CN118134729B - Intelligent forecasting method and system for urban flood control - Google Patents

Abstract

The intelligent forecasting method for urban flood control includes the following steps: collecting urban flood control data, extracting rainfall data, historical flood data and non-rainfall flood driving factors from the urban flood control data; constructing four flood forecasting models, including a flood forecasting statistical model, a flood forecasting physical model, a flood forecasting machine learning model and a flood forecasting historical similarity model; for each forecasting model, extracting data from urban flood control data and similar rainfall data to construct a training set, and training each forecasting model in sequence; outputting a flood forecast based on each trained forecasting model; and using a weighted method to synthesize the four flood forecasting results to obtain a comprehensive flood forecast, i.e., an intelligent forecast for urban flood control. The present invention improves the accuracy and timeliness of flood forecasting.

Description

Intelligent forecasting method and system for urban flood control

Technical Field

The invention relates to an intelligent forecasting method for urban flood control.

Background technique

In the field of smart cities and disaster prevention and mitigation, urban flood control has always been an important topic. Traditional urban flood forecasting methods mainly rely on historical experience and statistical models. With the development of artificial intelligence technology, neural networks are increasingly being used in the field of urban flood control. For example, convolutional neural networks are used to extract rainfall characteristics and combine terrain, pipe networks and other factors to predict waterlogging risks. Other technologies integrate meteorological radar data, hydrological models and machine learning algorithms to achieve accurate flood risk warnings at the block level.

Although there are many intelligent urban flood forecasting methods, there are still some problems that need to be solved urgently: first, there is a lack of in-depth analysis of the driving factors of flood formation, and the interpretability and applicability of the forecasting model need to be strengthened; second, the fusion and utilization of multi-source heterogeneous data are not sufficient, and the data quality and real-time performance need to be improved; third, the uncertainty assessment and dynamic correction mechanism of the forecasting model are not perfect, and the credibility of the forecast results needs to be improved; fourth, there is a lack of refined forecasting plans for complex urban environments, and the early warning capabilities for extreme rainstorms and waterlogging are insufficient.

Therefore, research and innovation are needed.

Summary of the invention

The purpose of the invention is to provide an intelligent forecasting method for urban flood control to solve the above problems existing in the prior art. On the other hand, an intelligent forecasting system for urban flood control is provided.

Technical solution, according to one aspect of the present application, provides a smart forecasting method for urban flood control, comprising the following steps:

Step S1, collecting urban flood control data, extracting rainfall data, historical flood data and non-rainfall flood driving factors from the urban flood control data; calculating the synchronization between each non-rainfall flood driving factor and the flood process, constructing a set of key non-rainfall flood driving factors, and calculating the mapping relationship between flood events and rainfall data based on historical flood data and its corresponding rainfall data, obtaining similar rainfall, and forming similar rainfall data;

Step S2, constructing four flood forecasting models, including a flood forecasting statistical model, a flood forecasting physical model, a flood forecasting machine learning model and a flood forecasting historical similarity model;

Step S3: for each forecast model, extract data from urban flood control data and similar rainfall data to construct a training set, and train each forecast model in sequence;

Step S4: output flood forecasts based on each trained forecast model; use a weighted method to integrate the four flood forecast results to obtain a comprehensive flood forecast, that is, an intelligent forecast for urban flood control.

According to another aspect of the present application, there is provided a smart forecasting system for urban flood control, comprising:

at least one processor; and

A memory communicatively connected to at least one of the processors; wherein,

The memory stores instructions that can be executed by the processor, and the instructions are used to be executed by the processor to implement the intelligent forecasting method for urban flood control described in any of the above technical solutions.

Beneficial effects: The intelligent forecasting method for urban flood control provides new technical support and solutions for urban flood control, improving the city's flood control capabilities and the efficiency of responding to floods. The relevant technical effects will be described in detail below in conjunction with the specific implementation methods.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 is a flow chart of the present invention.

FIG. 2 is a flow chart of step S1 of the present invention.

FIG. 3 is a flow chart of step S3 of the present invention.

FIG. 4 is a flow chart of step S4 of the present invention.

Detailed ways

As shown in FIG1 , the following technical solution is proposed.

According to one aspect of the present application, a smart forecasting method for urban flood control comprises the following steps:

Step S1, collecting urban flood control data, extracting rainfall data, historical flood data and non-rainfall flood driving factors from the urban flood control data; calculating the synchronization between each non-rainfall flood driving factor and the flood process, constructing a set of key non-rainfall flood driving factors, and calculating the mapping relationship between flood events and rainfall data based on historical flood data and its corresponding rainfall data, obtaining similar rainfall, and forming similar rainfall data;

Step S2, constructing four flood forecasting models, including a flood forecasting statistical model, a flood forecasting physical model, a flood forecasting machine learning model and a flood forecasting historical similarity model;

Step S3: for each forecast model, extract data from urban flood control data and similar rainfall data to construct a training set, and train each forecast model in sequence;

Step S4: output flood forecasts based on each trained forecast model; use a weighted method to integrate the four flood forecast results to obtain a comprehensive flood forecast, that is, an intelligent forecast for urban flood control.

Traditional flood forecasting focuses mainly on rainfall factors, while ignoring the impact of various non-rainfall driving factors such as underlying surface, reservoir scheduling, and tidal support. By introducing mathematical methods such as Fréchet distance and mutual information, the correlation between driving factors and flood processes is quantitatively characterized, and key non-rainfall factors with high synchronization with floods are screened out. The quantitative driving factor analysis method can reveal the internal mechanism of flood formation and discover the causal characteristics of typical flood events. It deepens the understanding of physical mechanisms and provides a solid foundation for subsequent model construction and scenario analysis.

By analyzing the correspondence between historical rainfall and floods, and using two complementary statistical indicators, the Pearson correlation coefficient and the DTW distance, rainfall events with similar characteristics to the current rainfall can be identified from historical data and used as reference samples for subsequent forecasts. This similarity analysis idea makes up for the limitations of a single physical model, and uses the empirical knowledge of historical data to assist forecasting, greatly improving the interpretability and confidence of the forecast. In addition, this method uses dual-index screening, which not only considers the correlation of rainfall, but also takes into account the similarity of rainfall time distribution, ensuring the representativeness of the selected samples and providing high-quality training data for subsequent model integration.

