A method and system for refined hydrological forecasting based on rain measurement radar
Patent Information
- Application Number
- CN202611156315.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-31
- Publication Date
- 2026-10-27
AI Technical Summary
[0004]本发明提供一种基于测雨雷达的精细化水文预报方法以及系统,旨在解决现有技术中无法适配快速产流响应、单AI模型功能过载、高分辨率雷达适配缺失数据优势未转化、模型耦合松散无功能分工与物理约束协同以及损失函数设计单一缺乏水文专属双重优化目标的问题
1.模型表达能力大幅提升,特征提取更精准,双LSTM分工架构让编码器专注特征提取、解码器专注预报,避免单LSTM“一肩挑”导致的功能干扰,历史降雨到径流的时序依赖特征提取精度提升,模型泛化能力显著增强。
Smart Images

Figure CN122983933A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of hydrological forecasting technology, and in particular to a refined hydrological forecasting method and system based on rainfall radar. Background Technology
[0002] Floods in small and medium-sized river basins are characterized by steep peak rises, short durations (2-3 hours), and concentrated impacts, placing stringent demands on the "spatiotemporal accuracy" and "response speed" of hydrological forecasts. Currently, the core technological bottlenecks in the field of hydrological forecasting are concentrated in three main areas: "model fusion methods," "data utilization scale," and "optimization target design." Existing technologies cannot simultaneously address these issues, resulting in the following problems: 1. The fusion scale is limited to the 1-hour level, which cannot adapt to rapid runoff response. Existing technologies do not combine high-resolution rain-measuring radar data to carry out minute-level fusion. They can only achieve simple superposition on the 1-hour scale, which cannot capture the 10-minute rapid runoff and confluence process caused by local heavy rainfall, resulting in a lag in early warning response. 2. Single AI model is overloaded and lacks expressive power. Existing technology requires a single AI model to simultaneously undertake the dual tasks of extracting historical rainfall and runoff features and forecasting output, resulting in inaccurate feature extraction, interference with forecasting logic, failure to fully explore the temporal dependencies in the data, and poor model generalization ability. 3. Lack of high-resolution radar adaptation and failure to convert data advantages: Existing technologies lack a dedicated preprocessing process for X-band dual-polarization phased array radars, which either dilutes the high resolution or fails to eliminate detection errors, thus failing to provide high-quality data support for minute-level fusion. 4. The model is loosely coupled, lacks functional division of labor and physical constraint collaboration. The existing technology of "hydrological model + single AI model" is a serial superposition. The AI model only performs simple error correction and does not form a deep collaboration with the hydrological model of "physical constraint + division of labor". This is prone to high-resolution but unreasonable forecast results. 5. The loss function design is simplistic and lacks a hydrology-specific dual optimization objective. Existing loss functions suffer from two core problems: First, the optimization direction is singular, either focusing solely on fitting observation data (which can easily lead to forecast results deviating from hydrological physical laws), or the optimization objective is unrelated to the core requirements of runoff over-resolution and forecast correction. They fail to construct a dual optimization structure of "observation accuracy + physical constraints," thus failing to provide suitable optimization guidance for the deep coupling of the dual AI model architecture and the hydrological model. Second, there is insufficient adaptation to extreme scenarios. The training weights for extreme flood events are not specifically strengthened, and insufficient attention is paid to extreme flood events, resulting in low forecast accuracy in extreme flood scenarios, making it difficult to meet the core requirement of accurate early warning for extreme events in flood prevention and disaster reduction.
[0003] The widespread adoption of X-band dual-polarization phased array rain-measuring radar has provided a data foundation for minute-level forecasts, and the encoder-decoder paradigm has provided technical support for the division of labor in processing time-series data. However, existing technologies have not deeply coordinated the "dual LSTM division of labor architecture," "high-resolution radar data," "traditional hydrological models," and "dual-weight loss functions," and have not formed a closed-loop collaborative system of "minute-level fusion + dual LSTM division of labor + dual-weight optimization + physical constraints + extreme flood event focusing." This is the core technological gap that this invention aims to overcome, and it is also the key difference from the simple splicing of existing technologies. Summary of the Invention
[0004] This invention provides a refined hydrological forecasting method and system based on rainfall radar, aiming to solve the problems in the existing technology, such as inability to adapt to rapid flow generation response, overload of single AI model functions, failure to transform the advantage of missing data in high-resolution radar adaptation, loose model coupling without functional division of labor and physical constraint coordination, and single loss function design lacking hydrological-specific dual optimization objectives.
