Flood forecasting method based on multi-climate-zone adaptive LSTM network

Abstract

The invention relates to the technical field of flood forecasting and artificial intelligence, and is particularly suitable for a hydrological disaster early warning scene in a multi-climate area. The invention provides a multi-factor collaborative modeling adaptive LSTM network flood forecasting method based on hydrological mechanism difference of multiple climate regions and the current situation of frequent occurrence of extreme events. According to the method, through multi-source data preprocessing, climate-topographic feature cooperative calculation, adaptive gating adjustment, cross-region migration optimization and forecast result correction, high-precision forecast of flood in different climate regions is realized; the core of the method is to construct a climate-terrain-extreme event linkage feature system and a dynamic gating mechanism, solve the problems of poor adaptability and inaccurate extreme event prediction of a traditional model, and provide technical support for intelligent water conservancy disaster prevention and reduction.

Description

Technical Field

[0001] This invention relates to the field of flood forecasting technology, and in particular to a flood forecasting method based on a multi-climate zone adaptive LSTM network. Background Technology

[0002] Flood forecasting is a core component of disaster prevention and mitigation systems, and its accuracy directly impacts the safety of people's lives and property and socio-economic stability. Hydrological mechanisms differ significantly across climate zones. Humid regions exhibit strong spatiotemporal heterogeneity in precipitation, semi-humid regions show complex dynamic changes in the evaporation-precipitation balance, and semi-arid regions experience frequent extreme precipitation events with pronounced hydrological abrupt changes. Traditional flood forecasting models often employ a uniform architecture, failing to adequately consider the hydrological characteristics arising from climate zone differences, resulting in insufficient model adaptability. While LSTM and its variants are used in flood forecasting, most focus only on a single climate zone or a single influencing factor, neglecting to construct a feature system that integrates climate, topography, and extreme events.

[0003] In practical applications, the synergistic effects of humidity gradient changes and evaporation-precipitation coupling, the attenuation effect of topographic slope on hydrological response, and the interference of spatiotemporal clustering characteristics of extreme events on forecast accuracy have not been effectively addressed, resulting in low forecast accuracy across climate zones, high rates of missed and false alarms for extreme flood events, and poor parameter adaptability when migrating models across regions.

[0004] The aforementioned problems severely restrict the reliability of flood forecasting and fail to meet the actual needs of disaster prevention and mitigation in multiple climate zones. There is an urgent need for a flood forecasting method that can dynamically adapt to the characteristics of different climate zones by coordinating multiple factors. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing a flood forecasting method based on a multi-climate zone adaptive LSTM network, which includes multi-source data acquisition and preprocessing steps, and further includes the following steps:

[0006] The climate-topography feature co-calculation steps integrate humidity gradient change rate, evaporation-precipitation coupling elasticity coefficient, topographic slope attenuation coefficient, and spatiotemporal clustering of extreme events to construct a multi-factor linkage feature system;

[0007] The adaptive LSTM gating weight adjustment step dynamically matches the hydrological characteristics of different climate zones based on the multi-factor linkage feature system, selects the optimal LSTM variant, and adjusts the gating weights.

[0008] The cross-climate zone model transfer optimization step optimizes the cross-regional model parameters through transfer learning, based on the multi-factor linkage feature system and the adjusted LSTM gate weights.

[0009] The flood forecast result correction step combines historical data with the multi-factor linkage feature system and the adjusted LSTM gating weights to achieve high-precision flood forecasting through error compensation.

[0010] Preferably, the multi-source data acquisition and preprocessing steps include collecting meteorological data, topographic data, and hydrological measured data. The meteorological data includes precipitation, temperature, and humidity; the topographic data includes watershed slope; and the hydrological measured data includes peak flow. Outliers are detected using the quantile method, and box plots are drawn to remove outliers exceeding the middle 50% range. Min-Max normalization is applied to the precipitation and water level data, Z-Score standardization is applied to the temperature data, and Box-Cox transformation is applied to the peak flow to mitigate right skewness.

[0011] In a further preferred embodiment, the climate-topography feature co-calculation step includes calculating the humidity gradient change rate, evaporation-precipitation coupling elasticity coefficient, topographic slope attenuation coefficient, and spatiotemporal clustering degree of extreme events based on the preprocessed data; the humidity gradient change rate is calculated by the ratio of the humidity difference between adjacent time periods to the time interval; the evaporation-precipitation coupling elasticity coefficient is calculated by the ratio of the evaporation change rate to the precipitation change rate; the topographic slope attenuation coefficient is calculated by taking the difference between the product of the tangent of the average slope of the watershed and the correction coefficient; and the spatiotemporal clustering degree of extreme events is calculated by the ratio of the average intensity of the number of extreme precipitation events within the time period to the watershed area over the time span.

[0012] Further optimized, the adaptive LSTM gating weight adjustment step includes selecting the optimal LSTM variant based on the climate-topography feature adaptation coefficient, selecting the ConvLSTM variant for humid regions, the BiLSTM variant for semi-humid regions, and the PeepholeLSTM variant for semi-arid regions; and dynamically adjusting the LSTM forget gate input gate weights through climate-topography feature parameters to enhance the ability to remember and capture key hydrological features of different climate zones.

[0013] In a further optimized approach, the climate-topography feature co-calculation step employs a feature adaptation coefficient to quantify the comprehensive impact. The formula for calculating the feature adaptation coefficient is as follows:

[0014] ;

[0015] in: The humidity gradient change rate, expressed in % / d, reflects the rate of change of the spatial distribution of humidity over a period of time. The evaporation-precipitation coupling elasticity coefficient is dimensionless and reflects the sensitivity of evaporation to changes in precipitation. This is the topographic slope attenuation coefficient, which is dimensionless and reflects the attenuation effect of topography on hydrological response. , representing the spatiotemporal clustering degree of extreme events, is dimensionless and reflects the spatiotemporal density of extreme precipitation events; This marks the starting point for historical data statistics. This marks the end point for historical data statistics. It is the integral variable.

[0016] A further preferred embodiment of the formula for calculating the weights of the forget gate and input gate in the adaptive LSTM gating weight adjustment step is:

[0017] ;

[0018] ;

[0019] in The forgetting gate weight is dimensionless and controls the degree to which historical information is retained. The input gate weights are dimensionless and control the intensity of new information input. The spatiotemporal gradient of the feature fitting coefficients is expressed in % / (d·km), reflecting the spatiotemporal variation trend of the fitting coefficients. The hyperbolic tangent activation function is used. It is an S-type activation function used to restrict the weight range to between 0 and 1.