Compared with a single model, this solution integrates multiple modeling paradigms such as statistical learning, physical mechanism, data-driven, and similar reasoning to depict the evolution of floods from different perspectives. Among them, the statistical model focuses on extracting historical laws from the data, the physical model focuses on simulating the flood generation and convergence mechanism, the machine learning model is good at mining the correlation patterns in complex data, and the similarity model focuses on the reasoning and prediction of analogous cases. This multi-perspective, multi-path model integration framework inherits the essence of traditional methods and absorbs new technologies of artificial intelligence. It can give full play to the forecasting effectiveness of various models, take advantage of their strengths and avoid their weaknesses, and make up for their weaknesses, thereby significantly improving the robustness and adaptability of the forecast.

Urban flood forecasting faces challenges such as rapid changes in rainfall and frequent updates of the underlying surface, which places high demands on the real-time and dynamic nature of the forecast model. To address this problem, the solution adopts a rolling update strategy, which continuously incorporates the latest monitoring data into the training samples through a moving time window, and dynamically corrects the model using the measured data during the validation period. At the same time, the solution also introduces an adversarial verification mechanism, which tests the robustness of the model by artificially constructing some extreme scenario data and adjusts the model parameters accordingly. This mode of incremental learning and active verification gives the forecast model the ability to continuously evolve and improve itself, enabling it to adapt to the ever-changing external environment and always maintain a high level of forecast accuracy and reliability.

Since the principles, structures, parameters, training data, etc. of each model are different, their forecast results will inevitably have certain deviations and inconsistencies. How to extract the optimal combined forecast from these seemingly plausible results is a major challenge in the field of integrated learning. To this end, the scheme adopts strategies such as game theory and Bayesian model averaging, optimizes model weights by setting Nash equilibrium objective functions, or dynamically adjusts weight probability distributions using real-time forecast performance. In addition, the scheme also uses technical means such as Monte Carlo discarding to evaluate the uncertainty of forecast results through random perturbations, and uses indicators such as confidence intervals to quantify forecast credibility. These integrated strategies and uncertainty analysis methods can fully tap the complementarity and collective wisdom of multiple models, weaken the one-sidedness of a single model, and make forecast results more robust and balanced.

In short, by constructing a set of key driving factors, we can reveal the flood formation mechanism and lay the foundation for forecasting; by introducing similarity analysis, we can make full use of historical experience and improve the interpretability of forecasts; by developing multi-model integration and integrating multidisciplinary methods, we can enhance the robustness of forecasts; by adopting incremental adversarial learning, we can adapt to dynamic environments and ensure the timeliness of forecasts; by applying Bayesian integration and uncertainty quantification, we can optimize decisions and improve the credibility of forecasts. It can solve the following technical problems existing in existing technologies: quantitative characterization and screening of non-rainfall factor influences; similarity measurement and matching in complex data environments; integrated optimization and automatic updating of multi-source heterogeneous models; robustness and adaptability of forecast models under dynamic change scenarios; interpretability, credibility and uncertainty control of forecast results.

According to one aspect of the present application, step S1 further comprises:

Step S11, collecting urban flood control data, including rainfall data, historical flood data and non-rainfall flood driving factors, wherein the non-rainfall flood driving factors include non-rainfall meteorological data;

Step S12, extracting each historical flood process and the non-rainfall flood driving factor corresponding to the historical flood process from the historical flood data, and constructing a set of non-rainfall flood driving factors for each historical flood process;

Step S13: for each historical flood process, the synchronization between each non-rainfall flood driving factor and the flood process is calculated using the Flechet distance, and a key non-rainfall flood driving factor set for each historical flood process is constructed;

Step S14: Based on the historical flood data and the corresponding rainfall data, a mapping relationship between flood events and rainfall is calculated, similar rainfall is identified, and similar rainfall data is generated.

According to one aspect of the present application, step S13 is further:

Step S13a, collecting data of each non-rainfall flood driving factor in sequence to obtain a data sequence of each non-rainfall flood driving factor;

Step S13b, for each historical flood process, calculating the Fleche distance and mutual information between each non-rainfall flood driving factor and the historical flood process, and identifying the non-rainfall flood driving factor and the flood process whose Fleche distance is less than a threshold and whose mutual information is greater than a threshold as having synchronization higher than a threshold;

Step S13c: for each historical flood process, non-rainfall flood driving factors with synchronization higher than a threshold are sequentially extracted to construct a set of key non-rainfall flood driving factors for each historical flood process.

According to one aspect of the present application, step S14 is further:

Step S14a, based on the historical flood process and its corresponding rainfall data, setting the Pearson correlation coefficient threshold and the DTW distance threshold;

Step S14b: for each historical flood process, use the Pearson correlation coefficient and DTW distance to respectively construct a mapping relationship between the historical flood process and rainfall data;

Step S14c: Calculate the Pearson correlation coefficient and DTW distance of all rainfall data and historical flood processes, and identify rainfall with a correlation coefficient greater than a Pearson threshold and a DTW distance less than the DTW threshold as similar rainfall.

In this embodiment, traditional flood forecasting mainly relies on rainfall and hydrological elements, while ignoring the influence of meteorological factors such as temperature, humidity, and wind speed. In fact, these non-rainfall factors affect the watershed runoff process by affecting the water vapor cycle and underlying surface conditions, and are one of the important contributors to flood formation. This step innovatively proposes the concept of non-rainfall driving factors and incorporates them as important candidate features into subsequent analysis. This strategy of using multiple sources of data helps to understand and characterize the cause mechanism of floods from a more comprehensive perspective, which is a beneficial expansion of the data dimension. In addition, the data types take into account multiple professional fields such as rainfall, hydrology, and meteorology, and the data forms include various types such as actual measurement, statistics, and forecasts, reflecting a strong systematic and comprehensive nature, which provides high-quality and diversified data support for modeling and analysis.

By extracting typical flood events from historical data and extracting relevant non-rainfall influencing factors around each flood, an event-driven factor set is formed. Different from the common time series modeling, this step constructs training samples based on "flood events", and each sample consists of a complete flood process and its influencing factors. This data organization form from an event perspective, on the one hand, respects the continuity and integrity of flood evolution and avoids the mechanical division of time segments, and on the other hand, highlights the individual characteristics of each flood, making it easier to explore the differences between different events. In addition, extracting driving factors around flood events follows the causal logic of hydrological processes, helps to depict the intrinsic connection between each factor and floods, and is a mechanism-oriented sample construction strategy. This event-driven approach can overcome the limitations of traditional time-driven methods and provide a data perspective that is more in line with the laws of flood occurrence, thereby improving the pertinence of subsequent analysis and modeling.

