flood can be grasped in advance before the flood occurs, guiding the implementation of emergency measures such as personnel evacuation, dike reinforcement and reservoir capacity regulation, so as to effectively reduce the flood risk.

[0003] Spatio-temporal feature fusion refers to combining time (time series) and space (position or geography) information in data analysis and modeling process to understand and model the dynamic behavior of the system in a more comprehensive way. In the modeling of many complex systems, especially in the fields of meteorology, environment and transportation, spatio-temporal feature fusion can effectively improve the accuracy of prediction and decision-making.

[0004] However, the existing flood disaster prediction method based on spatio-temporal feature fusion does not consider that the relationship between precipitation and runoff is influenced by complex meteorological, topographical and human activities, so that the existing flood disaster prediction method based on spatio-temporal feature fusion is difficult to accurately capture this nonlinear relationship, resulting in low prediction accuracy, especially under extreme precipitation and complex topographical conditions. Moreover, the existing flood disaster prediction method based on spatio-temporal feature fusion usually ignores the spatio-temporal features of precipitation, runoff and other data, so that the method cannot accurately reflect the actual influence of strong precipitation on flood in real time, and the traditional method cannot effectively consider the influence of precipitation difference in different regions and the influence of topography on water flow in the basin, resulting in large errors in the prediction results. At present, no effective solution has been proposed for the problems in the related art. SUMMARY

[0005] In view of the problems in the related art, the present application proposes a flood disaster prediction method based on spatio-temporal feature fusion to overcome the above technical problems existing in the prior art.

[0006] In order to achieve the above purpose, the specific technical scheme adopted by the present application is as follows:

[0007] The flood disaster prediction method based on spatio-temporal feature fusion comprises the following steps:

[0008] S1, collecting river structure parameters, future precipitation prediction parameters, historical precipitation data and historical runoff data, and performing data preprocessing in time sequence;

[0009] S2. Use data-driven algorithms to extract the spatiotemporal parameters of precipitation and runoff from historical precipitation and runoff data to generate a flood impact parameter set.

[0010] As a preferred embodiment, the step of using data-driven algorithms to extract the spatiotemporal parameters of precipitation and runoff from historical precipitation and runoff data to generate a flood impact parameter set includes the following steps:

[0011] S21. Normalize historical precipitation and runoff data, construct a sliding data structure based on time window, and segment the precipitation and runoff data into time windows.

[0012] As a preferred embodiment, the normalization of historical precipitation and runoff data, the construction of a sliding data structure based on a time-series window, and the segmentation of precipitation and runoff data into time windows include the following steps:

[0013] S211. Missing values ​​are filled and outliers are corrected for historical precipitation and runoff data, and normalization is performed using a uniform scale.

[0014] S212. Preset the time window length and sliding step size, and construct the time window structure based on the time window length and sliding step size;

[0015] S213. Segment the historical precipitation data and historical runoff data according to the time-series window structure to obtain the precipitation window sequence and runoff window sequence;

[0016] S214. Index and align the precipitation window sequence and runoff window sequence, and establish the time series window correspondence.

[0017] S22. Based on the time window segmentation results, window features are extracted from historical precipitation data and historical runoff data to obtain precipitation window parameters and runoff window parameters.

[0018] As a preferred embodiment, the step of constructing a window impact model using a long short-term memory network algorithm, training and optimizing the window impact model by inputting precipitation window parameters and runoff window parameters, and generating a flood impact parameter set through the optimized window impact model includes the following steps:

[0019] S221. Use the Long Short-Term Memory network algorithm and construct a window impact model based on precipitation window parameters and runoff window parameters;

[0020] S222. Use precipitation window parameters and runoff window parameters as input features, and construct the input layer of the window influence model;

[0021] S223, bring the precipitation window parameter and the runoff window parameter into the window influence model for training, and calculate the loss function of the window influence model;

[0022] S224, evaluate the performance of the window influence model through cross-validation, adjust the window influence model according to the performance evaluation result of the window influence model and the loss function, and generate a set of flood influence parameters according to the adjusted window influence model.

[0023] S23, use a long short-term memory network algorithm to construct a window influence model, and bring the precipitation window parameter and the runoff window parameter into the window influence model for training and optimization, and generate a set of flood influence parameters through the optimized window influence model;

[0024] S24, set a weight influence rule, and optimize and output the set of flood influence parameters based on the weight influence rule.

[0025] S3, preset a river bearing rule, and establish a river bearing model based on the river bearing rule and river structure parameters;

[0026] As a preferred solution, the preset river bearing rule and the establishment of the river bearing model based on the river bearing rule and the river structure parameters include the following steps:

[0027] S31, extract geographical environment data, riverbed structure data and hydro-meteorological data from the river structure parameters, and set river bearing capacity rules and flood bearing threshold parameters;

[0028] As a preferred solution, the extraction of geographical environment data, riverbed structure data and hydro-meteorological data from the river structure parameters, and the setting of river bearing capacity rules and flood bearing threshold parameters include the following steps:

[0029] S311, quantitatively process the geographical environment data, riverbed structure data and hydro-meteorological data in the river structure parameters;

[0030] S312, set river bearing capacity rules according to the quantified geographical environment data, riverbed structure data and hydro-meteorological data, and the river bearing capacity rules include river channel water carrying capacity rules and flood discharge capacity rules;

[0031] S313, set flood bearing threshold parameters according to the river bearing capacity rules, form a river bearing capacity benchmark, and set the river bearing capacity rules and the flood bearing threshold parameters based on the river bearing capacity benchmark.

[0032] S32, establish a river bearing model according to historical runoff data and river bearing capacity rules;

[0033] S33, quantize the river structure parameters, and bring them into the river bearing model combined with the river bearing capacity rules for training and optimization;

[0034] S34, verify the trained and optimized river bearing model, and output the verified river bearing model.

[0035] S4, bring the historical flood impact parameters in the flood impact parameter set into the river bearing model for training and optimization, and generate a flood forecast scheme according to the trained and optimized river bearing model;

[0036] As a preferred solution, the step of bringing the historical flood impact parameters in the flood impact parameter set into the river bearing model for training and optimization, and generating a flood forecast scheme according to the trained and optimized river bearing model comprises the following steps:

[0037] S41, data cleaning is performed on the historical flood impact parameters in the flood impact parameter set, and the cleaned historical flood impact parameters are brought into the river bearing model for training;

[0038] S42, in the training process of the river bearing model, the output results of the model are compared with the historical flood observation results by using the error evaluation function, and the training error value is calculated;

[0039] As a preferred solution, the step of comparing the output results of the model with the historical flood observation results by using the error evaluation function, and calculating the training error value in the training process of the river bearing model comprises the following steps:

[0040] S421, the output results of the river bearing model are compared with the normalized historical flood observation results, and a corresponding relationship between the output results of the river bearing model and the historical flood observation results is established;

[0041] S422, the difference between the output results of the river bearing model and the historical flood observation results is calculated based on the error evaluation function, and the training error value is obtained;

[0042] S423, the calculated training error value is statistically analyzed, and the analysis result is fed back to the river bearing model training process for updating.