[0020] Further preferred steps for cross-climate zone model transfer optimization include constructing a meteorological zoning feature encoder, implementing a classifier-driven model switching strategy to achieve fine-tuning of GRU model parameters from humid to semi-arid regions, and combining GAN-generated synthetic data to enhance the sample size in semi-arid regions and solve the problem of modeling with small sample data.

[0021] In a further preferred embodiment, the flood forecast result correction step includes calculating the correlation between the prediction error of historical data and the feature fitting coefficient forget gate weights to obtain the error compensation coefficient; correcting the original prediction value of the LSTM network using the error compensation coefficient; and achieving prediction bias correction by fusing the feature fitting coefficient gate weights and extreme event spatiotemporal clustering.

[0022] In a further preferred embodiment, the error compensation coefficient is calculated by combining the difference between historical measured data and original predicted data with the feature adaptation coefficient and forget gate weight of the corresponding time period. During the calculation process, a weighted summation method is used to highlight the impact of data in key time periods.

[0023] A further preferred formula for calculating the final peak flow forecast value in the flood forecast result correction step is:

[0024] ;

[0025] in: This is the predicted peak flow rate, in m³ / s. The original peak flow prediction value of the LSTM network is expressed in m³ / s. The error compensation coefficient is dimensionless and reflects the degree of compensation for historical prediction errors. The calculation formula is:

[0026] ;

[0027] in: For the first The measured peak flow rates for each historical period, in m³ / s; For the first The original peak flow forecast values ​​for each historical period, in m³ / s; For the first The feature fit coefficients for each historical period are dimensionless. For the first The forgetting gate weight for each historical period is dimensionless. This represents the total number of historical data samples.

[0028] The technical advantages of this invention are as follows: The core inventive technique lies in constructing a multi-factor linkage feature system encompassing climate, topography, and extreme events; designing a dynamic adaptive LSTM gating mechanism; and proposing a cross-climate zone migration optimization and error correction method. By integrating multi-dimensional influencing factors through feature adaptation coefficients, and dynamically responding to climate zone differences with gating weights, transfer learning and error compensation enhance cross-regional adaptability and forecast accuracy. This method effectively solves the core problems of traditional models' insufficient adaptation to hydrological mechanisms in multiple climate zones, weak ability to capture extreme events, and low cross-regional migration accuracy, providing an efficient and reliable technical path for flood forecasting in multiple climate zones. Attached Figure Description

[0029] Figure 1 This application presents a flood forecasting method based on a multi-climate zone adaptive LSTM network. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0031] Traditional technical solutions have the following technical problems: existing flood forecasting models do not take into account the combined effects of humidity gradient change rate, evaporation-precipitation coupling elasticity coefficient, topographic slope attenuation coefficient, and spatiotemporal clustering of extreme events. They adopt a unified architecture to adapt to multiple climate zones, resulting in poor adaptability of the model to the hydrological characteristics of different climate zones, low accuracy in predicting extreme flood events, and difficulty in parameter adaptation when migrating across regions.

[0032] Based on this, please refer to Figure 1This embodiment provides a flood forecasting method based on a multi-climate zone adaptive LSTM network, including multi-source data acquisition and preprocessing steps, and further including the following steps:

[0033] S1: Climate-topography feature collaborative calculation steps, integrating humidity gradient change rate, evaporation-precipitation coupling elasticity coefficient, topographic slope attenuation coefficient, and spatiotemporal clustering of extreme events to construct a multi-factor linkage feature system;

[0034] S2: Adaptive LSTM gate weight adjustment step, dynamically matching hydrological characteristics of different climate zones based on the multi-factor linkage feature system, selecting the optimal LSTM variant and adjusting the gate weight;

[0035] S3: Cross-climate zone model transfer optimization step, based on the multi-factor linkage feature system and the adjusted LSTM gating weights, optimizes the cross-regional model parameters through transfer learning;

[0036] S4: Flood forecast result correction step, combining historical data with the multi-factor linkage feature system and the adjusted LSTM gating weights, to achieve high-precision flood forecasting through error compensation.

[0037] The core of this technical solution is the construction of a multi-factor collaborative adaptive modeling framework. Multi-source data preprocessing provides a high-quality data foundation for subsequent calculations. The climate-topography feature collaborative calculation step extracts key parameters reflecting the essential differences between different climate zones. By integrating four types of core features, a multi-factor linkage system is formed, accurately depicting the comprehensive impact of climate and topography on hydrological processes. The adaptive gating adjustment step, based on the multi-factor linkage feature system, selects suitable LSTM variants for the differences in hydrological characteristics across different climate zones, while dynamically adjusting the gating weights to enable the network to focus on key hydrological information. The cross-climate zone model transfer optimization step utilizes the common patterns of the multi-factor linkage system and the adaptation experience of gating weights to reduce data dependence on the target region through transfer learning, solving the model adaptation problem caused by differences in data distribution across different climate zones. The flood forecast result correction step integrates historical data error patterns with current feature and weight information to achieve accurate compensation for prediction bias. Each step is interconnected, forming a complete technical closed loop from data processing to model optimization to result output, focusing entirely on improving multi-climate zone adaptability and forecast accuracy.

[0038] The algorithmic improvement of the adaptive LSTM gating weight adjustment step innovatively constructs a climate-topography feature adaptation coefficient quantification model, integrating four core features: humidity gradient change rate, evaporation-precipitation coupling elasticity coefficient, topographic slope attenuation coefficient, and spatiotemporal clustering of extreme events. It achieves accurate quantification of the synergistic influence of multiple factors through nonlinear mathematical formulas, replacing the traditional single-feature-driven gating adjustment logic. A climate-zone differentiated LSTM variant selection strategy is designed, matching ConvLSTM, BiLSTM, and PeepholeLSTM variants respectively to address the differences in hydrological characteristics in humid, semi-humid, and semi-arid regions, overcoming the adaptation limitations of the traditional unified LSTM architecture. A dynamic gating weight calculation method is proposed, based on the feature adaptation coefficient, the spatiotemporal clustering of extreme events, and the spatiotemporal gradient of the feature adaptation coefficient, using the hyperbolic tangent activation function tanh and the sigmoid activation function.