The synchronicity between non-rainfall factors and flood processes is quantified using Fleche distance and mutual information. Fleche distance is a measure of the similarity between two time series, which is more sensitive to the shape characteristics of the series, and is therefore more suitable for describing the consistency between factors and floods in terms of temporal dynamics. Mutual information is an indicator that measures the correlation between two random variables, which is more sensitive to nonlinear relationships, and is therefore more suitable for describing the consistency between factors and floods in terms of numerical correlation. Combining these two metrics, the key factors that best represent flood changes can be screened from the two dimensions of temporal consistency and numerical correlation, which considers both the fit between factors and floods in the shape of the time-course curve and the correlation between factors and floods in terms of numerical variation. This is a dual synchronicity characterization method based on spatiotemporal characteristics. This provides more reliable and comprehensive feature support for subsequent modeling and analysis. In addition, by setting a threshold to filter out factors with high synchronicity, redundant information can be removed, the complexity of subsequent analysis can be reduced, and good pertinence is reflected.

How to quickly find the group of rainfall data that is most similar to the current rainfall from the massive historical rainfall data and use it as a reference sample to guide subsequent forecasts is one of the technical problems that need to be solved. This step is not a simple comparison of rainfall, but a comprehensive consideration of the numerical correlation and time-course distribution similarity of rainfall. Among them, the Pearson correlation coefficient measures the linear correlation strength of the two sets of rainfall data, while the DTW distance measures the shape similarity of the two sets of rainfall time-course curves. These two indicators characterize the rainfall similarity from the perspectives of static values and dynamic changes, respectively, and can comprehensively examine the key characteristics of rainfall. In addition, by setting the thresholds of the two indicators respectively and using the "and" logic to filter out rainfall events that meet both numerical correlation and time-course similarity, the quality of similarity samples can be further improved, making them more representative and comparable. This similar rainfall identification method uses the idea of similarity reasoning and supplements sample information through analogy analysis, which can make up for the shortcomings of the simple data-driven method and improve the interpretability and reliability of the forecast.

In summary, this step introduces non-rainfall driving factors to expand the data dimension and deepen the understanding of flood mechanisms; constructs samples based on flood events, respects the integrity of the flood process, and conforms to hydrological laws; applies Fréchet distance and mutual information to quantify flood synchronization from multiple angles and provide reliable features; combines Pearson correlation coefficient and DTW distance to comprehensively characterize rainfall similarity and utilize similar reasoning ideas. This step effectively solves the following technical problems: how to select the most flood-predicting driving factors from multi-source data; how to design a sample data organization form that conforms to the laws of flood occurrence; how to quantitatively evaluate the correlation and synchronization between driving factors and flood processes; and how to quickly find the most similar rainfall events from historical data as references.

According to one aspect of the present application, step S3 further comprises:

Step S31, extracting the flood process, and dividing the urban flood control data, rainfall data and non-rainfall flood driving factors of each flood process into K continuous processes in chronological order, where K is a natural number greater than 1; at the initial stage, the current flood process is given by four flood forecasting models;

Step S32: taking each time period as a training step, extracting non-rainfall flood driving factors, rainfall data and corresponding historical flood processes for each time period, calculating the Frechet distance between the current flood process and the historical flood data, and screening out similar flood processes; using four flood forecasting models to forecast and output a comprehensive flood forecast, and then using the subsequent processes of similar flood processes as a validation set to validate the four flood forecasting models; until the four trained flood forecasting models are obtained;

Step S33: According to the preconfigured cycle, the model is verified adversarially. Adversarial samples are constructed by adding disturbances to the input non-rainfall flood driving factor data to test the model's prediction ability under extreme conditions. The model's vulnerability to adversarial samples is analyzed, and the model is modified and optimized in a targeted manner to improve the robustness of flood forecasting under harsh conditions.

According to one aspect of the present application, the S32 further includes assigning weights to the four flood forecasting models during the training process of each step length:

Step S32a, using game theory, setting Nash equilibrium as the coordination target, weighting the flood forecasts corresponding to the four flood forecast models, and synthesizing the flood forecast weights corresponding to the four flood forecast models to obtain a comprehensive flood forecast; or,

The four flood forecasting models are integrated by using the Bayesian model averaging method to set the prior probability distribution. According to the real-time prediction performance of each model, its posterior probability is dynamically updated through the Bayesian formula. The prediction results of the four flood forecasting models are weighted averaged according to their posterior probabilities to obtain the comprehensive flood forecast results.

Step S32b, comparing the comprehensive flood forecast with the historical flood data of the actual flood process, and adjusting the weight value based on the difference between the two;

Step S32c, assigning the adjusted weight values to the flood forecasts of the four flood forecast models to obtain a comprehensive flood forecast;

Step S32d, quantifying the uncertainty of the comprehensive flood forecast results, using the Monte Carlo discarding method to randomly discard model neurons and repeat the prediction multiple times to obtain the probability distribution of the flood forecast;

Step S32e, calculate the mean and confidence interval of the flood forecast distribution, and output a probabilistic flood forecast including the uncertainty level; evaluate the reliability of the flood forecast results, and use the confidence interval width as an indicator of the uncertainty size.

In this embodiment, the idea of rolling time window is adopted to construct training data. Traditional data-driven models usually randomly divide data into training set and test set, ignoring the time sequence of samples. This step fully considers the dynamic evolution characteristics of floods. By moving the analysis time window, each flood process is divided into several continuous stages, each of which contains a period of historical data and future data to be predicted. This time-sequential data organization form not only retains the evolution law of floods in different stages, but also takes into account the continuity of the whole process, which is more in line with the actual application scenario of flood forecasting. In addition, this method allows the model to obtain a forecast result at each stage, which can provide more dynamic and real-time forecast information for flood control decisions, reflecting strong practicality. Introducing the integrated forecast of four types of models in the initial stage can quickly give a benchmark forecast, providing a good starting point for subsequent training, which is a typical transfer learning strategy.

Regarding the problem of how to achieve incremental updates and adaptive corrections of model parameters. Flood forecasting faces challenges such as changes in data distribution and updates in mechanism cognition, and the model must be able to continuously learn and improve itself. A complete training-forecasting-verification process is constructed based on stage data. The model is trained and calibrated once at each time step. By feeding subsequent real data back to the model, it can capture new changes in the evolution of floods in a timely manner and continuously reduce forecast deviations. From the perspective of training data sources, a variety of heterogeneous data such as non-rainfall influencing factors, rainfall data, and similar historical floods are integrated, enabling the model to fully learn the internal mechanism of flood formation. From the perspective of training results, using the subsequent process of similar floods as a verification set can objectively evaluate the prediction ability of the model, in order to achieve better extrapolation performance for real floods. This mode of training, forecasting, rolling updates, and continuous verification gives the model the ability to continuously evolve, which not only improves the dynamics of the forecast, but also enhances the interpretability of the forecast. It is a personalized modeling strategy that is adapted to local conditions.