[0043] S43, preset a model updating rule, and update the river bearing model based on the model updating rule and the training error value;

[0044] S44, bring the historical flood impact parameters into the updated river bearing model to generate a corresponding flood forecast scheme.

[0045] S5, verify and optimize the flood forecast scheme according to the future precipitation prediction parameters, and output the verified and optimized flood forecast scheme as the flood disaster prediction method.

[0046] As a preferred solution, the step of verifying and optimizing the flood forecasting scheme according to the future precipitation prediction parameter and outputting the verified and optimized flood forecasting scheme as the flood disaster forecasting method comprises the following steps:

[0047] S51, preset a verification and optimization error threshold, and calculate a precipitation difference value of the future precipitation prediction parameter and historical precipitation data;

[0048] S52, compare the verification and optimization error threshold with the precipitation difference value, and generate a verification and optimization strategy based on the comparison result;

[0049] As a preferred solution, the step of comparing the verification and optimization error threshold with the precipitation difference value and generating a verification and optimization strategy based on the comparison result comprises the following steps:

[0050] S521, set the verification and optimization error threshold, and calculate the precipitation difference value according to the future precipitation prediction parameter and historical precipitation data;

[0051] S522, preset an optimization strategy threshold range and an optimization strategy set, compare the precipitation difference value with the verification and optimization error threshold, and judge whether the difference value is within the optimization strategy threshold range based on the comparison result;

[0052] If yes, a verification and optimization strategy of maintaining the existing flood forecasting scheme is generated;

[0053] If no, a verification and optimization strategy of adjusting the flood forecasting scheme is generated.

[0054] S53, verify and optimize the flood forecasting scheme according to the verification and optimization strategy;

[0055] S54, add the future precipitation prediction parameter into the verified and optimized flood forecasting scheme, and generate a flood disaster forecasting method output.

[0056] Compared with the prior art, the present application has the following beneficial effects:

[0057] 1. The application improves the accuracy of the flood forecasting scheme by combining spatio-temporal feature fusion and data-driven algorithms, and extracts the spatio-temporal parameters of river structure, precipitation and runoff in multiple dimensions to generate a set of flood impact parameters that can comprehensively and accurately reflect various factors affecting the occurrence of floods. Based on different time windows and sliding steps, an impact model is constructed through a long short-term memory network algorithm to realize deep modeling of complex flood change processes, thereby significantly improving the accuracy and timeliness of the forecast. Meanwhile, a training optimization method based on error evaluation function is used to compare with historical flood observation results, constantly optimize the river bearing model, ensure the response of the model to actual flood disasters is more accurate, and through cross-validation and training error analysis, the model is flexibly adjusted and optimized to improve its adaptability and reliability in actual flood disaster forecasting.

[0058] 2. When generating the flood forecasting scheme, the existing forecasting scheme is verified and optimized through future precipitation prediction parameters to ensure that the flood forecasting scheme remains highly reliable and accurate when facing different weather changes. By setting a verification and optimization error threshold and generating a verification and optimization strategy, the forecasting demand under different environmental conditions is flexibly adapted to further improve the execution effect of the disaster forecasting scheme. Meanwhile, through the automatic feature extraction and training optimization mechanism, the need for human intervention is reduced, the operability and generalizability of the scheme are improved, and through the construction of a multi-path graphical user interface test process and an abnormal detection mechanism, the efficiency and sustainability of the entire flood forecasting process are ensured to enhance the emergency response capability of the system to sudden flood situations.

[0059] 3. The application improves the reliability of flood forecasting through structured feature extraction and optimization adjustment, reduces the computational cost in the model training and updating process, and combines historical data and real-time prediction to achieve efficient flood disaster warning at a lower cost, which has a wide application prospect. The application uses spatio-temporal fusion of multiple data sources and long short-term memory network algorithm, so that the method is not only suitable for flood forecasting in a single basin, but also can be extended to flood prediction and warning in different basins and different climate zones to cope with various complex and dynamic flood disaster situations and provide a comprehensive solution for intelligent flood disaster management and emergency decision-making in a variable environment. BRIEF DESCRIPTION OF DRAWINGS

[0060] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0061] Figure 1is a method flowchart of a flood disaster forecasting method based on spatio-temporal feature fusion according to an embodiment of the present application. DETAILED DESCRIPTION

[0062] The specific embodiments of the present application are described in further detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present application but are not intended to limit the scope of the present application.

[0063] Therefore, the detailed description of the embodiments of the present application provided below in the accompanying drawings is not intended to limit the scope of the claimed present application, but only represents selected embodiments of the present application. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present application without making creative efforts are within the scope of protection of the present application.

[0064] The present application is further illustrated in conjunction with the accompanying drawings and specific embodiments, as shown in Figure 1 The flood disaster forecasting method based on spatio-temporal feature fusion according to an embodiment of the present application comprises the following steps:

[0065] S1, collecting river structure parameters, future precipitation prediction parameters, historical precipitation data and historical runoff data, and performing data preprocessing in time sequence;

[0066] Specifically, the river structure parameters are obtained by remote sensing images, GIS data, unmanned aerial survey and hydrogeological survey to obtain geographic environment information, riverbed structure information and hydro-meteorological information, and combined with the results of hydrological monitoring station and sonar detection to obtain data such as river channel shape, water level and flow rate and water carrying capacity; the future precipitation prediction parameters can be output based on a numerical weather prediction model and combined with meteorological radar and satellite observation for short-term precipitation prediction; the historical precipitation data mainly come from ground meteorological station monitoring records and reanalysis data sets, and the historical runoff data are obtained by real-time monitoring of hydrological stations or hydrological simulation model calculation.

[0067] For the above multi-source data, first, missing value filling and outlier correction are performed in the preprocessing link, and interpolation method, statistical screening or spatio-temporal kriging interpolation is used to improve data integrity and reliability; then, normalization and standardization methods are used to convert different dimensional data such as precipitation, runoff and riverbed elevation to a unified scale, and the consistency of multi-source data is ensured by time stamp alignment, and on this basis, the data sequence is segmented according to the set time window length and sliding step to form precipitation window sequence and runoff window sequence, and an index alignment relationship is established to realize the construction of time sequence window structure.

[0068] S2, using data-driven algorithm to extract precipitation spatio-temporal parameters and runoff spatio-temporal parameters of historical precipitation data and historical runoff data to generate a set of flood influence parameters;

[0069] In this embodiment of the application, the step of using a data-driven algorithm to extract the spatiotemporal parameters of precipitation and runoff from historical precipitation data and historical runoff data to generate a flood impact parameter set includes the following steps:

[0070] S21. Normalize historical precipitation and runoff data, construct a sliding data structure based on time window, and segment the precipitation and runoff data into time windows.