[0039] A sigmoid algorithm is used to construct a weighted calculation model that links the forget gate and input gate, enabling dynamic adaptive adjustment of weights based on climatic zone characteristics. The principle is that the feature adaptation coefficient integrates four core features to construct a comprehensive quantitative index that characterizes the essential differences in hydrological conditions across different climatic zones, providing a quantitative basis for variant selection and gating adjustment. The humidity gradient change rate reflects the dynamics of humidity spatial distribution, the evaporation-precipitation coupling elasticity coefficient reflects the water-heat balance relationship, the topographic slope attenuation coefficient quantifies the constraints of topography on hydrological response, and the spatiotemporal clustering of extreme events highlights the impact of extreme disturbances. The synergistic effect of these four features achieves accurate characterization of the linkage between climate, topography, and extreme events. Variants are selected based on the core hydrological contradictions in different climatic zones. In humid regions, precipitation exhibits strong spatiotemporal heterogeneity; ConvLSTM captures spatial heterogeneity through convolutional operations. Inter-dependent relationships; dynamic changes in the evaporation-precipitation balance in semi-humid regions, BiLSTM uses bidirectional time series modeling to capture the balance evolution law; significant hydrological abrupt changes in semi-arid regions, PeepholeLSTM enhances the sensitivity to abrupt change signals by connecting cell state with the gating peephole; gating weights achieve dynamic response through mathematical models, and forget gate weights integrate feature adaptation coefficients and the influence of extreme events to control the intensity of historical information retention; input gate weights and forget gates form a linkage balance, adjusting the intensity of new information input according to the characteristics of the climate zone; tanh and sigmoid functions ensure that the weight values ​​are limited to the 0-1 range, which conforms to the physical meaning of gating. The technical effects include achieving precise adaptation of the model architecture to the characteristics of different climate zones, solving the problem of insufficient adaptability caused by the one-size-fits-all approach of traditional models, and significantly improving the accuracy of capturing hydrological features of different climate zones. The dynamic gating mechanism enhances the ability to memorize and filter key hydrological information, and significantly improves the capture accuracy of core features such as the migration process of rainstorm centers in humid areas, the alternation of water and heat balance in semi-humid areas, and the abrupt changes in extreme precipitation in semi-arid areas, providing high-quality feature support for subsequent forecasts. It also enhances the model's adaptability to complex hydrological processes, reduces the impact of single feature perturbations on forecast results, and significantly enhances the model's robustness.

[0040] The algorithmic improvement of the cross-climate zone model transfer optimization step involves constructing a meteorological zoning feature encoder, employing a deep neural network structure, and extracting common and unique features of different climate zones through multi-layer convolutional and fully connected layers. This provides a feature adaptation basis for cross-regional parameter transfer, replacing the traditional direct parameter transfer mode. A classifier-driven model switching strategy is designed, training a binary classifier based on the unique features output by the feature encoder to automatically identify the climate type of the target region, achieving directional transfer from the pre-trained GRU model in the humid region to the target region, and optimizing the adaptability of the transfer starting point. A parameter transfer mechanism combining freezing and fine-tuning is proposed, freezing the parameters of the bottom feature extraction layer of the pre-trained model and only fine-tuning the top gating weights and fully connected layers. At the same time, a GAN generative adversarial network is introduced to generate synthetic data to supplement the data volume of small sample areas such as semi-arid regions. The principle is that the feature encoder extracts hierarchical representations of multi-dimensional features through deep learning. Common features retain the general laws of hydrological processes in different climate zones, ensuring that the transfer model has basic forecasting capabilities. Individual features distinguish the unique hydrological attributes of each climate zone, providing a clear direction for parameter fine-tuning and reducing the adaptation difficulty caused by differences in cross-regional feature distribution. The classifier-driven model switching is based on the principle of feature similarity matching. By identifying the feature differences between the target region and the pre-training region, the most suitable pre-trained model is selected as the starting point for transfer, avoiding parameter mismatch caused by blind transfer and reducing the convergence difficulty of transfer learning. Freezing the bottom-level parameters can retain the effective feature extraction capabilities obtained from training with a large amount of data in humid areas, and fine-tuning the top-level parameters can achieve rapid adaptation to the data distribution of the target region. The GAN learns the distribution law of real data through the generator and generates synthetic samples that are consistent with the statistical characteristics of real data. The discriminator optimizes the performance of the generator. The two train against each other to achieve high-quality sample expansion and solve the model overfitting problem caused by insufficient data in small sample areas. The technical benefits are: significantly improving the transfer adaptability of models across climate zones, solving the core problem of poor adaptability of traditional transfer methods; effectively expanding the training set size of small sample areas with synthetic data, reducing the risk of model overfitting, enhancing generalization ability, and significantly reducing cross-regional forecast errors; shortening the model training cycle in the target area, eliminating the need to train from scratch, and greatly improving training efficiency by utilizing pre-trained model parameters and synthetic data, thereby reducing the time cost and data dependence of multi-climate zone modeling.

[0041] The algorithm improvement for the flood forecast result correction step proposes a weighted error compensation coefficient calculation method. Based on the difference between historical measured and predicted data, a weighted sum is performed by combining the feature adaptation coefficient and forget gate weight for the corresponding time period. This highlights the impact of errors during critical periods such as extreme precipitation and sudden changes in water and heat balance, replacing the traditional equal-weighted error statistical method. A multi-factor fusion correction model is constructed, integrating feature adaptation coefficient, gate weight, spatiotemporal clustering of extreme events, and error compensation coefficient. The original LSTM prediction value is corrected through a nonlinear mathematical formula, breaking through the limitations of traditional single error correction. A hierarchical correction logic is designed. The first level compensates for systematic deviations through the synergistic compensation of feature adaptation coefficient, forget gate weight, and error compensation coefficient. The second level adjusts extreme scenario deviations through the spatiotemporal clustering of extreme events and input gate weight, achieving comprehensive deviation correction. The principle is that the error compensation coefficient strengthens the error patterns in key periods through a weighting mechanism. Periods with larger feature fit coefficients represent periods with significant climate-topographic influences, and the forget gate weight reflects the model's processing status of information in that period. The weighting factor formed by the product of these two factors ensures that errors in key periods such as extreme precipitation and strong topographic constraints dominate the compensation coefficient, ensuring that the compensation coefficient accurately matches the core error patterns. Multi-factor fusion correction is based on feature, parameter, and error linkage logic. The feature fit coefficient and gate weight reflect the climate-topographic characteristics and model processing status of the current forecast scenario, the error compensation coefficient quantifies historical error patterns, and the spatiotemporal clustering of extreme events highlights the particularity of extreme scenarios. The four factors work together to achieve accurate correction of the original prediction value, compensating for both systematic biases and specific biases under extreme scenarios. Hierarchical correction achieves accurate bias allocation through mathematical formulas. The first level of correction targets the common systematic biases, while the second level focuses on the specific biases of extreme scenarios, avoiding over-correction or under-correction caused by a single correction logic, and ensuring the rationality and accuracy of the correction results. The technical benefits include a significant improvement in the representativeness and reliability of the error compensation coefficient, providing a precise basis for correction; multi-factor fusion correction effectively reduces the original prediction bias, resulting in a significant improvement in the accuracy of the final peak flow forecast and a substantial reduction in the forecast error of extreme flood events, thus solving the problem of limited effectiveness of traditional correction methods; the correction model has strong scene adaptability, and can stably output high-precision forecast results in different climate zones and different hydrological scenarios, significantly enhancing the stability and reliability of the forecast results.