This embodiment also uses strategies such as cooperative game theory and Bayesian model averaging to optimize the integration weights of multiple models. Game theory can achieve Pareto optimality of multi-model forecasting from a dynamic non-cooperative perspective by setting the Nash equilibrium target. This equilibrium is conducive to improving the robustness of the forecast. The Bayesian average rule uses real-time forecast performance to adjust the model weights through posterior probability updates, so that the integration effect can adaptively approach the optimal model. This adaptability is conducive to improving the accuracy of the forecast. The core of the two integration strategies is to use the model mutual evaluation mechanism to achieve the survival of the fittest, dynamically select models with good forecasting effects to strengthen them, and thus maximize the overall forecasting effectiveness. This intelligent model selection and fusion method is crucial to improving the adaptability of the forecasting system in complex and changeable flood scenarios.

Finally, the comprehensive forecast results of the current stage are compared with the subsequent actual flood process to form a forecast deviation, and the weight parameters of each model are adjusted accordingly. This is actually a supervised learning process. By feeding back the real results to the model, it can learn from the error, self-calibrate, and continuously approach the optimal state. The adjusted model weights are reapplied to the current stage to generate an updated comprehensive forecast. This dynamic feedback control mechanism gives the model the ability to continuously self-correct. On the one hand, it can reduce the forecast deviation, and on the other hand, it can adapt to the new trend of flood evolution and maintain the effectiveness of the forecast. This feedback control idea draws on modern control theory and is a typical closed-loop optimization strategy. It is of great significance to improve the adaptability and robustness of the forecast model.

By introducing the concepts of probabilistic forecasting and uncertainty analysis. Traditional flood forecasting mostly uses deterministic models to give a fixed forecast value, ignoring the randomness of the flood itself and the uncertainty of the model. This step innovatively applies techniques such as Monte Carlo discarding, randomly perturbing the model structure, and repeating the prediction many times to obtain a probability distribution of the forecast results. This paradigm shift from determinism to probability, on the one hand, respects the inherent uncertainty of flood phenomena more, and on the other hand, it also provides quantitative indicators for the reliability of the model. Based on the mean and confidence interval of the forecast distribution, more abundant reference information can be provided for flood control decisions. The mean reflects the expected level of the forecast, while the confidence interval reflects the credibility of the forecast. The width of the confidence interval determines the size of the uncertainty, and its dynamic changes also reveal the changing trend of the model reliability. This way of perceiving and expressing uncertainty is a breakthrough in the traditional deterministic model, making the forecast results more comprehensive and transparent, and providing an important basis for risk assessment and management.

In some embodiments, no matter how well the model is trained, it is inevitable to encounter some extreme abnormal situations in practical applications. This step deliberately challenges and confuses the model by artificially constructing some adversarial non-rainfall driving factor data to test its predictive ability under harsh conditions. This adversarial verification process can help discover the weak links of the model and reveal the key factors that may affect the forecast effect. Based on the model vulnerability analysis, the model structure or parameters can be adjusted according to local conditions to improve the anti-interference ability. In addition, incorporating adversarial samples into training can play a role in data enhancement, enabling the model to learn more extreme patterns and enhance the extrapolation ability of the forecast. This mechanism of active verification and targeted optimization is a guarantee of the robustness and generalization of the model, and is crucial to improving the practicality and reliability of the forecast system. This idea draws on the concept of adversarial learning in the field of artificial intelligence and represents a new direction for model testing and optimization.

In short, the use of rolling time windows conforms to the law of flood dynamic evolution and improves the timeliness of forecasts; the construction of a training-forecast-verification process realizes incremental model updates and adaptive corrections to improve forecast continuity; the integration of game theory and Bayesian averaging optimizes model integration strategies to improve forecast adaptability; the introduction of forecast deviation feedback control realizes model self-correction and improves forecast accuracy; the application of the Monte Carlo discarding method quantifies forecast uncertainty and improves forecast transparency; adversarial verification and optimization are carried out to enhance model robustness and improve forecast reliability. These technologies solve the following technical problems: how to establish a model training paradigm that conforms to the law of flood dynamic evolution; how to realize the continuous evolution and self-improvement of model parameters; how to weigh the forecast results of multi-source heterogeneous models; how to reduce forecast deviations and improve prediction accuracy; how to characterize the inherent uncertainty of forecasts; how to evaluate and improve the robustness of models.

According to one aspect of the present application, step S4 is further:

Step S41, respectively calculating the four trained flood forecasting models to obtain the corresponding flood forecasts;

Step S42: Use a weighted method to obtain a final comprehensive flood forecast, that is, an intelligent forecast for urban flood control.

According to one aspect of the present application, step S41 is further:

Step S41a, extracting flood data and rainfall data that have occurred in the current period and their corresponding non-rainfall flood driving factors and inputting them into the flood forecast statistical model to obtain the flood forecast of the flood forecast statistical model;

Step S41b, extracting flood data, rainfall data, terrain data, land use data, and town layout planning that have occurred in the current period and inputting them into the flood forecasting physical model to obtain a flood forecast of the flood forecasting physical model;

Step S41c, extracting flood data, rainfall data, non-rainfall meteorological data, terrain data, land use data, and town layout planning that have occurred in the current period and inputting them into the flood forecasting machine learning model to obtain a flood forecast of the flood forecasting machine learning model;

Step S41d, extracting flood data and rainfall data that have occurred in the current period, and historical similar flood data corresponding to the current rainfall data, inputting them into the flood forecast historical similarity model, and obtaining the flood forecast of the flood forecast historical similarity model.

According to one aspect of the present application, the S42b further comprises:

Step S42b1, extracting and classifying flood types according to the comprehensive flood forecast, and calculating the similarity between the flood process and the historical flood process for each flood type;

Step S42b2: Calculate the similarity probability distribution between the comprehensive flood forecast and various historical floods based on the classification results and similarity measurement of the flood process;

Step S42b3: According to the similarity probability distribution of the comprehensive flood forecast, the weight value is adjusted, and different similarity measurement methods are weightedly integrated to correct the forecast result to obtain the final forecast result.

According to one aspect of the present application, step S42b1 is further:

Based on historical flood data, floods are classified into flood types, including slow-rise and slow-recession floods, steep-rise and steep-fall floods, multi-peak floods, and gentle floods.