[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A refined hydrological forecasting method based on rainfall radar includes the following steps: Data preprocessing and fusion: Collect real-time data, including raw detection data from X-band dual-polarization phased array rain measurement radar, transform the raw detection data from X-band dual-polarization phased array rain measurement radar into QPE gridded raw data, and perform data standardization and adaptation processing on the QPE raw data. Combine this with the collected topographic and hydrological data for preprocessing and data fusion. Model construction: A dual LSTM network with encoder-decoder paradigm is adopted to form a collaborative relationship with the XAJ model in terms of physical constraints, feature extraction and super-resolution correction. The model parameters are specially set to adapt to the hydrological forecasting scenario. Model training: The model is trained using historical data under the constraint of a dual-weight loss function, and the trained dual LSTM encoder-decoder deeply coupled with XAJ is output. Refined runoff forecast output: The collected real-time data is input into the trained dual LSTM encoder-decoder and XAJ deeply coupled model to output a refined runoff forecast product.
[0006] Preferably, the data preprocessing and fusion further includes the following steps: Radar QPE data acquisition and verification: The raw detection data of the X-band dual polarization phased array rain measurement radar is collected and, after quality control, gridded QPE data within the watershed area is generated. Topographic and hydrological data collection: Key topographic information is extracted from DEM data, including watershed boundaries, topographic slope, and river length and slope, for sub-watershed division and parameter calibration of the XAJ model. Collect measured data from hydrological stations, including runoff and water level data, as the true values and validation basis for model training; Organize watershed underlying surface data: perform regional statistics according to land use type and soil type to provide a basis for watershed adaptation of XAJ model physical parameters.
[0007] Preferably, the data fusion further includes the following steps: High-resolution data-specific aggregation: The raw radar QPE data acquired from radar QPE data is statistically aggregated by surface averaging to generate gridded QPE data; Time series data normalization: Mini-maximum normalization is performed on gridded QPE data and measured runoff data to eliminate dimensional differences and accelerate the convergence of subsequent model training; Missing data completion: Temporal continuity detection is performed on the aggregated data. For a small number of missing data, linear interpolation is used to complete them. If there are multiple consecutive missing steps, radar data and rain gauge data from similar terrain areas during the same period are used for collaborative completion to ensure the temporal integrity of the data and avoid affecting the model feature extraction effect.
[0008] Preferably, the XAJ model adopts a semi-distributed XAJ model structure. Based on the topography and underlying surface characteristics of the watershed, the study watershed is discretized into several sub-watersheds. Each sub-watershed performs runoff calculation independently, and then the runoff is aggregated into the whole watershed runoff process through the runoff confluence module.
[0009] Preferably, the encoder-decoder paradigm dual LSTM network includes an encoder LSTM and a decoder LSTM. The encoder LSTM is used for feature extraction and is free from prediction interference, while the decoder LSTM is used for super-resolution prediction and is subject to physical constraints.
[0010] Preferably, the encoder LSTM takes into account standardized data over a past period, including radar QPE data and measured runoff data within that period. Through the LSTM gating mechanism, it focuses on extracting minute-level temporal features of the dynamic evolution of rainfall-runoff, and outputs condensed encoded hidden states and cell states, encoding historical hydrological process information to provide accurate feature support for the decoder.
[0011] Preferably, the decoder LSTM inputs standardized data of the same duration as the encoder LSTM input, including radar QPE data and hourly runoff forecast data from the XAJ model. Nonlinear correction and spatiotemporal super-resolution are performed through the LSTM network. Based on the minute-level features extracted by the encoder LSTM, the bias of the XAJ model in the rapid runoff scenario is corrected, and the forecast result is output.
[0012] Preferably, the dual-weight loss function constraint includes the following formula: ; ; in Weights for losses based on different flow values. Forecast runoff values from a dual LSTM encoder-decoder + XAJ model. This is the measured runoff value. is the predicted runoff value of the XAJ model, and w1 and w2 are weighting coefficients.
