Basin flood runoff forecasting system and forecasting method based on water conservancy big data

Abstract

The present invention provides a basin flood runoff forecasting system and a forecasting method based on water conservancy big data, and relates to the technical field of flood runoff prediction. The present invention constructs a basin flood runoff forecasting system based on water conservancy big data by integrating the advantages of a physical model and a data-driven model. A bidirectional LSTM network is used for outlier interpolation, which can effectively improve data quality. The HEC-HMS model optimizes the accuracy of the runoff process in combination with a terrain correction factor. The LSTM network enhances the data-driven ability of flow prediction by fusing geographic spatial features with time series data. A model fusion module introduces an attention mechanism and water conservation constraints, realizes a dynamic weight balance between a physical model and a data-driven model, and comprehensively improves the reliability and consistency of the forecast results. The model verification module realizes continuous comparison and verification of the forecast results with measured data through a closed-loop feedback mechanism of total error and dynamic threshold.

Description

Technical Field

[0001] The present invention relates to the technical field of flood runoff prediction, and in particular to a basin flood runoff forecasting system and a forecasting method based on water conservancy big data. Background Art

[0002] As an important means of flood prevention and disaster reduction, the accuracy and timeliness of flood runoff forecasting are directly related to the scientific nature of decision-making and the effectiveness of emergency response. However, traditional flood forecasting methods have obvious limitations in data acquisition, model accuracy, and real-time performance, making it difficult to meet the needs of modern flood prevention and control for high-precision forecasting.

[0003] At present, flood forecasting mainly relies on a single modeling method of physical models (such as HEC-HMS) and data-driven models (such as LSTM networks). Although the physical model has good interpretability of the physical mechanism, the selection of initial parameters and the accuracy of terrain data have a great influence on the model results, and it is easy to have deviations under complex terrain and extreme rainfall conditions. The data-driven model has a strong ability to learn time series features and performs well in flow forecasting, but its neglect of physical mechanisms leads to unreliable predictions under extreme events. At the same time, it has a strong dependence on data quality. When faced with outliers and missing values ​​in measured data, the prediction accuracy drops significantly.

[0004] Existing technologies still have obvious deficiencies in model fusion and data processing. Traditional fusion methods are mostly based on simple weighted averages or empirical formulas, which are difficult to dynamically adapt to the characteristics of different hydrological conditions, resulting in the inability to give full play to the comprehensive performance of the model. At the same time, in the data preprocessing process, the identification and interpolation methods of outliers are relatively backward, which can easily introduce prediction errors. In addition, the verification and optimization process of a single model lacks an effective closed-loop feedback mechanism, making it difficult to achieve dynamic adjustment of model parameters and continuous improvement of forecast accuracy. These problems limit the overall performance and application effect of the flood forecasting system.

[0005] The above information disclosed in this Background section is only for enhancement of understanding of the background of the present disclosure and therefore it may contain information that does not constitute the prior art that is already known to one of ordinary skill in the art. Summary of the invention

[0006] The purpose of the present invention is to provide a basin flood runoff forecasting system and forecasting method based on water conservancy big data to solve the problems raised in the above background technology.

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

[0008] A basin flood runoff forecasting system based on water conservancy big data, including:

[0009] The data acquisition module is used to collect the original time series data of the monitoring stations and the geographical location information of the monitoring stations from the hydrological monitoring network of the basin to be forecasted, calculate the dynamic threshold by using the sliding window technology, identify and mark the outliers, and interpolate the outliers through the bidirectional LSTM network to obtain the cleaned time series data. The original time series data includes flow and rainfall;

[0010] The physical model module is used to input the cleaned time series data into the HEC-HMS model, calculate the flow of the basin, and adjust the SCS unit line by introducing the terrain correction factor so that the HEC-HMS model outputs the runoff process line;

[0011] The driving model module is used to extract geographic spatial features by monitoring the geographic location information of the site, splice it with the cleaned time series data, input it into the LSTM network model for training, constrain the network output through the gradient penalty term, and let the LSTM network model output data-driven traffic prediction results;

[0012] Model fusion module, which is used to fuse the runoff process line and flow prediction results through the attention mechanism, calculate the dynamic weight to balance the contribution of the two models, introduce water conservation constraints, and finally generate a comprehensive forecast result;

[0013] The model verification module is used to compare the forecast results with the measured data, calculate the total error, and evaluate the model performance through the total error. If it is lower than the threshold, the parameter optimization module is triggered to adjust the model parameters and re-execute the modeling process. After verification, the final flood forecast results are output.

[0014] Furthermore, the specific logic for identifying and marking outliers is:

[0015] At the monitoring site, with the current real-time moment as the center, slide the 30-day window forward to extract the corresponding original time series data and discretize it into equal time intervals. Align the sequence data of flow and rainfall according to the timestamps, and calculate the arithmetic mean and standard deviation of the sequence data of flow and rainfall in the window respectively. The formula is as follows: in, and Respectively represent Arithmetic mean and standard deviation of class sequence data, Indicates the category of sequence data, represents the flow sequence data, represents rainfall series data, Represents the total number of data points of the discretized sequence data, In the discretized sequence data, Class sequence data data points;

[0016] The coefficient of variation is obtained by dividing the standard deviation by the mean, multiplied by the fixed proportional coefficient and added to the dynamic threshold. The formula is: in, Indicates Dynamic thresholding for sequence-like data, is the coefficient of variation;

[0017] The data points in the discrete flow and rainfall sequence data segment are compared point by point with the absolute difference between the current value and the dynamic threshold to obtain the abnormal point. , mark the point as an outlier.

[0018] Furthermore, the specific logic for obtaining the cleaned time series data is as follows:

[0019] For each outlier point, the data of the 5 data points before and after it are intercepted. If the data exceeds the data boundary, the missing part is filled with zeros, and the zero-value placeholder is filled to form an 11-dimensional input vector sequence;

[0020] The input vector sequence is input into a pre-trained bidirectional LSTM network, the input dimension of the bidirectional LSTM network is 11, the output dimension is 1, and the bidirectional LSTM network outputs the interpolation value at the corresponding moment. After filling the abnormal point, the cleaned time series data is obtained.

[0021] Furthermore, the specific logic of the runoff process line output by the HEC-HMS model is:

[0022] The cleaned time series data is divided into a rainfall input sequence and a runoff output sequence. The rainfall sequence data is used as the rainfall input sequence, and the flow sequence data is used as the runoff output sequence.

[0023] Extract the geographical location information of the monitoring site, use the GIS database, take the monitoring site coordinates as the center and the radius of 3 kilometers as the monitoring area, and obtain the slope data in the monitoring area;

[0024] Extract all slope values ​​in the monitoring area from the GIS database, traverse all slope values, find the maximum value, and use the maximum slope value as slope characteristic data;

[0025] The terrain correction factor is calculated through the slope characteristic data, and the SCS unit line built into HEC-HMS is corrected based on the formula: in, represents the terrain correction factor, Indicates the angle value corresponding to the slope characteristic data, is the gain factor, and , and They represent the exponential coefficients of the unit line peak time and unit line attenuation segment of the corrected SCS unit line, and They represent the exponential coefficients of the unit line peak time and unit line attenuation segment of the original SCS unit line respectively;

[0026] The corrected SCS unit line, rainfall input sequence and runoff output sequence are input into the HEC-HMS model to output the runoff process line.