[0042] The technical solutions achieve the following effects: enabling high-precision flood forecasting across multiple climate zones, enhancing the ability to capture extreme events, improving the adaptability of cross-regional models, and providing reliable technical support for disaster prevention and mitigation.

[0043] Traditional technical solutions have the following technical problems: existing data preprocessing methods do not differentiate data based on the characteristics of flood forecasting data, outlier detection methods are easily affected by extreme values, and data standardization methods are singular, resulting in low quality of preprocessed data that cannot accurately reflect hydrological and meteorological characteristics and affect the accuracy of subsequent model building.

[0044] Based on this, the multi-source data acquisition and preprocessing steps include collecting meteorological data, topographic data, and hydrological measured data. Meteorological data includes precipitation, temperature, and humidity, with precipitation specifically including quantitative indicators such as rainfall amount and duration. Hydrological measured data includes peak flow and water level data from key watershed stations, such as real-time water levels and historical high water levels. Water level data and peak flow together reflect changes in hydrological processes. Topographic data includes watershed slope. The process involves using the quantile method to detect outliers and drawing box plots to remove outliers exceeding the middle 50% range. Topographic data, such as watershed slope, is also included. It is necessary to participate in outlier detection simultaneously, screen outliers in slope data of different sub-regions of the watershed, and remove extreme slope values ​​that exceed the middle 50% range, such as 0 slope or super-steep slope caused by measurement error. No additional normalization is required for the preprocessed slope data, as the physical meaning of the original slope needs to be preserved in the subsequent calculation of the terrain slope attenuation coefficient. Min-Max normalization is used for rainfall and water level data, Z-Score standardization is used for temperature data, and Box-Cox transformation is used for peak flow to alleviate right skew distribution.

[0045] This technical solution designs differentiated preprocessing strategies based on the characteristics of different types of data. Precipitation in meteorological data directly affects flood runoff; temperature and humidity indirectly affect the hydrological cycle by influencing the evaporation process; watershed slope in topographic data determines surface runoff velocity and confluence time; and peak flow in hydrologically measured data is the core objective of forecasting. The quantile method detects outliers by defining the middle 50% of the data, effectively avoiding interference from outliers caused by extreme precipitation or measurement errors, and is suitable for the skewed distribution characteristics commonly found in hydrological and meteorological data. Min-Max normalization maps rainfall and water level data to a fixed range, preserving the relative magnitude of the data and adapting to the large differences in their value ranges. Z-Score standardization ensures that temperature data follows a normal distribution with a mean of 0 and a standard deviation of 1, eliminating the influence of dimensions while preserving the distribution characteristics of the data. Box-Cox transformation alleviates the right-skewed distribution of peak flow data through power transformation, making the data closer to a normal distribution and improving the model's ability to fit extreme values. This differentiated preprocessing workflow can effectively remove data noise, unify data scale, and optimize data distribution characteristics, providing high-quality input data for subsequent feature calculation and model training, and ensuring that the model can capture the real hydrological and meteorological change patterns.

[0046] The technical effects achieved by this solution are: improving data quality, removing abnormal interference, optimizing data distribution characteristics, providing reliable data support for subsequent modeling, and indirectly improving forecast accuracy.

[0047] Traditional technical solutions have the following technical problems: existing feature calculation methods mostly consider a single meteorological or topographic factor, without constructing a feature system that links multiple factors, and cannot quantify the synergistic impact of various factors on the flood process. As a result, the extracted features cannot accurately reflect the hydrological response mechanism of different climate zones, affecting the model's ability to capture hydrological features.

[0048] Based on this, the climate-topography feature co-calculation steps include calculating the humidity gradient change rate, evaporation-precipitation coupling elasticity coefficient, topographic slope attenuation coefficient, and extreme event spatiotemporal clustering degree based on the preprocessed data; the humidity gradient change rate is calculated by the ratio of the humidity difference between adjacent time periods to the time interval, the evaporation-precipitation coupling elasticity coefficient is calculated by the ratio of the evaporation change rate to the precipitation change rate, the topographic slope attenuation coefficient is calculated by the difference between the product of the tangent of the average slope of the watershed and the correction coefficient, and the extreme event spatiotemporal clustering degree is calculated by the ratio of the average intensity of the number of extreme precipitation events within the time period to the watershed area over the time span.

[0049] The four core features extracted by this technical solution characterize the key driving mechanisms of flood formation from four dimensions: humidity dynamics, water and heat balance, topographic constraints, and extreme disturbances, forming a multi-factor linkage feature system.

[0050] The humidity gradient change rate is calculated as follows:

[0051] ,

[0052] in The humidity at the current time. The humidity level of the previous period. The time interval is used to directly reflect the dynamic trend of humidity spatial distribution. High humidity gradients in humid areas often indicate strong precipitation potential.

[0053] The calculation method for the evaporation-precipitation coupling elasticity coefficient is as follows:

[0054] ,

[0055] in This represents the change in evaporation. This is the original evaporation amount. This represents the change in precipitation. This represents the original precipitation. This characteristic reflects the impact of water and heat exchange balance on hydrological processes. In semi-humid regions, the dynamic change of this coefficient directly determines the increase or decrease in runoff.

[0056] The method for calculating the terrain slope attenuation coefficient is as follows:

[0057] ,

[0058] in The average slope of the watershed is used to quantify the attenuation effect of topography on the hydrological response. The greater the slope, the faster the surface runoff velocity, the more severe the hydrological response, and the smaller the attenuation coefficient.

[0059] The method for calculating the spatiotemporal clustering of extreme events is as follows:

[0060] ,

[0061] in This represents the number of extreme precipitation events within a given period. The average intensity of the event. For the drainage area, Over a time span, this feature reflects the spatiotemporal density of extreme precipitation events and is a major driving factor for floods in semi-arid regions.

[0062] The four features complement and work synergistically to comprehensively depict the differences in hydrological characteristics in different climate zones, providing a core basis for model adaptation.

[0063] The technical effects achieved by this solution are: to construct a multi-dimensional linkage feature system, accurately characterize flood driving factors, quantify the combined effects of climate and topography, and provide feature support for the model to adapt to different climate zones.

[0064] Traditional technical solutions have the following technical problems: existing LSTM model applications do not select appropriate variant models according to the hydrological characteristics of different climate zones, and the gating weights are set in a fixed manner, which cannot dynamically respond to changes in hydrological characteristics caused by differences in climate zones. As a result, the model has insufficient ability to capture key hydrological processes in specific climate zones, and the forecast accuracy is limited.