[0071] In this embodiment of the application, the normalization processing of historical precipitation data and historical runoff data, the construction of a sliding data structure based on a time-series window, and the segmentation of precipitation data and runoff data into time windows include the following steps:

[0072] S211. Missing values ​​are filled and outliers are corrected for historical precipitation and runoff data, and normalization is performed using a uniform scale.

[0073] Specifically, missing values ​​are filled in. For short-term, continuous missing data, linear interpolation is used for estimation. For time-series data with obvious periodicity, seasonal mean or sliding window mean can be used for filling. When the missing interval is long or there is spatial correlation between data from multiple sites, spatiotemporal kriging interpolation can be used in combination with observations from surrounding sites for estimation. At the same time, machine learning methods such as KNN nearest neighbor algorithm or prediction models based on long short-term memory network are used to fill gaps. For completely missing site data, reanalysis data or observations from neighboring sites are introduced to replace them.

[0074] In terms of outlier correction, unreasonable data, such as negative precipitation or records far exceeding historical extremes, are eliminated through physical constraints. The peak runoff is corrected based on the river's flood control capacity. At the same time, statistical methods such as the interquartile range method are used to identify and correct outliers. Local abrupt changes are smoothed using the sliding window mean or median. In terms of uniform scale processing, in order to eliminate the influence of different dimensions, normalization methods are usually used to transform the data. For non-negative data, Min-Max normalization can be used to compress the values ​​to the [0, 1] interval. For data with an approximately normal distribution, the Z-Score standardization method is used. For extreme precipitation data with a significant skewed distribution, logarithmic transformation is performed before normalization.

[0075] S212. Preset the time window length and sliding step size, and construct the time window structure based on the time window length and sliding step size;

[0076] Specifically, the length of the time window determines the amount of data processed in each time window, which is selected according to the occurrence period of the flood event, the duration of the precipitation process, and the hydrological response characteristics of the basin. For a basin that responds quickly to precipitation, a shorter time window (such as 1 day or 3 days) is selected; while for a basin that responds slowly, or in the case of seasonal floods, a longer window length (such as 7 days or 14 days) is selected, and the sliding step determines the interval at which the window moves on the time axis, usually selected as 1 day or 1 hour, depending on the time resolution and computational requirements of the forecast target. A smaller step can provide a more fine-grained time series, but will increase the amount of computation, while a larger step can reduce redundant data and is suitable for large-scale flood forecasting.

[0077] After determining the length of the time window and the sliding step, the historical precipitation and runoff data are segmented according to the time window to form a time window structure. The specific operation is to divide the original data into multiple time periods according to the window length, for example, if the window length is 7 days, the first window includes 7 days of data, the second window includes the next 7 days of data, and so on. Then extract the feature data of precipitation and runoff within each window, such as daily precipitation total, average flow, etc., and construct a feature vector to represent the spatio-temporal characteristics within the time window. Each window is assigned a unique time index to ensure that the model can accurately identify the position of each window and utilize the temporal dependence within the window. Through this windowing process, time series data can be converted into a structure suitable for machine learning model processing, improving the accuracy and timeliness of flood forecasting.

[0078] S213, segmenting the historical precipitation data and the historical runoff data according to the time window structure to obtain a precipitation window sequence and a runoff window sequence;

[0079] Specifically, the length of the time window and the sliding step are determined, and then the original historical precipitation data and historical runoff data are segmented in time sequence according to these parameters. Assuming that the historical data starts from a certain date, first, from the first time point, the data within the first time window length is taken to form the first precipitation window and the first runoff window, then the window is slid forward by a step according to the set sliding step, and the data is taken again from the new position to form the next window data. This process will continue until the entire historical data set is divided into multiple time windows, and the data in each time window forms a window sequence, the precipitation data forms a precipitation window sequence, and the runoff data forms a runoff window sequence. The data in each window sequence is arranged in time sequence, representing the spatio-temporal variation characteristics of precipitation and runoff in that time period.

[0080] S214, index alignment of the precipitation window sequence and the runoff window sequence, and establishment of a time window correspondence.

[0081] Specifically, after dividing the precipitation and runoff data into time window, the obtained precipitation window sequence and runoff window sequence are independent, each of which contains different time period data. In order to ensure that each precipitation window can be correctly matched with its corresponding runoff window, it is necessary to align them based on the time stamp. The specific method is to record the start time and end time of each window, and pair the precipitation window with the runoff window according to the time index. Assuming that the window length and sliding step of the precipitation window sequence and the runoff window sequence are the same, and the window start time and the window end time can be one-to-one corresponding. In this way, the time period of each precipitation window and runoff window can form a one-to-one corresponding relationship, ensuring that they are processed synchronously in the model training process.

[0082] In actual operation, an index mapping table or data structure (such as a dictionary, DataFrame, etc.) is established to store the corresponding relationship between each precipitation window and runoff window, and the table records the time range of each precipitation window and the time range of the corresponding runoff window. Through this method, it is ensured that the precipitation and runoff data can be correctly paired in subsequent model training and evaluation, avoiding prediction errors caused by time window mismatch.

[0083] S22, based on the time window segmentation result, window feature extraction is performed on the historical precipitation data and the historical runoff data to obtain precipitation window parameters and runoff window parameters;

[0084] In the embodiments of the present application, the long short-term memory network algorithm is used to construct a window influence model, and the precipitation window parameters and the runoff window parameters are brought into the window influence model for training and optimization, and the flood influence parameter set is generated through the trained and optimized window influence model, which includes the following steps:

[0085] S221, using the long short-term memory network algorithm, a window influence model is constructed according to the precipitation window parameters and the runoff window parameters;

[0086] Specifically, a training data set is constructed according to the precipitation window and the runoff window. Each precipitation window parameter and runoff window parameter represents the historical precipitation and flow information in a time window. These window sequences serve as input features, which will represent the past precipitation and flow states respectively.

[0087] Next, the input of the long short-term memory network is the time series data composed of these window parameters. Specifically, the precipitation window and the runoff window serve as the input sequence of the long short-term memory network, and the long short-term memory network model can learn the time series dependence of past precipitation and flow through its memory unit. In order to build a suitable window influence model, first, the precipitation and runoff data are mapped to the input layer of the long short-term memory network, and the input data is arranged in the order of time window. Each window parameter contains information of multiple time steps, and the long short-term memory network learns the dynamic relationship between precipitation and runoff according to these information.

[0088] During the model training process, the precipitation and runoff data are input into the long short-term memory network as training samples, and the network automatically adjusts its weights to learn the change rule of precipitation and runoff in different time windows. After training, the long short-term memory network model can predict the corresponding runoff change according to the input precipitation window sequence, or further predict the occurrence of flood, and in order to improve the accuracy of the model, the hyperparameters of the model such as learning rate, hidden layer unit number, etc. can be adjusted through optimization algorithm (such as Adam or SGD) to ensure the fitting ability of the long short-term memory network model to time series data. Finally, the long short-term memory network model obtained by training is used to predict future flood flow or provide probability prediction of flood occurrence.