[0027] Furthermore, the specific logic of obtaining training data for inputting into the LSTM network model for training is as follows:

[0028] For the monitoring area corresponding to the monitoring station, GIS tools are used to extract elevation data in the monitoring area, and the elevation standard deviation in the monitoring area is calculated as the elevation feature data;

[0029] Extract the normalized difference vegetation index data in the monitoring area and calculate the arithmetic mean as the vegetation characteristic data of the monitoring area;

[0030] From the soil type database, extract the soil type information of the area where the monitoring station is located according to the geographical coordinates of the monitoring station, perform one-hot encoding on the soil type, and convert it into a numerical vector as soil characteristic data;

[0031] The slope characteristic data, elevation characteristic data, vegetation characteristic data and soil characteristic data are associated with the data of each data point in the discrete rainfall sequence data to form an input vector as the input of the LSTM network model. The flow of the corresponding next data point is used as the label of the LSTM network model to train the LSTM network model.

[0032] Furthermore, the specific logic for the LSTM network model to output data-driven runoff prediction results is:

[0033] Input the input vector to the trained LSTM network model to obtain the output traffic sequence of the LSTM network model , and expressed as: in, Indicates the output traffic sequence data predicted by the LSTM network model. The predicted data for each data point;

[0034] The predicted flow change of adjacent data points is calculated based on the formula: in, Indicates The gradient of the data point;

[0035] The square sum of the gradients is used as a penalty term and added to the model's loss function. The formula is: in, represents the penalty loss function, represents the hyperparameter that controls the strength of the penalty term;

[0036] The penalty loss function is combined with the original loss function based on mean square error to form a new loss function, based on the formula: in, Represents the new loss function formed, The actual measurement Traffic data for each data point;

[0037] To minimize the comprehensive loss function As the goal, adjust the parameters of the LSTM model and obtain the traffic prediction results output by the LSTM network model.

[0038] Furthermore, the specific logic for generating comprehensive forecast results is:

[0039] The runoff hydrograph is defined as: ,in Indicates the first The observed flow rate of data points, Represents the total number of data points of the discretized sequence data;

[0040] The traffic prediction result output by the LSTM network model is defined as: ,in Indicates the traffic prediction result. The predicted flow rate for each data point;

[0041] The discretized flow sequence data is subjected to fast Fourier transform to obtain the frequency domain characteristics of the flow sequence data, specifically: in, represents frequency characteristic data, Indicates the extracted The amplitude of the characteristic frequency, is the total number of features;

[0042] The attention weight is calculated based on the formula: in, represents the attention weight, and represents the weight matrix and bias, express Activation function, Indicates normalizing the weights;

[0043] The formula for obtaining the comprehensive fusion result by fusing the runoff process line and the prediction result according to the attention weight is: in, represents the comprehensive fusion result, In the comprehensive fusion result, The flow of data points, The first column represents the fused runoff process line and the prediction result. The flow of data points, represents the water conservation constraint, Represents the flow adjustment, which is calculated through the water conservation constraint.

[0044] Furthermore, the specific formula for evaluating the model performance through the total error is: in, Represents the total error, if , the forecast system is used to output the final flood forecast result, otherwise the internal parameters of the physical model module, the driving model module and the model fusion module are readjusted. Represents the validation threshold.

[0045] The present invention further provides a basin flood runoff forecasting method based on water conservancy big data, the forecasting method is performed by the basin flood runoff forecasting system based on water conservancy big data, and the specific steps include:

[0046] The original time series data and geographical location information of the monitoring stations are collected from the hydrological monitoring network of the basin to be forecasted, and the dynamic threshold is calculated using the sliding window technology to identify and mark the outliers. The outliers are interpolated through the bidirectional LSTM network to obtain the cleaned time series data, and the original time series data includes flow and rainfall;

[0047] The cleaned time series data are input into the HEC-HMS model to calculate the flow of the basin. By introducing the terrain correction factor and adjusting the SCS unit line, the HEC-HMS model outputs the runoff process line.

[0048] By monitoring the geographic location information of the site, we extract the geographic spatial features, splice them with the cleaned time series data, input them into the LSTM network model for training, and constrain the network output through the gradient penalty term, so that the LSTM network model can output data-driven traffic prediction results;

[0049] The runoff process line and flow prediction results are fused through the attention mechanism, the dynamic weight is calculated to balance the contribution of the two models, the water conservation constraint is introduced, and finally the comprehensive forecast result is generated;

[0050] The forecast results are compared with the measured data, the total error is calculated, and the model performance is evaluated by the total error. If it is lower than the threshold, the parameter optimization module is triggered, the model parameters are adjusted and the modeling process is re-executed. After verification, the final flood forecast results are output.

[0051] Compared with the prior art, the present invention has the following beneficial effects:

[0052] The present invention integrates the advantages of physical models and data-driven models to construct a basin flood runoff forecasting system based on water conservancy big data. The bidirectional LSTM network adopted for outlier interpolation can effectively improve data quality. The HEC-HMS model optimizes the accuracy of the runoff process in combination with the terrain correction factor. The LSTM network enhances the data-driven ability of flow prediction by fusing geographic spatial features with time series data. The model fusion module introduces the attention mechanism and water conservation constraints, realizes the dynamic weight balance of the physical model and the data-driven model, and comprehensively improves the reliability and consistency of the forecast results. The model verification module realizes the continuous comparison and verification of the forecast results with the measured data through the closed-loop feedback mechanism of the total error and the dynamic threshold. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 It is a schematic diagram of the overall system structure of the present invention;

[0054] Figure 2 It is a schematic diagram of the overall method flow of the present invention. DETAILED DESCRIPTION

[0055] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with specific embodiments.

[0056] It should be noted that, unless otherwise defined, the technical terms or scientific terms used in the present invention should be understood by people with ordinary skills in the field to which the present invention belongs. The words "first", "second" and similar words used in the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. "Include" or "comprise" and similar words mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. "Connect" or "connected" and similar words are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0057] Example

[0058] See also Figure 1 , the present invention provides a technical solution:

[0059] A basin flood runoff forecasting system based on water conservancy big data, including:

[0060] The data acquisition module is used to collect the original time series data of the monitoring stations and the geographical location information of the monitoring stations from the hydrological monitoring network of the basin to be forecasted, calculate the dynamic threshold using the sliding window technology, identify and mark the outliers, and interpolate the outliers through the bidirectional LSTM network to obtain the cleaned time series data. The original time series data includes flow and rainfall.

[0061] The data collection module collects the original time series data of the monitoring stations in the hydrological monitoring network of the basin to be forecasted in real time, ensuring the integrity and timeliness of the data. Compared with the traditional data collection method, this solution uses sliding window technology to dynamically calculate the threshold, so as to more accurately identify and mark outliers. This method not only improves the accuracy of outlier detection, but also can adapt to the data fluctuation characteristics of different time periods and different monitoring stations.

[0062] Through the bidirectional LSTM network, outlier interpolation is performed. Traditional interpolation methods usually rely on simple linear interpolation or statistical interpolation, ignoring the time dependency and nonlinear relationship of time series data. The bidirectional LSTM network can simultaneously utilize the information before and after the time series, capture complex time series patterns, and thus achieve more accurate data interpolation.

[0063] Hydrological monitoring stations are deployed in the basins to be forecasted to collect real-time time series data such as flow and rainfall, and the data is remotely transmitted to the data center through Internet of Things technology or satellite communication; secondly, the geographical location information of the monitoring station is entered into the system through the built-in GPS module or manual mapping, and is automatically associated with the time series data and stored in the spatial database. During the data collection process, the system synchronously integrates the timestamp and spatial coordinate information to ensure temporal and spatial consistency, and provides standardized input for subsequent dynamic threshold calculations, outlier interpolation, and flood forecasting models.