[0013] Preferably, the refined runoff forecast output further includes the following steps: Real-time data rapid preprocessing: The raw radar data acquired in real time is rapidly processed according to a standardized data preprocessing and fusion process; Real-time model inference: The preprocessed real-time minute-level radar QPE data is input into the coupled model, and the model automatically executes the entire process of hourly runoff calculation of the XAJ model, encoder time series feature extraction and decoder super-resolution forecast; Forecast results output and formatting: Output minute-by-minute runoff data, peak flow forecast, and peak occurrence time forecast for the forecast period. The output data supports the interface formats of mainstream hydrological early warning systems and can be directly used for flood warning threshold comparison and scheduling decision reference.
[0014] A system for implementing the above-mentioned refined hydrological forecasting method based on rainfall radar includes the following modules: Data preprocessing and fusion module: collects real-time data, including raw detection data from X-band dual polarization phased array rain measurement radar, converts the raw detection data from X-band dual polarization phased array rain measurement radar into QPE gridded raw data, and performs data standardization and adaptation processing on the QPE raw data, in conjunction with the collection of topographic and hydrological data. Model module: It adopts a dual LSTM network in the encoder-decoder paradigm, which forms a collaborative relationship with the XAJ model in terms of physical constraints, feature extraction and super-resolution correction. The model parameters are specially set to adapt to the hydrological forecasting scenario. The model is trained on historical data under the constraint of dual weight loss function, and the trained dual LSTM encoder-decoder is output as a deeply coupled model with XAJ. The refined runoff forecast output module: It inputs the collected real-time data into the trained dual LSTM encoder-decoder and XAJ deeply coupled model to output refined runoff forecast products.
[0015] The beneficial effects of this invention are: 1. The model's expressive power is greatly improved, and feature extraction is more accurate. The dual LSTM architecture allows the encoder to focus on feature extraction and the decoder to focus on forecasting, avoiding functional interference caused by a single LSTM "doing everything". The accuracy of time-series dependent feature extraction from historical rainfall to runoff is improved, and the model's generalization ability is significantly enhanced.
[0016] 2. Breaking through the limitations of fusion scale and adapting to rapid runoff response, it combines minute-level fusion design with high-resolution radar to achieve deep fusion of "physical model + AI model" at the minute level for the first time. It can accurately capture the rapid runoff process caused by local heavy rainfall and solve the problem of delayed early warning.
[0017] 3. The composite loss function that integrates physical constraints and focuses on extreme floods has stronger adaptability and more reasonable results. The unique dual optimization design ensures that the forecast values fit the measured data (improving accuracy) and forces them not to deviate from the hydrological and physical laws (avoiding unreasonable over-scoring). It can also improve the model's forecast accuracy for extreme flood events. The dynamic weight adjustment rule allows the model to adapt to different terrains and watersheds, making it more robust.
[0018] 4. Data utilization is more efficient, the high-resolution advantage is fully transformed, the dedicated preprocessing process retains the high-resolution advantage of radar at 40 seconds / 30 meters, and minute-level fusion allows for the full extraction of fine spatiotemporal information in the data, providing a solid foundation for accurate forecasting.
[0019] 5. It boasts superior overall performance, strong operational adaptability, and innovative collaboration to achieve multiple requirements: minute-level resolution, rated forecast period, high precision, physical rationality, and real-time performance. It also features short inference time, supports interfaces with mainstream hydrological systems, and can be directly applied in operations; existing technologies cannot simultaneously meet these requirements. Attached Figure Description
[0020] Figure 1 This is a schematic diagram of the process of the present invention. Detailed Implementation
[0021] The present invention will be further described in detail below with reference to specific embodiments.
[0022] As a preferred embodiment 1, a refined hydrological forecasting method based on rainfall radar includes the following steps: Data preprocessing and fusion: Collect real-time data, including raw detection data from X-band dual-polarization phased array rain measurement radar, transform the raw detection data from X-band dual-polarization phased array rain measurement radar into QPE gridded raw data, and perform data standardization and adaptation processing on the QPE raw data. Combine this with the collected topographic and hydrological data for preprocessing and data fusion. Model construction: A dual LSTM network with encoder-decoder paradigm is adopted to form a collaborative relationship with the XAJ model in terms of physical constraints, feature extraction and super-resolution correction. The model parameters are specially set to adapt to the hydrological forecasting scenario. Model training: The model is trained using historical data under the constraint of a dual-weight loss function, and the trained dual LSTM encoder-decoder deeply coupled with XAJ is output. Refined runoff forecast output: The collected real-time data is input into the trained dual LSTM encoder-decoder and XAJ deeply coupled model to output a refined runoff forecast product.
