Flood forecasting method and system considering similarity and multi-dimensional classification system
By constructing a flood forecasting method based on the combined similarity of rainfall and runoff and a three-dimensional classification system, the problems of parameter calibration difficulties and insufficient generalization ability of data-driven models under complex underlying surface conditions in traditional methods are solved, and high-precision hourly flood forecasts are achieved.
Patent Information
- Application Number
- CN202511034982.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-25
- Publication Date
- 2025-11-07
AI Technical Summary
Traditional flood forecasting methods face difficulties in parameter calibration under complex underlying surface conditions, have low computational efficiency, and are unable to meet the needs of real-time forecasting. Furthermore, data-driven models have insufficient generalization ability in rare flood events and unstable performance in small samples.
A flood forecasting method based on rainfall-runoff joint similarity and a three-dimensional classification system is proposed. By acquiring historical data, a multi-dimensional classification system is established to automatically identify runoff change patterns, match the optimal parameters of the model, and generate hourly runoff forecasts for future periods.
It improves the accuracy and robustness of flood forecasting under complex hydrological conditions, adapts to the conditions of insufficient data types and limited historical data, and realizes 72-hour hourly runoff forecasting.
Smart Images
Figure CN120910722A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of runoff prediction, and particularly relates to a flood prediction method and system considering similarity and a multi-dimensional classification system. BACKGROUND
[0002] In recent years, with the intensification of global climate change, extreme rainfall events occur frequently, and the risk of flood disasters in river basins has increased significantly, which puts forward higher requirements for flood control and disaster reduction, water resources regulation and emergency management. Traditional flood prediction methods mainly include physical driving models and data-driven models: physical driving models (such as Xin'anjiang model, SWAT model, etc.) are based on hydrological physical processes and depend on accurate basin parameters (such as soil permeability, terrain slope, etc.), but under complex underlying surface conditions, parameter calibration is difficult, and the calculation efficiency is low, which is difficult to meet the real-time prediction demand. Data-driven models (such as neural networks, support vector machines, etc.) mine the rainfall-runoff relationship through historical data, which can adapt to nonlinear problems, but have the following limitations: It is difficult to associate the prediction results with specific hydrological mechanisms, and the interpretability is poor; the generalization ability is insufficient in rare flood events, and the small sample performance is unstable. Therefore, a flood prediction model that can combine hydrological physical processes and data mining technology is needed. SUMMARY
[0003] To solve the technical problems in the above background art, the application provides a flood prediction method and system considering similarity and a multi-dimensional classification system by constructing a "rainfall-runoff joint similarity" and a "three-dimensional classification system".
[0004] The application adopts the following technical scheme: a flood prediction method considering similarity and a multi-dimensional classification system, comprising the following steps: Obtain historical rainfall data and historical runoff data, and divide them into a historical sample library and a test sample library according to time sequence; Based on the historical sample library, analyze the relationship between the prediction factors and the runoff of the prediction section to determine the influence lag of the prediction factors , ; the prediction factors at least include rainfall and runoff; Integrate rainfall and rainfall change process, runoff and runoff change process to construct a rainfall-runoff joint similarity index between the historical sample library and the target sample library; Establish a multi-dimensional classification system, determine the runoff change mode based on the multi-dimensional classification system; construct a similarity flood prediction model, train the model similarity flood prediction model for different runoff change modes, and obtain the model parameters and prediction scheme corresponding to each runoff change mode; The flood forecasting model is constructed, the runoff change mode is automatically identified according to a multi-dimensional classification system, the optimal parameters and the forecasting scheme are automatically matched based on the identified runoff change mode, and the hourly runoff forecast in a future period is generated through rolling forecast.