[0064] In this embodiment, the specific logic for identifying and marking outliers is:

[0065] At the monitoring site, with the current real-time moment as the center, slide the 30-day window forward to extract the corresponding original time series data and discretize it into equal time intervals. Align the sequence data of flow and rainfall according to the timestamps, and calculate the arithmetic mean and standard deviation of the sequence data of flow and rainfall in the window respectively. The formula is as follows: in, and Respectively represent Arithmetic mean and standard deviation of class sequence data, Indicates the category of sequence data, represents the flow sequence data, represents rainfall series data, Represents the total number of data points of the discretized sequence data, In the discretized sequence data, Class sequence data data points;

[0066] The coefficient of variation is obtained by dividing the standard deviation by the mean, multiplied by the fixed proportional coefficient and added to the dynamic threshold. The formula is: in, Indicates Dynamic thresholding for sequence-like data, is the coefficient of variation;

[0067] The data points in the discrete flow and rainfall sequence data segment are compared point by point with the absolute difference between the current value and the dynamic threshold to obtain the abnormal point. , mark the point as an outlier.

[0068] The purpose of "equal time interval discretization" is to convert continuous data into discrete data points to facilitate subsequent processing and analysis. For example, a sensor records flow data once a minute, generating a series of discrete data points, so that each data point has a fixed time interval, which is convenient for time series analysis and modeling, and helps capture the periodicity and trend of the data.

[0069] The flow and rainfall series data are aligned by timestamp because the two types of data may have different recording times. For example, a flow sensor may record once a minute, while a rainfall sensor may record once every five minutes. By aligning, we can ensure that there is corresponding data at each time point in the analysis, which is convenient for calculating dynamic thresholds and subsequent data analysis.

[0070] The calculation of dynamic thresholds uses the mean and standard deviation, so that the thresholds can be automatically adjusted according to the actual data and adapt to different data fluctuations. The coefficients in the formula, such as 3 and 0.3, are usually determined based on statistical experience or optimization processes to ensure that the thresholds can effectively identify outliers without being overly disturbed by data fluctuations. The data characteristics of flow and rainfall are different. Flow is affected by terrain and river factors, while rainfall is mainly affected by meteorological conditions. Calculating dynamic thresholds separately can more accurately identify their respective outliers and avoid the data fluctuations of one party affecting the anomaly detection of the other party.

[0071] Dynamic thresholds are used to compare with actual data points, and data points outside a certain range are marked as outliers. This method is similar to the control limit in statistical process control. By dynamically adjusting the threshold, the flexibility and accuracy of outlier detection are improved, data quality is ensured, and a reliable foundation is provided for subsequent data analysis and model training.

[0072] By dynamically calculating the threshold and using the bidirectional LSTM network to interpolate outliers, it is possible to more accurately identify and process outliers in the data, maintaining the integrity and continuity of the data sequence. In addition, the dynamic threshold method can adapt to the data fluctuation characteristics of different time periods and different monitoring sites, ensuring the flexibility and reliability of outlier processing, thereby improving the prediction accuracy and reliability of the flood forecasting model, comparing the absolute difference between the data point and the dynamic threshold, identifying and marking outliers, and preparing for subsequent outlier interpolation. Through these steps, the integrity and accuracy of the data can be ensured, providing reliable input data for the flood forecasting model.

[0073] Furthermore, the specific logic for obtaining the cleaned time series data is as follows:

[0074] For each outlier point, the data of the 5 data points before and after it are intercepted. If the data exceeds the data boundary, the missing part is filled with zeros, and the zero-value placeholder is filled to form an 11-dimensional input vector sequence;

[0075] The input vector sequence is input into a pre-trained bidirectional LSTM network, the input dimension of the bidirectional LSTM network is 11, the output dimension is 1, and the bidirectional LSTM network outputs the interpolation value at the corresponding moment. After filling the abnormal point, the cleaned time series data is obtained.

[0076] For each marked anomaly point, a data segment of 5 data points before and after it is intercepted with it as the center (a total of 11 time points). If the window exceeds the time series boundary (such as near the first or last point), the missing part is filled with zero values ​​to generate an 11-dimensional input vector sequence. The 11-dimensional vector is input into the pre-trained bidirectional LSTM network, which predicts the reasonable interpolation value of the current anomaly point by capturing the contextual dependency of the previous and next time series. The input dimension of the bidirectional LSTM is aligned with the length of the data segment (11 dimensions), and the output dimension is 1 (single-point interpolation value). The interpolation value output by the model replaces the original anomaly value to form a complete time series data after cleaning, ensuring the continuity and physical rationality of the sequence.

[0077] Bidirectional LSTM can model the dynamic correlation of physical processes such as flood peak propagation and rainfall-runoff response by jointly analyzing the flow and rainfall characteristics before and after the anomaly. The interpolated values ​​are more in line with hydrological laws. Compared with traditional mean filling or linear interpolation, LSTM can adaptively adjust the repair strategy according to the context of the anomaly (such as the starting section or flood peak section of the rainfall event) to avoid global errors caused by local fluctuations. Zero-filling processing avoids information loss caused by data truncation in the boundary area. At the same time, the zero-value mask mechanism in model training is used to ensure the adaptability of the interpolation process to the boundary.

[0078] The bidirectional LSTM layer extracts the forward and reverse features of the time series, and the fully connected layer maps them into scalar outputs. The interpolated values ​​are denormalized and then used to replace the original abnormal points. If there are multiple abnormal points in the same window, point-by-point serial processing is used to perform a smoothness test on the interpolated data segment (such as threshold filtering of gradient mutations at adjacent points), and the logical consistency is cross-validated with the upstream and downstream site data.

[0079] The pre-trained bidirectional LSTM network is designed to learn the intrinsic laws of flow and rainfall time series using historical hydrological data. First, we collect high-quality historical data from multiple monitoring stations, which are cleaned to ensure that there are no missing values ​​and outliers. Then, we use a sliding window method with 11 data points as a sample, where the center point is the target value and the 5 data points before and after are used as context information. With this design, the model is able to capture the front-end and back-end dependencies of time series data, such as the lagged effect of rainfall on flow.

[0080] During the pre-training process, we use a bidirectional LSTM structure. The forward LSTM processes the time series from the past to the present, and the backward LSTM processes the time series from the future to the present. This structure enables the model to use both past and future contextual information to more accurately predict the target value. The model's loss function uses the mean square error (MSE) and the parameters are updated through the Adam optimizer. To prevent overfitting, we introduced L2 regularization and Dropout technology, and used the early stopping method to monitor the validation set loss to ensure that the model stops training at the best state.

[0081] Through pre-training, the model not only learns the dynamic characteristics of hydrological data, but also has context-awareness and can accurately predict data values ​​under normal circumstances. This provides a reliable basis for subsequent anomaly interpolation in actual data, ensuring that the interpolated values ​​are both consistent with the local context and have global physical consistency.

[0082] The physical model module is used to input the cleaned time series data into the HEC-HMS model, calculate the flow of the basin, and adjust the SCS unit line by introducing the terrain correction factor so that the HEC-HMS model outputs the runoff process line.

[0083] By inputting the cleaned time series data into the HEC-HMS model and dynamically adjusting the runoff parameters in combination with the terrain correction factor, the data-driven interpolation results can be deeply integrated with the physical mechanism. The terrain correction factor can calibrate the surface runoff response rate of the SCS unit line, making the runoff process line output by the model closer to the real basin characteristics, making up for the shortcomings of the simple data-driven method in physical consistency. The cleaned time series data (flow, rainfall) are normalized according to the HEC-HMS input format, and the terrain correction factor is extracted based on the basin DEM data. The calculation formula of the SCS curve number is adjusted by the correction factor, such as reducing the infiltration rate and increasing the runoff coefficient in the steep slope area, to generate spatially differentiated unit line parameters. The corrected unit line parameters and rainfall input are imported into the model to run the runoff hydrological simulation. The surface runoff, base flow and river channel evolution process are calculated for each sub-basin, and the runoff process line of the outlet section of the entire basin is superimposed to generate.