[0023] Preferably, the data preprocessing and fusion further includes the following steps: Radar QPE data acquisition and verification: The raw detection data of the X-band dual polarization phased array rain measurement radar is collected and, after quality control, gridded QPE data within the watershed area is generated. Topographic and hydrological data collection: Key topographic information is extracted from DEM data, including watershed boundaries, topographic slope, and river length and slope, for sub-watershed division and parameter calibration of the XAJ model. Collect measured data from hydrological stations, including runoff and water level data, as the true values and validation basis for model training; Organize watershed underlying surface data: perform regional statistics according to land use type and soil type to provide a basis for watershed adaptation of XAJ model physical parameters.
[0024] Preferably, the data fusion further includes the following steps: High-resolution data-specific aggregation: The raw radar QPE data acquired from radar QPE data is statistically aggregated by surface averaging to generate gridded QPE data; Time series data normalization: Mini-maximum normalization is performed on gridded QPE data and measured runoff data to eliminate dimensional differences and accelerate the convergence of subsequent model training; Missing data completion: Temporal continuity detection is performed on the aggregated data. For a small number of missing data, linear interpolation is used to complete them. If there are multiple consecutive missing steps, radar data and rain gauge data from similar terrain areas during the same period are used for collaborative completion to ensure the temporal integrity of the data and avoid affecting the model feature extraction effect.
[0025] Preferably, the XAJ model adopts a semi-distributed XAJ model structure. Based on the topography and underlying surface characteristics of the watershed, the study watershed is discretized into several sub-watersheds. Each sub-watershed performs runoff calculation independently, and then the runoff is aggregated into the whole watershed runoff process through the runoff confluence module.
[0026] Preferably, the encoder-decoder paradigm dual LSTM network includes an encoder LSTM and a decoder LSTM. The encoder LSTM is used for feature extraction and is free from prediction interference, while the decoder LSTM is used for super-resolution prediction and is subject to physical constraints.
[0027] Preferably, the encoder LSTM takes into account standardized data over a past period, including radar QPE data and measured runoff data within that period. Through the LSTM gating mechanism, it focuses on extracting minute-level temporal features of the dynamic evolution of rainfall-runoff, and outputs condensed encoded hidden states and cell states, encoding historical hydrological process information to provide accurate feature support for the decoder.
[0028] Preferably, the decoder LSTM inputs standardized data of the same duration as the encoder LSTM input, including radar QPE data and hourly runoff forecast data from the XAJ model. Nonlinear correction and spatiotemporal super-resolution are performed through the LSTM network. Based on the minute-level features extracted by the encoder LSTM, the bias of the XAJ model in the rapid runoff scenario is corrected, and the forecast result is output.
[0029] Preferably, the dual-weight loss function constraint includes the following formula: ; ; in Weights for losses based on different flow values. Forecast runoff values from a dual LSTM encoder-decoder + XAJ model. This is the measured runoff value. is the predicted runoff value of the XAJ model, and w1 and w2 are weighting coefficients.
[0030] Preferably, the refined runoff forecast output further includes the following steps: Real-time data rapid preprocessing: The raw radar data acquired in real time is rapidly processed according to a standardized data preprocessing and fusion process; Real-time model inference: The preprocessed real-time minute-level radar QPE data is input into the coupled model, and the model automatically executes the entire process of hourly runoff calculation of the XAJ model, encoder time series feature extraction and decoder super-resolution forecast; Forecast results output and formatting: Output minute-by-minute runoff data, peak flow forecast, and peak occurrence time forecast for the forecast period. The output data supports the interface formats of mainstream hydrological early warning systems and can be directly used for flood warning threshold comparison and scheduling decision reference.
[0031] As a preferred embodiment 2, a system for implementing the above-described refined hydrological forecasting method based on rainfall radar includes the following modules: Data preprocessing and fusion module: collects real-time data, including raw detection data from X-band dual polarization phased array rain measurement radar, converts the raw detection data from X-band dual polarization phased array rain measurement radar into QPE gridded raw data, and performs data standardization and adaptation processing on the QPE raw data, in conjunction with the collection of topographic and hydrological data. Model module: It adopts a dual LSTM network in the encoder-decoder paradigm, which forms a collaborative relationship with the XAJ model in terms of physical constraints, feature extraction and super-resolution correction. The model parameters are specially set to adapt to the hydrological forecasting scenario. The model is trained on historical data under the constraint of dual weight loss function, and the trained dual LSTM encoder-decoder is output as a deeply coupled model with XAJ. The refined runoff forecast output module: It inputs the collected real-time data into the trained dual LSTM encoder-decoder and XAJ deeply coupled model to output refined runoff forecast products.