[0005] In a further embodiment, the determination process of the influence lag time of the prediction factor is as follows: A quantity threshold of rain stations is preset , the actual quantity of rain stations in the prediction area is obtained : If , the influence lag time of each rain station is obtained by analyzing the relationship between the antecedent rainfall of each rain station and the runoff of the prediction section one by one . If , the prediction area is divided into polygonal regions by using the Thiessen polygon method, the polygonal regions are merged to obtain sub-regions according to the data correlation of each rain station, and then ; the areal rainfall of each sub-region is calculated, and the relationship between the areal rainfall of each sub-region and the runoff of the prediction section is analyzed one by one to obtain the corresponding influence lag time.
[0006] In a further embodiment, the construction process of the rainfall-runoff combined similarity index is as follows: The similarity of the antecedent rainfall and the rainfall process change between the historical sample library and the target sample library is evaluated by using the rainfall optimal similarity coefficient , and the rainfall similarity between the historical sample library and the target sample library is obtained; The similarity of the antecedent runoff size and the runoff change trend between the historical sample library and the target sample library is evaluated by using the runoff optimal similarity coefficient , and the runoff similarity between the historical sample library and the target sample library is obtained; The rainfall-runoff combined similarity index is calculated by using the weighted average method based on the rainfall optimal similarity coefficient and the runoff optimal similarity coefficient : ; wherein and are the influence weights of rainfall and runoff, respectively.
[0007] In further embodiments, the multi-dimensional classification system is a three-dimensional system of "rainfall grade-flow grade-rise grade"; correspondingly, the determination process of the runoff change pattern comprises one or more of rainfall grade division, flow grade division and rise grade division.
[0008] In further embodiments, the construction process of the similarity flood forecast model is as follows: For the current time sequence , the rainfall time sequence and the runoff time sequence are established respectively. The influence of the previous rainfall and the previous runoff on the forecast runoff is expressed, and the current characteristic sequence group of the forecast runoff is integrated based on the rainfall time sequence and the runoff time sequence ; In the historical sample library, the most similar group of historical characteristic sequence groups and the corresponding similarity change value are found by the rainfall-runoff joint similarity index. ; Based on the target sample library, the target sample of the current time sequence t is predicted based on the most similar group of historical characteristic sequence groups .
[0009] In further embodiments, the forecast scheme at least includes: forecasting the absolute value of the current runoff and / or forecasting the relative value of the current runoff .
[0010] In further embodiments, the similarity flood forecast model is adaptively trained based on the test sample library, and the model parameters corresponding to each runoff change pattern are obtained; wherein the model parameters at least include: the influence time sequence process of the previous rainfall, the influence time sequence process of the previous runoff, the rainfall weight, the runoff weight, the number of similarity samples, and the output form of the forecast result.
[0011] In further embodiments, the rainfall grade division process is as follows: considering the time lag effect, the corresponding forecast area surface rainfall that has an impact on the cross-section runoff is calculated, and the corresponding rainfall grade is divided according to the frequency curve of the surface rainfall in the specified time length of the forecast area; The flow grade division method is as follows: the flow threshold values of each grade of flood frequency are set in advance, and the corresponding flow grade is divided based on the flow threshold values of each grade. The step of dividing the rising range is as follows: dividing according to the frequency of the flow rate change rate per unit time; wherein the flow rate change rate per unit time is determined by the difference between the measured runoff of a previous specified time length and the measured runoff of the previous two specified time lengths; and dividing the rising range according to the multiple relationship of the base flow.
[0012] In further embodiments, the most similar group of historical feature sequences is determined based on the absolute value of the actual runoff : , wherein is the weight of each historical sample h , and is the measured runoff value of the historical sample at the time . The most similar group of historical feature sequences is determined based on the relative value of the actual runoff : , is the measured runoff value of the historical sample at the time h .