[0084] In this embodiment, the specific logic of the HEC-HMS model output runoff process line is:

[0085] The cleaned time series data is divided into a rainfall input sequence and a runoff output sequence. The rainfall sequence data is used as the rainfall input sequence, and the flow sequence data is used as the runoff output sequence.

[0086] Extract the geographical location information of the monitoring site, use the GIS database, take the monitoring site coordinates as the center and the radius of 3 kilometers as the monitoring area, and obtain the slope data in the monitoring area;

[0087] Extract all slope values ​​in the monitoring area from the GIS database, traverse all slope values, find the maximum value, and use the maximum slope value as slope characteristic data;

[0088] The terrain correction factor is calculated through the slope characteristic data, and the SCS unit line built into HEC-HMS is corrected based on the formula: in, represents the terrain correction factor, Indicates the angle value corresponding to the slope characteristic data, is the gain factor, and , and They represent the exponential coefficients of the unit line peak time and unit line attenuation segment of the corrected SCS unit line, and They represent the exponential coefficients of the unit line peak time and unit line attenuation segment of the original SCS unit line respectively;

[0089] The corrected SCS unit line, rainfall input sequence and runoff output sequence are input into the HEC-HMS model to output the runoff process line.

[0090] The SCS unit line is dynamically corrected by slope characteristics, and the spatial heterogeneity of terrain is integrated into the traditional hydrological model, so that the simulation results are more in line with the physical characteristics of the actual basin. For example, the terrain correction factor value of the monitoring station in the mountain area is significantly greater than that of the plain station. The corrected unit line can accurately reflect the steepening characteristics of the flood wave. If the interpolation value in the data cleaning link deviates significantly from the HEC-HMS simulated runoff, the rationality of the interpolation value can be reversely verified to locate potential anomalies, forming a two-way error correction mechanism of "data cleaning-physical verification".

[0091] The coordinates of the monitoring stations are extracted from the hydrological database, and the slope extraction tool (such as SlopeAnalysis) is called through ArcGIS Engine to calculate the maximum value of the raster slope (θ) within a radius of 3 kilometers, with an accuracy of 0.1°. The rainfall and flow series are time-aligned to ensure that the HEC-HMS input timestamp is consistent (such as 1 hour). The historical flood event data is selected, and the Nash efficiency coefficient (NSE) is used as the optimization target, and the least squares method is used for fitting. value.

[0092] In hydrological forecasting, data quality and consistency are crucial. By cleaning and segmenting the flow time series data and processing rainfall and runoff data separately, the model can learn the relationship between rainfall and runoff more clearly. In addition, extracting the geographical location information of the monitoring station and obtaining slope data through the GIS database can provide the model with more accurate terrain features, especially in complex terrain areas. This approach can effectively reflect the impact of terrain on hydrological processes, thereby improving the reliability of the model. Only on the basis of accurate data and clear terrain features can subsequent model operation and parameter adjustment obtain reliable results.

[0093] Topographic characteristics, especially slope, have an important influence on runoff generation and confluence in hydrological processes. In the traditional HEC-HMS model, the parameters of the SCS unit line are usually calculated by empirical formulas, which cannot fully reflect the differences under different terrain conditions. By introducing a terrain correction factor and dynamically adjusting the peak time and attenuation coefficient of the SCS unit line based on slope characteristic data, the model can better fit the actual terrain conditions. This method is particularly suitable for basins with complex terrain, such as mountainous or hilly areas, so that the model can show high prediction accuracy under different terrain conditions, thereby better adapting to the complex natural environment.

[0094] Traditional HEC-HMS model parameters usually rely on empirical formulas or simple calibration methods, which are difficult to fully reflect the actual situation of the basin. By dynamically adjusting the SCS unit line parameters based on terrain characteristics, the model can better simulate the actual flood process. For example, the runoff velocity in steep slope areas is faster, and the terrain correction factor can accelerate the peak time of the unit line and respond to rainfall input more quickly. At the same time, through dynamic adjustment and real-time optimization of the gain coefficient, the model can adapt to different rainfall intensities and terrain conditions, thereby significantly improving the accuracy and stability of flood prediction. This parameter optimization method makes the model more flexible and efficient in practical applications.

[0095] The driving model module is used to extract geographic spatial features by monitoring the geographic location information of the site, splice it with the cleaned time series data, input it into the LSTM network model for training, constrain the network output through the gradient penalty term, and let the LSTM network model output data-driven traffic prediction results.

[0096] The deep fusion strategy of multi-source geographic data and time series data can significantly improve the LSTM network model's ability to represent complex hydrological processes. By coupling static geographic features such as slope, elevation, vegetation, and soil type with dynamic rainfall sequences, the model can simultaneously capture the spatial heterogeneity of the basin and the temporal dependence of the hydrological response. For example, the standard deviation of elevation reflects the degree of topography in the monitoring area, which directly affects the spatial allocation efficiency of precipitation runoff; the average value of the normalized difference vegetation index (NDVI) represents the state of surface vegetation coverage, which in turn affects the soil infiltration rate and evapotranspiration intensity; the unique hot encoding of soil type quantifies the differentiated regulatory effects of different textures on surface runoff and soil flow. The introduction of these static features enables the model to break through the limitations of traditional hydrological time series prediction that only relies on historical flow / rainfall data, especially in the case of data missing or extreme events (such as landslides caused by heavy rains leading to sudden changes in terrain), and compensate for local noise or anomalies in the time series through the spatial information of geographic features, thereby enhancing the robustness of the prediction.

[0097] First, a spatial analysis of the 3 km radius area around the monitoring site was conducted using GIS tools: the standard deviation of the raster elevation value was calculated using the digital elevation model (DEM) as a quantitative indicator of terrain complexity; the NDVI band was extracted from the remote sensing image, and the arithmetic mean in the monitoring area was calculated after eliminating cloud interference to characterize the spatiotemporal stability of vegetation coverage. Soil type data needs to be matched with high-precision geographic coordinates. For example, the classification code of the soil system where the monitoring point is located (such as clay loam, sandy soil, etc.) is extracted from the 1:50,000 soil database, and converted into a multidimensional vector through unique hot encoding to avoid information loss caused by single numericalization. The discretized rainfall series data is dynamically bound to the above-mentioned geographic features according to the timestamp (such as every hour) - each rainfall data point is associated with the slope, elevation, vegetation, and soil feature vectors at the corresponding time, forming an input unit that integrates spatiotemporal attributes. During the training process, the LSTM network takes the input units at the current moment and their historical sequences (such as the past 24 hours) as input, outputs the flow forecast value at the next moment, optimizes the network weights through back propagation, and finally forms a joint prediction model that can simultaneously analyze the "spatial geographic driving mechanism" and "temporal hydrological response law".

[0098] By introducing geographical static features, the model is essentially injected with a priori physical constraint information. For example, when rainfall falls in a clay area, the soil feature vector will automatically strengthen the model's recognition of the "low infiltration rate-high surface runoff" pattern; if the vegetation coverage in a certain area drops sharply (NDVI decreases), the model can quickly adjust the estimate of the runoff coefficient in combination with historical data. This spatiotemporal fusion training method not only improves the accuracy of the model in conventional scenarios, but also enables it to adapt to sudden changes in terrain or land use (such as reduced vegetation coverage caused by urbanization), providing a high-quality data foundation for subsequent collaborative verification and dynamic coupling with the HEC-HMS physical model.