[0032] like Figure 1 As shown, in a preferred embodiment 3, a specific solution is as follows: (I) Overall Architecture of the Technical Solution This technical solution consists of four key components, forming a closed loop of "data support - model coupling - optimized training - forecast output". Each component is designed to address the shortcomings of existing technologies, and three core innovations are integrated throughout: 1. Data preprocessing and fusion: Dedicated to adapting to high-resolution radar, generating high-quality data at the 10-minute level, providing support for minute-level fusion; 2. Deep coupling of dual LSTM encoder-decoder + XAJ: Through the division of labor between encoder and decoder, the collaboration of "feature extraction-super-resolution prediction" is achieved, combined with XAJ physical constraints; 3. Training a composite loss function that integrates physical constraints and focuses on extreme floods: With the dual optimization of "observation accuracy + physical rationality" as the core objective, the model accurately balances runoff super-resolution accuracy and hydrological physical rationality through dual collaborative optimization, while effectively improving the model's generalization ability and watershed adaptability; at the same time, an extreme event weighting mechanism is introduced to significantly improve the model's forecast accuracy for extreme flood events by strengthening the training weight of extreme flood samples.
[0033] 4. Refined forecast output: Outputs 10-minute runoff data to directly meet the flood early warning needs of small and medium-sized watersheds.
[0034] (II) Specific Steps 1. Data Preprocessing and Fusion Core objective: To transform raw data from X-band dual-polarization phased array precipitation radar into 10-minute-level high-resolution gridded quantitative precipitation estimation (QPE) data adapted to the coupled model, eliminating noise and terrain interference, retaining the advantages of high spatiotemporal resolution, providing high-quality data support for over-resolution and correction, and solving the problem of insufficient data adaptation in existing technologies.
[0035] Inputs: Raw data from X-band dual-polarization phased array rain-measuring radar (time resolution 40 seconds, range resolution 30 meters), digital elevation model (DEM) data of the study watershed, watershed boundary vector data, measured runoff data from hydrological stations, and watershed underlying surface data (land use type, soil type).
[0036] Output: 10-minute normalized gridded QPE data for watershed partitions, standardized measured runoff data, watershed topographic parameters, and XAJ model initial parameter dataset.
[0037] Specific implementation steps: Radar QPE data acquisition and verification: The raw detection data (time resolution 40 seconds, range resolution 30 meters) of the X-band dual polarization phased array rain measurement radar were collected and, after quality control, gridded QPE data within the watershed area were generated.
[0038] Topographic and hydrological data collection: Key topographic information is extracted from DEM data, including watershed boundaries, topographic slope, river length and slope, for sub-watershed division and parameter calibration of the XAJ model. Collect measured data from hydrological stations, including hourly / 10-minute runoff flow and water level data, as the true values and validation basis for model training; Organize watershed underlying surface data: Perform regional statistics according to land use type (arable land, forest land, etc.) and soil type (loam, clay, etc.) to provide a basis for watershed adaptation of XAJ model physical parameters.
[0039] Data standardization and adaptation: High-resolution data-specific aggregation: The original radar QPE data with a time resolution of 40 seconds is aggregated by surface average statistical aggregation according to "10-minute time step + watershed partition spatial granularity" to generate watershed partition 10-minute gridded QPE data - which retains the advantages of radar's high spatiotemporal resolution, adapts to the timing requirements of model calculation, and avoids data redundancy. Time series data normalization: Mini-maximum normalization is performed on 10-minute radar QPE data and measured runoff data to eliminate dimensional differences and accelerate the convergence of subsequent model training; Missing data completion: Temporal continuity detection is performed on the aggregated data. For a small number of missing data (such as missing data at a single time step), linear interpolation is used to complete the missing data. If there are multiple consecutive missing data steps, radar data and rain gauge data from similar terrain areas at the same time are used for collaborative completion to ensure the temporal integrity of the data and avoid affecting the model's feature extraction effect.
[0040] 2. Construction of a Dual LSTM Encoder-Decoder Deeply Coupled with the XAJ Model Core objective: To construct a deeply coupled system of "dual LSTM division of labor + XAJ physical constraints" to achieve minute-level fusion, super-resolution and forecasting collaboration, and solve the problems of single LSTM overload and fusion scale limitation in existing technologies.