[0013] A flood forecasting system that takes into account both similarity and a multi-dimensional classification system, for implementing the flood forecasting method described above, comprises: A first module configured to obtain historical rainfall data and historical runoff data, and divide the historical sample library and the test sample library according to time sequence; A second module configured to analyze the relationship between the prediction factors and the prediction section runoff based on the historical sample library, to determine the influence lag corresponding to the prediction factors , ; the prediction factors at least include rainfall and runoff; A third module configured to construct a rainfall-runoff joint similarity index between the historical sample library and the target sample library by comprehensively considering rainfall and rainfall change process, runoff and runoff change process; A fourth module configured to establish a multi-dimensional classification system, determine the runoff change mode based on the multi-dimensional classification system, construct a similarity flood forecasting model, train the model similarity flood forecasting model for different runoff change modes respectively, and obtain the model parameters and the prediction scheme corresponding to each runoff change mode; A fifth module configured to construct a flood forecasting model, automatically identify the runoff change mode according to the multi-dimensional classification system, automatically match the optimal model parameters and the prediction scheme based on the identified runoff change mode, and generate the hourly runoff prediction in the future period through rolling prediction.
[0014] The application provides a flood forecasting method considering similarity and a multi-dimensional classification system. The application requires relatively less data, i.e., rainfall and runoff, and can adapt to adverse forecasting conditions of insufficient data types and limited historical data. The application comprehensively considers the differences in flood characteristics under different meteorological and hydrological conditions, proposes a "rainfall grade-flow grade-rise grade" classification framework and a conditional judgment method under different runoff change modes, establishes a forecasting scheme for different runoff change modes, and integrates a flood forecasting model capable of automatically identifying runoff change modes and automatically matching optimal parameters of the model. The model generates a 72-hour runoff forecast for the future through rolling forecasting, and improves the flood forecasting accuracy and robustness under complex hydrological conditions. BRIEF DESCRIPTION OF DRAWINGS
[0015] Figure 1 is a flowchart of Example 1.
[0016] Figure 2 is a division diagram of a sub-area in Example 1.
[0017] Figure 3 is a model parameter and forecasting scheme diagram in Example 1.
[0018] Figure 4 is a flood forecasting effect diagram in Example 1.
[0019] Figure 5 is an architecture diagram of Example 2. DETAILED DESCRIPTION
[0020] The application will be further described below in combination with the drawings and examples of the specification.
[0021] Example 1 The example provides a flood forecasting method considering similarity and a multi-dimensional classification system. Taking flood forecasting of the Hongfeng section of MaoTiao River Basin as an example, a 72-hour runoff forecast for the future is generated.
[0022] As shown in Figure 1 , historical rainfall data and historical runoff data are obtained, and historical sample library and test sample library are divided according to time sequence. It should be noted that the historical rainfall data and historical runoff data described in the example are at least 10 years of data, and the data is hourly data. The first 7 years of historical data are divided into a historical sample library, and the last 3 years are divided into a test period. The test period should include wet, normal and dry years. If the condition is not met, the test period can be appropriately extended.
[0023] Based on the historical sample library, the relationship between the forecasting factor and the runoff of the forecasting section is analyzed to determine the influence lag of the forecasting factor , The forecasting factor At least including: rainfall and runoff; By combining rainfall and rainfall variation processes, and runoff flow and runoff variation processes, a joint similarity index between the historical sample database and the target sample database is constructed. A multi-dimensional classification system is established, and runoff change patterns are determined based on the multi-dimensional classification system; a similarity flood forecasting model is constructed, and the model similarity flood forecasting model is trained for different runoff change patterns to obtain the model parameters and forecasting schemes corresponding to each runoff change pattern; A flood forecasting model is constructed, and runoff change patterns are automatically identified based on a multi-dimensional classification system. The optimal parameters and forecasting schemes of the model are automatically matched based on the identified runoff change patterns, and hourly runoff forecasts for future periods are generated through rolling forecasts.
[0024] Based on the above examples, this embodiment obtained hourly rainfall and flow data at various rain gauge stations (Hongfeng, Maiweng, Huangmaocun, Qingyutang, Qiyanqiao, Pingzhai, and Guangshun) along the Hongfeng section of the Maotiao River basin from 9:00 on July 23, 2015 to 23:00 on June 30, 2024. The period from 13:00 on July 23, 2015 to 23:00 on December 31, 2020 was set as a historical sample database, and runoff forecasting and performance testing were conducted at the Hongfeng section during the flood season (May to October each year) from 00:00 on January 1, 2021 to 00:00 on July 1, 2024.