[0099] In this embodiment, the specific logic of obtaining the training data input into the LSTM network model for training is:

[0100] For the monitoring area corresponding to the monitoring station, GIS tools are used to extract elevation data in the monitoring area, and the elevation standard deviation in the monitoring area is calculated as the elevation feature data;

[0101] Extract the normalized difference vegetation index data in the monitoring area and calculate the arithmetic mean as the vegetation characteristic data of the monitoring area;

[0102] From the soil type database, extract the soil type information of the area where the monitoring station is located according to the geographical coordinates of the monitoring station, perform one-hot encoding on the soil type, and convert it into a numerical vector as soil characteristic data;

[0103] The slope characteristic data, elevation characteristic data, vegetation characteristic data and soil characteristic data are associated with the data of each data point in the discrete rainfall sequence data to form an input vector as the input of the LSTM network model. The flow of the corresponding next data point is used as the label of the LSTM network model to train the LSTM network model.

[0104] The standard deviation of elevation reflects the complexity of the terrain and affects the distribution of runoff; the mean NDVI indicates vegetation cover, which is involved in infiltration and evaporation; and soil type affects runoff generation. Combining these characteristics with rainfall can enhance the model's understanding of spatial heterogeneity and temporal dynamics. This is particularly important in flood prediction because different terrains and surface conditions respond very differently to rainfall.

[0105] By integrating multivariate geospatial features with rainfall sequence data to construct the input of the LSTM model, the spatiotemporal coupling modeling capability of hydrological forecasting can be significantly enhanced. The introduction of geographic feature data (such as elevation standard deviation, vegetation index, and soil type) essentially transforms the spatial heterogeneity of the basin's underlying surface into quantifiable model inputs, allowing the LSTM network to simultaneously perceive the long-term impact of static factors such as topography, vegetation coverage, and soil permeability on hydrological processes in time series forecasting. For example, the elevation standard deviation can reflect the complexity of the terrain in the monitoring area - areas with steep slopes and undulating terrain will accelerate runoff convergence, while areas with low standard deviations may prolong water flow diffusion time due to flat terrain; the NDVI mean is directly related to vegetation cover density, and high vegetation cover areas weaken surface runoff intensity through canopy interception and root water absorption; the unique hot encoding of soil type can distinguish the key effects of textures such as sand and clay on infiltration rate. The association of these features with dynamic rainfall data enables the model to spontaneously learn the rainfall-runoff response patterns under geographic spatial constraints during training, rather than simply relying on the statistical correlation of time series, thereby improving the prediction stability of extreme events (such as differences in runoff production in areas with different soil types after heavy rain).

[0106] Feature fusion is achieved through georeferencing and spatiotemporal alignment of multi-source data. First, the spatial analysis scope (such as a circular buffer with a radius of 3 kilometers) is defined based on the coordinates of the monitoring site. The elevation values ​​of all grid points are extracted from the digital elevation model using GIS tools, and their standard deviations are calculated to quantify the degree of spatial variation of the terrain. At the same time, NDVI data is extracted from multi-temporal remote sensing images. After pixel-level quality control (such as cloud mask) is used to ensure data reliability, the regional mean is calculated to characterize the overall state of vegetation coverage. The soil type needs to be matched with high-precision geographic coordinates. For example, the soil type of the monitoring point is extracted from a soil database with a spatial resolution of 100 meters, and converted into a multidimensional vector through one-hot encoding (such as clay is encoded as [1,0,0] and sand is [0,1,0]) to avoid misjudgment of category relationships caused by traditional numerical encoding. Discrete rainfall data is divided into continuous sequences according to the time window, and the rainfall of each data point is spliced ​​with the slope, elevation, vegetation, and soil feature vectors corresponding to that moment to form a multidimensional input unit that integrates spatiotemporal attributes. During the training process, the LSTM network takes the input unit sequence of historical time periods (such as the past 24 hours) as the basis, dynamically captures long-term and short-term dependencies through a gating mechanism, outputs the flow prediction value for the next moment, and iteratively optimizes the weight parameters through a loss function, ultimately forming an intelligent prediction model that can simultaneously analyze the "spatial geographic driving mechanism" and the "temporal hydrological response process."

[0107] This embodiment makes up for the prediction bias of the traditional LSTM model caused by ignoring spatial attributes, and also complements the HEC-HMS physical model. For example, when the sensor data is abnormal or the historical sequence is short, the geographical features can be used as prior knowledge to constrain the rationality of the model output - if the soil type in a certain area is highly permeable sand, even if a short-term heavy rainfall is input, the model will automatically suppress the over-prediction of peak flow based on the "high infiltration-low runoff" mode learned in training; on the contrary, if the vegetation cover is significantly reduced (such as a decrease in the mean NDVI), the model can dynamically adjust the runoff coefficient in combination with historical data to warn of possible flood risks in advance. This spatiotemporal fusion training method essentially constructs a "data-driven virtual watershed", making the subsequent joint verification and parameter transfer with the HEC-HMS physical model (such as using the runoff process predicted by LSTM as the boundary condition of the physical model) more scientific and reliable.

[0108] Furthermore, the specific logic for the LSTM network model to output data-driven runoff prediction results is:

[0109] Input the input vector to the trained LSTM network model to obtain the output traffic sequence of the LSTM network model , and expressed as: in, Indicates the output traffic sequence data predicted by the LSTM network model. The predicted data for each data point;

[0110] The predicted flow change of adjacent data points is calculated based on the formula: in, Indicates The gradient of the data point;

[0111] The square sum of the gradients is used as a penalty term and added to the model's loss function. The formula is: in, represents the penalty loss function, represents the hyperparameter that controls the strength of the penalty term;

[0112] The penalty loss function is combined with the original loss function based on mean square error to form a new loss function, based on the formula: in, Represents the new loss function formed, The actual measurement Traffic data for each data point;

[0113] To minimize the comprehensive loss function As the goal, adjust the parameters of the LSTM model and obtain the traffic prediction results output by the LSTM network model.

[0114] By designing a gradient penalty term to constrain the output smoothness of the LSTM model, the physical rationality and temporal continuity of the runoff prediction results can be effectively improved. When the traditional LSTM model only uses the mean square error (MSE) as the loss function, it is easy to cause non-physical mutations in the predicted flow (for example, the flow of adjacent data points increases or decreases suddenly) due to training data noise or model overfitting. This fluctuation is often contrary to the actual runoff production and confluence process in hydrological scenarios. Natural runoff changes are usually affected by buffering effects such as soil infiltration and river channel regulation, showing a relatively gentle upward or downward trend. The gradient penalty term essentially injects the prior knowledge of "flow continuity" into the model by constraining the sum of squares of differences between adjacent predicted values, so that LSTM not only pursues single-point prediction accuracy during training, but also meets the physical law of smooth transition of hydrological process lines. For example, when rainfall input suddenly increases, the model will still give priority to outputting a gradual flow growth that conforms to the surface storage-runoff mechanism, rather than generating abnormal spikes similar to pulse signals, which is of key significance for subsequent joint simulations with the HEC-HMS physical model (such as flood evolution simulation).

[0115] First, the input vector (integrated geographic features and rainfall sequence) is input into the LSTM network to obtain the predicted flow sequence. Then, each data point in the sequence is traversed, the difference between adjacent predicted values ​​is calculated, and the sum is multiplied by the hyperparameter to form a gradient penalty term. This penalty term is superimposed with the standard MSE loss to form a comprehensive loss function. The value of the hyperparameter needs to be determined through cross-validation: if it is too small, the model may not be able to effectively suppress non-physical fluctuations; if it is too large, it will over-smooth the real flow changes (such as covering up the reasonable flood peak caused by heavy rain). It can usually be tuned based on the balance between the Nash efficiency coefficient (NSE) on the validation set and the smoothness of the predicted curve. During the back-propagation process, the optimizer simultaneously updates the network weights to minimize, forcing the LSTM to capture complex nonlinear relationships while spontaneously learning an output mode that conforms to the characteristics of hydrological dynamics.