[0041] Inputs: 10-minute gridded QPE data output from step 1, standardized measured runoff data, watershed topographic parameters, and initial parameters of the XAJ model.
[0042] Output results: Preliminary results of 10-minute runoff forecasts over 36 time steps (3 hours) (including 10-minute flow data). Specific implementation steps: (1) XAJ model module: physical mechanism foundation construction A semi-distributed XAJ model structure is adopted. Based on the topography and underlying surface characteristics of the watershed, the watershed is discretized into several sub-watersheds. Each sub-watershed performs runoff calculation independently, and then the runoff is aggregated into the whole watershed runoff process through the runoff confluence module.
[0043] Input data: 10-minute radar QPE surface average rainfall data output from step 1, as well as initial condition parameters such as watershed evapotranspiration capacity and soil moisture content.
[0044] Core calculations: Through the four modules of the XAJ model (evapotranspiration estimation, runoff generation calculation, runoff segmentation, and confluence calculation), it outputs hourly runoff forecast results for the entire watershed (including surface runoff, groundwater runoff, and total runoff), providing strict physical constraint anchors for the subsequent LSTM module—ensuring that the over-resolution and correction results do not deviate from the hydrological cycle.
[0045] (2) Dual LSTM encoder-decoder module (core division of labor design): Model Structure: A dual LSTM network in the encoder-decoder paradigm is adopted, forming a collaborative relationship with the XAJ model of "physical constraints + feature extraction + super-resolution correction". The model parameters are specifically set to adapt to the hydrological forecasting scenario, and the specific parameter settings are shown in the table below: Table 1. Specific parameters of the dual LSTM encoder-decoder module
[0046] 1) Encoder LSTM: Focuses on feature extraction, without prediction interference. Input data: 10-minute standardized data from the past 6 hours (radar QPE data (P) and measured runoff data (Q) at 36 time steps); Core computation: Through a 3-layer LSTM gating mechanism, it focuses on extracting the temporal features of the dynamic evolution of rainfall-runoff without participating in any forecast calculations, thus avoiding functional interference; Output: Condensed encoded hidden states (h0) and cell states (c0) encode historical hydrological process information, providing accurate feature support for the decoder.
[0047] 2) Decoder LSTM: Focuses on super-resolution prediction, subject to physical constraints. Initial conditions: The encoder's output state is used as the initial condition to ensure feature coherence; Input data: 10-minute radar QPE data for the next 6 hours (36 time steps P) +1 ~P +36 ) and hourly runoff forecast data of the XAJ model (XAJ_Q); Core computing: Simultaneously implementing two major functions, and subject to XAJ physical constraints: Minute-level super-resolution and accurate forecasting: Nonlinear correction and spatiotemporal super-resolution are performed through LSTM network. Based on minute-level features extracted by encoder, the bias of XAJ model in rapid flow generation scenario is corrected, and 10-minute-level forecast results are output. Output: Preliminary runoff forecast data at the 10-minute level over 36 time steps (3-hour lead time).
[0048] 3. Model training and watershed optimization under dual-weight loss function constraints: Core objective: To improve the generalization ability and watershed adaptability of the dual LSTM architecture through dual optimization of the dual-weight loss function, while balancing super-resolution accuracy and physical rationality, and solving the problems of single loss function and insufficient generalization ability in existing technologies.
[0049] Input data: preliminary runoff forecast results, measured runoff data, XAJ model runoff data, and watershed topography and underlying surface characteristics data output from step 2.
[0050] Output: The trained dual LSTM encoder-decoder deeply coupled model with XAJ.
[0051] Specific implementation steps: (1) Data set partitioning: The training set and validation set are divided proportionally. The training set covers different seasons, weather systems and hydrological years to ensure the representativeness of the samples; (2) Construction and training of a composite loss function that integrates physical constraints and focuses on extreme floods: To address the shortcomings of existing loss functions that focus on a single objective, a dual-weighted mean square error loss function is constructed that balances "observation accuracy" and "physical rationality," as shown in the following formula:
[0052]
[0053] in Weights for losses based on different flow values. The predicted runoff values are from the LSTM-XAJ module. This is the measured runoff value. Here, w1 and w2 are the predicted runoff values from the XAJ model, and w1 and w2 are weighting coefficients. This design effectively ensures that the runoff forecast results accurately fit the observed data, guaranteeing forecast accuracy, while strictly adhering to the laws of hydrophysical processes and not deviating from core hydrological constraints such as water balance. Furthermore, it significantly enhances the model's ability to detect and forecast extreme flood events.