[0065] Based on this, the adaptive LSTM gating weight adjustment step includes selecting the optimal LSTM variant according to the climate-topography feature adaptation coefficient, selecting the ConvLSTM variant in humid areas, the BiLSTM variant in semi-humid areas, and the PeepholeLSTM variant in semi-arid areas; and dynamically adjusting the LSTM forget gate input gate weights through climate-topography feature parameters to enhance the ability to remember and capture key hydrological features of different climate zones.

[0066] This technical solution leverages the differences in core hydrological characteristics across different climate zones to achieve precise matching and dynamic optimization of LSTM variant models and gating weights. In humid regions, the core hydrological characteristic is strong spatiotemporal heterogeneity of precipitation and frequent migration of rainstorm centers. The ConvLSTM variant, by incorporating convolution operations into the gating computation, effectively captures the spatial dependencies of radar precipitation data and accurately models the spatiotemporal propagation process of precipitation. In semi-humid regions, the core hydrological characteristic is significant dynamic changes in the evaporation-precipitation balance and frequent alternation between wet and dry periods. The BiLSTM variant, through LSTM networks in both forward and reverse directions, utilizes historical and future time-series information to comprehensively capture the dynamic evolution of the evaporation-precipitation coupling relationship. In semi-arid regions, the core hydrological characteristic is frequent extreme precipitation events and strong abrupt changes in hydrological processes. The PeepholeLSTM variant, by increasing the peephole connection between cell states and gating units, enhances the sensitivity of gating units to hydrological abrupt changes and improves the ability to capture extreme events. Regarding gating weight adjustment, driven by climate-topography characteristic parameters, humid regions enhance the capture of new precipitation events by increasing the input gate weight and retain historical information on the migration of rainstorm centers by adjusting the forget gate weight; semi-humid regions adapt to the alternating changes in evaporation-precipitation balance by dynamically balancing the input gate and forget gate weights; and semi-arid regions achieve rapid adaptation to hydrological changes by increasing the forget gate's forgetting rate of information during non-extreme periods and enhancing the input gate's response strength to extreme precipitation signals.

[0067] Based on the correlation between hydrological characteristics and climatic factors, the annual precipitation (P), aridity index (K), and mean humidity gradient change rate (%) were used. A quantitative classification system for three climate zones was constructed. The humid zone is defined as having an annual precipitation P ≥ 1600 mm, an aridity index K ≤ 0.5, and a mean humidity gradient change rate of [missing value]. The annual precipitation is ≥0.8% / d, with the core characteristic being strong spatial and temporal heterogeneity. In the semi-humid region, the annual precipitation is 800mm≤P<1600mm, the aridity index is 0.5<K≤1.0, and the average humidity gradient change rate is 0.4% / d≤ <0.8% / d, the core characteristic is significant dynamic changes in the evaporation-precipitation balance; in semi-arid regions, annual precipitation is 200mm≤P<800mm, aridity index is 1.0<K≤2.0, and the average humidity gradient change rate is <0.8%. <0.4% / d, the core characteristic is the frequent occurrence of extreme precipitation events.

[0068] Within the same climate zone, further subdivisions are made based on more refined hydroclimatic indicators. Humid zones are divided into rainy and humid types. Rainy types have an annual precipitation ≥2000mm, precipitation concentration (the proportion of the maximum precipitation in the three months of the year to the total annual precipitation) ≥40%, and the spatiotemporal concentration of extreme events δ≥0.6, with frequent and intense migration of rainstorm centers. Humid types have an annual precipitation ≤1600mm and <2000mm, precipitation concentration ≤25% and <40%, and the spatiotemporal concentration of extreme events ≤0.3 and <0.6, with relatively uniform precipitation distribution. Semi-humid regions are divided into transitional and arid types. The transitional type has annual precipitation between 1200mm and 1600mm, an evaporation-precipitation coupling elasticity coefficient β ≥ 0.8, a wet-dry cycle ≤ 3 months, and mild fluctuations in hydrothermal balance. The arid type has annual precipitation between 800mm and 1200mm, an evaporation-precipitation coupling elasticity coefficient 0.5 ≤ β < 0.8, a wet-dry cycle > 3 months, and evaporation is sensitive to changes in precipitation. Semi-arid regions are divided into fluctuating and extreme types. The fluctuating type has annual precipitation between 400mm and 800mm, a spatiotemporal clustering of extreme events 0.2 ≤ δ < 0.4, and a hydrological process abrupt change frequency ≤ 5 times / year, with relatively mild precipitation fluctuations. The extreme type has annual precipitation between 200mm and 400mm, a spatiotemporal clustering of extreme events δ ≥ 0.4, a hydrological process abrupt change frequency > 5 times / year, and extreme precipitation dominates flood processes.

[0069] The technical effects achieved by this solution are: to achieve accurate matching between model variants and climate zones, to dynamically optimize the gating weight configuration, to enhance the model's ability to capture hydrological characteristics of different climate zones, and to improve model adaptability.

[0070] Traditional technical solutions have the following technical problems: existing feature quantification methods cannot effectively integrate the synergistic effects of multiple factors, and a single feature cannot fully reflect the comprehensive effect of climate-topography on flood processes, resulting in insufficient representational ability of model input features and affecting the model's adaptation accuracy to hydrological responses in different climate zones.

[0071] Based on this, the feature adaptation coefficient is used to quantify the comprehensive impact in the climate-topography feature co-calculation step. The formula for calculating the feature adaptation coefficient is as follows:

[0072]

[0073] in The humidity gradient change rate, expressed in % / d, reflects the rate of change of the spatial distribution of humidity over a period of time. The evaporation-precipitation coupling elasticity coefficient is dimensionless and reflects the sensitivity of evaporation to changes in precipitation. This is the topographic slope attenuation coefficient, which is dimensionless and reflects the attenuation effect of topography on hydrological response. , representing the spatiotemporal clustering degree of extreme events, is dimensionless and reflects the spatiotemporal density of extreme precipitation events; This marks the starting point for historical data statistics. This marks the end point for historical data statistics. It is the integral variable.