[0089] S222, input the precipitation window parameter and the runoff window parameter as input features, and construct the input layer of the window influence model;

[0090] Specifically, the precipitation and runoff data are divided according to the time window. Each time window contains a continuous precipitation and flow data in days. For example, if the time window length is set to 7 days, each window will contain the precipitation and flow data of the past 7 days. In the preprocessing stage, the precipitation and runoff data are usually normalized or standardized to eliminate the dimension difference, and then the precipitation and runoff data of each time window are combined into a feature matrix, each matrix contains two columns corresponding to precipitation and runoff. When constructing the input layer of the long short-term memory network model, these window data are spliced into a three-dimensional tensor with the shape of (sample number, time step length, feature number), where the time step length is 7 days and the feature number is 2 (precipitation and runoff). The input of each sample is a data matrix containing 7 days of precipitation and 7 days of runoff, and the long short-term memory network learns the time series dependence between precipitation and runoff through this structure to make flood prediction or flow prediction.

[0091] S223, input the precipitation window parameter and the runoff window parameter into the window influence model for training, and calculate the loss function of the window influence model;

[0092] Specifically, the preprocessed time-series data is organized into a format suitable for model input. Each training sample consists of a precipitation window and a runoff window. The precipitation window contains precipitation data from the past T time steps, and the runoff window contains runoff observations for the corresponding time period. This data is converted into a three-dimensional tensor input to the Long Short-Term Memory (LSTM) network. The tensor shape is (number of samples, time step, number of features), where the feature dimensions are arranged sequentially as precipitation and runoff. Model training employs supervised learning, using historical runoff windows as labeled data. Predicted runoff is calculated through forward propagation, and mean squared error (MSE) or mean absolute error (MAE) is used as the loss function to quantify the deviation between the predicted runoff and the actual observed values. During training, the Adam optimizer is used for backpropagation, and the weight parameters of the LSM network are iteratively updated using the gradient descent algorithm. An early stopping mechanism is also used to prevent overfitting.

[0093] S224. Evaluate the performance of the window impact model through cross-validation, adjust the window impact model based on the performance evaluation results and loss function, and generate a flood impact parameter set based on the adjusted window impact model.

[0094] Specifically, the process of evaluating the impact of cross-validation windows on model performance first requires dividing the dataset into multiple subsets. Each subset is used as a validation set, while the remaining subsets are used as training sets for multiple training and evaluation cycles. The specific steps are as follows: historical precipitation and runoff data are divided into K folds. During each training cycle, K-1 folds are used as the training set, and the remaining fold is used as the validation set. After each training cycle, performance metrics (such as mean squared error and mean absolute error) on the validation set are calculated, and finally, the average of the K rounds' results is calculated to obtain the comprehensive evaluation result of the model.

[0095] Based on the model performance evaluation results obtained from cross-validation, if the model's prediction accuracy is unsatisfactory, it can be optimized by adjusting model hyperparameters (such as the number of hidden units in the Long Short-Term Memory network, the learning rate, etc.), increasing training data, or changing the feature selection strategy. The loss function (such as mean squared error or mean absolute error) plays a crucial role in the training process, reflecting the difference between the model's predictions and the true values. By monitoring changes in the loss function, the model's structure and hyperparameters can be further adjusted until the loss function approaches its minimum, indicating that the model has achieved a good fit.

[0096] After adjusting the window impact model, the trained model is used to generate a flood impact parameter set. These parameter sets typically include predicted future runoff, peak flow, flood duration, etc., reflecting the relationship between precipitation and runoff within different time windows. By inputting precipitation and runoff windows, the model can output flood impact parameters for future periods, which can be used for flood warning and risk assessment.

[0097] S23, using a long short-term memory network algorithm to construct a window influence model, and bringing the precipitation window parameters and the runoff window parameters into the window influence model for training and optimization, and generating a flood influence parameter set through the optimized window influence model;

[0098] Specifically, the historical precipitation and runoff data are divided according to a preset time window to form a plurality of precipitation windows and runoff windows. For example, if the time window length is 7 days, each precipitation window and runoff window contains precipitation and flow data of the past 7 days. The data in each window forms a feature vector representing the precipitation and flow characteristics of the time period. Then, a long short-term memory network model is constructed, and the input layer of the model is composed of the feature data of the precipitation window and the runoff window. Usually, the two features are spliced into an input matrix. The input data shape of each sample is (time step T, feature number 2), where T represents the time window length (for example, 7 days), and the feature number is 2, representing the precipitation and runoff, respectively.

[0099] During the training process, the precipitation window and the runoff window are used as input, and the actual runoff is used as label data. The long short-term memory network is trained using a supervised learning method. By forward propagating the long short-term memory network model, the corresponding runoff can be predicted according to the input precipitation and runoff data, and then the difference between the predicted result of the model and the actual value is evaluated by a loss function (such as mean square error MSE). In each round of training, the optimization algorithm (such as Adam or SGD) adjusts the weights of the LSTM network, updates the model parameters through backpropagation, and gradually improves the prediction accuracy.

[0100] Through multiple training and optimization, the long short-term memory network model can finally accurately capture the time series relationship between precipitation and runoff, and predict the future runoff or flood occurrence when new precipitation data is input. After training, the optimized window influence model can generate a flood influence parameter set according to new precipitation data. The parameter set includes predicted flood flow, flow peak, flood duration, etc., which is used for flood warning and risk assessment.

[0101] S24, setting a weight influence rule, and optimizing and outputting the flood influence parameter set based on the weight influence rule.

[0102] Specifically, according to historical data and expert experience, the influence degree of each parameter (such as precipitation, basin area, soil moisture, etc.) on the occurrence of flood is determined, for example, precipitation is the main factor affecting flood flow, so a higher weight is given; and some auxiliary factors (such as soil moisture) can be adjusted according to their actual influence on runoff, then the weight is adjusted by using a machine learning model, the model can automatically optimize the weight through the training process to find the optimal relationship between each input feature and the influence of flood, after obtaining the weight rule, the next step is to optimize the flood influence parameter set. The optimization goal is usually to calculate the contribution of each influence parameter to the flood warning model according to these weight rules, and by adjusting the input features and parameter settings of the model, the accuracy and stability of the prediction are maximized. For example, if the prediction result of precipitation is more critical, increase the weight of precipitation related parameters; if some input features have less contribution to the influence of flood, reduce their weight or consider removing them.

[0103] Finally, the optimized flood influence parameter set can be used to generate the final flood warning and flow prediction. After weight optimization, the parameter set can more accurately reflect the influence of different features on the occurrence of flood and the change of flow, and provide more reliable flood warning and decision support data. These parameter sets include predicted flood flow, flow peak, flood duration, etc., which provide scientific basis for disaster prevention and mitigation and water resource management.