[0116] Driven by both mathematical constraints and physical laws, the interpretability of the prediction results and the stability of the system are significantly enhanced. In extreme event prediction scenarios (such as short-term heavy rainfall), the model can not only reflect the trend of rapid increase in flow, but also avoid prediction shocks caused by local noise (such as sensor acquisition errors), making the output runoff process line more consistent with the simulation results of the HEC-HMS physical model in morphology, laying a foundation for data consistency for subsequent model coupling (such as using LSTM prediction results as boundary conditions of the physical model). In addition, the introduction of gradient penalty terms can also alleviate the overfitting problem under small sample training - when there is less historical data, the model tends to generate more conservative and physically intuitive prediction curves, rather than blindly fitting random fluctuations in the training set, which is particularly important for flood prediction at newly built monitoring sites or in data-scarce areas.

[0117] The model fusion module is used to fuse the runoff process line and flow prediction results through the attention mechanism, calculate the dynamic weight to balance the contribution of the two models, introduce water conservation constraints, and finally generate a comprehensive forecast result.

[0118] The model fusion module fuses the runoff process line and runoff prediction results through the attention mechanism, calculates the dynamic weight to balance the contribution of the two models, introduces water conservation constraints, and finally generates a comprehensive forecast result. This step can significantly improve the prediction accuracy, reliability and adaptability of the overall solution.

[0119] Runoff process lines are usually simulated by physical models (HEC-HMS), which are based on detailed hydrological physical processes and reflect the impact of the basin's topography, soil, vegetation and other underlying surface characteristics as well as meteorological conditions such as rainfall on runoff. The runoff prediction results may come from machine learning models (LSTM), which learn time series patterns from historical data and predict future runoff. The two models have their own strengths. The physical model has a stable physical foundation, while the data-driven model can flexibly capture complex time series patterns.

[0120] The model fusion module uses an attention mechanism to dynamically adjust the output weights of the physical model and the data-driven model. The attention mechanism calculates the weight of each model based on the current data point, historical data, and other relevant features (such as rainfall intensity, watershed topography, etc.). For example, during periods of heavy rainfall, the data-driven model may have higher prediction accuracy because it has learned the relationship between heavy rainfall and runoff; while during periods of stable rainfall, the physical model may provide more reliable prediction results.

[0121] Align the runoff process line and the runoff prediction results to ensure their correspondence in data points and space. Then, input these data into the attention network to calculate the weights of the two models for each data point. The weights are calculated based on multiple features, such as the current rainfall intensity, the topographic characteristics of the basin, historical prediction errors, etc. After calculating the dynamic weights, the outputs of the two models are weighted averaged according to the weights to obtain the comprehensive forecast results. At the same time, water conservation constraints are introduced to adjust the comprehensive forecast results to ensure that they comply with the basic principles of water conservation. This may involve post-processing of the comprehensive forecast results, such as adjusting the predicted values ​​of certain data points to ensure the overall water balance.

[0122] Define a comprehensive loss function, including prediction error and water conservation deviation. By optimizing this loss function and adjusting the parameters of the attention network, the fused prediction results are optimized in terms of accuracy and rationality.

[0123] In this embodiment, the specific logic for generating the comprehensive forecast result is:

[0124] The runoff hydrograph is defined as: ,in Indicates the first The observed flow rate of data points, Represents the total number of data points of the discretized sequence data;

[0125] The traffic prediction result output by the LSTM network model is defined as: ,in Indicates the traffic prediction result. The predicted flow rate for each data point;

[0126] The discretized flow sequence data is subjected to fast Fourier transform to obtain the frequency domain characteristics of the flow sequence data, specifically: in, represents frequency characteristic data, Indicates the extracted The amplitude of the characteristic frequency, is the total number of features;

[0127] The attention weight is calculated based on the formula: in, represents the attention weight, and represents the weight matrix and bias, express Activation function, Indicates normalizing the weights;

[0128] The formula for obtaining the comprehensive fusion result by fusing the runoff process line and the prediction result according to the attention weight is: in, represents the comprehensive fusion result, In the comprehensive fusion result, The flow of data points, The first column represents the fused runoff process line and the prediction result. The flow of data points, represents the water conservation constraint, Represents the flow adjustment, which is calculated through the water conservation constraint.

[0129] The model fusion module combines the runoff process line and LSTM prediction results through the attention mechanism, and introduces the water conservation constraint to generate a comprehensive forecast result, thereby significantly optimizing the overall forecast performance. Specifically, this step can integrate the advantages of the physical model and the data-driven model, and use the attention mechanism to dynamically adjust the contribution weights of the two to ensure that the forecast results are consistent with physical laws and adapt to actual time series changes. In addition, the introduction of water conservation constraints further improves the rationality and accuracy of the results, ensuring the comprehensiveness and consistency of the forecast.

[0130] The outputs of the two models are aligned to ensure that the timestamps and data points match. Subsequently, the traffic sequence is subjected to a fast Fourier transform to extract frequency domain features that can reflect the periodicity and change patterns of the traffic sequence. Next, the attention weights are calculated using these frequency domain features. Through the nonlinear transformation of the neural network, the weights are ensured to dynamically reflect the importance of different time points and subsequences. The introduction of the attention mechanism enables the model to capture and balance the contribution of each data point and feature in a more detailed manner.

[0131] By combining attention weights and water conservation constraints, the comprehensive forecast results not only maintain the stability and rationality of the physical model, but also incorporate the LSTM model's flexible capture of time series to ensure the maximum accuracy and consistency of the forecast results. This process optimizes the attention network and loss function to ultimately generate high-quality comprehensive forecast results, providing a reliable basis for subsequent hydrological forecasts and flood warnings.

[0132] The calculation of flow adjustment is closely related to the water conservation constraint. According to the water conservation requirement, the overall flow of the comprehensive forecast result must be consistent with the total amount of the LSTM prediction result. Specifically, in the fusion process, the initial fusion result is first obtained through the attention weight. However, due to the possible systematic deviation between the physical model and the data-driven model, the total flow after direct weighting may not be strictly equal to the sum of the flow prediction output by the LSTM network model. The adjustment amount needs to be introduced and corrected by the constraint condition. In actual calculation, the objective function of minimizing the sum of squares of the adjustment can be used, such as the flow adjustment value solved by the Lagrange multiplier method, to ultimately ensure that the water conservation is achieved while the distribution of the adjustment amount is reasonable.

[0133] The acquisition of weight matrix and bias depends on the training process. In the construction of the attention network, the frequency domain features are first used as input and nonlinearly transformed through the fully connected layer. Among them, the weight matrix and bias are the parameters to be learned by the model, and their initialization can adopt random initialization or pre-training strategy. In the training stage, the back propagation algorithm is combined with the optimizer (such as Adam), and the weighted sum of the prediction error and the water conservation deviation is used as the loss function to gradually adjust the value of the weight matrix and bias, so that the attention weight can adaptively weigh the contribution of the physical model and the LSTM model. After the training is completed, the weight matrix and bias are fixed for inference of new data.

[0134] and The core purpose of the structural design is to extract nonlinear dependencies in frequency domain features and generate reasonable weight distribution. The role of the ReLU activation function is to introduce nonlinear expression capabilities, so that the model can capture the complex frequency domain features in the traffic sequence. The nonlinear transformation here enhances the model's sensitivity to high-frequency or low-frequency features. The softmax function maps the transformed eigenvalues ​​to probability distributions, thereby ensuring that the fusion weights have clear physical meanings - for example, a higher attention weight indicates a greater contribution to the LSTM prediction results, whereas a lower attention weight indicates a dominant contribution from the physical model. This structural design not only retains the dynamic characteristics of the frequency domain features, but also ensures the stability and interpretability of the fusion process through probabilistic weights, ultimately improving the reliability and adaptability of the prediction results.