[0054] Training execution: The Adam optimizer is used with a learning rate of 1e-4, a batch size of 512, and 130 training epochs. The optimization objective is the two-weighted mean squared error. The DC value, peak error (PeakE), and loss function value of the validation set are monitored in real time. Training is stopped if there is no decrease for 10 consecutive epochs (early stopping mechanism) to avoid overfitting.
[0055] (3) Watershed adaptability calibration: XAJ parameter calibration: Adjust core physical parameters according to the underlying surface type; Dynamic adjustment of dual-weight coefficients: The above weight adjustment rules are executed according to terrain type to ensure that the model is adapted to the characteristics of different watersheds; Verification after calibration ensures accuracy for operational applications.
[0056] 4. Refined runoff forecast output Core objective: To output refined runoff data at the 10-minute level to meet the real-time and accuracy requirements of flood early warning in small and medium-sized watersheds.
[0057] Input: Real-time radar raw data, historical hydrological data, and trained coupled model.
[0058] Output: 3-hour lead time and 10-minute level refined runoff forecast products.
[0059] Specific implementation steps: (1) Real-time data rapid preprocessing: The real-time collected radar raw data is rapidly processed according to the standardized process in step 1: including quality control, terrain occlusion correction, 10-minute level aggregation, and normalization, to ensure that the format and quality of the real-time input data are consistent with those of the training data; (2) Real-time model inference: The preprocessed real-time 10-minute radar QPE data is input into the coupled model. The model automatically executes the entire process of hourly runoff calculation of XAJ model, time series feature extraction of encoder and super-resolution forecast of decoder without manual intervention. (3) Forecast results output and formatting: Output 10-minute runoff data (time resolution 10 minutes) within the 3-hour forecast period, flood peak flow forecast value, and flood peak occurrence time forecast value; The output data supports the interface format of mainstream hydrological early warning systems and can be directly used for flood early warning threshold comparison and scheduling decision reference.
[0060] Field measurements show that the coefficient of determination (DC) of this technology is ≥0.81 (3-hour forecast period), and the peak error (PeakE) is ≤9.91m. 3 / s, both are superior to the traditional XAJ model (DC=0.76-0.79, PeakE=9.65-14.62m). 3 / s).
Claims
1. A refined hydrological forecasting method based on rainfall radar, characterized in that, Includes the following steps: Data preprocessing and fusion: Collect real-time data, including raw detection data from X-band dual-polarization phased array rain measurement radar, convert the raw detection data from X-band dual-polarization phased array rain measurement radar into QPE gridded raw data, and perform data standardization and adaptation processing on the QPE raw data. Combine this with the collected topographic and hydrological data for preprocessing and data fusion. Model construction: A dual LSTM network with an encoder-decoder paradigm is adopted to form a collaborative relationship with the XAJ model in terms of physical constraints, feature extraction and super-resolution correction. The model parameters are specially set to adapt to the hydrological forecasting scenario. Model training: The model is trained using historical data under the constraint of a dual-weight loss function, and the trained dual LSTM encoder-decoder deeply coupled with XAJ is output. Refined runoff forecast output: The collected real-time data is input into the trained dual LSTM encoder-decoder and XAJ deeply coupled model to output a refined runoff forecast product.
2. The refined hydrological forecasting method based on rainfall radar according to claim 1, characterized in that, The data preprocessing and fusion also includes the following steps: Radar QPE data acquisition and verification: The raw detection data of the X-band dual polarization phased array rain measurement radar is collected and, after quality control, gridded QPE data within the watershed area is generated. Topographic and hydrological data collection: Key topographic information is extracted from DEM data, including watershed boundaries, topographic slope, and river length and slope, for sub-watershed division and parameter calibration of the XAJ model. Collect measured data from hydrological stations, including runoff and water level data, as the true values and validation basis for model training; Organize watershed underlying surface data: perform regional statistics according to land use type and soil type to provide a basis for watershed adaptation of XAJ model physical parameters.