[0025] Based on this, the influence lag time of the forecast factors in this embodiment The determination process is as follows: Preset the threshold number of rain gauge stations Obtain the actual number of rainfall stations within the forecast area. : like ,but The relationship between the antecedent rainfall and the predicted cross-sectional runoff at each rainfall station was analyzed using Pearson correlation analysis or grey relational analysis to obtain the lag time of each influence. This embodiment uses... Taking a value of 6 as an example, if... Then, the relationship between the previous rainfall at each rainfall station and the runoff at the forecast section is analyzed to determine the impact lag time of the previous rainfall at each rainfall station.
[0026] like The forecast area was divided into sections using the Thiessen polygon method. A polygonal region is defined, and the data from each rainfall station is correlated. Merging the polygonal regions yields... Each sub-region, then Calculate the surface rainfall of each sub-region, and analyze the relationship between the surface rainfall of each sub-region and the runoff of the forecast section one by one by using Pearson correlation analysis or the grey correlation degree method to obtain the influence lag of each sub-region.
[0027] As shown in Figure 2 , there are currently 7 rainfall sites, so the data of each rainfall site is analyzed in correlation considering the geographical distribution, control area, runoff yield characteristics, and the propagation lag of rainfall and runoff, etc. of each rainfall site. Further, the 7 polygons are merged into 4 sub-regions, and the surface rainfall of each sub-region is calculated. The 4 sub-regions are: sub-region one is the region where Hongfeng and Maion rainfall sites are located, sub-region two is the region where Huangmaocun and Qingyutang rainfall sites are located, sub-region three is the region where Qiyangqiao and Pingzhai rainfall sites are located, and sub-region four is the region where Guangshun rainfall site is located.
[0028] The Pearson correlation coefficient analysis method is used to study the influence lag of the surface rainfall of the four sub-regions upstream of the Hongfeng section, the previous runoff, and the runoff of the Hongfeng section. According to the size of the correlation coefficient, the influence lag of the previous runoff is 6h, the influence lag of the rainfall of sub-region one and sub-region two is 3h, the influence lag of the rainfall of sub-region three is 9h, and the influence lag of the rainfall of sub-region four is 12h.
[0029] In further embodiments, the construction process of the rainfall-runoff combined similarity index is as follows: The rainfall optimal similarity coefficient is used to evaluate the similarity of the previous rainfall and the change of the rainfall process between the historical sample library and the target sample library, and the rainfall similarity between the historical sample library and the target sample library is obtained. The runoff optimal similarity coefficient is used to evaluate the similarity of the previous runoff and the change trend of the runoff between the historical sample library and the target sample library, and the runoff similarity between the historical sample library and the target sample library is obtained. The rainfall optimal similarity coefficient and the runoff optimal similarity coefficient are calculated by using the weighted average method. : ; in the formula, and are the influence weights of rainfall and runoff, respectively, which can be rated by the traversal method.
[0030] Further, the calculation process of the rainfall optimal similarity coefficient and the runoff optimal similarity coefficient is the same, and represents the rainfall optimal similarity coefficient or the runoff optimal similarity coefficient , then: ; wherein, is the rainfall station in the historical sample library and the target sample library, is the correlation coefficient, is the rainfall station optimal similarity coefficient of rainfall or optimal similarity coefficient of runoff, is the optimal similarity coefficient of rainfall or the optimal similarity coefficient of runoff . It is further understood that if the rainfall station is 5, the optimal similarity coefficients of 5 rainfall stations need to be integrated.
[0031] Suppose the runoff on July 17 needs to be predicted, the data included in the current target sample library is the measured runoff / measured rainfall on July 14, July 15 and July 16, and the predicted rainfall on July 17. And the runoff on July 17 is predicted by subsequent technical means.