[0074] This technical solution utilizes feature adaptation coefficients. A multi-factor nonlinear linkage mechanism is constructed to achieve precise quantification of the combined impact of climate and topography. (Formula Part 1) Focusing on the synergistic effect of the rate of change of humidity gradient and the elastic coefficient of evaporation-precipitation coupling, The weighting of the impact of drastic changes in humidity gradient on hydrological processes is amplified. The term incorporates the hydrothermal balance effect, and the denominator contains... The term introduces terrain attenuation constraints. The term achieves numerical smoothing through a logarithmic function, avoiding the influence of... and Numerical overflow caused by an excessively large product; Part Two By incorporating the spatiotemporal cumulative effect of the spatiotemporal clustering of extreme events through definite integral operations, the integration interval... Covering the complete historical statistical period, numerator This reflects the interaction of humidity gradient, hydrothermal coupling, and topographic constraints at any given time. The immediate synergistic effect, denominator The square root function is used to adjust the intensity of the impact of extreme events, thus preventing extreme values ​​from dominating the calculation results. The unit % / d ensures the accuracy of the quantification of the rate of humidity change. The unit % / (d·km) reflects the spatiotemporal variation intensity of the feature fit coefficient. Each parameter is dimensionless or unit coordinated to ensure that the feature fit coefficients output by the formula as a whole are consistent. As a dimensionless quantity, it facilitates subsequent gating weight calculation and model input normalization. This formula, through complex mathematical operations, comprehensively integrates the linear and nonlinear relationships of four independent features, forming a comprehensive quantitative index capable of characterizing the essential differences in hydrological features across different climatic zones.

[0075] The technical effects achieved by this solution are: quantifying the synergistic influence of multiple factors, improving feature representation capabilities, providing quantitative basis for model variant selection and gating adjustment, and enhancing the model's adaptability to different climate zones.

[0076] Traditional technical solutions have the following technical problems: the gating weights of existing LSTM models are mostly fixed values ​​or obtained only through simple parameter optimization, and are not associated with climate-topography features. They cannot dynamically respond to changes in hydrological features in different climate zones, resulting in insufficient gating mechanism for screening and memorizing key hydrological information, which affects the model's forecasting performance.

[0077] Based on this, the formulas for calculating the weights of the forget gate and the input gate in the adaptive LSTM gating weight adjustment step are as follows:

[0078]

[0079]

[0080] in The forgetting gate weight is dimensionless and controls the degree to which historical information is retained. The input gate weights are dimensionless and control the intensity of new information input. The spatiotemporal gradient of the feature fitting coefficients is expressed in % / (d·km), reflecting the spatiotemporal variation trend of the fitting coefficients. The hyperbolic tangent activation function is used. It is an S-type activation function used to restrict the weight range to between 0 and 1.

[0081] This technical solution constructs a dynamic correlation mechanism between gating weights and multi-factor features through two interconnected formulas, enabling adaptive adjustment of the gating weights. (Forget gate weight) In the calculation formula, The influence of the direct fusion feature adaptation coefficient on the spatiotemporal clustering of extreme events. As a comprehensive quantitative indicator of climate and topography, it determines the basic weight for retaining historical information. The impact of extreme events is mapped to the interval [-1, 1] using the hyperbolic tangent function, and then... Adjust the intensity of historical information retention after multiplication; Spatiotemporal gradient of feature adaptation coefficients introduced in the term , Reflecting the moderating weights in non-climate-terrain-dominated scenarios, Integrating the synergistic effects of extreme events and spatiotemporal trends, The function maps the input gate weights to the [0,1] interval, enabling fine-tuning of the forget gate weights. Input gate weights The calculation formula is based on Based on, through Establish a fundamental balance between the forget gate and the input gate. The term introduces a secondary collaborative adjustment of extreme events and feature fit coefficients, when and When the value is large, the exponent term approaches 0. Approaching 1, it strengthens the input of new information; when and When the value is small, the exponent term approaches 1. Approaching To maintain a dynamic balance of gating weights. and Activation function ensures and The value of is always limited to the interval [0,1], which conforms to the physical meaning of gating weights and can be adjusted according to different climate zones. , , The value is dynamically adjusted to achieve precise adaptation to different hydrological characteristics, such as humid areas. When the value is high, Tendency to retain more historical spatiotemporal information Strengthen the monitoring of new precipitation events; semi-arid regions When the value is high, Quickly forget non-extreme information. The focus is on capturing signals of extreme precipitation.

[0082] The technical effects achieved by this solution are: to realize dynamic adaptive adjustment of gating weights, to strengthen the screening and memorization of key hydrological information, to improve the model's adaptability to the hydrological characteristics of different climate zones, and thus to improve forecast accuracy.

[0083] Traditional technical solutions have the following technical problems: existing cross-regional model transfer methods do not take into account the significant differences in data distribution across different climate zones, the direct transfer of model parameters has poor adaptability, and there is a problem of insufficient data sample size in semi-arid regions and other areas, resulting in low cross-regional forecast accuracy and weak model generalization ability.

[0084] Based on this, the cross-climate zone model transfer optimization steps include constructing a meteorological zone feature encoder, implementing a classifier-driven model switching strategy to fine-tune GRU model parameters from humid to semi-arid regions, and combining GAN-generated synthetic data to enhance the sample size in semi-arid regions, solving the problem of modeling with small sample data. This technical solution constructs a cross-regional transfer optimization system of "feature encoding - model switching - parameter transfer - sample enhancement" to comprehensively solve the model adaptation problem of different climate zones. The meteorological zone feature encoder adopts a deep neural network structure, inputting four core features from a multi-factor linkage feature system. Through multi-layer convolution and fully connected layers, it extracts common and unique features of different climate zones. Common features are used to maintain the basic performance of model transfer, while unique features are used to distinguish the hydrological differences between different climate zones, providing a feature adaptation basis for model parameter transfer. The classifier-driven model switching strategy, based on the unique features output by the feature encoder, trains a binary classifier to determine the climate type of the target region and automatically selects a pre-trained GRU model in the humid region as the transfer starting point, ensuring that the transferred model has a basic feature extraction capability similar to that of the target region. During the transfer tuning of GRU model parameters, the parameters of the bottom feature extraction layer of the pre-trained model are frozen, and only the parameters of the top gating weights and fully connected layers are fine-tuned. A small amount of measured data from the target region is used to adjust the model to adapt to the data distribution of the target region. This retains the effective feature extraction capability trained on a large amount of data in humid areas while quickly adapting to the hydrological characteristics differences in semi-arid areas. To address the problem of insufficient sample size in semi-arid areas, GANs are used to generate synthetic data. The generator learns the generation patterns of hydrological and meteorological data based on the distribution of real data, generating synthetic samples with statistical characteristics consistent with real data. The discriminator optimizes the generator performance by distinguishing between real and synthetic samples. Finally, the generated synthetic data is mixed with real data, effectively expanding the training sample size and alleviating the model overfitting problem caused by small sample data.

[0085] The technical effects achieved by this solution are: improving the cross-regional model transfer adaptability, solving the problem of small sample modeling, enhancing the model generalization ability, and improving the accuracy of cross-climate zone forecasts.