[0104] S3, presetting a river bearing rule, and establishing a river bearing model based on the river bearing rule and river structure parameters;

[0105] In the embodiment of the present application, the preset river bearing rule, and the river bearing model is established based on the river bearing rule and the river structure parameters, which includes the following steps:

[0106] S31, extracting geographical environment data, riverbed structure data and hydro-meteorological data in the river structure parameters, and setting river bearing capacity rules and flood bearing threshold parameters;

[0107] As a preferred solution, the extraction of geographical environment data, riverbed structure data and hydro-meteorological data in the river structure parameters, and the setting of river bearing capacity rules and flood bearing threshold parameters include the following steps:

[0108] S311, quantitatively processing the geographical environment data, riverbed structure data and hydro-meteorological data in the river structure parameters;

[0109] S312, setting river bearing capacity rules according to the quantified geographical environment data, riverbed structure data and hydro-meteorological data, and the river bearing capacity rules include river channel water carrying capacity rules and flood discharge capacity rules;

[0110] Specifically, according to the quantized geographic environment data, riverbed structure data and hydro-meteorological data, river bearing capacity rules are set, and these data are comprehensively analyzed to evaluate the carrying capacity and flood discharge capacity of the river. The river bearing capacity rules include river channel water carrying capacity rules and flood discharge capacity rules. The river channel water carrying capacity is mainly affected by the riverbed structure, flow rate and topography, and when setting, the maximum flow rate, flow rate limit and river channel capacity monitoring of the river channel need to be calculated. The flood discharge capacity is related to the design of the spillway, flood peak and flood discharge channel, and when setting, the flood warning, real-time monitoring and drainage system capacity need to be considered. Through numerical simulation, historical data analysis and real-time meteorological data, the river channel water carrying capacity and flood discharge capacity rules are quantized and dynamically adjusted to cope with different hydro-meteorological conditions, thereby effectively reducing the impact of flood disasters.

[0111] S313, according to the river bearing capacity rules, set the flood bearing threshold parameter to form the river bearing capacity benchmark, and set the river bearing capacity rules and the flood bearing threshold parameter based on the river bearing capacity benchmark.

[0112] Specifically, according to the river bearing capacity rules, the flood bearing threshold parameter is set and the river bearing capacity benchmark is formed, which needs to first clarify the physical limit of the river channel and the warning demand. By analyzing the quantized geographic environment data (such as watershed area, slope, land use type), riverbed structure data (such as cross-section shape, hydraulic radius, roughness coefficient) and hydro-meteorological data (such as historical rainfall, flow rate, air temperature, etc.), and systematically constructing the carrying capacity model of the river, the flood bearing threshold parameter such as the maximum safe water level, the maximum design flow rate and the maximum flood discharge capacity is set as the criterion for measuring whether the flood pressure exceeds the safety limit of the river channel.

[0113] Subsequently, these key threshold values are integrated into the river bearing capacity benchmark to define the safe operation boundary of the river channel under normal and extreme climate conditions. This benchmark can include different levels of warning levels, such as normal, alert, super alert and danger, which correspond to different flow rate or water level intervals. Based on this benchmark, further dynamic adjustment of the river bearing capacity rules is made, such as starting the flood discharge plan when the water level reaches a certain threshold, or reducing the reservoir capacity in advance when the predicted rainfall exceeds a certain cumulative amount. In addition, the flood bearing threshold parameter can also be adjusted in detail according to different river sections, seasons and upstream and downstream linkage relationships.

[0114] S32, establish a river bearing model according to historical runoff data and river bearing capacity rules;

[0115] Specifically, a river carrying capacity model is established based on historical runoff data and river carrying capacity rules. Historical runoff data, including flow rate, flow velocity, water level, and other information, are collected and analyzed. The river's maximum carrying capacity is determined by analyzing the trend of historical runoff data and extreme events, and is matched with the river carrying capacity rules.

[0116] Next, the river's water delivery capacity and flood discharge capacity are set according to the river carrying capacity rules. For example, based on the width, depth, slope, and other characteristics of the river channel, the maximum carrying capacity of the river under different flow rates and water levels is calculated to determine the maximum safe flow rate, maximum water level, and flood discharge threshold. These rules are adjusted based on actual hydrological data, flow data, and climate conditions in the river basin to ensure that the model can reasonably assess the carrying pressure of the river under extreme precipitation or flood conditions. In establishing the river carrying capacity model, appropriate mathematical and computational methods, such as hydrodynamic models, statistical regression analysis, or machine learning models, are used to learn from historical data and calculate the river carrying capacity under different flow rates and water levels. These models can simulate and predict the behavior of rivers under different environmental conditions, helping to analyze whether the river will be overloaded or overflow under various conditions, and thus providing a basis for flood control and disaster reduction decisions. Finally, the river carrying capacity model is formed through the training and verification of the model based on the historical runoff data of the river and the river carrying capacity rules, and is dynamically updated.

[0117] S33, quantize the river structure parameters and bring them into the river carrying capacity model for training and optimization combined with the river carrying capacity rules;

[0118] Specifically, the river structure parameters are quantized and brought into the river carrying capacity model for training and optimization combined with the river carrying capacity rules. The geometric data and hydraulic characteristics of the river, such as river width, depth, slope, roughness, flow velocity, and bed sand particle size, are collected. These parameters are obtained through field measurement, remote sensing technology, or hydrological monitoring equipment, and are standardized to eliminate dimensional differences. Then, the hydraulic characteristics, such as the water section, hydraulic radius, and water delivery capacity, are calculated using these parameters, and are calibrated with historical hydrological data.

[0119] On this basis, combined with the river bearing capacity rules, threshold values such as the maximum safe water level, the maximum flow, and the maximum flow rate are set, and these rules are converted into model constraints. For example, a hard constraint condition (such as data rejection or resampling when the flow exceeds the maximum bearing flow) is set, and a soft constraint (such as a penalty for water level or flow exceeding the threshold value) is added to the loss function to ensure that the predicted values output by the model meet the actual river bearing capacity. During the training process, an integrated method such as gradient boosting tree is used in combination with the hydrodynamic model for hybrid modeling. During model training, cross-validation and sensitivity analysis are used to select the optimal features, and hyperparameter optimization is used to improve the generalization ability of the model. Finally, the model is continuously iterated and calibrated to accurately predict the bearing capacity of the river under different hydrological conditions and output the key parameters required for flood warning and flood control decision-making.

[0120] S34, verifying the river bearing model optimized by training and outputting the river bearing model that passes the verification.

[0121] S4, bringing the historical flood impact parameters in the flood impact parameter set into the river bearing model for training optimization, and generating a flood forecasting scheme according to the river bearing model optimized by training;

[0122] In the embodiments of the present application, the step of bringing the historical flood impact parameters in the flood impact parameter set into the river bearing model for training optimization, and generating a flood forecasting scheme according to the river bearing model optimized by training includes the following steps:

[0123] S41, data cleaning is performed on the historical flood impact parameters in the flood impact parameter set, and the cleaned historical flood impact parameters are brought into the river bearing model for training;

[0124] S42, in the training process of the river bearing model, the output results of the model are compared with the historical flood observation results by using an error evaluation function, and a training error value is calculated;

[0125] In the embodiments of the present application, the step of comparing the output results of the river bearing model with the normalized historical flood observation results, and establishing a corresponding relationship between the output results of the river bearing model and the historical flood observation results includes the following steps:

[0126] S421, comparing the output results of the river bearing model with the normalized historical flood observation results, and establishing a corresponding relationship between the output results of the river bearing model and the historical flood observation results;

[0127] Specifically, to compare the output results of the river bearing model with the normalized historical flood observation results, it is necessary to normalize the historical flood data first, convert the flood data of different years and different river basins to the same dimension range, such as through minimum maximum normalization, so that the comparison of data has consistency. Historical flood observation results usually include flood peak, maximum water level, flow, duration, etc. These indicators need to be matched with the output results of the river bearing model, such as predicted flood flow, water level, etc.