[0032] The prediction factors in this embodiment are two, which are rainfall and runoff. In other embodiments, there may be other categories of prediction factors, so it can be defined that there are prediction factors, and the calculation process of the corresponding optimal similarity coefficient is as follows: wherein, is the shape coefficient of the prediction factor , and is the value coefficient of the prediction factor .
[0033] wherein, the calculation formula of the shape coefficient is as follows: ; the calculation formula of the value coefficient is as follows: ; is the local outlier factor, and the calculation formula is as follows: .
[0034] wherein, is the average value of the mth influencing lag of the prediction factor in the target sample, is the average value of the mth influencing lag of the prediction factor in the historical sample, is the corresponding value of the prediction factor in the target sample at the mth influencing lag, is the corresponding value of the prediction factor in the target sample at the mth influencing lag, For the first in the historical sample The corresponding values of each forecast factor at m lag times. It should be noted that the specific rules for the mean and corresponding values here are as follows: if the current forecast factor is rainfall, then the mean is the average rainfall value and the corresponding value is rainfall; if the current forecast factor is runoff, then the mean is the average runoff value and the corresponding value is runoff.
[0035] In a further embodiment, the multi-dimensional classification system is a three-dimensional system of "rainfall level - flow rate level - runoff rate increase level"; correspondingly, the process of determining the runoff change pattern includes one or more of the following: rainfall level classification, flow rate level classification, and runoff rate increase level classification. In other words, based on the number of samples and runoff characteristics under different runoff categories, only rainfall level classification, or only rainfall level and flow rate level classification, can be used. When the number of samples under a certain level is small, samples from adjacent levels can be merged.
[0036] Furthermore, the rainfall classification process is as follows: Considering the time lag effect, calculate the corresponding forecast area areal rainfall that affects the cross-sectional runoff. The corresponding rainfall levels are determined based on the frequency curve of areal rainfall within a specified time period in the forecast area. Using the example above, the lag times of previous rainfall in the four sub-regions are 2 hours, 4 hours, 6 hours, and 8 hours, respectively. The areal rainfall in the study area... The calculation formula is as follows: In the formula, , , and These represent the areal rainfall in sub-regions of zone 1, zone 2, zone 3, and zone 4, respectively.
[0037] Furthermore, taking 1 hour as an example, the hourly rainfall corresponding to 90% / 50% / 10% flood frequencies is used. This serves as a dividing line. A 90% flood frequency indicates coverage of the vast majority of minor flood events, a 50% flood frequency indicates common floods, and a 10% flood frequency indicates avoidable extreme events.
[0038] like The rainfall level is then classified into light rain / no rain levels; if The rainfall level will then be classified as moderate rain; if The rainfall level will then be classified as heavy rain; if If the rainfall intensity is greater than 10 mm / h, it is basically a rainstorm flood, and the sample size is small. At this time, rainfall can basically cause runoff, so there is no need to further classify the flow rate and the rate of increase.
[0039] The method for dividing the flow level is as follows: the threshold values of the flow level of each grade of flood frequency are set in advance, and the corresponding flow level is divided based on the threshold values.
[0040] For example, the threshold values of the flow level of each grade of flood frequency are set as 90%, 50% and 10% respectively, the flow when the flood frequency is greater than or equal to 90% is considered as small flow, the flow when the flood frequency is between 50% and 90% is considered as medium flow, the flow when the flood frequency is between 10% and 50% is considered as large flow, and the flow when the flood frequency is less than or equal to 10% is considered as big flow. The division standard is determined according to the sample quantity and the characteristics of runoff yield and concentration.