Cluster analysis was performed on historical flood events and comprehensive flood forecast processes to obtain the central flood processes of various flood types;

The similarity between each flood process and various central flood processes is calculated and classified. For slow-rising and slow-falling floods, the DTW distance is used to calculate the similarity; for steep-rising and steep-falling floods, the Euclidean distance is used to calculate the similarity; for multi-peak floods, the flood peak characteristic weighted Euclidean distance is used to calculate the similarity; for gentle floods, the area similarity coefficient is used.

In this embodiment, the input features required by different models are extracted in a targeted manner. The four types of models have different focuses on modeling ideas and required data: statistical models mainly use time series data such as rainfall and floods to establish forecast relationships through correlation and causal analysis; physical models rely more on geographic information data such as terrain and land use, and obtain forecast results through mechanism deduction and numerical simulation; machine learning models comprehensively use multi-source heterogeneous data such as meteorology, hydrology, geography, and planning, and achieve forecasts through data mining and pattern recognition; similarity models focus on using historical flood data and give forecasts through analogy analysis and similarity reasoning. This step fully considers the characteristics of different models, and selects the feature subsets that best reflect their respective advantages according to local conditions. It is a typical personalized and differentiated data preparation strategy. This approach meets the modeling needs of different models on the one hand, and avoids the interference of invalid redundant data on the other hand, which can maximize the forecasting efficiency of each model. In addition, the data types cover multiple professional fields such as rainfall, floods, meteorology, terrain, land use, and urban layout, reflecting a strong comprehensiveness and systematicity.

How to weigh the forecast results of multiple models to form a comprehensive and reliable final forecast. Due to differences in the basic assumptions, structural frameworks, parameter settings, etc. of different models, their forecast results are often different, and may even have large deviations. How to extract a comprehensive forecast with good consistency from these seemingly contradictory results is a common challenge faced by multi-model integration. In response to this problem, this step proposes a model correction method based on similarity analysis and probability fusion. First, the current flood is classified into a typical category according to the comprehensive forecast, and its similarity with the historical flood of this type is calculated. This process uses the idea of similar reasoning. By comparing and analyzing similar floods in history, the rationality of the comprehensive forecast can be preliminarily judged. Then, based on the similarity analysis, the probability distribution of the comprehensive forecast and various floods is obtained, and the weight coefficients of each model are dynamically adjusted. This is a typical case of combining probability theory with model integration. The probability distribution reflects the possibility of different flood types, and it is used as a reference to optimize the model combination strategy, which can enhance the adaptability of the forecast. Finally, the adjusted weights are used to perform weighted averaging of the multi-model forecasts, and the results are error corrected to obtain the final corrected forecast. This idea of combining fusion and correction not only plays the indexing role of probability weights, but also takes into account the correction of forecast deviations, thereby maximizing the reliability of the comprehensive forecast.

At the same time, this embodiment adopts differentiated similarity measurement strategies for different types of floods. Traditional similar flood identification mostly adopts fixed distance measurement, such as Euclidean distance, Manhattan distance, etc., ignoring the differences in the cause mechanism and evolution law of different flood types. In response to this problem, this step proposes a classified similarity analysis method. First, based on the cluster analysis of historical flood data, typical flood categories such as slow rise and slow retreat type, steep rise and steep fall type, multi-peak type, and gentle type are summarized, and the central flood process of each type is determined. This is a semi-supervised learning idea. Then, the similarity between the comprehensive forecast flood and various central floods is calculated, and the category to which it belongs is judged accordingly. In the similarity calculation, different measurement methods are adopted for different categories: for slow rise and slow retreat type floods, DTW distance is used to highlight the overall similarity of the time series; for steep rise and steep fall type floods, Euclidean distance is used to highlight the local similarity of the waveform; for multi-peak floods, flood peak feature weighting is used to highlight the similarity of key features; for gentle floods, area similarity coefficient is used to highlight the similarity of cumulative total amount. This similarity analysis based on different flood types, on the one hand, respects the characteristics of different flood types, and on the other hand, reflects the guiding role of hydrological mechanism knowledge in data mining. It is a fusion of qualitative experience and quantitative calculation, which can more accurately characterize the similarities of floods and provide a more reliable reference for subsequent model optimization.

In general, by extracting model input features in a personalized way, we can meet differentiated needs and give full play to the strengths of each model; apply similarity analysis to initially judge the rationality of forecasts, draw on similar reasoning ideas, and integrate historical experience; dynamically adjust model weights, introduce probability distributions, and optimize integration strategies; give equal weight to weighted average and error correction, coordinate weight fusion and deviation correction, and form reliable forecasts; for different flood types, use differentiated similarity metrics, integrate qualitative experience and quantitative analysis. The following technical problems have been solved: how to take into account the characteristics of different models and extract the optimal input features; how to use historical flood experience to initially judge the rationality of forecast results; how to weigh the forecast results of different models to form a consistent forecast; how to reduce forecast deviations and improve prediction accuracy; how to enhance the pertinence and interpretability of similarity analysis.

In this application, non-rainfall driving factors include: Northern Hemisphere Polar Vortex Area Index (NHPVA), Northern Hemisphere Polar Vortex Intensity Index (NHPVI), Northern Hemisphere Polar Vortex Center Meridional Position Index (NHPVCLON), Northern Hemisphere Polar Vortex Center Latitudinal Position Index (NHPVCLAT), Western Pacific Subtropical High Area Index (WPSHA), Western Pacific Subtropical High Intensity Index (WPSHI), Western Pacific Subtropical High Ridgeline Position Index (WPSHRP), Western Pacific Subtropical High Westward Extension Ridge Point Index (WPSHWRP), Western Pacific Subtropical High Northern Limit Position Index (WP SHNBP), Eurasian zonal circulation index (EZC), Eurasian meridional circulation index (EMC), East Asian trough position index (EATP), East Asian trough intensity index (EATI), Nino1+2 area SST index, Nino3 area SST index, Nino4 area SST index, Nino3.4 area SST index, trough (T), ridge (R), high-level jet (HJ), shear line (SL), low vortex (VO), cyclone (CL), front (FS), typhoon (TY), cold air (CA) and low-level jet (LJ).

It should be noted that during training, game theory or Bayesian averaging methods are used to adjust weights so that the forecast model can fit better. When using this model, the current flood process can be compared with the historical flood process for similarity, and then weights are given based on the similarity, and verification and comparison are performed to provide more accurate data.