3. The refined hydrological forecasting method based on rainfall radar according to claim 2, characterized in that, The data fusion also includes the following steps: High-resolution data-specific aggregation: The raw radar QPE data acquired from radar QPE data is statistically aggregated by surface averaging to generate gridded QPE data; Time series data normalization: Mini-maximum normalization is performed on gridded QPE data and measured runoff data to eliminate dimensional differences and accelerate the convergence of subsequent model training; Missing data completion: Temporal continuity detection is performed on the aggregated data. For a small number of missing data, linear interpolation is used to complete them. If there are multiple consecutive missing steps, radar data and rain gauge data from similar terrain areas during the same period are used for collaborative completion to ensure the temporal integrity of the data and avoid affecting the model feature extraction effect.
4. The refined hydrological forecasting method based on rainfall radar according to claim 1, characterized in that, The XAJ model adopts a semi-distributed XAJ model structure. Based on the topography and underlying surface characteristics of the watershed, the study watershed is discretized into several sub-watersheds. Each sub-watershed performs runoff calculation independently, and then the runoff is aggregated into the whole watershed runoff process through the runoff confluence module.
5. The refined hydrological forecasting method based on rainfall radar according to claim 1, characterized in that, The encoder-decoder paradigm dual LSTM network includes an encoder LSTM and a decoder LSTM. The encoder LSTM is used for feature extraction and is free from prediction interference, while the decoder LSTM is used for super-resolution prediction and is subject to physical constraints.
6. The refined hydrological forecasting method based on rainfall radar according to claim 5, characterized in that, The encoder LSTM takes into account standardized data from a past period, including radar QPE data and measured runoff data within that period. Through the LSTM gating mechanism, it focuses on extracting minute-level temporal features of the dynamic evolution of rainfall-runoff, and outputs condensed encoded hidden states and cell states, encoding historical hydrological process information to provide accurate feature support for the decoder.
7. A refined hydrological forecasting method based on rainfall radar according to claim 6, characterized in that, The decoder LSTM takes into account standardized data of the same duration as the encoder LSTM input, including radar QPE data and hourly runoff forecast data from the XAJ model. The LSTM network performs nonlinear correction and spatiotemporal super-resolution. Based on minute-level features extracted by the encoder LSTM, the bias of the XAJ model in rapid runoff scenarios is corrected, and the forecast results are output.
8. The refined hydrological forecasting method based on rainfall radar according to claim 1, characterized in that, The dual-weighted loss function constraint includes the following formula: ; ; in Weights for losses based on different flow values. Forecast runoff values from a dual LSTM encoder-decoder + XAJ model. This is the measured runoff value. is the predicted runoff value of the XAJ model, and w1 and w2 are weighting coefficients.
9. A refined hydrological forecasting method based on rainfall radar according to claim 8, characterized in that, The refined runoff forecast output also includes the following steps: Real-time data rapid preprocessing: The raw radar data acquired in real time is rapidly processed according to a standardized data preprocessing and fusion process; Real-time model inference: The preprocessed real-time minute-level radar QPE data is input into the coupled model, and the model automatically executes the entire process of hourly runoff calculation of the XAJ model, encoder time series feature extraction and decoder super-resolution forecast; Forecast results output and formatting: Output minute-by-minute runoff data, peak flow forecast, and peak occurrence time forecast for the forecast period. The output data supports the interface formats of mainstream hydrological early warning systems and can be directly used for flood warning threshold comparison and scheduling decision reference.
10. A system for implementing the refined hydrological forecasting method based on rainfall radar as described in any one of claims 1 to 9, characterized in that, Includes the following modules: Data preprocessing and fusion module: collects real-time data, including raw detection data from X-band dual polarization phased array rain measurement radar, converts the raw detection data from X-band dual polarization phased array rain measurement radar into QPE gridded raw data, and performs data standardization and adaptation processing on the QPE raw data, in conjunction with the collection of topographic and hydrological data. Model module: It adopts a dual LSTM network in the encoder-decoder paradigm, which forms a collaborative relationship with the XAJ model in terms of physical constraints, feature extraction and super-resolution correction. The model parameters are specially set to adapt to the hydrological forecasting scenario. The model is trained on historical data under the constraint of dual weight loss function, and the trained dual LSTM encoder-decoder is output as a deeply coupled model with XAJ. The refined runoff forecast output module: It inputs the collected real-time data into the trained dual LSTM encoder-decoder and XAJ deeply coupled model to output refined runoff forecast products.