[0041] For further example, when the rainfall intensity ranges from 7mm / h to 10mm / h and from 4mm / h to 7mm / h, it belongs to heavy rain and moderate rain, and the sample quantity is large. With the extension of rainfall time and the accumulation of rainfall, the soil begins to saturate, and the runoff yield area gradually expands. The flood rise rate caused by this rainfall level is basically around 100m³ / s. Most of the samples belong to the case that the rainfall time is short and the rainfall accumulation is small, and the flood order is about 200m³ / s, and the flood frequency is greater than or equal to 90%. Combined with the flood rise rate, the flow level of the previous day is divided into greater than or equal to 100m³ / s and less than 100m³ / s.
[0042] The steps for dividing the rise level are as follows: dividing according to the frequency of the flow change rate per unit time; wherein the flow change rate per unit time is determined by the difference between the measured runoff of a specified time length and the measured runoff of the previous two specified time lengths; and dividing the rise level according to the multiple relationship of the base flow. For example, dividing according to 1 times the base flow, 2 times the base flow and n times the base flow to reflect the relative intensity of the flood rise.
[0043] For example, when the rainfall intensity is not higher than 4mm / h, it belongs to light rain or no rain. When the rainfall intensity is not higher than 4mm / h and the previous flow level is not less than 500m³ / s, it belongs to large flood and the scene is less, so the rise level is not further divided. When the rainfall intensity is not higher than 4mm / h and the previous flow level is less than 100m³ / s, it indicates that there is no large rainfall in the previous period, and there will be no large runoff amplitude, so the runoff rise level is not further divided. When the rainfall intensity is not higher than 4mm / h and the previous flow level is between 100m³ / s and 500m³ / s, since the base flow of the Hongfeng section control basin is not higher than 50m³ / s when it is raining, the previous runoff rise is divided according to the backwater, the rise is not higher than 50m³ / s, the rise is between 50m³ / s and 100m³ / s, and the rise is not less than 100m³ / s.
[0044] In another embodiment, the construction process of the similarity flood forecasting model is as follows: For the current time sequence , the rainfall time sequence and runoff time series The influence of antecedent rainfall and antecedent runoff on forecast runoff is expressed. Among them, the rainfall time series is expressed as: , wherein, is the influence of the areal rainfall of the sub-region on the forecast runoff time lag, is the time series of the areal rainfall of the sub-region .
[0045] is the antecedent runoff time series of the current time series , is the influence of the antecedent runoff on the forecast runoff time lag.
[0046] Based on the rainfall time series and the runoff time series , the current characteristic sequence group of the forecast runoff is integrated : .
[0047] In the historical sample library, the most similar group of historical characteristic sequence groups and the corresponding similar quantity change value to the current characteristic sequence group are found by the rainfall-runoff combined similarity index; the calculation method can refer to the calculation method of the rainfall-runoff combined similarity index.
[0048] Further, the closer the optimal similarity coefficient is to the rainfall-runoff combined similarity index, the more similar the two samples are.
[0049] The target sample of the current time series is predicted by the most similar group of historical characteristic sequence groups : , is the weight coefficient of the historical characteristic sequence group .
[0050] Therefore, after determining the target sample , the prediction scheme at least includes: predicting the absolute value of the current runoff and / or predicting the relative value of the current runoff .
[0051] Further, the absolute value of the current runoff is predicted according to the actual runoff value of the most similar group of historical characteristic sequence groups . : , wherein, is the weight of each historical sample h , is the measured runoff value of the historical sample at time; forecast the current runoff relative value based on the actual runoff change value of the most similar group of historical feature sequences : , is the measured runoff value of the historical sample h at time.
[0052] Adaptively train the similarity flood forecasting model based on the test sample library to obtain model parameters corresponding to each runoff change pattern. Adopt the traversal method to train the similarity forecasting model for different runoff change patterns, and determine the optimal model parameters and forecasting scheme under different runoff change patterns with the minimum average relative error or absolute error in the test period as the criterion, as shown in Figure 3 .
[0053] Among them, the model parameters at least include: the influence time sequence process of the previous rainfall, the influence time sequence process of the previous runoff, the rainfall weight, the runoff weight, the number of similar samples, and the output form of the forecasting result.