[0135] The model verification module is used to compare the forecast results with the measured data, calculate the total error, and evaluate the model performance through the total error. If it is lower than the threshold, the parameter optimization module is triggered to adjust the model parameters and re-execute the modeling process. After verification, the final flood forecast results are output.

[0136] In this embodiment, the specific formula for evaluating the model performance through the total error is: in, Represents the total error, if , the forecast system is used to output the final flood forecast result, otherwise the internal parameters of the physical model module, the driving model module and the model fusion module are readjusted. Represents the validation threshold.

[0137] The introduction of the model verification module builds a closed-loop feedback mechanism, which strengthens the adaptability and reliability of the forecast process from the system level. By comparing the total amount of error between the comprehensive forecast results and the measured data, the module can quantitatively evaluate the accuracy of the model output and dynamically trigger the parameter optimization process based on the preset threshold. This design effectively solves the problem of accumulated prediction deviations caused by dynamic environmental changes or data distribution offsets in traditional hydrological models. Specifically, when RM exceeds the threshold, the system will automatically trace back to the parameter optimization links of the physical model, data-driven model and fusion module, and adjust the internal parameters of the model (such as the weight matrix of LSTM, the frequency domain feature weights in the attention network, etc.) through back-propagation algorithms, gradient descent and other mechanisms, while ensuring that the water conservation constraints are not violated during the optimization process. This process is iterated repeatedly until RM meets the accuracy requirements, thus forming a dynamic closed loop of "prediction-verification-optimization", which significantly improves the model's adaptability to complex hydrological conditions.

[0138] The model verification module accesses the measured hydrological data in real time, aligns the data points with the forecast sequence output by the fusion module, and calculates the total deviation based on the root mean square error formula (RM). This error indicator not only reflects the global accuracy of the forecast results, but also amplifies the impact of significant deviations through the square term, allowing the system to quickly identify key abnormal periods. If RM does not meet the standard, the parameter optimization module will comprehensively analyze the contribution of each submodule: for example, in view of the possible insufficient fit of the runoff process line in the physical model, its parameters will be adjusted to enhance the sensitivity to terrain and rainfall intensity; for the LSTM model, its ability to capture time series features will be optimized through gradient updates; at the same time, the attention weight generation network (W_f, b_f) in the model fusion module will also be retrained to ensure that the mapping relationship between frequency domain features and dynamic weights is adaptively updated as the data distribution evolves. The entire process is implemented through an automated pipeline, reducing the cost of manual intervention.

[0139] This embodiment realizes the self-calibration and continuous evolution of the forecasting system. By directly feeding back the verification results to the model parameter space, the system can actively correct the forecast deviation caused by input data noise, extreme weather events or changes in underlying surface conditions. For example, in the early stage of the flood season, when the snowmelt runoff pattern differs from the training data, the model can quickly adjust the attention weight allocation strategy through multiple iterative optimizations to balance the contribution ratio of the physical mechanism model and the data-driven model. At the same time, the water conservation constraint runs through the entire optimization process as a hard condition, avoiding the physical unreasonable phenomena that may be caused by the simple pursuit of error minimization. Ultimately, this closed-loop architecture enables the forecasting system to continuously improve the accuracy of the description of complex hydrological processes on the basis of ensuring the physical consistency of the forecast results, providing reliable support for flood control decisions with both stability and sensitivity.

[0140] To determine the verification threshold, it is necessary to set a basic threshold interval in combination with the allowable range of flow deviation corresponding to the flood warning level, and then dynamically calibrate the threshold through the statistical characteristics of historical flood samples. For example, the sliding window method is used to calculate the RM distribution of the prediction results of the verification set, and the upper limit of the confidence interval is taken as the initial reference value. In actual deployment, a feedback mechanism is introduced. When the RM value of multiple consecutive forecast periods continues to be lower than the current threshold, the threshold contraction algorithm is automatically triggered. On the contrary, if the threshold is frequently exceeded, the hydrological conditions are analyzed in combination with expert experience, and the threshold is re-evaluated or the model structure is improved. Finally, cross-validation is used to ensure that the threshold setting can avoid prediction failure due to being too loose, and will not cause invalid parameter optimization due to being too strict.

[0141] See also Figure 2 The present invention also provides a basin flood runoff forecasting method based on water conservancy big data, the forecasting method is executed by the basin flood runoff forecasting system based on water conservancy big data, and the specific steps include:

[0142] Step 1: Collect the original time series data of the monitoring stations and the geographical location information of the monitoring stations from the hydrological monitoring network of the basin to be forecasted, calculate the dynamic threshold using the sliding window technology, identify and mark the outliers, and interpolate the outliers through the bidirectional LSTM network to obtain the cleaned time series data. The original time series data includes flow and rainfall;

[0143] Step 2: Input the cleaned time series data into the HEC-HMS model to calculate the flow of the basin. By introducing the terrain correction factor and adjusting the SCS unit line, the HEC-HMS model outputs the runoff process line;

[0144] Step 3: Extract geographic spatial features from the monitoring site’s geographic location information, combine it with the cleaned time series data, and input it into the LSTM network model for training. Constrain the network output through the gradient penalty term, so that the LSTM network model can output data-driven traffic prediction results.

[0145] Step 4: The runoff process line and flow prediction results are integrated through the attention mechanism, the dynamic weight is calculated to balance the contribution of the two models, the water conservation constraint is introduced, and finally the comprehensive forecast result is generated;

[0146] Step 5: Compare the forecast results with the measured data, calculate the total error, and evaluate the model performance through the total error. If it is lower than the threshold, the parameter optimization module is triggered, the model parameters are adjusted and the modeling process is re-executed. After verification, the final flood forecast results are output.

[0147] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters in the formula are set by technicians in this field according to actual conditions.

[0148] The above embodiments may be implemented in whole or in part by software, hardware, firmware or any other combination thereof. When implemented by software, the above embodiments may be implemented in whole or in part in the form of a computer program product. Those skilled in the art may appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein may be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed by hardware or software methods depends on the specific application and design constraints of the technical solution.

[0149] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, and may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0150] The above description is only a specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application.

Claims

1. A basin flood runoff forecasting system based on water conservancy big data, characterized in that: include: The data acquisition module is used to collect the original time series data of the monitoring stations and the geographical location information of the monitoring stations from the hydrological monitoring network of the basin to be forecasted, calculate the dynamic threshold by using the sliding window technology, identify and mark the outliers, and interpolate the outliers through the bidirectional LSTM network to obtain the cleaned time series data. The original time series data includes flow and rainfall; The physical model module is used to input the cleaned time series data into the HEC-HMS model, calculate the flow of the basin, and adjust the SCS unit line by introducing the terrain correction factor so that the HEC-HMS model outputs the runoff process line; The driving model module is used to extract geographic spatial features by monitoring the geographic location information of the site, splice it with the cleaned time series data, input it into the LSTM network model for training, constrain the network output through the gradient penalty term, and let the LSTM network model output data-driven traffic prediction results; Model fusion module, which is used to fuse the runoff process line and flow prediction results through the attention mechanism, calculate the dynamic weight to balance the contribution of the two models, introduce water conservation constraints, and finally generate a comprehensive forecast result; The model verification module is used to compare the forecast results with the measured data, calculate the total error, and evaluate the model performance through the total error. If it is lower than the threshold, the parameter optimization module is triggered to adjust the model parameters and re-execute the modeling process. After verification, the final flood forecast results are output.