[0086] Traditional technical solutions have the following technical problems: existing forecast result correction methods mostly use simple error statistics correction, without combining the correlation between model input features and model internal parameters. The correction logic lacks physical meaning support, resulting in limited correction effect and inability to effectively compensate for prediction deviations caused by insufficient feature capture or improper parameter adaptation.

[0087] Based on this, the flood forecast result correction steps include calculating the correlation between the prediction error of historical data and the feature fitting coefficients and forget gate weights to obtain the error compensation coefficient; correcting the original prediction value of the LSTM network using the error compensation coefficient; and integrating the feature fitting coefficients, gate weights, and spatiotemporal clustering of extreme events to correct prediction bias. This technical solution constructs a precise correction system from error correlation and coefficient calculation to multi-factor fusion correction, achieving dynamic compensation for prediction bias. The calculation of the error compensation coefficient is based on the deep correlation between the prediction error and key features and model parameters. The historical prediction error is obtained by the difference between the measured peak flow and the original prediction value during historical periods. The forgetting gate weight reflects the combined effects of climate and topography during this period. The correlation analysis of these three factors—reflecting the model's processing strategy for historical information during that period—can accurately identify the error patterns of the model under different climate-topography conditions, such as high humidity in humid areas. Errors in the time period mostly stem from insufficient capture of spatiotemporal features, particularly in semi-arid regions. Errors during the correction period often stem from insufficient response to extreme events. The correction process utilizes not only error compensation coefficients... It quantifies historical error patterns and dynamically integrates the feature adaptation coefficients of current predictions. Gating weights and Spatiotemporal clustering of extreme events This forms a multi-dimensional correction factor, which is used to differentiate corrections based on the error characteristics of different forecast scenarios. For example, when and When both are high, it indicates that the current scene is significantly affected by climate and terrain and extreme events exist. The correction process will focus on strengthening the weight of the error compensation coefficient, while also combining... and The value of the adjustment range is adjusted to ensure that prediction bias is fully compensated in extreme scenarios; when lower and When the error is relatively small, it indicates that the hydrological process in the current scenario is relatively stable. The correction process will appropriately reduce the error compensation weight to avoid prediction distortion caused by overcorrection.

[0088] The technical effects achieved by this solution are: accurately compensating for prediction bias, improving the accuracy of forecast results, and enhancing the stability and reliability of forecast results.

[0089] Traditional technical solutions have the following technical problems: existing error compensation coefficient calculation methods do not take into account the differences in the importance of data in different time periods, and adopt simple averaging or equal weighting summation methods, which result in the error patterns in key time periods not being fully reflected, the representativeness of the compensation coefficients being insufficient, and affecting the correction effect.

[0090] Based on this, the error compensation coefficient is calculated by combining the difference between historical measured data and original predicted data with the feature adaptation coefficient and forget gate weight of the corresponding time period. The weighted summation method is used in the calculation process to highlight the impact of data in key time periods.

[0091] This technical solution optimizes the calculation of error compensation coefficients through a weighted summation mechanism, thereby improving the representativeness and reliability of the compensation coefficients. The calculation process first calculates the prediction error for each historical period. ,in For the first Measured peak flow for a historical period The corresponding original predicted values ​​are used; then weighting factors are constructed. ,in For the first Feature fit coefficients for each historical period For the first The forgetting gate weights for each historical period are used to effectively identify key periods. The larger size indicates that the climate and topography had a significant impact during this period. This reflects the model's processing status for that time period; the time periods with a larger product of the two factors represent key time periods that have a greater impact on the overall forecast error. Finally, a weighted summation formula is used. The error compensation coefficient is calculated with the numerator being the weighted sum of errors and the denominator being the sum of squares of the weighting factors, achieving normalization to ensure a reasonable range for the coefficient's values. This weighted calculation method assigns greater weight to critical time periods, making the error compensation coefficient more reflective of the main error patterns. For example, the weighting factors are larger for periods of heavy rainfall in humid regions and periods of extreme precipitation in semi-arid regions, where the errors dominate the compensation coefficient calculation. This ensures the compensation coefficient accurately matches the error characteristics of these critical scenarios, providing a more targeted basis for error compensation in subsequent correction steps.

[0092] The technical effect achieved by this solution is to improve the representativeness and reliability of the error compensation coefficient, provide support for accurate correction, and further improve the accuracy of forecast results.

[0093] Traditional technical solutions have the following technical problems: existing final forecast calculation methods do not fully integrate the synergistic effects of multiple factors, the formula logic is simple, and it cannot effectively integrate multi-dimensional information such as feature information, model parameters, and error compensation, resulting in the final forecast results not being able to fully utilize existing information and the forecast accuracy being limited.

[0094] Based on this, the formula for calculating the final peak flow forecast value in the flood forecast result correction step is as follows:

[0095]

[0096] in This is the predicted peak flow rate, in m³ / s. The original peak flow prediction value of the LSTM network is expressed in m³ / s. The error compensation coefficient is dimensionless and reflects the degree of compensation for historical prediction errors. The calculation formula is:

[0097]

[0098] in For the first The measured peak flow rates for each historical period, in m³ / s; For the first The original peak flow forecast values ​​for each historical period, in m³ / s; For the first The feature fit coefficients for each historical period are dimensionless. For the first The forgetting gate weight for each historical period is dimensionless. This represents the total number of historical data samples.

[0099] The final forecast calculation formula of this technical solution integrates six key elements—original forecast value, multi-factor characteristics, model parameters, and error compensation—through a multi-level fusion mechanism to achieve accurate correction of the original forecast value. In the formula… The terms constitute the first correction factor. This reflects the combined impact of climate and terrain on the current scene. This reflects the model's processing status of historical information. By quantifying historical error compensation patterns, the product of these three factors enables coordinated correction of climate-topography features, model parameters, and error compensation, which can effectively compensate for systematic biases caused by insufficient feature capture or improper parameter adaptation. The term constitutes the second correction factor. Highlighting the impact of extreme events, Magnifying the significant role of climate-topography, denominator The correction magnitude is adjusted by controlling the input gate weights to avoid correction distortion caused by excessive input of new information. This factor primarily compensates for prediction bias in extreme scenarios. Error Compensation Coefficient The calculation formula, through weighted summation and normalization, constructs a quantitative correlation between historical errors and key features and model parameters. To weight the error sum and highlight the impact of errors during critical periods, the denominator... To normalize the compensation coefficients by using the sum of squares of the weighting factors, ensuring... The numerical range is reasonable and physically meaningful. The dimensions of all parameters in the formula are strictly consistent. and The units are all m³ / s. , , , , All quantities are dimensionless, ensuring the consistency of units and the rationality of physical meaning in the calculation results, and the final output can accurately reflect the forecast value of the actual flood process.