[0054] Finally, integrate all types of models, and establish runoff change pattern recognition conditions according to the rainfall grade, flow grade, and rising range grade in the model, as shown in Table 1, to realize automatic recognition of the runoff change pattern and automatic matching of the optimal model parameters and forecasting scheme.
[0055] Table 1 By continuously adding new rainfall and runoff data, the rolling forecast generates a 72-hour flood forecast every hour in the future, as shown in Figure 4 .
[0056] Embodiment 2 With reference to Figure 5 , the embodiment discloses a flood forecasting system considering similarity and multi-dimensional classification system, which is used to realize the flood forecasting method in embodiment 1, and comprises: A first module is configured to obtain historical rainfall data and historical runoff data, and divide the historical sample library and the test sample library according to the time sequence; A second module is configured to analyze the relationship between the forecasting factor and the forecasting section runoff based on the historical sample library, so as to determine the influence lag corresponding to the forecasting factor , ; the forecasting factor at least including: rainfall and runoff; a third module configured to construct a rainfall-runoff joint similarity index between the historical sample library and the target sample library by integrating rainfall and rainfall variation process, runoff and runoff variation process; a fourth module configured to establish a multi-dimensional classification system, determine a runoff variation mode based on the multi-dimensional classification system, construct a similarity flood forecasting model, train the model similarity flood forecasting model for different runoff variation modes respectively, and obtain model parameters and a forecasting scheme corresponding to each runoff variation mode; a fifth module configured to construct a flood forecasting model, automatically identify a runoff variation mode according to the multi-dimensional classification system, automatically match optimal model parameters and a forecasting scheme based on the identified runoff variation mode, and generate a per-hour runoff forecast in a future period through rolling forecast.
Claims
1. A flood forecasting method that takes into account both similarity and multidimensional classification systems, characterized by, The method comprises the following steps: obtaining historical rainfall data and historical runoff data, and dividing the historical sample library and the test sample library according to time sequence; Based on the historical sample library analysis, the relationship between the prediction factor and the prediction section runoff is analyzed to determine the influence lag time corresponding to the prediction factor , ; the prediction factor at least includes: rainfall and runoff; comprehensively constructing the rainfall-runoff combined similarity index between the historical sample library and the target sample library according to rainfall and rainfall change process and runoff flow and runoff change process; establishing a multi-dimensional classification system, determining the runoff change mode based on the multi-dimensional classification system, constructing a similarity flood forecasting model, training the model similarity flood forecasting model for different runoff change modes, and obtaining the model parameters and the forecasting scheme corresponding to each runoff change mode; constructing a flood forecasting model, automatically identifying the runoff change mode according to the multi-dimensional classification system, automatically matching the optimal model parameters and the forecasting scheme based on the identified runoff change mode, and generating the hourly runoff forecast in the future period through rolling prediction.
2. The flood forecasting method according to claim 1, characterized in that, the prediction factor corresponds to an influence lag The determination process is as follows: pre-setting a quantity threshold of rain gauge sites , acquiring an actual quantity of rain gauge sites in a forecast area : If Then The relationship between the antecedent rainfall and the predicted runoff of each rainfall station is analyzed one by one to obtain the corresponding influence lag time. If , the forecast area is divided into polygonal regions by using the Thiessen polygon method, the polygonal regions are merged according to the data correlation of each rain gauge site to obtain sub-regions, and ; the areal rainfall of each sub-region is calculated, and the relationship between the areal rainfall of each sub-region and the forecast section runoff is analyzed one by one to obtain the corresponding influence lag. 3.The flood forecasting method of claim 1, wherein, The construction process of the rainfall-runoff combined similarity index is as follows: By the rainfall optimal similarity coefficient The similarity of the antecedent rainfall and the rainfall process change between the historical sample library and the target sample library is evaluated to obtain the rainfall similarity between the historical sample library and the target sample library. Utilizing runoff optimal similarity coefficient The similarity of the runoff between the historical sample library and the target sample library is obtained by evaluating the similarity of the runoff size and the runoff change trend between the historical sample library and the target sample library. Based on the optimal similarity coefficient of rainfall and runoff Calculate the joint similarity index of rainfall-runoff by weighted average method : ; where, and are the influence weights for rainfall and runoff, respectively.