In another embodiment of the present application, the calculation process of the Fréchet distance and the mutual information is as follows:

Calculation of Fréchet distance, given the time series data of flood driving factors and flood processes X={x1, x2, ..., xn} and Y={y1, y2, ..., yn}, where n is the length of the sequence. The calculation steps of Fréchet distance are as follows:

Calculate the Euclidean distance from each element xi in X to all elements in Y and get the distance matrix D{X→Y}:

D _X→Y (i, j) = sqrt((xi - yj) ² ), i, j = 1, 2, ..., n;

For each element xi in X, find the nearest element y{j _i } in Y, that is: j _i = argmin _j D _X→Y (i, j);

Calculate the distance from xi to y{ _ji } _dX→Y (i) = DX _→Y (i, j_i).

Similarly, the Euclidean distance matrix D _Y→X from each element yj in Y to all elements in X is calculated, and the nearest neighbor element xi _j of yj in X _is found and the distance d _Y→X (j) is calculated.

Take the maximum value of the distances in the two directions to obtain the final Fréchet distance: d _F (X, Y) = max(max _i d _X→Y (i), max _j d _Y→X (j)); the smaller the Fréchet distance, the more similar the shapes of the two time series are.

Mutual information is used to measure the mutual dependence between two random variables. For discrete random variables X and Y, mutual information is defined as: I(X;Y) =∑ _x∈X ∑ _y∈Y P(x,y)log ₂ (P(x,y)/(P(x)P(y));

Where P(x) and P(y) are the marginal probability distributions of X and Y respectively, and P(x, y) is the joint probability distribution. The steps for calculating mutual information are as follows:

Discretize the time series data X and Y of flood driving factors and flood processes. You can use equal-width binning or equal-frequency binning to convert continuous variables into discrete variables.

Calculate the empirical probability distribution P(x) and P(y) of discretized X and Y. That is, count the frequency of occurrence of each discrete value and then divide it by the total number of samples.

Calculate the empirical joint probability distribution P(x, y) of X and Y. For each pair of discrete value combinations (x, y), count the frequency of their co-occurrence and divide it by the total number of samples. Substitute the above probability distribution into the mutual information formula and calculate I(X; Y). The larger the mutual information, the stronger the correlation between the two factors.

Uncertainty quantification of comprehensive flood forecasting: Randomly discard a portion of neurons in the trained model and estimate the uncertainty of the results by repeating the prediction multiple times. For the comprehensive flood forecasting model, the implementation steps of MC-DO are as follows:

A dropout layer is added after multiple fully connected layers of the comprehensive flood forecasting model. Dropout means that during the training process, the output of a part of neurons is randomly set to zero with a certain probability p.

Use the model with dropout to fit the training data until the model converges. During training, a portion of neurons are randomly dropped when each batch of data passes through the model. This is equivalent to training a set of models with shared parameters.

In the test phase, for each input sample x, T random discard forward predictions are performed. Each time a prediction is made, a portion of the neurons in the model are randomly discarded to generate a prediction result y*t. This is equivalent to sampling T models from the model set for prediction.

Step 4: Calculate the mean and variance of T prediction results {y*1, y*2, ..., y*T}: E(y*) = (1/T)∑ _t=1 ^T (y*t); Var (y*) = (1/T)∑ _t=1 ^T (y*t- E(y*)) ² ; the mean reflects the expectation of the prediction, and the variance reflects the uncertainty of the prediction.

Assuming that the prediction results follow a normal distribution and given a confidence level α, calculate the confidence interval:

[E(y*)-(z _α /2)sqrt(Var (y*)), E(y*)+(z _α /2)sqrt(Var (y*))]

Where z _α /2 is the α/2 quantile of the standard normal distribution. The confidence interval gives the uncertainty range of the prediction results. By randomly discarding neurons to approximate the Bayesian neural network, the uncertainty of the model is estimated at a low computational cost. The wider the confidence interval, the more uncertain the model's predictions are.

Adversarial validation is used to evaluate the robustness of machine learning models. Its goal is to generate some adversarial samples with slight perturbations, so that the model's predictions on these samples change significantly, thereby detecting the model's vulnerability. The general steps of adversarial validation are as follows:

Select a trained flood forecast model fθ, where θ is the model parameter.

Randomly select a sample x from the test set and predict its flood label y*= fθ according to the model.

Construct an adversarial objective function J(x') so that when the input is the perturbed sample x', the model's prediction result is significantly different from the original prediction result y*. Commonly used objective functions include cross entropy loss, maximum confidence loss, etc.

Use optimization algorithms such as gradient ascent to maximize the objective function and generate adversarial perturbations δ:

maxJ(x+δ),s,t,||δ|| _p ≤ ε;

Where ||δ|| _p represents the Lp norm, and ε is the size limit of the perturbation. Usually p is taken to be infinity, then ||δ|| _{infinity≤ε} means that each element of the perturbation does not exceed ε.

Superimpose the perturbation on the original sample to obtain the adversarial sample: x _adv = x +δ.

Input the adversarial sample into the model and get its predicted label y* _adv = fθ(x _adv ), and compare the difference between y* _adv and y*. If the difference is large, it means that the model is sensitive to perturbations and is not robust enough.

Repeat the above steps to evaluate the average adversarial robustness of the model on a large number of samples. Robustness indicators can be the prediction accuracy of adversarial samples, average prediction loss, etc. Analyze the adversarial vulnerabilities of the model, improve the model structure or training method, and improve the model robustness. Common methods include adversarial training, gradient regularization, input preprocessing, etc.

Adversarial verification helps build a more secure and reliable flood forecasting system by simulating potential adversarial attacks and proactively identifying model weaknesses.

The process of step S42b1 is specifically as follows:

According to the causes, processes and characteristics of floods, floods can be divided into the following categories: slow-rising and slow-falling floods, steep-rising and steep-falling floods, multi-peak floods and gentle floods. For the above four types of floods, an adaptive similarity calculation method is designed:

Cluster analysis is performed on historical flood events and comprehensive flood forecast processes to obtain the central flood processes of various floods. Common clustering algorithms such as K-means and hierarchical clustering can be used.

Calculate the similarity between each flood process and various central flood processes and classify them. The similarity calculation adopts the following method:

For slow-rise and slow-recession floods, the Dynamic Time Warping (DTW) distance is used:

Step 1: Normalize the water level series Q1={q11, q12, ..., q1n} and Q2={q21, q22, ..., q2m} of the two flood processes to eliminate the dimension effect.

Step 2: Construct a distance matrix D∈R ^n×m between two sequences, where D(i, j) represents the Euclidean distance between q1i and q2j.

Step 3: Find an optimal matching path from D(1, 1) to D(n, m) on the distance matrix so that the cumulative distance on the path is minimized. This path can be solved by dynamic programming:

M(i, j) = D(i, j) + min{M(i-1, j-1), M(i-1, j), M(i, j-1)}; where M(i, j) represents the minimum cumulative distance from (1, 1) to (i, j).