2. The basin flood runoff forecasting system based on water conservancy big data according to claim 1 is characterized by: The specific logic for identifying and marking outliers is: At the monitoring site, with the current real-time moment as the center, slide the 30-day window forward to extract the corresponding original time series data and discretize it into equal time intervals. Align the sequence data of flow and rainfall according to the timestamps, and calculate the arithmetic mean and standard deviation of the sequence data of flow and rainfall in the window respectively. The formula is as follows: in, and Respectively represent Arithmetic mean and standard deviation of class sequence data, Indicates the category of sequence data, represents the flow sequence data, represents rainfall series data, Represents the total number of data points of the discretized sequence data, In the discretized sequence data, Class sequence data data points; The coefficient of variation is obtained by dividing the standard deviation by the mean, multiplied by the fixed proportional coefficient and added to the dynamic threshold. The formula is: in, Indicates Dynamic thresholding for sequence-like data, is the coefficient of variation; The data points in the discrete flow and rainfall sequence data segment are compared point by point with the absolute difference between the current value and the dynamic threshold to obtain the abnormal point. , mark the point as an outlier.

3. The basin flood runoff forecasting system based on water conservancy big data according to claim 2 is characterized by: The specific logic for obtaining the cleaned time series data is as follows: For each outlier point, the data of the 5 data points before and after it are intercepted. If the data exceeds the data boundary, the missing part is filled with zeros, and the zero-value placeholder is filled to form an 11-dimensional input vector sequence; The input vector sequence is input into a pre-trained bidirectional LSTM network, the input dimension of the bidirectional LSTM network is 11, the output dimension is 1, and the bidirectional LSTM network outputs the interpolation value at the corresponding moment. After filling the abnormal point, the cleaned time series data is obtained.

4. The basin flood runoff forecasting system based on water conservancy big data according to claim 3 is characterized by: The specific logic of the HEC-HMS model output runoff process line is: The cleaned time series data is divided into a rainfall input sequence and a runoff output sequence. The rainfall sequence data is used as the rainfall input sequence, and the flow sequence data is used as the runoff output sequence. Extract the geographical location information of the monitoring site, use the GIS database, take the monitoring site coordinates as the center and the radius of 3 kilometers as the monitoring area, and obtain the slope data in the monitoring area; Extract all slope values ​​in the monitoring area from the GIS database, traverse all slope values, find the maximum value, and use the maximum slope value as slope characteristic data; The terrain correction factor is calculated through the slope characteristic data, and the SCS unit line built into HEC-HMS is corrected based on the formula: in, represents the terrain correction factor, Indicates the angle value corresponding to the slope characteristic data, is the gain factor, and , and They represent the exponential coefficients of the unit line peak time and unit line attenuation segment of the corrected SCS unit line, and They represent the exponential coefficients of the unit line peak time and unit line attenuation segment of the original SCS unit line respectively; The corrected SCS unit line, rainfall input sequence and runoff output sequence are input into the HEC-HMS model to output the runoff process line.

5. The basin flood runoff forecasting system based on water conservancy big data according to claim 4 is characterized by: The specific logic for obtaining training data for inputting into the LSTM network model for training is: For the monitoring area corresponding to the monitoring station, GIS tools are used to extract elevation data in the monitoring area, and the elevation standard deviation in the monitoring area is calculated as the elevation feature data; Extract the normalized difference vegetation index data in the monitoring area and calculate the arithmetic mean as the vegetation characteristic data of the monitoring area; From the soil type database, extract the soil type information of the area where the monitoring station is located according to the geographical coordinates of the monitoring station, perform one-hot encoding on the soil type, and convert it into a numerical vector as soil characteristic data; The slope characteristic data, elevation characteristic data, vegetation characteristic data and soil characteristic data are associated with the data of each data point in the discrete rainfall sequence data to form an input vector as the input of the LSTM network model. The flow of the corresponding next data point is used as the label of the LSTM network model to train the LSTM network model.

6. The basin flood runoff forecasting system based on water conservancy big data according to claim 5 is characterized by: The specific logic for the LSTM network model to output data-driven runoff prediction results is: Input the input vector to the trained LSTM network model to obtain the output traffic sequence of the LSTM network model , and expressed as: in, Indicates the output traffic sequence data predicted by the LSTM network model. The predicted data for each data point; The predicted flow change of adjacent data points is calculated based on the formula: in, Indicates The gradient of a data point; The square sum of the gradients is used as a penalty term and added to the model's loss function. The formula is: in, represents the penalty loss function, represents the hyperparameter that controls the strength of the penalty term; The penalty loss function is combined with the original loss function based on mean square error to form a new loss function, based on the formula: in, Represents the new loss function formed, The actual measurement Traffic data for each data point; To minimize the comprehensive loss function As the goal, adjust the parameters of the LSTM model and obtain the traffic prediction results output by the LSTM network model.

7. The basin flood runoff forecasting system based on water conservancy big data according to claim 2 is characterized by: The specific logic for generating comprehensive forecast results is: The runoff hydrograph is defined as: ,in Indicates the first The observed flow rate of data points, Represents the total number of data points of the discretized sequence data; The traffic prediction result output by the LSTM network model is defined as: ,in Indicates the traffic prediction result. The predicted flow rate for each data point; The discretized flow sequence data is subjected to fast Fourier transform to obtain the frequency domain characteristics of the flow sequence data, specifically: in, represents frequency characteristic data, Indicates the extracted The amplitude of the characteristic frequency, is the total number of features; The attention weight is calculated based on the formula: in, represents the attention weight, and represents the weight matrix and bias, express Activation function, Indicates normalizing the weights; The formula for obtaining the comprehensive fusion result by fusing the runoff process line and the prediction result according to the attention weight is: in, represents the comprehensive fusion result, In the comprehensive fusion result, The flow of data points, The first column represents the fused runoff process line and the prediction result. The flow of data points, represents the water conservation constraint, Represents the flow adjustment, which is calculated through the water conservation constraint.

8. The basin flood runoff forecasting system based on water conservancy big data according to claim 7 is characterized by: The specific formula for evaluating model performance through total error is: in, Represents the total error, if , the forecast system is used to output the final flood forecast result, otherwise the internal parameters of the physical model module, the driving model module and the model fusion module are readjusted. Represents the validation threshold.

9. A basin flood runoff forecasting method based on water conservancy big data, characterized by: The forecasting method is performed by a basin flood runoff forecasting system based on water conservancy big data as described in any one of claims 1 to 8, and the specific steps include: The original time series data and geographical location information of the monitoring stations are collected from the hydrological monitoring network of the basin to be forecasted, and the dynamic threshold is calculated using the sliding window technology to identify and mark the outliers. The outliers are interpolated through the bidirectional LSTM network to obtain the cleaned time series data, and the original time series data includes flow and rainfall; The cleaned time series data are input into the HEC-HMS model to calculate the flow of the basin. By introducing the terrain correction factor and adjusting the SCS unit line, the HEC-HMS model outputs the runoff process line. By monitoring the geographic location information of the site, we extract the geographic spatial features, splice them with the cleaned time series data, input them into the LSTM network model for training, and constrain the network output through the gradient penalty term, so that the LSTM network model can output data-driven traffic prediction results; The runoff process line and flow prediction results are fused through the attention mechanism, the dynamic weight is calculated to balance the contribution of the two models, the water conservation constraint is introduced, and finally the comprehensive forecast result is generated; The forecast results are compared with the measured data, the total error is calculated, and the model performance is evaluated by the total error. If it is lower than the threshold, the parameter optimization module is triggered, the model parameters are adjusted and the modeling process is re-executed. After verification, the final flood forecast results are output.

Images