4. The flood forecasting method of claim 1, wherein, The multi-dimensional classification system is a three-dimensional system of "rainfall grade-flow rate-rise grade"; correspondingly, the determination process of the runoff change mode comprises one or more of rainfall grade division, flow rate division and rise grade division.
5. The flood forecasting method according to claim 1, characterized in that, The construction process of the similarity flood forecasting model is as follows: For the current timing , respectively, the rainfall time series and runoff time series Express the influence of the previous rainfall and the previous runoff on the forecast runoff, and based on the rainfall time series and runoff time series Integration of the current feature sequence group of the forecast runoff ; find a most similar historical feature sequence group from the historical sample library through a rainfall-runoff combined similarity index most similar group historical feature sequence group and a corresponding similarity change value ; most similar to group history feature sequence group , predict the target sample at current time sequence t based on the target sample library .
6. The flood forecasting method of claim 1, wherein, The prediction scheme comprises at least predicting a current runoff absolute value and / or predicting a current runoff relative value .
7. The flood forecasting method of claim 1, wherein, The model parameters at least include the influence time sequence process of the previous rainfall, the influence time sequence process of the previous runoff, the rainfall weight, the runoff weight, the number of similar samples, and the output form of the forecasting result.
8. The flood forecasting method of claim 4, wherein, The rain grade classification process is as follows: considering the time delay effect, the corresponding forecast area surface rainfall which has influence on the cross section runoff is calculated , and the corresponding rain grade is obtained according to the frequency curve of the surface rainfall in the specified time period of the forecast area. The flow rate division method is as follows: the flow rate thresholds of each grade of flood frequency are set in advance, and the corresponding flow rate grades are divided based on the flow rate thresholds; The rise grade division steps are as follows: division is performed according to the unit time flow rate change rate frequency; wherein the unit time flow rate change rate is determined by the difference between the measured runoff of a previous specified time length and the measured runoff of two previous specified time lengths; the rise grade is divided according to the multiple relationship of the base flow.
9. The flood forecasting method of claim 6, wherein, According to the most similar Group history characteristic sequence group The actual runoff value forecasts the current runoff absolute value : , wherein, is the weight of each historical sample h , and is the measured runoff value of the historical sample at the time . based on the most similar group history feature sequence group actual runoff change value forecasts current runoff relative value : , is a historical sample h at the time of the measured runoff value.
10. A flood forecasting system that takes into account both similarity and multi-dimensional classification systems for implementing the flood forecasting method according to any one of claims 1 to 9, characterized in that, The method comprises the following steps: The first module is configured to obtain historical rainfall data and historical runoff data, and divide the historical sample library and the test sample library according to time sequence; The second module is configured to analyze the relationship between the prediction factor and the prediction section runoff based on the historical sample library to determine the influence lag corresponding to the prediction factor. , The prediction factor at least includes rainfall and runoff. The third module is configured to comprehensively construct the rainfall-runoff combined similarity index between the historical sample library and the target sample library according to rainfall and rainfall change process and runoff flow and runoff change process; The fourth module is configured to establish a multi-dimensional classification system, determine the runoff change mode based on the multi-dimensional classification system, construct a similarity flood forecasting model, train the model similarity flood forecasting model for different runoff change modes, and obtain the model parameters and the forecasting scheme corresponding to each runoff change mode; The fifth module is configured to construct a flood forecasting model, automatically identify the runoff change mode according to the multi-dimensional classification system, automatically match the optimal model parameters and the forecasting scheme based on the identified runoff change mode, and generate the hourly runoff forecast in the future period through rolling prediction.
Citation Information
Cited By
Scene flood forecasting method considering similar incoming water coupling deep learning
CN122022085A