[0128] In the comparison process, first, ensure that the output of the river bearing model is consistent with the time range and spatial position of the historical flood observation results. Then, calculate the relative error or similarity of each flood event, usually by calculating the mean square error (MSE) or mean absolute error (MAE) and other indicators to evaluate the matching degree of the model prediction results and the historical observation data, and by establishing a regression model or similarity function to explore the corresponding relationship between the output of the river bearing model and the historical flood observation results. For example, linear regression or nonlinear regression method is used to analyze the relationship between the output results and the observation results, and to identify which model output parameters (such as flow, water level) can best reflect the actual situation of historical floods.

[0129] S422, calculate the difference between the output results of the river bearing model and the historical flood observation results based on the error evaluation function, and obtain the training error value;

[0130] Specifically, the specific formula for measuring the deviation between the model prediction results and the actual observation results using the mean square error algorithm is:

[0131]

[0132] Where y i represents the historical flood observation results (true value), represents the output results of the river bearing model (predicted value), n is the sample number, and MSE measures the average square difference between the predicted value and the actual value, and the smaller the value, the more accurate the model prediction.

[0133] S423, statistics and analysis of the calculated training error value, and feedback the analysis results to the river bearing model training process for updating.

[0134] Specifically, the error values are summarized, usually using mean square error (MSE), mean absolute error (MAE) and root mean square error (RMSE) and other indicators, and for each round of training, record these error values and calculate their mean, standard deviation and other statistical quantities, so as to evaluate the training effect of the model.

[0135] In the analysis of error results, the following steps are taken:

[0136] Error distribution analysis: Analyze the distribution of errors to check for systematic biases (e.g., model predictions are consistently low or high). By plotting error distribution or box plots, identify outliers or bias patterns visually.

[0137] Overfitting and underfitting detection: Compare the errors of the training and validation sets to check for overfitting (low training error and high validation error) or underfitting (both training and validation errors are high). Error sensitivity analysis: Analyze the contribution of each feature to the error to understand which input parameters have a greater impact on the model output. Use sensitivity analysis methods (such as SHAP value analysis) to identify important features.

[0138] Based on the results of statistical analysis, use the feedback information to optimize the training process of the river bearing model. For example, if it is found that some features contribute more to the error, perform feature selection on the model, remove redundant features, or fine-tune important features. If overfitting is detected, adjust the regularization term, increase the training data, or modify the network structure to alleviate the overfitting problem; if underfitting exists, try to increase the complexity of the model or prolong the training time.

[0139] S43, preset model updating rules, and update the river bearing model based on the model updating rules and the training error value;

[0140] Specifically, define clear updating strategies and standards, including the following aspects:

[0141] Error threshold setting: According to the training error of the model (such as mean square error MSE, mean absolute error MAE, root mean square error RMSE, etc.), set an error threshold, when the error of the model reaches or exceeds the threshold, trigger the model update. The threshold is set according to the initial performance of the model, and is adjusted gradually as the training progresses.

[0142] Training rounds and early stopping rules: To avoid overfitting and underfitting, set the maximum training rounds and early stopping rules. When the training error continues to decrease, continue training; but if the validation set error starts to rise or the training error remains stable for several rounds, trigger the early stopping mechanism to prevent overfitting and save the current best model.

[0143] Learning rate adjustment: dynamically adjust the learning rate according to the change of training error. If the training error decreases slowly or stagnates, reduce the learning rate to fine-tune the model parameters; if the error decreases faster, you can appropriately increase the learning rate to speed up the model convergence.

[0144] Regularization term adjustment: If overfitting occurs, the model complexity can be limited by increasing the regularization term (such as L1, L2 regularization) to improve the generalization ability of the model. The regularization coefficient is dynamically adjusted according to the error evaluation index to balance the fitting degree and generalization ability.

[0145] Data augmentation and feature selection: If the training error is large and difficult to reduce, use data augmentation method to generate more training samples, or remove redundant features through feature selection to improve the efficiency and accuracy of the model.

[0146] Based on the above preset model update rules, the training error value is used as a feedback signal to help automatically adjust the model. After each round of training, the error value is calculated and compared with the preset threshold. If the error exceeds the threshold, the learning rate, regularization coefficient or other hyperparameters are adjusted according to the rules, and the model is retrained.

[0147] S44, the historical flood impact parameter is brought into the updated river bearing model to generate the corresponding flood forecasting scheme.

[0148] Specifically, the historical flood impact parameters (such as peak time and duration, disaster-causing water level / flow threshold, inundation depth and range, disaster-affected population and key infrastructure exposure, economic loss coefficient, emergency handling trigger line, etc.) are brought into the updated river bearing model to generate the flood forecasting scheme, and the "impact- working condition-response" mapping closed loop is realized: First, standardize the impact parameters and align them with the same period hydrology-hydrodynamic observation (rainfall, inflow, tide level, gate working condition) time, construct "impact label-boundary condition-control variable" sample pair, use the latest parameter set of the model to perform ensemble rolling prediction on multiple scenario boundaries (ensemble rainfall, upstream inflow, tide superposition and gate scheduling scheme) within the future prediction window, output hourly flow-water level-flow velocity field and its confidence interval; Then, the model output is converted into impact indicators (estimated inundation depth / area, threatened population and facilities, potential economic loss) through exposure and vulnerability functions, and compared with the warning / super warning / danger threshold in the river bearing capacity benchmark, to automatically determine the warning level and trigger period; On this basis, a scheme matrix is generated: containing recommended scheduling sequence (flood discharge gate opening degree and time, reservoir capacity pre-discharge target, flood diversion / flood detention opening and closing threshold), water level-flow-arrival time table of key cross sections and sensitive points, zoned warning and graded control list of prone areas (sealed roads, temporary pump station start-stop, embankment patrol intensity), personnel relocation and material preposition list; At the same time, the uncertainty is explained (subjective and objective probability, source of scenario divergence, key sensitive parameters) and the reanalysis update mechanism is given: When a new batch of real-time rainfall data / working condition data arrives, trigger data assimilation and short-term reforecast, rolling correction of scheme timing and intensity, and record the scheme-observed deviation to backfill the model, forming an executable, traceable, confidence interval and trigger condition flood forecasting scheme.