[0100] The technical solution achieves the following effects: it fully integrates key information from multiple dimensions, enables accurate correction of the original forecast values, significantly improves the final forecast accuracy, and ensures the reliability and rationality of the forecast results.

[0101] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A flood forecasting method based on a multi-climate zone adaptive LSTM network, comprising multi-source data acquisition and preprocessing steps, characterized in that, It also includes the following steps: The climate-topography feature co-calculation steps integrate humidity gradient change rate, evaporation-precipitation coupling elasticity coefficient, topographic slope attenuation coefficient, and spatiotemporal clustering of extreme events to construct a multi-factor linkage feature system; The adaptive LSTM gating weight adjustment step dynamically matches the hydrological characteristics of different climate zones based on the multi-factor linkage feature system, selects the optimal LSTM variant, and adjusts the gating weights. The cross-climate zone model transfer optimization step optimizes the cross-regional model parameters through transfer learning, based on the multi-factor linkage feature system and the adjusted LSTM gate weights. The flood forecast result correction step combines historical data with the multi-factor linkage feature system and the adjusted LSTM gating weights to achieve high-precision flood forecasting through error compensation.

2. The flood forecasting method based on a multi-climate zone adaptive LSTM network according to claim 1, characterized in that, The multi-source data acquisition and preprocessing steps include collecting meteorological data, topographic data, and hydrological measured data. The meteorological data includes precipitation, temperature, and humidity; the topographic data includes watershed slope; and the hydrological measured data includes peak flow. The quantile method is used to detect outliers, and a box plot is drawn to remove outliers exceeding the middle 50% range. Min-Max normalization is applied to the precipitation and water level data, Z-Score standardization is applied to the temperature data, and Box-Cox transformation is applied to the peak flow to mitigate right skewness.

3. The flood forecasting method based on a multi-climate zone adaptive LSTM network according to claim 1, characterized in that, The steps for co-calculating climate-topography features include calculating the rate of change of humidity gradient, the elasticity coefficient of evaporation-precipitation coupling, the attenuation coefficient of topographic slope, and the spatiotemporal clustering degree of extreme events based on preprocessed data; the rate of change of humidity gradient is calculated by the ratio of the difference in humidity between adjacent time periods to the time interval; the elasticity coefficient of evaporation-precipitation coupling is calculated by the ratio of the rate of change of evaporation to the rate of change of precipitation; the attenuation coefficient of topographic slope is calculated by taking the difference between the product of the tangent of the average slope of the watershed and the correction coefficient; and the spatiotemporal clustering degree of extreme events is calculated by the ratio of the average intensity of the number of extreme precipitation events within the time period to the time span of the watershed area.

4. The flood forecasting method based on a multi-climate zone adaptive LSTM network according to claim 1, characterized in that, The adaptive LSTM gating weight adjustment steps include selecting the optimal LSTM variant based on the climate-topography feature adaptation coefficient, selecting the ConvLSTM variant for humid regions, the BiLSTM variant for semi-humid regions, and the PeepholeLSTM variant for semi-arid regions; and dynamically adjusting the LSTM forget gate input gate weights through climate-topography feature parameters to enhance the ability to remember and capture key hydrological features of different climate zones.

5. The flood forecasting method based on a multi-climate zone adaptive LSTM network according to claim 1, characterized in that, In the climate-topography feature co-calculation step, the feature fit coefficient is used to quantify the comprehensive impact. The formula for calculating the feature fit coefficient is as follows: ; in: The humidity gradient change rate, expressed in % / d, reflects the rate of change of the spatial distribution of humidity over a period of time. The evaporation-precipitation coupling elasticity coefficient is dimensionless and reflects the sensitivity of evaporation to changes in precipitation. This is the topographic slope attenuation coefficient, which is dimensionless and reflects the attenuation effect of topography on hydrological response. , representing the spatiotemporal clustering degree of extreme events, is dimensionless and reflects the spatiotemporal density of extreme precipitation events; This marks the starting point for historical data statistics. This marks the end point for historical data statistics. It is the integral variable.

6. The flood forecasting method based on a multi-climate zone adaptive LSTM network according to claim 1, characterized in that, The formulas for calculating the weights of the forget gate and input gate in the adaptive LSTM gating weight adjustment step are as follows: ； ； in The forgetting gate weight is dimensionless and controls the degree to which historical information is retained. The input gate weights are dimensionless and control the intensity of new information input. The spatiotemporal gradient of the feature fitting coefficients, in units of % / ( This reflects the spatiotemporal variation trend of the adaptation coefficient; The hyperbolic tangent activation function is used. It is an S-type activation function used to restrict the weight range to between 0 and 1.

7. The flood forecasting method based on a multi-climate zone adaptive LSTM network according to claim 1, characterized in that, The steps for cross-climate zone model transfer optimization include constructing a meteorological zoning feature encoder and implementing a classifier-driven model switching strategy to fine-tune the GRU model parameters from humid to semi-arid regions. By combining GANs to generate synthetic data, the sample size in semi-arid regions is increased, solving the problem of modeling with small sample data.

8. The flood forecasting method based on a multi-climate zone adaptive LSTM network according to claim 1, characterized in that, The steps for correcting flood forecast results include calculating the correlation between the prediction error of historical data and the feature fitting coefficients and forget gate weights to obtain the error compensation coefficients; correcting the original prediction values ​​of the LSTM network using the error compensation coefficients; and integrating the feature fitting coefficients, gate weights, and spatiotemporal clustering of extreme events to correct prediction bias.

9. The flood forecasting method based on a multi-climate zone adaptive LSTM network according to claim 8, characterized in that, The error compensation coefficient is calculated by combining the difference between historical measured data and original predicted data with the feature adaptation coefficient and forget gate weight of the corresponding time period. The weighted summation method is used in the calculation process to highlight the impact of data in key time periods.

10. The flood forecasting method based on a multi-climate zone adaptive LSTM network according to claim 1, characterized in that, The formula for calculating the final peak flow forecast value in the flood forecast result correction step is as follows: ； in: This is the predicted peak flow rate, in cubic meters per second (m³). 3 / s; The raw peak flow prediction values ​​for the LSTM network are in meters. 3 / s; The error compensation coefficient is dimensionless and reflects the degree of compensation for historical prediction errors. The calculation formula is: ； in: For the first Measured peak flow for each historical period, in cubic meters per second. 3 / s; For the first The original peak flood discharge forecast for each historical period, in cubic meters per second. 3 / s; For the first The feature fit coefficients for each historical period are dimensionless. For the first The forgetting gate weight for each historical period is dimensionless. This represents the total number of historical data samples.

Images