Step 4: The cumulative distance M(n, m) of the optimal matching path is the DTW distance of the two flood processes, which is used to measure their similarity.

For steep rise and fall floods, use the Euclidean distance:

Step 1: Normalize the water level series Q1 and Q2 of the two flood processes to eliminate the dimension effect.

Step 2: Pad the two sequences to the same length L. Missing values can be filled using linear interpolation.

Step 3: Calculate the Euclidean distance between corresponding elements of two sequences: d(Q1, Q2) = sqrt{sum _i=1 ^L (q1i -q2i) ² }

The smaller the Euclidean distance, the more similar the two flood processes are.

For multi-peak floods, the weighted Euclidean distance of flood peak characteristics is used:

Step 1: Extract features of the two flood processes Q1 and Q_2 to obtain their respective peak water levels {p11, ..., p1k1 and {p21, ..., p2k2}, and peak occurrence times {t11, ..., t1k1} and {t21, ..., t2k2}.

Step 2: Calculate the flood peak characteristic distance matrix Dp∈R ^k1×k2 } of the two flood processes, where Dp(i, j) = wp|p1i- p2j| + wt|t1i - t2j|; wp and wt are the weights of the flood peak water level and occurrence time, respectively.

Step 3: The Hungarian algorithm solves the optimal match of the flood peak feature distance matrix and obtains the weighted sum of the flood peak matching distances as the similarity measure of the two flood processes.

For mild floods, use the area similarity factor:

Step 1: Normalize the water level series Q1 and Q2 of the two flood processes to eliminate the dimension effect.

Step 2: Calculate the areas A1 and A2 enclosed by the two water level series and the time axis, which can be approximated by the trapezoidal area method.

Step 3: Calculate the area similarity coefficient of the two flood processes: s(Q1, Q2) = (2min(A1, A2))/(A1 +A2);

The area similarity coefficient ranges from [0, 1]. The closer it is to 1, the more similar the two flood processes are.

Step S3c: Based on the classification results and similarity measures of the flood process, the similarity probability distribution between the comprehensive forecast flood and various historical floods is calculated.

Step S3d: Based on the similarity probability distribution of the comprehensive flood forecast, the forecast results obtained by different similarity measurement methods are weighted and integrated to generate the final probability forecast.

According to another aspect of the present application, in step S31, the process of constructing the training set is further as follows:

Step S31a, extracting the collected flood process and dividing it into N time periods;

Step S31b, extracting key non-rainfall flood driving factor data for N time periods respectively;

Step S31c, input the key non-rainfall flood driving factor data, hydrological data and non-rainfall meteorological data of N time periods into four flood forecasting models for calculation, and use the linear weighted method to obtain a comprehensive flood forecast, extract N flood processes, and record them as training sets.

Step S32a, extracting one flood process from the N flood processes, calculating the Fleche distance between the flood process and all historical flood data in turn, and identifying the historical flood data whose Fleche distance between the two is less than a threshold as a similar flood process;

Step S32b, sequentially calculating similar flood processes corresponding to N flood processes;

Step S32c, extracting subsequent flood processes of similar floods as a verification set.

We extracted 500 historical flood processes in a city and divided each flood process into K segments with a step length of 10 minutes. The last segment of less than 10 minutes is a separate segment. The corresponding urban flood control data, rainfall data and non-rainfall flood driving factors are also divided into corresponding K segments.

Four flood forecasting models were trained and validated using each of the 50 historical flood processes. The specific operations are as follows:

For the a-th historical flood process, a=1, 2, ..., 50, the current flood process is calculated using four flood forecasting models based on the first section (i.e., the first 10 minutes) of urban flood control data, rainfall data, and non-rainfall flood driving factors, and the model parameters are trained to minimize the error between the calculation process and the actual process of the first section of the a-th historical flood process;

Calculate the Fréchet distance between the current flood process and the other 499 historical flood data, select the most similar historical flood process, and use the second section of that flood (i.e. the second 10 minutes) as the validation set;

And so on, going through each section of the a-th historical flood process;

In this way, four flood forecasting models are trained and verified using each of the 50 historical flood processes.

The model is further verified by adding disturbances to the input non-rainfall flood driving factor data to construct adversarial samples, and the model's predictive ability under extreme conditions is tested. The model's vulnerability to adversarial samples is analyzed, and targeted model corrections and optimizations are performed to improve the robustness of flood forecasts under harsh conditions.

According to another aspect of the present application, there is provided a smart forecasting system for urban flood control, characterized in that it includes:

at least one processor; and

a memory communicatively connected to at least one of the processors; wherein,

The memory stores instructions that can be executed by the processor, and the instructions are used to be executed by the processor to implement any of the above-mentioned intelligent forecasting methods for urban flood control.

The preferred embodiments of the present invention are described in detail above; however, the present invention is not limited to the specific details in the above embodiments. Within the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all belong to the protection scope of the present invention.

Claims (5)

1. The intelligent forecasting method for urban flood control is characterized by comprising the following steps:

step S1, urban flood control data are collected, and rainfall data, historical flood data and non-rainfall flood driving factors are extracted from the urban flood control data; calculating synchronicity between each non-rainfall flood driving factor and the flood process, constructing a key non-rainfall flood driving factor set, calculating a mapping relation between flood events and rainfall data based on historical flood data and corresponding rainfall data thereof, obtaining similar rainfall, and forming similar rainfall data;

s2, constructing four flood forecast models, including a flood forecast statistical model, a flood forecast physical model, a flood forecast machine learning model and a flood forecast historical similarity model;

Step S3, aiming at each forecast model, respectively extracting data from urban flood control data and similar rainfall data to construct a training set, and training each forecast model in sequence;

s4, outputting flood forecast based on each forecast model after training; synthesizing four flood forecast results by adopting a weighting method to obtain a comprehensive flood forecast, namely, an intelligent forecast of urban flood control;

The step S1 is further:

Step S11, urban flood control data are collected, wherein the urban flood control data comprise rainfall data, historical flood data and non-rainfall flood driving factors, and the non-rainfall flood driving factors comprise non-rainfall meteorological data;

step S12, extracting each historical flood process from the historical flood data and a non-rainfall flood driving factor corresponding to the historical flood process, and constructing a non-rainfall flood driving factor set of each historical flood process;

step S13, calculating synchronicity between each non-rainfall flood driving factor and the flood process by adopting the Frechet distance according to each historical flood process, and constructing a key non-rainfall flood driving factor set of each historical flood process;