[0149] S5, verifying and optimizing the flood forecasting scheme according to the future precipitation prediction parameter, and outputting the flood disaster forecasting method based on the verified and optimized flood forecasting scheme.

[0150] In the embodiment of the present application, the step of verifying and optimizing the flood forecasting scheme according to the future precipitation prediction parameter, and outputting the flood disaster forecasting method based on the verified and optimized flood forecasting scheme comprises the following steps:

[0151] S51, setting a verification and optimization error threshold, and calculating a precipitation difference value between the future precipitation prediction parameter and the historical precipitation data;

[0152] Specifically, a reasonable error threshold is set according to actual needs. Generally, the error threshold can be set by historical data analysis and early warning needs, for example, the maximum allowed error can be set according to the standard deviation of historical precipitation data, or a smaller error threshold can be set according to the early warning sensitivity. For the calculation of the precipitation difference, the time scale of the future precipitation prediction data and the historical precipitation data needs to be aligned first. Then, the difference between the prediction value and the historical observation value is quantified by calculating the absolute difference. The prediction accuracy is evaluated by these difference values, and compared with the preset error threshold to determine whether the model needs to be optimized.

[0153] S52, comparing the verification and optimization error threshold with the precipitation difference value, and generating a verification and optimization strategy based on the comparison result;

[0154] In the embodiment of the present application, the step of comparing the verification and optimization error threshold with the precipitation difference value, and generating a verification and optimization strategy based on the comparison result comprises the following steps:

[0155] S521, setting a verification and optimization error threshold, and calculating a precipitation difference value between the future precipitation prediction parameter and the historical precipitation data;

[0156] S522, setting an optimization strategy threshold range and an optimization strategy set, comparing the precipitation difference value with the verification and optimization error threshold, and judging whether the difference value is within the optimization strategy threshold range based on the comparison result;

[0157] If yes, a verification and optimization strategy of maintaining the existing flood forecasting scheme is generated;

[0158] If no, a verification and optimization strategy of adjusting the flood forecasting scheme is generated.

[0159] S53, verifying and optimizing the flood forecasting scheme according to the verification and optimization strategy;

[0160] S54, adding the future precipitation prediction parameter into the verified and optimized flood forecasting scheme to generate the flood disaster forecasting method output.

[0161] Specifically, future precipitation prediction parameters are obtained, which are usually derived from meteorological forecast models or historical data analysis, covering information such as precipitation amount and intensity at different time periods. These precipitation prediction data are standardized to align with other river monitoring data such as flow and historical water levels. Ensure that the time scale of precipitation data is consistent with the time steps of the flood model, usually hourly or daily.

[0162] Then, the future precipitation prediction parameters are used as input variables, combined with the optimized river bearing model, to generate future flood scenario predictions through hydrodynamic models or flow simulation methods. Using these predicted precipitation data, water level, flow and velocity predictions under different scenarios are generated. Uncertainty analysis is performed on these prediction results to calculate disaster thresholds such as maximum water level and maximum flow under different scenarios, and compare them with river bearing capacity benchmarks such as flood warning lines to generate disaster predictions of different warning levels.

[0163] Next, flood disaster prediction output is generated using flood prediction results and future precipitation parameters, including the following key points:

[0164] Predict flood water level and flow: Based on precipitation prediction and river bearing model output, generate water level and flow predictions at different time periods to determine whether flood warning or flood discharge will be triggered.

[0165] Risk assessment: Combine geographic information, river carrying capacity and disaster thresholds to calculate potential flood impact areas, predict affected population, infrastructure exposure risk and economic loss.

[0166] Emergency response plan: Output emergency response plans according to different flood scenarios, such as evacuating people in advance, opening flood discharge channels, reinforcing embankments, etc.

[0167] Warning level setting: Generate flood warning levels based on water level, flow and historical flood data comparison, and clearly define the time and scope of the warning.

[0168] Finally, through these steps, future precipitation prediction parameters are effectively integrated into the flood prediction method.

[0169] In summary, by utilizing the above-mentioned technical solutions of this invention, the present invention improves the accuracy of flood forecasting schemes by combining spatiotemporal feature fusion and data-driven algorithms. Furthermore, by extracting spatiotemporal parameters of river structure, precipitation, and runoff from multiple dimensions, the generated flood impact parameter set comprehensively and accurately reflects various factors influencing flood occurrence. Moreover, based on different time windows and sliding step sizes, an impact model is constructed using a long short-term memory network algorithm to achieve deep modeling of complex flood change processes, thereby significantly improving the accuracy and timeliness of forecasts. Simultaneously, a training and optimization method based on an error evaluation function is employed to continuously optimize the river bearing capacity model by comparing it with historical flood observation results, ensuring that the model's response to actual flood disasters is more accurate. Finally, through cross-validation and training error analysis, the model is flexibly adjusted and optimized, enhancing its adaptability and reliability in actual flood disaster forecasting.

[0170] Furthermore, when generating flood forecasting schemes, this invention verifies and optimizes existing forecasting schemes using future precipitation prediction parameters. This ensures that the flood forecasting schemes maintain high reliability and accuracy in the face of different meteorological changes. By setting verification and optimization error thresholds and generating verification and optimization strategies, it flexibly adapts to forecasting needs under different environmental conditions, further improving the execution effect of disaster forecasting schemes. At the same time, through automated feature extraction and training optimization mechanisms, it reduces the need for human intervention, improves the operability and scalability of the schemes, and ensures the efficiency and continuity of the entire flood forecasting process by constructing a multi-path graphical user interface testing process and anomaly detection mechanism, thereby enhancing the system's emergency response capability to sudden flood situations.

[0171] Furthermore, this invention improves the reliability of flood forecasting through structured feature extraction and optimization, reduces computational costs during model training and updates, and achieves efficient flood disaster early warning at a lower cost by combining historical data with real-time predictions. It has broad application prospects. Moreover, this invention employs spatiotemporal fusion of multiple data sources and long short-term memory network algorithms, making this method not only applicable to flood forecasting in a single river basin, but also extendable to flood forecasting and early warning in different river basins and different climate zones. This provides a comprehensive solution to cope with various complex and dynamic flood disaster scenarios and is suitable for intelligent flood disaster management and emergency decision-making in variable environments.

[0172] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A flood disaster forecasting method based on spatio-temporal feature fusion, characterized in that, The method comprises the following steps: S1, collecting river structure parameters, future precipitation prediction parameters, historical precipitation data and historical runoff data, and performing data preprocessing in time sequence; S2, using a data-driven algorithm to extract precipitation spatiotemporal parameters and runoff spatiotemporal parameters of the historical precipitation data and the historical runoff data to generate a flood impact parameter set; S3, presetting a river bearing rule, and establishing a river bearing model based on the river bearing rule and the river structure parameters; S4, bringing historical flood impact parameters in the flood impact parameter set into the river bearing model for training and optimization, and generating a flood forecast scheme according to the river bearing model after training and optimization; S5, verifying and optimizing the flood forecast scheme according to the future precipitation prediction parameters, and outputting the flood disaster prediction method as the flood forecast scheme after verification and optimization.