Step S14, calculating the mapping relation between flood events and rainfall based on historical flood data and rainfall data corresponding to the historical flood data, identifying similar rainfall, and forming similar rainfall data;

The step S13 is further:

step S13a, sequentially collecting data of each non-rainfall flood driving factor to obtain a data sequence of each non-rainfall flood driving factor;

Step S13b, calculating the French distance and mutual information between each non-rainfall flood driving factor and the historical flood process according to each historical flood process, and determining that the synchronism of the non-rainfall flood driving factors and the flood processes is higher than the threshold value when the French distance is smaller than the threshold value and the mutual information is larger than the threshold value;

Step S13c, sequentially extracting non-rainfall flood driving factors with synchronism higher than a threshold value according to each historical flood process, and constructing a key non-rainfall flood driving factor set of each historical flood process;

the step S14 is further:

Step S14a, setting a Pearson correlation coefficient threshold and a DTW distance threshold based on a historical flood process and rainfall data corresponding to the historical flood process;

step S14b, respectively constructing mapping relations between the historical flood processes and rainfall data by adopting Pearson correlation coefficients and DTW distances according to each historical flood process;

Step S14c, calculating Pearson correlation coefficients and DTW distances of all rainfall data and the historical flood process, and identifying rainfall with the correlation number larger than the Pearson threshold and the DTW distance smaller than the DTW threshold as similar rainfall;

The step S3 is further:

s31, extracting flood processes, and dividing urban flood control data, rainfall data and non-rainfall flood driving factors of each flood process into K sections of continuous processes according to time sequence, wherein K is a natural number larger than 1; at the beginning, the current flood process is given out through four flood forecast models;

Step S32, taking each time period as a training step length, sequentially extracting non-rainfall flood driving factors, rainfall data and corresponding historical flood processes of the rainfall driving factors and the rainfall data for each time period, calculating the French distances of the current flood process and the historical flood data, and screening out similar flood processes; forecasting by adopting four flood forecasting models, outputting comprehensive flood forecasting, and then verifying the four flood forecasting models by taking the follow-up process of the similar flood process as a verification set; until four trained flood forecast models are obtained;

Step S33, according to a preconfigured period, performing countermeasure verification on the model, constructing a countermeasure sample by adding disturbance to input non-rainfall flood driving factor data, and testing the prediction capability of the model under extreme conditions; analyzing vulnerability of the model to the countermeasure sample, and pertinently carrying out model correction and optimization to improve the robustness of flood forecast under severe conditions;

The step S4 is further:

step S41, respectively calculating four trained flood forecast models to obtain respective corresponding flood forecast;

s42, comprehensively obtaining a final comprehensive flood forecast, namely, an intelligent forecast of urban flood control by adopting a weighting method;

the step S41 is further:

step S41a, extracting flood data, rainfall data and corresponding non-rainfall flood driving factors which have occurred in the current period, and inputting the flood data and the rainfall data into a flood forecast statistical model to obtain a flood forecast of the flood forecast statistical model;

step S41b, extracting flood data, rainfall data, topography data, land utilization data and town layout planning which are generated in the current period, and inputting a flood forecast physical model to obtain a flood forecast of the flood forecast physical model;

step S41c, extracting flood data, rainfall data, non-rainfall meteorological data, topographic data, land utilization data and town layout planning which have occurred in the current period, and inputting a flood forecast machine learning model to obtain a flood forecast of the flood forecast machine learning model;

Step S41d, extracting flood data, rainfall data and historical similar flood data corresponding to the current rainfall data, which are generated in the current period, and inputting a flood forecast historical similarity model to obtain a flood forecast of the flood forecast historical similarity model.

2. The intelligent forecasting method for urban flood control according to claim 1, wherein S32 further comprises weighting four flood forecasting models during each step of training:

Step S32a, adopting a game theory method, setting Nash equilibrium as a coordination target, giving weights to flood forecast corresponding to the four flood forecast models, and synthesizing flood forecast weight values corresponding to the four flood forecast models to obtain a comprehensive flood forecast; or alternatively

Integrating four flood forecast models by adopting a Bayesian model averaging method, and setting prior probability distribution; dynamically updating the posterior probability through a Bayesian formula according to the real-time prediction performance of each model; weighting and averaging the prediction results of the four flood prediction models according to posterior probability thereof to obtain a comprehensive flood prediction result;

Step S32b, comparing the comprehensive flood forecast with the history flood data actually generated in the flood process, and adjusting the weight value based on the difference of the comprehensive flood forecast and the history flood data;

Step S32c, giving the adjusted weight values to the flood forecast of the four flood forecast models to obtain a comprehensive flood forecast;

Step S32d, carrying out uncertainty quantification on the comprehensive flood forecast result, adopting a Monte Carlo discarding method to randomly discard model neurons and repeatedly predicting for a plurality of times to obtain probability distribution of flood forecast;

Step S32e, calculating the mean value and the confidence interval of flood forecast distribution, and outputting a probabilistic flood forecast containing uncertainty level; and evaluating the reliability of the flood forecast result, and taking the width of the confidence interval as an index of uncertainty.

3. The intelligent forecasting method for urban flood control according to claim 1, wherein S42 is further:

Step S42b1, extracting flood types according to comprehensive flood forecast, classifying, and calculating the similarity degree of the flood process and the historical flood process according to each flood type;

step S42b2, calculating the similarity probability distribution of the comprehensive forecast flood and various historical floods according to the classification result and the similarity measurement of the flood process;

Step S42b3, according to the similarity probability distribution of the comprehensive forecasting flood, adjusting the weight value, weighting and fusing different similarity measurement methods, correcting the forecasting result, and obtaining a final forecasting result.

4. A smart prediction method for urban flood control according to claim 3, wherein said step S42b1 is further:

classifying floods based on historical flood data to obtain flood types, including slow-rise slow-fall floods, steep-rise steep-fall floods, multimodal floods and gentle floods;

performing cluster analysis on the historical flood event and the comprehensive forecast flood process to obtain a central flood process of various floods;

Calculating the similarity of each flood process and each central flood process, and classifying, wherein for slow-swelling slow-annealing type floods, the similarity is calculated by using a DTW distance; for steep rising and falling type floods, calculating the similarity by using Euclidean distance; for multi-modal floods, calculating similarity by using the flood peak characteristic weighted Euclidean distance; for gentle floods, area similarity coefficients are used.

5. An intelligent forecasting system for urban flood control, comprising:

at least one processor; and

A memory communicatively coupled to at least one of the processors; wherein,

The memory stores instructions executable by the processor for execution by the processor to implement the smart prediction method for urban flood control of any one of claims 1 to 4.