2. The flood disaster forecasting method based on spatio-temporal feature fusion according to claim 1, characterized in that, The method of using a data-driven algorithm to extract precipitation spatiotemporal parameters and runoff spatiotemporal parameters of the historical precipitation data and the historical runoff data to generate a flood impact parameter set comprises the following steps: S21, performing normalization processing on the historical precipitation data and the historical runoff data, constructing a sliding data structure based on a time sequence window, and performing time window segmentation on the precipitation data and the runoff data; S22, performing window feature extraction on the historical precipitation data and the historical runoff data based on the time window segmentation result to obtain precipitation window parameters and runoff window parameters; S23, using a long short-term memory network algorithm to construct a window impact model, bringing the precipitation window parameters and the runoff window parameters into the window impact model for training and optimization, and generating a flood impact parameter set through the window impact model after training and optimization; S24, setting a weight influence rule, and optimizing and outputting the flood impact parameter set based on the weight influence rule. 3.The flood disaster forecasting method based on spatio-temporal feature fusion according to claim 1, characterized in that, The method of presetting a river bearing rule and establishing a river bearing model based on the river bearing rule and the river structure parameters comprises the following steps: S31, extracting geographical environment data, riverbed structure data and hydro-meteorological data from the river structure parameters, and setting a river bearing capacity rule and a flood bearing threshold parameter; S32, establishing a river bearing model according to the historical runoff data and the river bearing capacity rule; S33, quantifying the river structure parameters, and bringing the river structure parameters into the river bearing model for training and optimization in combination with the river bearing capacity rule; S34, verifying the river bearing model after training and optimization, and outputting the river bearing model that passes the verification.

4. The flood disaster forecasting method based on spatio-temporal feature fusion according to claim 1, characterized in that, The method of bringing historical flood impact parameters in the flood impact parameter set into the river bearing model for training and optimization, and generating a flood forecast scheme according to the river bearing model after training and optimization comprises the following steps: S41, performing data cleaning on the historical flood impact parameters in the flood impact parameter set, and bringing the cleaned historical flood impact parameters into the river bearing model for training; S42, in the training process of the river bearing model, comparing the model output result with the historical flood observation result through an error evaluation function, and calculating a training error value; S43, presetting a model updating rule, and updating the river bearing model based on the model updating rule and the training error value; S44, the historical flood impact parameter is brought into the updated river bearing model to generate a corresponding flood forecasting scheme.

5. The flood disaster forecasting method based on spatio-temporal feature fusion according to claim 1, characterized in that, The verification and optimization of the flood forecasting scheme according to the future precipitation prediction parameter, and the verification and optimization of the flood forecasting scheme as the flood disaster forecasting method output include the following steps: S51, a verification and optimization error threshold is preset, and a precipitation difference value of the future precipitation prediction parameter and the historical precipitation data is calculated; S52, the verification and optimization error threshold is compared with the precipitation difference value, and a verification and optimization strategy is generated based on the comparison result; S53, the verification and optimization strategy is used to verify and optimize the flood forecasting scheme; S54, the future precipitation prediction parameter is added to the verification and optimization of the flood forecasting scheme to generate the flood disaster forecasting method output.

6. The flood disaster forecasting method based on spatio-temporal feature fusion according to claim 2, characterized in that, The normalization processing of the historical precipitation data and the historical runoff data, the construction of the sliding data structure based on the time window, and the time window segmentation of the precipitation data and the runoff data include the following steps: S211, the historical precipitation data and the historical runoff data are filled with missing values and corrected with abnormal values, and normalized with a unified scale; S212, a time window length and a sliding step are preset, and a time window structure is constructed based on the time window length and the sliding step; S213, the historical precipitation data and the historical runoff data are segmented according to the time window structure to obtain a precipitation window sequence and a runoff window sequence; S214, the precipitation window sequence and the runoff window sequence are indexed and aligned, and a time window corresponding relationship is established.

7. The flood disaster forecasting method based on spatio-temporal feature fusion according to claim 6, characterized in that, The window influence model is constructed using the long short-term memory network algorithm, the precipitation window parameter and the runoff window parameter are brought into the window influence model for training and optimization, and the flood impact parameter set is generated through the trained window influence model, which includes the following steps: S221, the long short-term memory network algorithm is used, and the window influence model is constructed according to the precipitation window parameter and the runoff window parameter; S222, the precipitation window parameter and the runoff window parameter are used as input features, and the input layer of the window influence model is constructed; S223, the precipitation window parameter and the runoff window parameter are brought into the window influence model for training, and the loss function of the window influence model is calculated; S224, the performance of the window influence model is evaluated through cross-validation, the window influence model is adjusted according to the performance evaluation result of the window influence model and the loss function, and the flood impact parameter set is generated according to the adjusted window influence model.

8. The flood disaster forecasting method based on spatio-temporal feature fusion according to claim 3, characterized in that, The geographic environment data, riverbed structure data and hydro-meteorological data in the river structure parameter are extracted, and the river bearing capacity rule and the flood bearing threshold parameter are set, which include the following steps: S311, the geographic environment data, riverbed structure data and hydro-meteorological data in the river structure parameter are quantitatively processed; S312, the river bearing capacity rule is set according to the quantized geographic environment data, riverbed structure data and hydro-meteorological data, and the river bearing capacity rule includes the river water carrying capacity rule and the flood discharge capacity rule; S313, setting a flood bearing threshold parameter according to the river bearing capacity rule, forming a river bearing capacity benchmark, and setting the river bearing capacity rule and the flood bearing threshold parameter based on the river bearing capacity benchmark.

9. The flood disaster forecasting method based on spatio-temporal feature fusion according to claim 4, characterized in that, The step of comparing the model output result with the historical flood observation result through the error evaluation function and calculating the training error value in the river bearing model training process comprises the following steps: S421, comparing the output result of the river bearing model with the normalized historical flood observation result, and establishing a corresponding relationship between the output result of the river bearing model and the historical flood observation result; S422, calculating the difference between the output result of the river bearing model and the historical flood observation result based on the error evaluation function to obtain a training error value; S423, statistically analyzing the calculated training error value, and feeding back the analysis result to the river bearing model training process for updating.

10. The flood disaster forecasting method based on spatio-temporal feature fusion according to claim 5, characterized in that, The step of comparing the verification optimization error threshold value with the precipitation difference value and generating a verification optimization strategy based on the comparison result comprises the following steps: S521, setting a verification optimization error threshold value, and calculating a precipitation difference value according to the future precipitation prediction parameter and the historical precipitation data; S522, presetting an optimization strategy threshold value range and an optimization strategy set, comparing the precipitation difference value with the verification optimization error threshold value, and judging whether the difference value is within the optimization strategy threshold value range based on the comparison result; If yes, a verification optimization strategy of maintaining the existing flood forecasting scheme is generated; If no, a verification optimization strategy of adjusting the flood forecasting scheme is generated.

Images